Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

default · ~99.6k rows (showing the first 120k)

Split (2)

train · ~72.4k rows (showing the first 93.1k)

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1611.00350	2	1611	2019-01-19T16:55:50	Adversarial Influence Maximization	[ "cs.SI", "cs.LG", "stat.ML" ]	We consider the problem of influence maximization in fixed networks for contagion models in an adversarial setting. The goal is to select an optimal set of nodes to seed the influence process, such that the number of influenced nodes at the conclusion of the campaign is as large as possible. We formulate the problem as a repeated game between a player and adversary, where the adversary specifies the edges along which the contagion may spread, and the player chooses sets of nodes to influence in an online fashion. We establish upper and lower bounds on the minimax pseudo-regret in both undirected and directed networks.	cs.SI	cs	Adversarial Influence Maximization Justin Khim∗ Varun Jog† Po-Ling Loh†‡ January 19, 2019 Abstract We consider the problem of influence maximization in fixed networks for contagion models in an adversarial setting. The goal is to select an optimal set of nodes to seed the influence process, such that the number of influenced nodes at the conclusion of the campaign is as large as possible. We formulate the problem as a repeated game between a player and adversary, where the adversary specifies the edges along which the contagion may spread, and the player chooses sets of nodes to influence in an online fashion. We establish upper and lower bounds on the minimax pseudo-regret in both undirected and directed networks. 1 Introduction Many data sets in contemporary scientific applications possess some underlying network struc- ture [32]. Popular examples include data collected from social media websites such as Facebook and Twitter [1, 28], or electrocortical recordings gathered from a network of firing neurons [35]. An important application of network science arises in marketing, where researchers have studied the importance of word-of-mouth advertising for decades [23]. More recently, methods have been pro- posed by marketing researchers to quantify the importance of word-of-mouth marketing in online social networks in both theory and practice [40, 37]. Subsequent empirical studies suggest that word-of-mouth marketing has a significant effect in online social networks [3, 34]. At the same time, computer scientists have analyzed the problem of viral marketing from an optimization-theoretic perspective [15, 27, 11], where the goal is to select an optimal set of influencers to encourage product adoption in an online social network. This has led to rigorous theoretical guarantees that hold for stochastic models of word-of-mouth advertising inspired by physics and epidemiology, and the scope of the spread is quantified using a notion known as influence [24]. In social networks, edges represent potential interactions between individuals, and the problem of influence maximization corresponds to identifying subsets of individuals on which to impress an idea so that information spreads as widely as possible subject to an advertising budget. Formally, the influence of a subset of nodes is defined as the expected number of influenced individuals in a network at the conclusion of a spread, starting from an initial configuration where only the specified nodes are influenced. Even when the influence function is assumed to be com- putable for any subset using a black-box method in unit time, it is not clear whether influence maximization may be performed (exactly or approximately) in polynomial time, since searching over all subsets of k nodes is exponential in the number of nodes. Accordingly, the body of work in theoretical computer science has mostly focused on specific spreading models that give rise to ∗Department of Statistics, University of Pennsylvania, Philadelphia, PA 19104. †Department of Electrical & Computer Engineering, University of Wisconsin, Madison, WI 53706. ‡Department of Statistics, University of Wisconsin, Madison, WI 53706. 1 nice properties such as submodularity, implying that a greedy algorithm for influence maximization leads to a constant-factor approximation of the optimal set [24, 25, 8]. Other related work includes predicting when knowledge becomes viral; limiting the spread of information through carefully po- sitioned interventions [14, 16]; or competitive settings of influence maximization, e.g. competing for votes or market share [7, 19, 18]. A significant shortcoming in the analysis of stochastic spreading models is the fact that the parameters characterizing the spread of influence are generally assumed to be known, allowing for approximate evaluation of the influence function (either by analytic methods or simulation). However, such an assumption is not always practical. In the case of independent cascade models or linear threshold models, where parameters correspond to edge weights in the network, one might even question a scientist's prior knowledge of the precise network structure. To address these issues, some authors have studied the interesting question of accurately learning the influence function itself in a stochastic spreading model based on observing multiple rounds of infection [29, 26, 21]. Another approach involves a notion of "robust influence maximization," where the parameters are only specified to lie in fixed confidence sets, and the goal is to obtain a set of source vertices that approximately maximizes the true influence function, possibly in a worst-case sense [12, 20]. Robust influence maximization methods may also be model-dependent, meaning that a robust algorithm designed for the independent cascade model may lead to a severely non-optimal solution if the influence spread actually follows linear threshold model. Indeed, the parameters describing different models, as well as the nature of uncertainties permitted in them, may be completely different. Further, it is unclear that popular models of influence are good apprixmations of real-world behavior [17, 22]. In this paper, we take a rather different approach toward the problem of unknown spreading parameters that also avoids assumptions about a particular spreading mechanism. As discussed in more detail in Section 2, we only assume knowledge of an underlying fixed graph representing the paths along which a influence may spread, where the case of no prior knowledge corresponds to a complete graph. We formulate the influence maximization problem as an online game, where a "player" must make sequential decisions about the next seed set to choose based on observing the behavior of the spread in previous "rounds" of the game. Here, a round represents a particular instance of an influence process initialized from the specific seed nodes from beginning to end. We allow an "adversary" to choose the path of influence on each round in a completely arbitrary manner, as long as the process may only spread along edges of the graph -- in particular, this setting subsumes the stochastic models usually adopted in the influence maximization literature, while allowing for much more general spreading mechanisms (e.g., information does not necessarily propagate in an i.i.d. manner over all rounds of the game). Note that the adversary's strategy may be so arbitrary as to be "unlearnable." Thus, instead of simply trying to maximize the aggregate number of influenced vertices across all rounds, we seek to develop player strategies that bound the "regret" of the player, defined as the difference between the total number of vertices influenced using the player's strategy and the number of vertices that would have been influenced if the player had adopted the best constant choice of source set in hindsight. Such notions are taken from the literature on multi- armed bandits and online learning theory [5, 10], and adapted to the present setting. Our main contribution is to derive upper and lower bounds on the pseudo-regret for various adversarial and player strategies. We study both directed and undirected networks, where in the latter setting, contagion is allowed to spread in both directions when an edge is chosen by the adversary. Furthermore, we derive lower bounds for the minimax pseudo-regret when the underlying network is a complete graph, where the supremum is taken over all adversarial strategies and the infimum is taken over all player strategies. Our upper and lower bounds match up to constant factors in the case of directed networks. Notably, the bounds also agree with the usual rate for 2 pseudo-regret in multi-armed bandits, showing that no new information is gained by the player by exploiting network structure. On the other hand, a gap exists between our upper and lower bounds for undirected networks, leaving open the possibility that the player may leverage the additional information from the network to incur less regret. Additionally, the constant factor on the upper bound may be slightly improved, providing further evidence that graph structure may be exploited. Finally, we demonstrate how to extend our upper bounds to the setting where the player is allowed to choose multiple source vertices on each round. The proposed multi-source player strategy augments the source set sequentially using the single-source strategies as a subroutine, and is based on a general online greedy algorithm proposed by Streeter and Golovin [36]. The remainder of our paper is organized as follows: In Section 2, we provide some important background on online learning theory and formally define the adversarial spreading model and notions of regret to be studied in our paper. In Section 3, we present upper and lower bounds for pseudo-regret in the adversarial setting. We conclude the paper with a selection of open research questions in Section 4. All proofs, as well as a more technical discussion of related work, is contained in the appendices. Notation. For a set A, let 2A denote the power set of A. When we want to specify that we are taking the expectation with respect to a particular distribution p of some random variable X, we write EX∼p. In particular, we often write ES∼p to mean the expectation taken over the player's actions for a fixed set of adversarial actions, which is the same as the conditional expectation with respect to the adversary's actions. Similarly, we write EA to indicate the conditional expectation with respect to a fixed set of player actions. 2 Background and preliminaries We begin by formally defining the repeated game between the player and adversary and the types of strategies we will analyze in our paper. Next, we introduce the notions of regret we will study, and then connect our setting to related work in the learning theory literature. 2.1 Adversarial repeated games Consider a fixed graph G = (V, E) on n vertices, which may be directed or undirected. The adversarial influence maximization problem may be described as follows: Repeatedly over T rounds, the player selects an influence seed set St ⊆ V , with St = k, for t = 1, . . . , T . At the same time, the adversary designates a subset of edges At ⊆ E to be "open." A node is considered to be influenced at time t if and only if it is an element of St or is reachable from St via a path of open edges. Note that in the context of influence spreading, the open edges correspond to ties over which influence propagates in that round -- importantly, influence only has an opportunity to be transmitted between individuals that interact in the network, but may not necessarily spread over a particular connection on a specific round. In the case when G is an undirected graph, designating an edge to be open allows an influence campaign to spread in both directions. Furthermore, in the directed case, edges may exist in both directions between a given pair of nodes, in which case the adversary may designate both, one, or neither of the edges to be open. For an open edge set A ⊆ E and influence seed set S ⊆ V , we define f (A, S) to be the fraction of vertices in the graph lying in the influenced set. To connect our model to the canonical setting of influence maximization, note that [25] proposed a very general class of influence models called triggering models, which include the independent 3 cascade and the linear threshold models as special cases. At the beginning of the influence campaign, each node chooses a random "triggering" subset of neighbors according to a particular rule, and the incoming edges from those neighbors are designated to be "active." A vertex becomes influenced during the course of the process if and only if a path of active edges exists connecting that vertex to a vertex in the seed set. Thus, triggering models correspond to a special case of our framework, in which the edge sets are chosen in an i.i.d. manner from round to round, and the probability distribution over the edges is determined by the probability rule through which edges are assigned to be active (e.g., according to the linear threshold or independent cascade models). Next, we describe the classes of strategies A = {At} and S = {St} available to the adversary and player. We assume that the adversary is oblivious of the player's actions; i.e., at time t = 0, the adversary must decide on the (possibly random) strategy A. We use A to denote the set of oblivious adversary strategies and Ad to denote the set of deterministic adversary strategies. Turning to the classes of player strategies, we allow the player to choose his or her action at time t based on the feedback provided in response to the joint actions made by the player and adversary on preceding time steps. Although the player knows the edge set E of the underlying graph, we assume that the player only observes the status of edges (i, j) such that either i or j is in the reach of St (in the undirected case), and the player observes the status of every edge (i, j) such that i is in the reach of St (in the directed case). In other words, whereas the player cannot observe the subset of all edges that would have propagated influence in the network, he or she will know which edges transmitted influence if reached by the influence cascade initialized using his or her seed set. Formally, we write I (At,St) to denote the set of edges with status known to the player (i.e., all edges in the subgraph induced by At belonging to connected components containing nodes in St), and we denote I t = (I (A1,S1), . . . , I (At,St)). If At is chosen via a stochastic model such as the independent cascade model with discrete time steps for influence campaign t, our setup technically allows the player knowledge of the status of an edge between two vertices u and v if both were actually influenced by some other vertex w. Realistically we would not want the status of edge (u, v) to be returned as feedback, and we could enforce this by positing a model of how each influence campaign proceeds. However, this distinction does not affect our results or algorithms, and so we do not further restrict the feedback I (At,St). The player can only make decisions based on the feedback observed in previous rounds, so any allowable player strategy {St} has the property that St is a function of I t−1 (possibly with additional randomization). We denote the class of all player strategies by P, and denote the subclass of all deterministic player strategies by Pd, meaning that St is a deterministic function of I t−1. Note that strategies St ∈ Pd may still be random, due to possible randomization of the adversary, but conditioned on I t−1, the choice of St is deterministic. 2.2 Minimax regret The player wishes to devise a strategy that maximizes the aggregate number of influenced nodes up to time T . Using the notation from the previous section, we define the regret of the player to be RT (A,S) = TXt=1 f (At,S∗) − TXt=1 f (At,St), (1) where S∗ = arg max S:S=k TXt=1 f (At, S) 4 is the optimal fixed set that the player would have chosen in hindsight with full knowledge of the adversary's strategy. Note that the regret RT (A,S) may be a random quantity due to randomness in both the adversary's or player's strategies. Accordingly, we will seek to control the pseudo-regret RT (A,S) := max f (At, S) − S:S=k(EA,S" TXt=1 f (At,St)#) , (2) TXt=1 where the expectation in equation (2) is taken with respect to potential randomization in both A and S. As in the standard learning theory literature [9], recall that the expected regret and pseudo-regret are generally related via the inequality RT (A,S) ≤ E[RT (A,S)], although if A ∈ Ad, we have RT (A,S) = E[RT (A,S)]. Our interest in the pseudo-regret rather than the expected regret is purely motivated by the fact that the former quantity is often easier to bound than the latter and that this simplification is common in the literature on bandits. Finally, we introduce the scaled regret and the analogous quantity Rα(A,S) = α TXt=1 S:S=k(EA,S"α α R T (A,S) = max f (At,S∗) − TXt=1 f (At,St), (3) TXt=1 f (At, S) − f (At,St)#) . TXt=1 Note that α = 1 corresponds to the unscaled version. Our interest in the expression (3) is again for theoretical purposes, since we may obtain convenient upper bounds on the scaled pseudo-regret in the case α = 1 − 1 e using an online greedy algorithm. Note that when k > 1, the benchmark greedy algorithms used for influence maximization in the stochastic spreading setting are also only guaranteed to achieve a(cid:0)1 − 1 e(cid:1)-approximation of the truth, so in some sense, the scaled regret (3) only requires the player to perform comparably well in relation to the appropriately scaled optimal strategy. 3 Main results In this section, we provide upper and lower bounds for the pseudo-regret. Specifically, we focus on the quantity inf S∈P sup A∈A α R T (A,S), where the supremum is taken over the class of adversarial strategies, and the infimum is taken over the class of player strategies based on the feedback model we have described. In other words, we wish to characterize the hardness of the influence maximization problem in terms of the player's best possible strategy measured with respect to the worst-case game. A rough outline of our approach is as follows: We establish upper bounds by presenting par- ticular strategies for the player that ensure an appropriately bounded regret under all adversarial strategies. For lower bounds, the general technique is to provide an ensemble of possible actions for the adversary that are difficult for the player to distinguish in the influence maximization problem, which forces the player to incur a certain level of regret. 5 3.1 Undirected graphs We begin by deriving regret upper bounds for undirected graphs. We initially restrict our attention to the case k = 1. The proposed player strategy for k > 1, and corresponding regret bounds, builds upon the results in the single-source setting. 3.1.1 Upper bounds for a single source Consider a randomized player strategy that selects St = {i} with probability pi,t. The paper [9] suggests a method based on the Online Stochastic Mirror Descent (OSMD) algorithm, which is specified by loss estimates {ℓi,t} and learning rates {ηt}, as well as a Legendre function F . Here, we comment on the losses, and in order to avoid excessive technicalities, we defer additional details of the OSMD algorithm to the appendix. The most basic loss estimate, which follows from standard bandit theory and ignores all infor- mation about the graph, is ℓi,t pi,t 111St={i}, (4) i,t = bℓnode bℓnode where ℓi,t = 1 − f (At,{i}) is the loss incurred if the player were to choose St = {i}. Importantly, is always computable for any choice the player makes at time t and is an unbiased estimate of i,t ℓi,t. On the other hand, if St = {i} and another node j is influenced (i.e., in the connected component formed by the open edges of At), the player also knows the loss that would have been incurred if St = {j}, since f (At,{i}) = f (At,{j}). This motivates an alternative loss estimate that is nonzero even when St 6= {i}. In particular, we may express 1 ℓi,t = ℓt i,j, nXj6=i where ℓt edges of At. We then define i,j is the indicator that i and j are in different connected components formed by the open Zij , pi,t + pj,t ℓt i,j 1 i,t = is always computable by the player, since the value of ℓt bℓsym where Zij = 111St∩{i,j}6=∅. Note that bℓsym known. We call bℓsym Theorem 1 (Symmetric loss, OSMD). Suppose the player uses the strategy S sym OSMD corresponding to OSMD with the symmetric loss ℓsym and appropriate parameters. Then the pseudo-regret satisfies the bound nXj6=i is also an unbiased estimate for ℓi,t. The estimator bℓsym i,j is known by the player whenever St is the symmetric loss. Now, we state the following regret bounds: i,t i,t i,t RT (A,S sym OSMD) ≤ 2 sup A∈A 1 4√T n. Remark 1. It is instructive to compare the result of Theorem 1 with analogous regret bounds for generic multi-armed bandits. When the OSMD algorithm is run with the loss estimates (4), standard 2√T n. Thus, using the symmetric loss, which leverages the analysis establishes an upper bound of 2 graphical nature of the problem, produces slight gains. 3 6 3.1.2 Lower bounds We now establish lower bounds for the pseudo-regret in the case k = 1. This furnishes a better understanding of the hardness of the adversarial influence maximization problem. The general approach for deriving lower bounds is to produce a strategy for the adversary that forces the player to incur a certain level of regret regardless of which strategy is chosen. The intrinsic difficulty of online influence maximization may vary widely depending on the topology of the underlying graph, and methods for deriving lower bounds may also differ accordingly. In the case of a complete graph, we have the following result: Theorem 2. Suppose G = Kn is the complete graph on n ≥ 3 vertices. Then the pseudo-regret satisfies the lower bound 2 243 √T ≤ inf S∈P sup A∈A RT (A,S). Remark 2. Clearly, a gap exists between the lower bound derived in Theorem 2 and the upper bound appearing in Theorem 1. It is unclear which bound, if any, provides the proper minimax rate. However, note that if the lower bound were tight, it would imply that the proportion of vertices that the player misses by picking suboptimal source sets is constant, meaning the number of additional vertices the optimal source vertex influences is linear in the size of the graph. This differs substantially from the pseudo-regret of order √n known to be minimax optimal for the standard multi-armed bandit problem (and arises, for instance, in the case of directed graphs, as discussed in the next section). 3.1.3 Upper bounds for multiple sources We now turn to the case k > 1, where the player chooses multiple source vertices at each time step. As discussed in Section 2, we are interested in bounding the scaled pseudo-regret R T (A,S) with α = 1 − 1 e , since it is difficult to maximize the influence even in an offline setting, and the greedy algorithm is only guaranteed to provide a(cid:0)1 − 1 Our proposed player strategy is based on an online greedy adaptation of the strategy used in the single-source setting, and the full details are given in the appendix. We then have the following result concerning the scaled pseudo-regret: e(cid:1)-approximation of the truth. α Theorem 3 (Symmetric loss, multiple sources). Suppose k > 1 and the player uses the strategy S sym,k OSMD corresponding to the Online Greedy Algorithm with single-source strategy S sym OSMD. Then the scaled pseudo-regret satisfies the bound R (1−1/e) T (A,S sym,k OSMD) ≤ 2 sup A∈A 1 4 k√T n. Comparing Theorem 3 to Theorem 1, we see an additional factor of k in the pseudo-regret upper bound. Similar results may be derived when alternative single-source strategies are used as subroutines in the Online Greedy Algorithm. 3.2 Directed graphs We now derive upper and lower bounds for the pseudo-regret in the case of directed graphs, when k = 1. 7 3.2.1 Upper bounds The symmetric loss does not have a clear analog in the case of directed graphs. However, we may still use the node loss estimate for multi-armed bandit problems, given by equation (4). This leads to the following upper bound: Theorem 4. Suppose the player uses the strategy S node loss bℓnode and appropriate parameters. Then the pseudo-regret satisfies the bound 2√T n. 3 OSMD corresponding to OSMD with the node RT (A,S node OSMD) ≤ 2 sup A∈A Remark 3. In the case k > 1, we may again use the Online Greedy Algorithm used in Section 3.1.3 to obtain a player strategy composed of parallel runs of a single-source strategy. If the player uses the single-source strategy S node OSMD, we may obtain the scaled pseudo-regret bound R (1−1/e) T (A,S node,k OSMD) ≤ 2 sup A∈A 3 2 k√T n. 3.2.2 Lower bounds Finally, we provide a lower bound for the directed complete graph on n vertices. (This refers to the case where all edges are present and bidirectional.) We have the following result: Theorem 5. Suppose G is the directed complete graph on n vertices. Then the pseudo-regret satisfies the lower bound 1 48√6 √T n ≤ inf S∈P sup A∈A RT (A,S). Notably, the lower bound in Theorem 5 matches the upper bound in Theorem 4, up to constant factors. Thus, the minimax pseudo-regret for the influence maximization problem is Θ(√T n) in the case of directed graphs. In the case of undirected graphs, however (cf. Theorem 2), we only obtained a pseudo-regret lower bound of Ω(√T ). This is due to the fact that in undirected graphs, one may learn about the loss of other nodes at time t besides the loss at St. In contrast, it is possible to construct adversarial strategies for directed graphs that do not provide information regarding the loss incurred by choosing a source vertex other than St. Finally, we remark that a different choice of G might affect the lower bound, since influence maximization is easier for some graph topologies than others. However, Theorem 5 shows that the case of the complete graph is always guaranteed to incur a pseudo-regret that matches the general upper bound in Theorem 4, implying that this is the minimax optimal rate for any class of graphs containing the complete graph. 4 Discussion We have proposed and analyzed player strategies that control the pseudo-regret uniformly across all possible oblivious adversarial strategies. For the problem of single-source influence maximization in complete networks, we have also derived minimax lower bounds that establish the fundamental hardness of the online influence maximization problem. In particular, our lower and upper bounds 8 match up to constant factors in the case of directed complete graphs, implying that our proposed player strategy is in some sense optimal. Our work inspires a number of interesting questions for future study. An important open question concerns closing the gap between upper and lower bounds on the minimax pseudo-regret in the case of undirected graphs, to determine whether the feedback available in the influence maximization setting actually makes the online game easier than a standard bandit setting. Furthermore, our lower bounds only hold in the case of complete graphs and single-source influence maximization, and it would be worthwhile to obtain lower bounds that hold for other network topologies and seed sets containing multiple nodes. Our results only address a small subset of problems that may be posed and answered concerning a bandit theory of adversarial influence maximization with edge-level feedback. References [1] L. A. Adamic and E. Adar. Friends and neighbors on the Web. Social Networks, 25(3):211 -- 230, 2003. [2] N. Alon, N. Cesa-Bianchi, C. Gentile, S. Mannor, Y. Mansour, and O. Shamir. Nonstochastic multi-armed bandits with graph-structured feedback. SIAM Journal on Computing, 46(6):1785 -- 1826, 2017. [3] S. Aral and D. Walker. Creating social contagion through viral product design: A randomized trial of peer influence in networks. Management Science, 57(9):1623 -- 1639, 2011. [4] J.-Y. Audibert, S. Bubeck, and G. Lugosi. Regret in online combinatorial optimization. Math- ematics of Operations Research, 39(1):31 -- 45, 2013. [5] P. Auer, N. Cesa-Bianchi, Y. Freund, and R. E. Schapire. The nonstochastic multiarmed bandit problem. SIAM journal on computing, 32(1):48 -- 77, 2002. [6] G. Bartók, D. P. Foster, D. Pál, A. Rakhlin, and C. Szepesvári. Partial monitoring -- Classification, regret bounds, and algorithms. Mathematics of Operations Research, 39(4):967 -- 997, 2014. [7] S. Bharathi, D. Kempe, and M. Salek. Competitive influence maximization in social networks. Internet and Network Economics, pages 306 -- 311, 2007. [8] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier. Maximizing social influence in nearly op- timal time. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 946 -- 957. SIAM, 2014. [9] S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed bandits. Foundations and Trends in Machine Learning, 5(1):1 -- 122, 2012. [10] N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, New York, NY, 2006. [11] W. Chen, L. V. Lakshmanan, and C. Castillo. Information and influence propagation in social networks. Synthesis Lectures on Data Management, 5(4):1 -- 177, 2013. 9 [12] W. Chen, T. Lin, Z. Tan, M. Zhao, and X. Zhou. Robust influence maximization. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016. [13] W. Chen, Y. Wang, Y. Yuan, and Q. Wang. Combinatorial multi-armed bandit and its exten- sion to probabilistically triggered arms. Journal of Machine Learning Research, 17(50):1 -- 33, 2016. [14] J. Cheng, L. Adamic, P. A. Dow, J. Kleinberg, and J. Leskovec. Can cascades be predicted? In Proceedings of the 23rd International Conference on WWW, pages 925 -- 936. ACM, 2014. [15] P. Domingos and M. Richardson. Mining the network value of customers. In Proceedings of the seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 57 -- 66. ACM, 2001. [16] K. Drakopoulos, A. Ozdaglar, and J. N. Tsitsiklis. When is a network epidemic hard to eliminate? Mathematics of Operations Research, 42(1):1 -- 14, 2016. [17] S. Goel, D. J. Watts, and D. G. Goldstein. The structure of online diffusion networks. In Proceedings of the 13th ACM conference on electronic commerce, pages 623 -- 638. ACM, 2012. [18] M. Grabisch, A. Mandel, A. Rusinowska, and E. Tanimura. Strategic influence in social net- works. Mathematics of Operations Research, 2017. [19] X. He and D. Kempe. Price of anarchy for the N-player competitive cascade game with submodular activation functions. In International Conference on Web and Internet Economics, pages 232 -- 248. Springer, 2013. [20] X. He and D. Kempe. Robust influence maximization. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016. [21] X. He, K. Xu, D. Kempe, and Y. Liu. Learning influence functions from incomplete observations. In Advances in Neural Information Processing Systems, pages 2073 -- 2081, 2016. [22] L. Hu, B. Wilder, A. Yadav, E. Rice, and M. Tambe. Activating the"breakfast club": Modeling influence spread in natural-world social networks. arXiv preprint arXiv:1710.00364, 2017. [23] D. Katz and R. L. Kahn. The social psychology of organizations, volume 2. Wiley New York, 1978. [24] D. Kempe, J. Kleinberg, and E. Tardos. Maximizing the spread of influence through a social network. In Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '03, pages 137 -- 146, New York, NY, USA, 2003. ACM. [25] D. Kempe, J. Kleinberg, and É. Tardos. Influential nodes in a diffusion model for social networks. In Automata, languages and programming, pages 1127 -- 1138. Springer, 2005. [26] S. Lei, S. Maniu, L. Mo, R. Cheng, and P. Senellart. Online influence maximization. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '15, pages 645 -- 654, New York, NY, USA, 2015. ACM. [27] J. Leskovec, L. A. Adamic, and B. A. Huberman. The dynamics of viral marketing. ACM Transactions on the Web (TWEB), 1(1):5, 2007. 10 [28] D. Liben-Nowell and J. Kleinberg. The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology, 58(7):1019 -- 1031, 2007. [29] H. Narasimhan, D. C. Parkes, and Y. Singer. Learnability of influence in networks. In Advances in Neural Information Processing Systems, pages 3186 -- 3194, 2015. [30] G. L. Nemhauser and L. A. Wolsey. Best algorithms for approximating the maximum of a submodular set function. Mathematics of operations research, 3(3):177 -- 188, 1978. [31] G. Neu and G. Bartók. An efficient algorithm for learning with semi-bandit feedback. International Conference on Algorithmic Learning Theory, pages 234 -- 248. Springer, 2013. In [32] M. E. J. Newman. The structure and function of complex networks. SIAM review, 45(2):167 -- 256, 2003. [33] J. Olkhovskaya, G. Neu, and G. Lugosi. Online influence maximization with local observations. arXiv preprint arXiv:1805.11022, 2018. [34] S. Seiler, S. Yao, and W. Wang. Does online word of mouth increase demand? (And how?) Evidence from a natural experiment. Marketing Science, 2017. [35] O. Sporns. The human connectome: A complex network. Annals of the New York Academy of Sciences, 1224(1):109 -- 125, 2011. [36] M. Streeter and D. Golovin. An online algorithm for maximizing submodular functions. Tech- nical report, pages 1 -- 35, 2007. [37] M. Trusov, R. E. Bucklin, and K. Pauwels. Effects of word-of-mouth versus traditional mar- keting: findings from an internet social networking site. Journal of Marketing, 73(5):90 -- 102, 2009. [38] S. Vaswani, L. Lakshmanan, and M. Schmidt. Influence maximization with bandits. arXiv preprint arXiv:1503.00024, pages 1 -- 12, 2015. [39] Q. Wang and W. Chen. Improving regret bounds for combinatorial semi-bandits with proba- bilistically triggered arms and its applications. In Advances in Neural Information Processing Systems, pages 1161 -- 1171, 2017. [40] D. J. Watts and P. S. Dodds. Influentials, networks, and public opinion formation. Journal of consumer research, 34(4):441 -- 458, 2007. [41] Z. Wen, B. Kveton, and M. Valko. Online influence maximization under independent cascade model with semi-bandit feedback. Advances in Neural Information Processing Systems, 2017. 11 A Related work Here, we comment more thoroughly on important relationships between our problem setting and various online games existing in the learning theory literature. A key difference between the graph contagion setting and the standard multi-armed bandit setting is that in the latter case, the only information available to the player on each round is the reward obtained as a consequence of his or her actions. On the other hand, slightly more information is available to the player in our setting, since the player may often deduce additional information about which vertices would have been influenced for a different choice of source vertices, based on observing the scope of the influence process for a particular choice of source vertices. As a concrete example, the player knows that exactly the same set of nodes would have been influenced if he or she had chosen to influence a different seed node in the same connected component of the subgraph induced by the influenced nodes and adversarially chosen edges. Online games with partial monitoring [6] or graph-based feedback [2] generalize the bandit setting to repeated games in which the player may observe feedback corresponding to various subsets of other actions in addition to or instead of observing the feedback corresponding to his or her actions. Although such games resemble our problem setting, the possible actions available of the player in our case correspond to subsets of nodes of size k, leading to a rather complicated feedback graph that is additionally affected by the adversary's actions. Another online game with a similar flavor is the combinatorial prediction setting [4], where the player is allowed to pull a subset of arms on each round, and observes a loss equal to the sum of losses of the pulled arms in the case of bandit feedback or a subvector of losses corresponding to the pulled arms, in the case of semi-bandit feedback [31]. Our problem may be cast as a type of combinatorial prediction game with a feedback graph that varies from round to round and is unknown to the player. Note that the combinatorial game with edge semi-bandit feedback has been studied recently in the influence maximization literature [13, 38, 41, 39, 33], but these results only apply to stochastic adversaries, rather than the more general non-stochastic framework we study in this paper. Edge semi-bandit feedback refers to the fact that in a directed graph, the player receives feedback about the transmission status of different subsets of edges, corresponding to the outgoing edges from the nodes he or she chooses to seed on each round. B Proofs We now outline the proofs of our main results. B.1 Upper bounds for adversarial models In this section, we prove our upper bounds. To this end, we describe the OSMD algorithm, which generates a sequence of probability distributions {pt} to be employed by the player on successive rounds. Let ∆n ⊆ Rn denote the probability simplex. Online Stochastic Mirror Descent (OSMD) with loss estimates {bℓi,t} Given: A Legendre function F defined on Rn, with associated Bregman divergence DF (p, q) = F (p) − F (q) − (p − q)T∇F (q), 12 and a learning rate η > 0. Output: A stochastic player strategy {St}. Let p1 ∈ arg minp∈∆n F (p). For each round t = 1, . . . , T : (1) Draw a vertex St from the distribution pt. (2) Compute the vector of loss estimates bℓt = {bℓi,t}. (3) Set wt+1 = ∇F ∗(cid:16)∇F (pt) − ηbℓt(cid:17), where F ∗ is the convex conjugate of F . (4) Compute the new distribution pt+1 = arg minp∈∆n DF (p, wt+1). the reader to [9]. We will use the following result: In general, the OSMD algorithm is defined with respect to a compact, convex set K ⊆ Rn. The updates are characterized by noisy estimates of the gradient of the loss function, which we may conveniently define to be bℓt in the present scenario. For more details and generalizations, we refer 1. The strategy S corresponding to the OSMD algorithm with loss estimates bℓ, learning rate η > 0, Proposition 1 (Theorem 5.10 of [9]). Let the loss functions {ℓi,t} be nonnegative and bounded by and Legendre function Fψ, where ψ is a 0-potential, satisfies the pseudo-regret bound E" i,t (ψ−1)′(pi,t)# . bℓ2 sup A∈A RT (A,S) ≤ supp∈∆n Fψ(p) − Fψ(p1) η + η 2 TXt=1 nXi=1 We formally define 0-potentials and the associated Legendre functions in Appendix C. In our analysis, we take ψ(x) = 1 . The pseudo- regret bound in Proposition 1 may then be analyzed and bounded accordingly in various settings of interest. Details for the proof of Theorem 1 are also provided in Appendix C. x2 , yielding the Legendre function Fψ(x) = −2Pn i=1 x1/2 i B.2 Lower bounds for adversarial models. We now turn to establishing the lower bounds. The proofs of Theorems 2 and 5 are based on the same general strategy, which is summarized in the following proposition. To unify our results with standard bandit notation ([9]), we use the shorthand to denote the reward incurred at time t when the player chooses St = {i}. Then Xi,t = f (At,{i}) RT (A,S) = max 1≤i≤n EA,S TXt=1 (Xi,t − XSt,t). Proposition 2. Consider a deterministic player strategy S ∈ Pd. Let A0,A1, . . . ,An be stochastic adversarial strategies such that for each Ai, the set of edges played at time t is independent of the 13 past actions of the adversary. Let P0, P1, . . . , Pn denote the corresponding measures on the feedback I T , allowing for possible randomization only in the strategy of the adversary. Let Ei denote the expectation with respect to Pi. Suppose r ≤ min j6=i Ei [Xi,t − Xj,t] , ∀1 ≤ t ≤ T, (5) (6) (7) (8) and Then nXi=1 2n! ≤ n −r D rT n − 1 rT n − 1 n −r D KL (P0, Pi) ≤ D. TXt=1 Ei 1 n nXi=1 2n! ≤ inf S∈P (Xi,t − XSt,t). sup A∈A RT (A,S). In particular, if the bounds (5) and (6) hold uniformly for all choices of S ∈ Pd, then Remark 4. We remark briefly about the roles of the strategies {Ai} appearing in Proposition 2. In practice, the strategies are chosen to be similar, except selecting i as the source node is slightly more advantageous when the adversary uses strategy Ai. The strategy A0 is a baseline strategy that treats all nodes identically. Thus, the lower bound provided by Proposition 2 is the product of the cost of an incorrect choice of the source vertex, given by r, and a factor that determines how easy it is to distinguish the adversary strategies from each other, which depends on D. Proof. Proof of Proposition 2. We follow the method used in the proof of Theorem 3.5 in [9]. We first show how to obtain the bound (8) from the set of uniform bounds (7). Note that for any S ∈ P, we have sup A∈A RT (A,S) = sup A∈A max 1≤i≤n EA,S (Xi,t − XSt,t) TXt=1 TXt=1 TXt=1 Aj ≥ max 1≤j≤n max 1≤i≤n = max 1≤j≤n max 1≤i≤n ES E (Xi,t − XSt,t) ES Ej (Xi,t − XSt,t) ≥ max 1≤i≤n ≥ ES" 1 n ES Ei nXi=1 (Xi,t − XSt,t) TXt=1 (Xi,t − XSt,t)# , TXt=1 Ei where we have used the fact that the maximum is at least as large as the average in the final inequality. Since any player strategy in P lies in the convex hull of deterministic player strategies, a uniform bound (7) over Pd implies that inequality (8) holds, as well. 14 We now turn to the proof of inequality (7). The idea is to show that on average, the player incurs a certain loss whenever the wrong source vertex is played, and this event must happen sufficiently often. We first write Ei TXt=1 (Xi,t − XSt,t) = TXt=1 =Xj6=i EiXj6=i (Xi,t − Xj,t)111{St={j}} TXt=1 Ei [Xi,t − Xj,t] Ei(cid:2)111St={j}(cid:3) . In the last equality, we have used the assumption that the adversary's action at each time is independent of the past to conclude that the difference in rewards Xi,t− Xj,t (which depends on the adversary's action at time t) is independent of the indicator 111St={j} (which depends on the sequence of feedback received up to time t − 1). Using the bound (5), it follows that Ei TXt=1 (Xi,t − XSt,t) =Xj6=i rEi[Tj], where Ti = {t : St = {i}} denotes the number of times vertex i is selected as the source. qi,T = Ti Now let UT denote a vertex drawn according to the distribution qT = (q1,T , . . . , qn,T ), where T . The derivations above imply that Pi {UT = j} = rT (1 − Pi {UT = i}) , Ei so TXt=1 (Xi,t − XSt,t) = rTXj6=i nXi=1 1 n Ei (Xi,t − XSt,t) = rT 1 − TXt=1 Pi {UT = i} ≤ P0 {UT = i} +r 1 1 n 2 By Pinsker's inequality, we have Pi {UT = i}! . nXi=1 KL (P′0, P′i), (9) where P′i denotes the distribution of UT under the adversarial strategy Ai. By Jensen's inequality, we therefore have Finally, the chain rule for KL divergence implies that 1 n + 1 n 1 n Pi {UT = i} ≤ nXi=1r 1 nXi=1 KL(cid:0)P′0, P′i(cid:1) = KL (P0, Pi) +XI T 2 KL (P′0, P′i) ≤ 1 n +vuut 1 2n KL (P′0, P′i). (10) nXi=1 P0(cid:8)I T(cid:9) KL(cid:0)P′0{·I T}, P′i{·I T}(cid:1) . (11) Note that conditional on I T , the distribution of UT is the same under P′0 and P′i, since the player uses a deterministic strategy. Thus, equation (11) implies that Combining inequalities (9), (10), and (12), we arrive at the desired result (7). nXi=1 KL(cid:0)P′0, P′i(cid:1) = nXi=1 KL (P0, Pi) ≤ D. (12) To prove Theorems 2 and 5, it thus remains to find an appropriate set of strategies {A0,A1, . . . ,An} and verify the bounds (5) and (6). Details for the proofs are provided in Appendix E.1. 15 C Additional online upper bound proofs In this Appendix, we provide proofs for the pseudo-regret of player strategies based on the OSMD algorithm. We begin with some preliminaries. C.1 Preliminaries We first describe the function Fψ. Recall that a continuous function F : D → R is a Legendre function if F is strictly convex, F has continuous first partial derivatives on D, and lim x→D\D k∇F (x)k = ∞. The analysis in this paper concerns a very specific type of Legendre function associated to a 0- potential, as described in the following definition: Definition 1. A function ψ : (−∞, a) → R+ is called a 0-potential if it is convex, continuously differentiable, and satisfies the following conditions: We additionally define the associated function Fψ on (0,∞)n by ψ(x) = 0, lim x→−∞ ψ′ > 0, Fψ(x) = ψ(x) = ∞, lim x→a Z 1 0 ψ−1(s)ds ≤ ∞. nXi=1Z xi ψ−1(s)ds. 0 In particular, we will consider the 0-potential ψ(x) = (−x)−q. Then ψ−1(x) = −x− 1 q , so Fψ(x) = − q q − 1 nXi=1 q−1 q x i . Specifically, we will consider the case q = 2 (the same analysis could be performed with respect to q > 1, and then the final bound could be optimized over q). To employ Proposition 1, we need to bound two summands. The following simple lemma bounds the first term: Lemma 1. When ψ(x) = 1 x2 , we have the bound Fψ(p) − Fψ(p1) ≤ 2√n, ∀p ∈ ∆n. Proof. Proof. Since Fψ(p) ≤ 0 and kp1k1 = 1, Hölder's inequality implies that Fψ(p) − Fψ(p1) ≤ 2 nXi=1 p1/2 1,i ≤ 2n 1 2 . This completes the proof of the lemma. All that remains is to analyze the loss-specific term appearing in Proposition 1 and choose η appropriately. 16 C.2 Proof of Theorem 1 We first prove the following lemma: Lemma 2. We have the inequality i,t )2 E" (bℓsym (ψ−1)′(pi,t)# ≤ nXi=1 √2n, ∀1 ≤ t ≤ T. (13) Proof. Proof. Let Ft denote the sigma-field of all actions up to time t. We have nXi=1 i,t )2 E" (bℓsym (ψ−1)′(pi,t)(cid:12)(cid:12)(cid:12)(cid:12)Ft−1# (a) (b) = 2 p3/2 i,t i,t )2Ft−1i Eh(bℓsym nXi=1 pi,t!1/2 nXi=1(cid:16)pi,tEh(bℓsym ≤ 2 nXi=1 i,t )2Ft−1i(cid:17)2!1/2 = 2 nXi=1(cid:16)pi,tEh(bℓsym i,t )2Ft−1i(cid:17)2!1/2 , (14) where we have used the facts that (ψ−1)′(x) = 1 establish (a), and applied Hölder's inequality to obtain (b). 2 x−3/2 and pt is measurable with respect to Ft−1 to We now inspect the conditional expectation more closely. We have = 1 n2 i,t )2Ft−1i = E 1 Eh(bℓsym nXj6=i EXj6=iXk6=i EXj6=iXk6=i EXj6=i 1 n2 1 n2 + = 1 pi,t + pj,t ℓt i,jZij 2(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 , 1 (pi,t + pj,t)(pi,t + pk,t) 1 (pi,t + pj,t)(pi,t + pk,t) 1 (pi,t + pj,t)2 (ℓt i,j)2Zj(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 , i,jℓt ℓt ℓt i,jℓt i,kZijZik(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 i,kZi(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 where the third equality is due to the fact that ZijZik is 1 only when i is the source vertex or j = k 17 is the source vertex. Using the fact that ℓt i,j is bounded by 1, we then obtain Zi (pi,t + pj,t)(pi,t + pk,t)(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 (pi,t + pj,t)2(cid:12)(cid:12)(cid:12)(cid:12)Ft−1 n2Xj6=i pi,t Zj + 1 (pi,t + pj,t)(pi,t + pk,t) pi,t + pj,t (pi,t + pj,t)(pi,t + pk,t) i,t )2Ft−1i ≤ Eh(bℓsym ≤ ≤ = ≤ 1 + 1 n2 1 n2 EXj6=iXk6=i EXj6=i n2Xj6=iXk6=i n2Xj6=iXk6=i n2Xj6=iXk6=i nXk=1 1 n 1 1 1 pi,t + pk,t 1 . pi,t + pk,t pj,t (pi,t + pj,t)2 Combining this result with the bound (14), we have nXi=1 1 2 2 n = 1/2 p2 i,t i,t )2 (ψ−1)′(pi,t)(cid:12)(cid:12)(cid:12)(cid:12)Ft−1# ≤ 2 pi,t + pk,t!2 nXi=1 pi,t E" (bℓsym nXk=1 (pi,t + pj,t)(pi,t + pk,t) n nXi=1 nXj=1 nXk=1 pi,t + pj,t nn nXj=1 nXi=1 nXk=1 nXi=1 nXi=1 +Xk6=i nXk=1 nXi=1 ai + ak ai + ak ai + ak ai + ai n2 2 pi,t ≤ ai ai ai = = 1 2 . , + 1 2Xk6=i ak ai + ak 1 2 (15) Now, we have the useful equation for any nonnegative sequence {ai}n ai i=1. This may be seen via the following algebraic manipulations: n 2 n 2 + + 1 2Xk6=i 2Xk6=i 1 ai ai + ak ai + ak ai + ak n(n − 1) + 2 = = = = n 2 n2 2 . 18 Appealing to equation (15), we may replace the double sum by n2 2 and simplify the bound: nXi=1 i,t )2 E" (bℓsym (ψ−1)′(pi,t)(cid:12)(cid:12)(cid:12)(cid:12)Ft−1# ≤ 2 2 n(cid:18) n3 2 (cid:19) 1 = √2n. Taking an additional expectation and using the tower property, we arrive at the desired inequality. Combining Lemmas 1 and 2 with Proposition 1, we then have RT (A,S sym OSMD) ≤ sup A∈A 2√n η + ηTr n 2 . Optimizing over η, we take η = 2 3 4 T − 1 2 , which establishes the desired bound. D Adversarial influence maximization with multiple sources In this Appendix, we prove results concerning multiple influence sources. First, we need to give the precise algorithmic details of the online greedy algorithm. We assume the player is allowed to choose source vertices sequentially at time t and observes the corresponding edge feedback immediately after each selection. The algorithm, inspired by [36], is outlined below: Online Greedy Algorithm Given: A single-source player strategy S 1. Output: A k-source player strategy S k = {St}1≤t≤T . For each t = 1, . . . , T , choose St = {v1,t, . . . , vk,t} sequentially, as follows: (1) Select v1,t according to the single-source strategy S 1. (2) For each i > 1, select vi,t according to the single-source strategy S 1, based on the edge feedback I (At,{v1,t, . . . , vi,t}) \ I (At,{v1,t, . . . , vi−1,t}). In other words, the Online Greedy Algorithm runs the player's strategy for single-source selection k times in parallel, with losses computed marginally for each successively chosen vertex. The "greedy" component of the algorithm corresponds to the fact that the player makes a selection of the set of ith source vertices in the best possible way based on the information available (i.e., according to the single-source strategy that is designed to incur a small pseudo-regret). Note that the feedback I (At,{v1,t, . . . , vi,t}) \ I (At,{v1,t, . . . , vi−1,t}) is indeed computable by the player when choosing the ith vertex at round t, since the player has already observed I (At,{v1,t, . . . , vi−1,t}) after the first i − 1 source nodes are selected. t=1 ft(St). Thus, F (S) is the total reward for strategy S = {St}. In the stochastic setting, when T = 1, many Fix an adversarial strategy A, and define the functions ft(St) = f (At,St) and F (S) =PT 19 influence-maximization analyses exploit the submodularity of ft under certain stochastic assump- In the bandit setting, we wish to establish an analogous result for F , in order to tions on At. establish regret bounds when the player chooses source vertices according to a greedy algorithm. Since S ∈ (2V )T , the function F is not technically a set function. However, we may identify each player strategy S with an element of S∗ ∈ 2(V T ), and define F ∗(S∗) = F (S). Here, V T :=(cid:8)vT =(cid:0)v(1), v(2), . . . , v(T )(cid:1) v(i) ∈ V, for 1 ≤ i ≤ T(cid:9) , k } ∈ 2(V T ) corresponds to the strategy that selects the source nodes {u1(t), . . . , uk(t)} and S∗ = {uT in round t. 1 , . . . , uT In more detail, let S(i) t = {st(1), . . . , st(i)} ⊆ V denote the set of the first i seed vertices in round t, where S(0) t = ∅. Then, we can write ft(St) = kXi=1 One can then write the total reward as ft(cid:16)S(i) t (cid:17) − ft(cid:16)S(i−1) t (cid:17) . F (S) = If we define ft(cid:16)S(i) t (cid:17) − ft(cid:16)S(i−1) t (cid:17) t (cid:17) − ft(cid:16)S(i−1) t (cid:17) = kXi=1 TXt=1 i }) − F ∗({uT ft(cid:16)S(i) TXt=1 kXi=1 f∗i (S∗) = F ∗({uT TXt=1 = 1 , . . . , uT 1 , . . . , uT i−1}) ft ({uj(t) : j ≤ i}) − ft ({uj(t) : j ≤ i − 1}) and F ∗(S∗) = Pk i=1 f∗i (S∗), then we indeed get the desired equality F (S) = F ∗(S∗) while also switching our summation for submodularity to be over the ith vertices as opposed to the tth round. We first show that F ∗ is a monotone, submodular function: Lemma 3. The function ft(St) = f (At,St) is monotone and submodular, for every fixed At. Proof. Proof. It is trivial to see that ft is monotone, so we focus on proving submodularity. Our goal is to show that for a fixed At, and for any St ⊆ S′t and u ∈ V \S′t, we have ft(S′t ∪ {u}) − ft(S′t) ≤ ft(St ∪ {u}) − ft(St). (16) ZS ′ t∪{u},v − ZS ′ Let ZSt,v denote the indicator of an open path between a source node s ∈ St and v ∈ V , where by convention, ZSt,v = 1 if v ∈ St. Note that ft(St) = 1 we have nPv∈V ZSt,v. We will show that for v /∈ S′t∪{u}, (17) Summing over v ∈ (S′t ∪ {u})c, using ZSt∪{u},v − ZSt,v ≥ 0 for v ∈ S′t \ St, and dividing by n will yield the desired inequality (16). We have three cases to consider: In the first case, an open path exists from some s ∈ S′t to v. Then the left side of inequality (17) is equal to 0, while the right hand side is at least 0 by monotonicity. In the second case, an open path does not exist from any s ∈ S′t to v, but an open path exists from u to v. Then both sides of inequality (17) are equal to 1. Finally, if no open path exists from s ∈ S′t ∪ {u} to v, then both sides of inequality (17) are equal to 0. This completes the proof. t,v ≤ ZSt∪{u},v − ZSt,v. 20 Proposition 3. The function F ∗ is monotone and submodular. Proof. Proof. The properties are essentially immediate from Lemma 3. Let P and Q be elements of(cid:0)2V(cid:1)T such that P∗ ⊆ Q∗. Then F ∗(P∗) = TXt=1 ft(Pt) ≤ TXt=1 ft(Qt) = F ∗(Q∗), proving monotonicity. Similarly, if S ∈ (2V )T , we have F ∗(S∗ ∪ Q∗) − F ∗(Q∗) = (ft(St ∪ Qt) − ft(Qt)) TXt=1 TXt=1 (ft(St ∪ Pt) − ft(Pt)) ≤ = F ∗(S∗ ∪ P∗) − F ∗(P∗), proving submodularity. By the standard greedy approximation ([24, 30]), we then have (cid:18)1 − 1 e(cid:19) max S ∗≤K F ∗(S∗) ≤ F ∗(G∗), where G∗ is a set of cardinality K ≥ 1 constructed via a sequential greedy algorithm. However, this result is not immediately applicable to the online bandit setting, since we do not have direct access to F ∗. Thus, we can only hope to obtain an approximate greedy maximizer eG∗, and we wish to derive theoretical guarantees for F ∗(eG∗). Proposition 4 (Theorem 6 from [36]). Let f : 2V → R be a monotone, submodular function such that f (∅) = 0. Consider a set D ⊆ V and a sequence of error tolerances {ǫi}, and suppose {Gǫ i} is constructed in an approximate greedy manner, such that Gǫ i−1 ∪ {gi}, where Our result relies on the following general proposition: i = Gǫ 0 = ∅ and Gǫ i−1 ∪ {gi}) − f (Gǫ i−1) + ǫi. f (Gǫ max d∈D i−1 ∪ {d}) − f (Gǫ i−1) ≤ f (Gǫ Then for any K ≥ 1, we have (cid:18)1 − 1 e(cid:19) max S ∗∈DK f (S∗) − f (Gǫ K) ≤ ǫi, KXi=1 where DK consists of subsets of D containing at most K elements. Proposition 4 ensures that for submodular functions, successive errors {ǫi} in a sequential greedy algorithm only accumulate additively. The proof is provided in [36], but we include a proof in Appendix D.2 for completeness. 21 D.1 Proof of Theorem 3 Suppose A ∈ A . We will apply Proposition 4 with f = EAF ∗, V = V T , and K = k. Note that EAF ∗ inherits monotonicity and submodularity from F ∗. Also let D = {(v, . . . , v) : v ∈ V } ⊆ 2(V T ) denote the diagonal set of 2(V T ). For a (non-random) k-source strategy S∗ with S∗t = k for all t, we use the notation S∗ = {S∗1 , . . . ,S∗k}, where S∗i corresponds to the set of ith vertices chosen during the T rounds. Proposition 4 immediately gives (cid:18)1 − 1 S ∗∈D e(cid:19) max EAF ∗(S∗) − EAF ∗(Gǫ k) kXi=1 EA[F ∗(Gǫ max di∈D ≤ i−1 ∪ {di}) − F ∗(Gǫ i−1 ∪ {gi})], i} are chosen in an approximate greedy manner, and ǫi are upper bounded by the i} to be the We now take an expectation with respect to possible randomization in the player's strategy, to where the sets {Gǫ regret for the ith instance of the single-source algorithm. In particular, we consider {Gǫ choice of ith vertices S∗i corresponding to the player's choice under the strategy S 1. obtain R (1−1/e) T (A,S) ≤ (a) = (b) = di∈D EA(cid:2)F ∗(Gǫ EA(cid:2)F ∗(Gǫ ES(cid:20)max kXi=1 ES[1:i](cid:20)max kXi=1 ES[1:i−1](cid:20)ESi max kXi=1 di∈D di∈D EA(cid:2)F ∗(Gǫ i−1 ∪ {di}) − F ∗(Gǫ i−1 ∪ {gi})(cid:3)(cid:21) . i−1 ∪ {di}) − F ∗(Gǫ i−1 ∪ {di}) − F ∗(Gǫ i−1 ∪ {gi})(cid:3)(cid:21) i−1 ∪ {gi})(cid:3)(cid:21) Here, ES[1:i] denotes the expectation with respect to the first i vertices played, and the equality in (a) holds because the set of ith vertices played depends only on the sets of the first i vertices played. The equality in (b) holds because the set Gi−1, and hence the choice of di, does not depend on the selection of ith vertices. Furthermore, the inner expression is simply the pseudo-regret of strategy S 1. By Theorem 1, this is bounded by 2 4√T n. Summing up, we obtain the desired result. 1 D.2 Proof of Proposition 4 We begin with two supporting lemmas: Lemma 4. For any P ⊆ V and Q ⊆ D, we have f (P ∪ Q) ≤ f (P) + Q max v∈D [f (P ∪ {v}) − f (P)]. Proof. Proof. We proceed by induction on Q. The case Q = 1 is immediate. Now suppose the statement is true for all Q ≤ k, where k ≥ 1. Let c ∈ D, and suppose Q ⊆ D has cardinality k. 22 Then f (P ∪ (Q ∪ {c})) (a) (b) ≤ f (P ∪ {c}) + Q max d∈D [f ((P ∪ {c}) ∪ {d}) − f (P ∪ {c})] ≤ f (P) + max d∈D = f (P) + Q ∪ {c} max d∈D [f (P ∪ {d}) − f (P)] + Q max d∈D [f (P ∪ {d}) − f (P)], [f (P ∪ {d}) − f (P)] where (a) follows from the induction hypothesis and (b) follows from the induction hypothesis and submodularity. This completes the induction and proves the lemma. Lemma 5. Let δi := f (Gǫ i) − f (Gǫ i−1). For any Q ⊆ D, we have f (Q) ≤ f (Gǫ i−1) + Q(δi + ǫi). Proof. Proof. Using Lemma 4 and monotonicity of f , we have f (Q) ≤ f (Gǫ ≤ f (Gǫ ≤ f (Gǫ = f (Gǫ i−1 ∪ Q) i−1) + Q max d∈D i−1) + Q(cid:0)f (Gǫ i−1) + Q(δi + ǫi), [f (Gǫ i) − f (Gǫ i−1 ∪ {d}) − f (Gǫ i−1) + ǫi(cid:1) i−1)] completing the proof. We now define ∆i := maxS ∗∈DK f (S∗) − f (Gǫ f (S∗) ≤ f (Gǫ i−1). By Lemma 5, we have i−1) + K(δi + ǫi). max S ∗∈DK Subtracting f (Gǫ i−1), we obtain so ∆i ≤ K(δi + ǫi) = K(∆i − ∆i+1 + ǫi), ∆i+1 ≤ ∆i(cid:18)1 − 1 K(cid:19) + ǫi. Applying this inequality recursively, we see that ∆K+1 ≤ ∆1 KYi=1(cid:18)1 − 1 K(cid:19) + KXi=1 ǫi = ∆1(cid:18)1 − 1 K(cid:19)K + KXi=1 ǫi ≤ ∆1(cid:18) 1 e(cid:19) + ǫi. KXi=1 Rearranging and using the fact that f (∅) = 0 completes the proof. E Additional online lower bound proofs The main goal of this Appendix is to prove Theorems 2 and 5. Some of the computations are rather lengthy and are therefore included in Appendix E.2. 23 E.1 Proofs of theorems We first present the main components of the proofs, followed by detailed calculations involving the Kullback-Leibler divergence. E.1.1 Proof of Theorem 2 Let the adversarial strategies {Ai} be defined as follows: For each strategy, the adversary chooses a random subset of vertices, and opens all edges between vertices in the subset. For Ai, with 1 ≤ i ≤ n, the adversary includes vertex i with probability c n, and includes all other vertices with probability c n (1 − δ) each, where δ ∈ (0, 1/2) is a small constant. Finally, for A0, the adversary includes all vertices independently with probability c n (1 − δ). Successive actions of the adversary are i.i.d. across time steps. We now derive the following lemmas, which will be used in Proposition 2: Lemma 6. For any i 6= j and 1 ≤ t ≤ T , we have Ei[Xi,t − Xj,t] = (n − 2)c2 n3 (1 − δ)δ. Proof. Proof. Let Ct be the clique chosen by the adversary at time t. Note that if i, j ∈ Ct or i, j /∈ Ct, the difference in rewards is 0. Thus, the only cases of interest in computing the expectation are when exactly one of i or j is in Ct. Then 1 = Ei [Xi,t − Xj,t] = Ei(cid:2)(Ct − 1)111i∈Ct111j /∈Ct − (1 − Ct)111i /∈Ct111j∈Ct(cid:3) (1 − δ)i nih c (1 − δ)i (1 − δ)i −h1 − n(cid:16) c n(cid:17)h1 − (n − 2)(1 − δ)(cid:17)(cid:16) c n(cid:16) c (n − 2)(1 − δ)(cid:17)h1 − n(cid:17)2 (1 − δ)(cid:16)h1 − (n − 2)(cid:16) c 1 n (n − 2)c2 c n c n 1 c n − = n n c ni (1 − δ)(cid:17) = n3 (1 − δ)δ, where the second equality uses the fact that c Ct. Lemma 7. Let S ∈ Pd be a deterministic player strategy, and let Ti = {t : St = {i}}. Then we have the upper bound n (n − 2)(1 − δ) other vertices are expected to be in nXi=1 KL (P0, Pi) ≤ c(c + 1) n − c T δ2. The proof of Lemma 7 is provided in Appendix E.2.1. Thus, by Proposition 2, we have inf S∈P sup A∈A RT (A,S) ≥ T n3 (n − 2)c2 n(cid:17)2 n − 1 6 (cid:16) c (1 − δ)δ n − 1 δ − δ2r T n − δr T 2nr c n − c ! , 2nr c(c + 1) n T n − c ≥ (c + 1)! 24 where the second inequality uses the fact that n ≥ 3 and δ < 1/2. Finally, we optimize over δ and c. Since we have a quadratic equation in δ, we take yielding δ = inf S∈P sup A∈A RT (A,S) ≥ = ≥ , T c(c + 1) n − 1 2n r 2n n(cid:17)2 1 6 (cid:16) c √T(cid:16) c √T(cid:16) c n ≥ 2 T r n − c 4(cid:18) n − 1 n(cid:17)2(cid:18) n − 1 n2(cid:17)pn(n − c), 1 12√2 1 27√3 c(c + 1)! n (cid:19)2r 2n T r n − c n (cid:19)2s n(n − c) c(c + 1) c+1 ≥ 2 3 when n ≥ 3, and c 3 when c ≥ 2. The 3 , yielding the desired lower bound. Note that for this choice where the second inequality uses the bounds n−1 final expression is optimized at c = 2n of c, we indeed have δ < 1/2 when T ≥ 2. E.1.2 Proof of Theorem 5 Let the adversarial strategies {Ai} be defined as follows: For each strategy, the adversary indepen- dently designates every vertex to be a source, sink, or neither. The adversary then opens directed edges from all source vertices to all sink vertices. For Ai, with 1 ≤ i ≤ n, the adversary designates vertex i to be a source vertex with probability c n , and all other vertices to be source vertices with probability c n . Finally, for A0, the adversary designates all vertices to be source vertices with probability c n (1 − δ), and sink vertices with probability d n (1 − δ). All vertices are designated to be sink vertices with probability d n. Successive actions of the adversary are i.i.d. across time steps. We now derive the following lemmas, which will be used in Proposition 2: Lemma 8. For any i 6= j and 1 ≤ t ≤ T , we have Ei[Xi,t − Xj,t] = (n − 1)cd n3 δ. Proof. Proof. We compute the expectation of each term separately. Let Bt and Ct denote the n if i /∈ Bt; otherwise, Xi,t = 1+Ctn source and sink vertices at time t, respectively. Note that Xi,t = 1 . Hence, nEi[Xi,t] = E [111i /∈Bt + (1 + Ct)111i∈Bt] d c n(cid:17) +(cid:18)1 + (n − 1) n(cid:19)(cid:16) c n(cid:17) =(cid:16)1 − = 1 + (n − 1)cd . n2 The computation for Xj,t is similar: nEi[Xj,t] = E(cid:2)111j /∈Bt + (1 + Ct)111j∈Bt(cid:3) =(cid:16)1 − = 1 + (1 − δ)(cid:17) +(cid:18)1 + (n − 1) c n (n − 1)cd d n(cid:19) c n (1 − δ) n2 (1 − δ). 25 Taking the difference between these expectations proves the lemma. Lemma 9. Let S ∈ Pd be a deterministic player strategy, and let Ti = {t : St = {i}}. Then we have the upper bound nXi=1 KL (P0, Pi) ≤ c(n − d) n(n − c − d) T δ2. Essentially, the Kullback-Leibler divergence is of order 1 n , because playing a suboptimal vertex provides no information about which vertex is optimal. This is unlike the case of the undirected graph, where the optimal vertex is always more likely to be contained in the feedback that the player receives, and the KL divergence does not decay with n. The proof of Lemma 9 is provided in Appendix E.2.2. By Proposition 2, we then have inf S∈P sup A∈A RT (A,S) ≥ (n − 1)cd n3 Finally, we optimize over δ, c, and d. We take δT n − 1 n − δr T 2ns c(n − d) n(n − c − d)! . δ = to obtain 1 2(cid:18) n − 1 n (cid:19)r 2n T s n(n − c − d) c(n − d) , inf S∈P sup A∈A RT (A,S) ≥ = ≥ (n − 1)cd 4n3 T s n(n − c − d) Tr 2n n2s (1 − c/n − d/n) (cid:18)n − 1 n (cid:19)2 √nT(cid:18) n − 1 n (cid:19)3 cd n2s (1 − c/n − d/n) √nT (c/n) (1 − d/n) c(n − d) (c/n) (1 − d/n) cd , 1 2√2 1 16√2 where the last inequality uses the bound n−1 2. Finally, using the fact that the function n ≥ 1 f (x, y) = xys 1 − x − y when (x, y) =(cid:0) 1 x(1 − y) 6 , 2 3(cid:1), we obtain the bound √T n, 1 48√6 achieves its maximum value of 1 3√3 when c = n 6 and d = 2n 3 . inf S∈P sup A∈A RT (A,S) ≥ 26 E.2 Proofs of KL bounds In this Appendix, we derive the required upper bounds on the KL divergence between adversarial strategies. We begin by proving a useful technical lemma. Recall that Pi denotes the distribution of the edge feedback I T under strategy Ai, and S ∈ Pd is a fixed deterministic player strategy. Also recall that Ti = {t : St = {i}} denotes the number of times vertex i is chosen by the player. i . For i(v) to be the KL divergence i denote the distribution of the edge feedback I t under strategy Ai, so Pi = PT Let Pt each pair of nodes i and v and any 1 ≤ t ≤ T , define the function KLt between the edge feedback, conditioned on any I t−1 such that St = {v}: i{·I t−1}(cid:1) . i(v) = KL(cid:0)Pt 0{·I t−1}, Pt KLt i(v) is indeed a well-defined function of v, since conditioned on I t−1, the player's Note that KLt action St is deterministic. Hence, the randomness in I t is purely due to the stochastic action of the adversary at time t. Lemma 10. If KLt i(v) is independent of t, we have KL (P0, Pi) = KLi(i)E0[Ti] +Xj6=i KLi(j)E0[Tj]. (18) If in addition KLi(i) is independent of i, for 1 ≤ i ≤ n, and KLi(j) is constant for all nonzero pairs i 6= j, we have KL (P0, Pi) = KLi(i)T + KLi(j)(n − 1)T. (19) Proof. Proof. Note that equation (19) follows immediately from equation (18) by summing over i To derive equation (18), we use the chain rule for KL divergence: nXi=1 and using the fact thatPn i=1 KL (P0, Pi) = P0(cid:8)I t−1(cid:9) KL(cid:0)Pt 0(cid:8)·I t−1(cid:9) , Pt i(cid:8)·I t−1(cid:9)(cid:1) P0{I t−1}KLt i(v) P0{St = {v}}KLi(v) = P0{St = {i}}KLi(i) + TXt=1Xj6=i P0{St = {j}}KLi(j), E0[Ti] = T . = (a) = TXt=1 XI t−1 TXt=1 nXv=1 XI t−1:St={v} TXt=1 nXv=1 TXt=1 E0[Ti] = E0" TXt=1 111St={i}# = TXt=1 P0{St = {i}} 27 using the assumption that KLt that i(v) is independent of t in the equation (a). Now we simply recognize to obtain the desired equality. E.2.1 Proof of Lemma 7 Note that KLt i(v) is independent of t, since the adversary's actions are i.i.d. across time steps. Furthermore, KLi(i) is clearly independent of i and KLi(j) is constant for all pairs i 6= j, so equation (19) of Lemma 10 holds. We first compute an upper bound for KLi(i). Let X denote the size of the connected component containing i on a particular time step, based on the edges played by the adversary. Then KLi(i) = KL(P0(X), Pi(X)), where we abuse notation slightly and write Pi(X) to denote the distribution of X under adversarial strategy Ai. Also let Y be the indicator variable that i is in the clique selected by the adversary. By the chain rule for the KL divergence, KL(P0(X), Pi(X)) ≤ KL(P0(X, Y ), Pi(X, Y )). We will derive an upper bound for the latter quantity. In particular, the range of (X, Y ) is {(1, 0), (1, 1)} ∪ {(m, 1) : 2 ≤ m ≤ n}. This leads to the following expression for KL(P0(X, Y ), Pi(X, Y )): (cid:16)1 − c n n + c n (cid:19) n (1 − δ) 1 − c c n (1 − δ)(cid:17) log(cid:18) 1 − c (1 − δ)(cid:17)n−1(cid:19) log c n (1 − δ)(cid:0)1 − c +(cid:18) c (1 − δ)(cid:16)1 − n(cid:0)1 − c nXm=2(cid:18) n − 1 m − 1(cid:19) c (1 − δ)(cid:17)m−1(cid:16)1 − (1 − δ)(cid:16) c × log c n (1 − δ)(cid:1)n−m n (1 − δ)(cid:1)m−1(cid:0)1 − c n (1 − δ)(cid:0) c n (1 − δ)(cid:1)n−m n (1 − δ)(cid:1)m−1(cid:0)1 − c n(cid:0) c (cid:19) (1 − δ)(cid:17) log(cid:18) 1 − c n (1 − δ) 1 − c nXm=1(cid:18) n − 1 m − 1(cid:19) c (1 − δ)(cid:17)m−1(cid:16)1 − (1 − δ)(cid:16) c (1 − δ)(cid:17) log(cid:18) 1 − c (cid:19) + n (1 − δ) 1 − c c n c n c n c n c n + n n n n n n c (1 − δ) log (1 − δ) . =(cid:16)1 − =(cid:16)1 − ! n (1 − δ)(cid:1)n−1 n (1 − δ)(cid:1)n−1 (1 − δ)(cid:17)n−m ! (1 − δ)(cid:17)n−m log(1 − δ) Applying the inequality log(1 + x) ≤ x twice, we then obtain KL(P0(X), Pi(X)) ≤(cid:16)1 − n c n (1 − δ)(cid:17) cδ n − 1 − c − (1 − δ)(cid:19) n (cid:18) n − c(1 − δ) n − c c n cδ (1 − δ)δ . (20) = = cδ2 n − c 28 The computation for KLi(j) is similar. Let X denote the size of the connected component containing j, and let C denote the clique chosen by the adversary. Define the random variable Y = 0, 1, 2, if j /∈ C if j ∈ C and i /∈ C if i, j ∈ C. Again, it suffices to obtain a bound on KL(P0(X, Y ), Pi(X, Y )). The range of (X, Y ) is {(1, 0), (1, 1)} ∪ {(m, 1) : 2 ≤ m ≤ n − 1} ∪ {(m, 2) : 2 ≤ m ≤ n}. Further note that P0(1, 0) = Pi(1, 0), so we may ignore this term when computing the KL divergence. We then have following expression for KL(P0(X, Y ), Pi(X, Y )): n c n (1 − δ)(cid:17)n−m−1 (cid:16) c n = = n n c n c n c + + n c c n n n (1 − δ)(cid:1)n−2! log (cid:0) c n (1 − δ)(cid:1)(cid:0)1 − c n (1 − δ)(cid:1)n−1 (1 − δ)(cid:17)(cid:16)1 − (1 − δ)(cid:17)n−1 (cid:0) c n(cid:1)(cid:0)1 − c n (1 − δ)(cid:1)(cid:0)1 − c n−1Xm=2(cid:18) n − 2 m − 1(cid:19) c (1 − δ)(cid:17)m−1(cid:16)1 − (1 − δ)(cid:17)(cid:16) c (1 − δ)(cid:16)1 − ! × log c n (1 − δ)(cid:1)n−m−1 n (1 − δ)(cid:1)m−1(cid:0)1 − c n (1 − δ)(cid:1)(cid:0) c n (1 − δ)(cid:0)1 − c n (1 − δ)(cid:1)m−1(cid:0)1 − c n(cid:1)(cid:0) c n (1 − δ)(cid:0)1 − c n (1 − δ)(cid:1)n−m−1 nXm=2(cid:18) n − 2 m − 2(cid:19)(cid:16) c (1 − δ)(cid:17)2(cid:16) c (1 − δ)(cid:17)m−2(cid:16)1 − (1 − δ)(cid:17)n−m n (1 − δ)(cid:1)n−m! × log (cid:0) c n (1 − δ)(cid:1)n−m n (1 − δ)(cid:1)m−2(cid:0)1 − c n (1 − δ)(cid:1)2(cid:0) c n(cid:1)(cid:0) c (cid:0) c n (1 − δ)(cid:1)m−2(cid:0)1 − c n (1 − δ)(cid:1)(cid:0) c n−1Xm=1(cid:18) n − 2 m − 1(cid:19) c (1 − δ)(cid:17)m−1(cid:16)1 − (1 − δ)(cid:17)(cid:16) c (1 − δ)(cid:16)1 − × log c n (1 − δ)(cid:1)m−1(cid:0)1 − c n (1 − δ)(cid:1)(cid:0) c n (1 − δ)(cid:0)1 − c n (1 − δ)(cid:1)n−m−1 n (1 − δ)(cid:0)1 − c n(cid:1)(cid:0) c n (1 − δ)(cid:1)m−1(cid:0)1 − c n (1 − δ)(cid:1)n−m−1 nXm=2(cid:18) n − 2 m − 2(cid:19)(cid:16) c (1 − δ)(cid:17)2(cid:16) c (1 − δ)(cid:17)m−2(cid:16)1 − (1 − δ)(cid:17)n−m n (1 − δ)(cid:1)n−m! × log (cid:0) c n (1 − δ)(cid:1)2(cid:0) c n (1 − δ)(cid:1)m−2(cid:0)1 − c n (1 − δ)(cid:1)n−m n (1 − δ)(cid:1)m−2(cid:0)1 − c n(cid:1)(cid:0) c (cid:0) c n (1 − δ)(cid:1)(cid:0) c (1 − δ)(cid:17) log(cid:18) 1 − c (cid:19) +(cid:16) c (1 − δ)(cid:17)2 (1 − δ)(cid:16)1 − n! −(cid:16) c (1 − δ)(cid:17) cδ (1 − δ)(cid:16)1 − 1 − c (1 − δ)δ(cid:18) n − c(1 − δ) − (1 − δ)(cid:19) n(cid:17)2 =(cid:16) c n(cid:17)2 =(cid:16) c (1 − δ)δ2 n n − c KL(P0(X), Pi(X)) ≤ (1 − δ)(cid:17)2 n (1 − δ) 1 − c n log (1 − δ) . n − c . c n n n c n c n c n n n c n + n c n (1 − δ)(cid:17)n−m−1 ! n c n 29 We once again use the inequality log(1 + x) ≤ x to obtain δ (21) Combining inequalities (20) and (21) with equation (19) of Lemma 10, we obtain the bound nXi=1 KL (P0, Pi) ≤ ≤ completing the proof. E.2.2 Proof of Lemma 9 n(cid:17)2 δ2T +(cid:16) c T δ2, c n − c c(c + 1) n − c (1 − δ)δ2 n n − c (n − 1)T Pt Note that KLt i(v) is independent of t, since the adversary's actions are i.i.d. across time steps. Furthermore, KLi(i) is clearly independent of i and KLi(j) is constant for all pairs i 6= j, so equation (19) of Lemma 10 holds. 0{·I t−1} and n under both Note that when St = {j}, the distribution of the feedback It is the same under Pt i{·I t−1}, since the vertex i is chosen to be a sink vertex with the same probability d A0 and Ai. Hence, KLi(j) = 0. To compute KLi(i), let X denote the size of the influenced component containing i when St = {i}, and define the random variable 0, 1, 2, if i is a sink vertex if i is a source vertex otherwise. As in the proof of Lemma 7, we will upper-bound KL(P0(X, Y ), Pi(X, Y )), leading to an upper bound on KL(P0(X), Pi(X)). The range of (X, Y ) is Y = {(1, 0), (1, 2)} ∪ {(m, 1) : 2 ≤ m ≤ n}. We then have the following expression for KL(P0(X, Y ), Pi(X, Y )): d n log(cid:18) d/n n ! n d n c n (1 − δ) − n(cid:19) log 1 − c d/n(cid:19) +(cid:18)1 − n (1 − δ) − d n − d 1 − c nXm=2(cid:18) n − 1 m − 1(cid:19) c (1 − δ)(cid:18) d n(cid:19)n−m n(cid:19)m−1(cid:18)1 − d ! , × log c n(cid:1)n−m n(cid:1)m−1(cid:0)1 − d n (1 − δ)(cid:0) d n(cid:1)m−1(cid:0)1 − d n(cid:0) d n(cid:1)n−m log(cid:18) n − c − d + cδ n − c − d (cid:19) nXm=2(cid:18) n − 1 m − 1(cid:19)(cid:18) d log(cid:18) n − c − d + cδ n − c − d (cid:19) + n(cid:19)m−1(cid:18)1 − (1 − δ) log(1 − δ) c n c n n c (1 − δ) log(1 − δ). d n(cid:19)n−m + + n − c − d + cδ n n − c − d + cδ = ≤ Using the inequality log(1 + x) ≤ x, we then have KL(P0(X), Pi(X)) ≤ cδ(n − c − d + cδ) n(n − c − d) c n − (1 − δ)δ = c(n − d) n(n − c − d) δ2. Applying Lemma 10 completes the proof. 30
1704.05735	2	1704	2017-09-14T05:48:09	Collaborative Filtering with Social Local Models	[ "cs.SI" ]	Matrix Factorization (MF) is a very popular method for recommendation systems. It assumes that the underneath rating matrix is low-rank. However, this assumption can be too restrictive to capture complex relationships and interactions among users and items. Recently, Local LOw-Rank Matrix Approximation (LLORMA) has been shown to be very successful in addressing this issue. It just assumes the rating matrix is composed of a number of low-rank submatrices constructed from subsets of similar users and items. Although LLORMA outperforms MF, how to construct such submatrices remains a big problem. Motivated by the availability of rich social connections in today's recommendation systems, we propose a novel framework, i.e., Social LOcal low-rank Matrix Approximation (SLOMA), to address this problem. To the best of our knowledge, SLOMA is the first work to incorporate social connections into the local low-rank framework. Furthermore, we enhance SLOMA by applying social regularization to submatrices factorization, denoted as SLOMA++. Therefore, the proposed model can benefit from both social recommendation and the local low-rank assumption. Experimental results from two real-world datasets, Yelp and Douban, demonstrate the superiority of the proposed models over LLORMA and MF.	cs.SI	cs	Collaborative Filtering with Social Local Models Huan Zhao, Quanming Yao1, James T. Kwok and Dik Lun Lee Department of Computer Science and Engineering Hong Kong University of Science and Technology {hzhaoaf, qyaoaa, jamesk, dlee}@cse.ust.hk Hong Kong 7 1 0 2 p e S 4 1 ] I S . s c [ 2 v 5 3 7 5 0 . 4 0 7 1 : v i X r a Abstract-Matrix Factorization (MF) is a very popular method for recommendation systems. It assumes that the underneath rating matrix is low-rank. However, this assumption can be too restrictive to capture complex relationships and interactions among users and items. Recently, Local LOw-Rank Matrix Approximation (LLORMA) has been shown to be very successful in addressing this issue. It just assumes the rating matrix is composed of a number of low-rank submatrices constructed from subsets of similar users and items. Although LLORMA outperforms MF, how to construct such submatrices remains a big problem. Motivated by the availability of rich social connections in today's recommendation systems, we propose a novel framework, i.e., Social LOcal low-rank Matrix Approx- imation (SLOMA), to address this problem. To the best of our knowledge, SLOMA is the first work to incorporate social connections into the local low-rank framework. Furthermore, we enhance SLOMA by applying social regularization to submatrices factorization, denoted as SLOMA++. Therefore, the proposed model can benefit from both social recommendation and the local low-rank assumption. Experimental results from two real-world datasets, Yelp and Douban, demonstrate the superiority of the proposed models over LLORMA and MF. 1 Index Terms-Recommendation system, Collaborative Filter- ing, Matrix factorization, Local low-rank, Social network I. INTRODUCTION Recommendation System (RS) has become an indispensable tool in the big data era. It tackles information overload by helping users to get interesting items based on their previous behaviors. Collaborative Filtering (CF), a state-of-the-art RS technique, tries to predict users' ratings (or preferences) on unseen items based on similar users or items. Among various CF-based methods, Matrix Factorization (MF) is most popular due to its good performance and scalability [1], [2], [3], [4]. MF is based on the assumption that users' preferences to items are controlled by a small number of latent factors. Thus, the large user-item rating matrix can be decomposed into two smaller matrices, representing user-specific and item-specific latent factors, respectively. In other words, the rating matrix is of low-rank. Despite the success of MF in RS, the assumption that the rating matrix is low-rank (termed global low-rank) is problematic because the rating matrix in a real-world scenario is very large and composed of diverse rating behaviors, making the low-rank assumption of the rating matrix improper. Lee et al. [5] proposed a novel framework called Local LOw-Rank Matrix Approximation (LLORMA) to alleviate the global low- rank problem. Based on the observation that there tend to be groups of users who are interested in small sets of items, LLORMA takes the view that the rating matrix is composed of a number of low-rank submatrices (termed local low-rank), which is illustrated in the right part of Figure 1. Extensive experiments have demonstrated the effectiveness of LLORMA in recommendation systems [5], [6], [7], [8]. Besides RS, sub- sequent works have been performed in different domains based on the local low-rank assumption, e.g., image processing [9], [10], multi-label classifications [11], document analysis [12], demonstrating the efficacy of this framework. In the local low-rank approach, the construction of the submatrices is a fundamental problem. Lee et al. [5] proposed to firstly choose some random anchor points, i.e., user-item pairs, from the rating matrix. Then, for each anchor point, one submatrix is constructed by selecting the remaining points that are close to the anchor point based on some distance metrics. However, this method leads to several problems. Firstly, since anchor points are randomly selected, it is hard to explain the meaning of the obtained submatrices, and thus the recommendation results are also unexplainable. Secondly, since the submatrices are constructed around the anchor points using a distance threshold, they are very sensitive to the distance threshold. Thus, finding a good distance threshold value is a challenging task. Finally, the computation and space costs are both high, because we need to compute and store the pair-wise similarities of users and items in all submatrices. To overcome the weaknesses of LLORMA, we propose a novel framework, Social LOcal low-rank Matrix Approxi- mation (SLOMA), which incorporates the social connections among users into the local low-rank framework. While so- cial connections have been exploited effectively in MF-based RSs [13], [14], [15], [16], [17], to the best of our knowledge, SLOMA is the first work to utilize social connections under the local low-rank framework. As in most social network research, SLOMA assumes that the social graph embeds a number of social groups, within which users have similar preferences to and influences on each other. The basic idea of SLOMA is illustrated in the left part of Figure 1. There are three social groups underlying the social network,2 based on which we can construct three submatrices out of the rating matrix 1Quanming Yao is the Corresponding author. 2Note that the groups can overlap in reality, here we put them separately for simplicity. Fig. 1. An illustration of SLOMA. Left-hand side: an example social network, where underlying social groups are circled. Right-hand side: the core idea of LLORMA, that is, the rating matrix is composed of a number of submatrices which are low-rank. SLOMA enhances LLORMA with the social groups underlying the social network to construct meaningful submatrices. Therefore, SLOMA can enjoy the advantages of both social recommendation and LLORMA, leading to better recommendation performance. satisfying the low-rank assumption. The social connections help SLOMA in solving the important submatrix construction problem. Specifically, instead of randomly selecting anchor points as in LLORMA, SLOMA can select influential users, termed connectors, and pick the connectors' friends within a certain distance in the social graph (i.e., within several hops) to construct the submatrices, which are meaningful because they can be treated as social groups. See the example in Figure 1. Thus, SLOMA solves all the three problems faced by LLORMA, that is, it selects meaningful anchors and build meaningful submatrices, its distance measure is intuitive, and there is no need to keep the pairwise similarities of users and items for each submatrix, which avoids high computation and space costs. In summary, the contributions of our work are as follows. • To the best of our knowledge, SLOMA is the first work that incorporates social connections into the local low- rank framework, which can enjoy the advantages from both sides, i.e., social recommendation and local low- rank framework. • SLOMA addresses the submatrice construction problem in LLORMA by exploiting social connections. These submatrices have better interpretability (explained by social homophily theory) and lead to superior prediction accuracy over LLORMA. • In addition to SLOMA, we also propose SLOMA++ that incorporates social regularization techniques into each local model to further improve the performance. • We conduct extensive experiments on two real-world datasets to show the effectiveness of our proposed methods. We not only demonstrate the effectiveness of SLOMA and SLOMA++, but also give insights into so- cial recommendation under the local low-rank framework. The rest of the paper is organized as follows. MF, LLORMA and social recommendation are introduced in Section II. We then elaborate our SLOMA model in Section III, and the experiments as well as the analysis in Section IV and V, respectively. Finally, we conclude our work in Section VI. II. BACKGROUND A. MF and LLORMA MF has been one of the most popular approaches for rating prediction in recommendation systems. Based on the assumption that users' preferences to items are governed by a small number of latent factors, the rating matrix O ∈ Rm×n can be approximated by the product of l-rank matrices, R ≈ UV(cid:62), (1) where U ∈ Rm×l and V ∈ Rn×l with l (cid:28) min(m, n), repre- senting, respectively, the latent features of users' preferences and items. The approximating process can be completed by solving the following optimization problem: (Oij − uiv(cid:62) (cid:88) (2) j )2, min U,V 1 2 (i,j)∈Ω where ui and vj are the ith and jth rows of U and V, representing the latent feature vectors of user ui and item vj, respectively. · F denotes the Frobenius norm, and Ω is the index set of the observations, and Oij is the observed rating of user ui to item vj. In order to avoid overfitting, two regularization terms, λ 2 Equation (2). In the literature, this method is also termed Regularized SVD (RegSVD) [1]. (cid:1), are added to (cid:0)U2 F + V2 F Instead of assuming the rating matrix to be low-rank, Lee et al. proposed a novel framework LLORMA that assumes the rating matrix is local low-rank [5], i.e., it is composed of a number of submatrices which are of low-rank. After obtaining the submatrices, MF is applied to each submatrix independently, and then a weighted ensemble scheme is de- signed to approximate the rating matrix. The key component of LLORMA is the construction of submatrices, which is depicted in Algorithm 1. Algorithm 1 Submatrice Construction (LLORMA) Input: Observed index set Ω, number of submatrices q, distance threshold d1, d2; Select a point (it, jt) randomly from Ω; 1: for t = 1, 2, ..., q do 2: 3: Obtain a subset of users U t =(cid:8)uk : d(uit, uk) ≤ d1 (cid:9); Obtain a subset of items V t =(cid:8)vk : d(vjt, vk) ≤ d2 (cid:9); 7: return (cid:8)Mt(cid:9)q Obtain Mt; // The tth submatrix based on U t and V t. 4: 5: 6: end for t=1 In Line 2, a point (it, jt), termed anchor point, is chosen randomly.3 d(·) is a distance function for any two users or items in the rating matrix. d1 and d2 are two distance thresholds for selecting users or items according to Line 3 and 4. A submatrix is then constructed from the selected users and items that are within the distance threshold from the anchor point. In LLORMA, the distance d is based on the cosine similarity between any two latent features of users or items. It is reported that such measurement gives the best predicting performance [5]. Specifically, the arc-cosine between users uit and uk is (cid:18) (cid:19) d(it, k) = arccos uituk uit · uk . (3) Besides, the similarity between items are also computed based on arc-cosine distance at Equation (3). After obtaining q submatrices, MF is applied to each subma- trice independently, leading to q groups of user-specific and item-specific latent features. Finally, the rating matrix R is approximated by the weighted ensemble of product of these q groups of user-specific and item-specific latent feature vectors as follows q(cid:88) t=1 ij(cid:80)q wt s=1 ws ij (cid:2)ut j)(cid:62)(cid:3), i(vt Rij = (4) i and vt where wt ij is the weight computed based on the distance between a point (ui, vj) and an anchor point (uit, vjt) in every submatrix. ut j represent the latent feature vectors of ui and vj, respectively, from the tth submatrix. It means that the prediction of a point (ui, vj) in the rating matrix is a weighted average of the predicted ratings from all submatrices where ui and vj both occur. Experiments on real-world datasets show that LLORMA improves prediction accuracy compared to RegSVD. Besides recommendation [5], [6], [7], [8], the local low-rank assumption has been demonstrated to be effective 3There are more complex methods for picking up anchor points, but random choosing gives the best predicting performance as reported in [5]. Thus, we focus on random choosing here. in many other domains, including image processing [9], [10], multi-label classifications [11], and document analysis [12]. B. Social Recommendation in MF Social recommendation has become more and more popular with the proliferation of online social networks such as Face- book, Twitter and Yelp. Previous works have been performed to incorporate social connections into the MF framework [13], [14], [15], [17], [18]. Most of these works are based on the social homophily theory [19] that people with similar preferences tend to be connected as friends. Thus, the latent feature vectors of friends should be closer after factorizing the rating matrix. One typical method is to add the social regularization term to the objective function (2), which is defined in the following: Sijui − uj2 2, (5) m(cid:88) (cid:88) i=1 j∈F (i) β 2 where β > 0 is the weight of social regularization term, F(i) is the set of ui's friends, and Sij is the similarity between ui and uf . The more similar two users are, the closer their latent feature vectors. We note that there are no previous work on social recom- mendation under the local low-rank framework. The major dif- ference between our SLOMA and LLORMA is that LLORMA uses random anchor points to construct submatrices, while we use social connections. Thus, we can obtain meaningful submatrices and better prediction ability because each sub- matrix can represent the behavior of a socially connected user group. Moreover, we enhance SLOMA to SLOMA++ by incorporating social regularization into every local model, and demonstrate with experiments that recommendation per- formance can be further improved. III. SOCIAL LOCAL MATRIX APPROXIMATION In this section, we first describe the main problem facing submatrix construction in LLORMA, which also motivates our work. Then, we elaborate the proposed SLOMA framework. A. Motivation Despite the success of LLORMA, the main weakness of LLORMA is its submatrix construction method (Algorithm 1). We identify three problems below. First, the anchor points are chosen randomly, which do not have any reasonable justification and cannot be interpreted. Second, the constructed submatrices are neither stable nor meaningful. Specifically, the cosine similarity in Equation (3) is used in LLROMA, which is computed in Lines 3 and 4 of Algorithm 1. From Lemma III.1 (proof can be found in Appendix A), we can see that different pairs of (U, V) may be obtained from solving the optimization problem in Equation (2). However, the cosine similarity is not robust to the transformation in Lemma III.1. Using the same anchor point with the pair (U, V) or ( U, V), different submatrices can be constructed. As a result, the constructed submatrices are not consistent and then become meaningless. Third, to predict a missing rating in the original big rating matrix, a weighted average scheme is used to combine the ratings in the submatrices according to Equation (4). This suffers from the same problem facing submatrix construction, as the weights are also computed based on cosine similarity. Lemma III.1. Given any matrices U and V which are an optimal solution to (2), then U = UQ and V = VQ−1 are also an optimal solution, if matrix Q is invertible. Motivated by these problems, we propose a novel frame- work integrating social connections with the local low-rank framework. Based on the social homophily theory [19] that people with similar preferences tend to be connected as friends, we use the social connections among users as an explicit indicator of similarity between users' preferences in constructing the submatrices. Since each submatrix contains socially connected users with similar preferences, it satisfies the low-rank property. Moreover, we argue and demonstrate with experiments that the submatrices obtained in our model are better than those in LLORMA in terms of overall predic- tion ability. The core idea of our proposed model is illustrated in Figure 1. B. SLOMA In this section, we give a formal and mathematical elabo- ration of SLOMA and design the optimization approaches. We first introduce the notations in this work. Let U = {u1, u2, ..., um} and V = {v1, v2, ..., vn} be the sets of users and items, respectively. Let G = (U,E) represents a social network graph, where the vertex set U = {u1, u2, ..., um} represents the users and the edge set E = {e1, e2, ..., ep} represents the social connections among all users. The weight of every edge of G is set to 1, representing the existence of friendship between two users.4 Then, for any two users ui, uj, d(ui, uj) is the distance between ui and uj, which can be computed from the social graph. In this paper, we propose a novel framework to integrate social recommendation with the local low-rank assumption. The key challenge is how to construct submatrices from the rating matrix. We develop several approaches, including heuristic and systematic approaches. After constructing the submatrices, we apply MF to them and obtain a number of user-specific and item-specific latent features in different submatrices. It these submatrices have overlaps, i.e., a user and an item can occur in more than one submatrices, which aligns with our intuition that users tend to participate in multiple groups in social networks. is obvious that Overall, SLOMA consists of the following steps: • Identify q social groups from the social graph G, and then construct q submatrices from the rating matrix O based on those groups. • Apply MF to all the submatrices, i.e., {Mt}q t=1, inde- pendently, and obtain q groups of user-specific and item- specific latent features. 4We assume that the social graph is undirected and unweighted. • Predict the missing ratings in R using the ensemble of predictions from all submatrices. C. Construction of Submatrices In SLOMA, we assume that there are groups of users un- derlying the social network, whose preferences are similar due to the fact that they are socially connected. To identify these groups, we develop both heuristic and systematic approaches in this work. Heuristic Approaches. For the heuristic approaches, we build the social groups based on the fact that users' influences to each other can propagate through the networks. We observe that users in a social group can affect each other and the amount of influence can vary. Intuitively, those with more friends tend to have larger influence than others. Based on this, we construct the submatrices by first selecting a number of influential users, i.e., the connectors, and their friends within a fixed number of hops, e.g., three, in the social network. Then, for each group of users, we select the items they rate and then the submatrix is constructed from this user-item group. Note that there are several methods for selecting the connectors with different overall performance. We discuss this in Section III-E and analyze the experimental results in Section V-D. After selecting q connectors, construction of the submatrix is shown in Algorithm 2. The shortest path in Line 3 can be obtained by algorithms such as Dijkstra's algorithm. number of submatrices q, distance threshold d; Algorithm 2 Heuristic Submatrix Construction. Input: Observed index set Ω, social graph G = (U,E), 1: Obtain q users as connectors from U; 2: for t = 1, 2, . . . , q do 3: Obtain D(uit, uk): the shortest distance between uit and all the other users uks in G; Obtain a subset of users U t =(cid:8)uk : D(uit, uk) ≤ d(cid:9); Obtain a subset of items V t =(cid:8)vj : (i, j) ∈ Ω,∀ui ∈ U t(cid:9); // All items rated by users in U t. 8: return (cid:8)Mt(cid:9)q Obtain Mt; // The tth submatrix based on U t and V t t=1 4: 5: 6: 7: end for Systematic Approaches. The systematic approaches are based on methods for overlapping community detection, which have been intensively investigated in recent years. We can see that social groups are equivalent to communities in social networks. In the literature, a community is a group of people who have more interactions within the group than those outside it [20], and hence users in the same community have more characteristics in common than with users outside of it. This leads to the low-rank property of the submatrix constructed from one community. Naturally, communities can overlap with each other, and many works have been done for detecting such communities, e.g., line graph partitioning [21], clique percola- tion [22], eigenvector methods [23], egonet analysis [24], [25] and low-rank models [26]. In this work, we adopt BIGCLAM [26], which can scale to large datasets and has good empirical performance. The process is shown in Algorithm 3. We can see that the only difference from Algorithm 2 is how we construct the social groups (Line 1). Algorithm 3 Systematic Submatrix Construction Input: Observed index set Ω, social graph G = (U,E), 1: Apply BIGCLAM [26] to G to obtain the set C of q number of submatrices q; communities; 2: for t = 1, 2, . . . , q do 3: Obtain a subset of users U t ∈ C; // U t is the tth community. Obtain a subset of items V t =(cid:8)vj : (i, j)∈ Ω,∀ui∈ U t(cid:9); // All items rated by users in U t. Obtain Mt; // The tth submatrix based on U t and V t 4: 5: 6: end for 7: return (cid:8)Mt(cid:9)q t=1 Note that different methods lead to different numbers of users in the submatrices, i.e., they may cover different numbers of users in the social graph, which will influence the overall prediction ability of SLOMA. In Sections V-D and V-E, we give detailed comparisons and analysis for these different submatrix construction methods, and show that the heuristic approaches outperform the systematic ones. D. Matrix Factorization in Social Local Models After obtaining q submatrices, we apply MF to each sub- matrix independently according to Equation (2). Note that in LLORMA, each submatrix corresponds to a local model. Similarly, each submatrix in SLOMA corresponds to a social local model. In the following sections, we use the terms "local model" and "submatrix" interchangeably. By training each social local model, we can obtain q groups of user latent feature matrices, U1, . . . , Uq, and item latent feature matrices, V1, . . . , Vq. Then, the overall rating of user ui given to item vj is predicted according to the following ensemble method: q(cid:88) t=1 Rij = 1 q ut i(vt j)(cid:62), (6) i and vt where ut j, respectively, are the user-specific and item- specific latent feature vectors in the tth submatrix where they occur, and q is the number of submatrices where user ui and item vj both occur. Equation (6) means that the rating of ui given to vj is predicted by the average of predicted ratings that ui gives to vj in all submatrices. E. Different Methods to Select Connectors The construction of social groups in SLOMA is similar to overlapping community detection using seed set expan- sion [27], [28]. There tend to be a small number of so- called long-tail users, i.e., those with few friends, which are difficult to be selected into any of the social groups, leading to incomplete coverage of SLOMA. We define two kinds of coverage in this work: i) User Coverage is the percentage of unique users in U selected by all of the social groups; ii) Rating Coverage is the percentage of unique ratings in Ω selected by all of the submatrices. Rating coverage is dependent on user coverage because when constructing each submatrix, we usually select all of the items rated by that group of users. Therefore, in this paper, we use coverage to represent any of these two concepts depending on the context. In the local low-rank framework, the coverage of the submatrices is very important for the overall prediction performance. In LLORMA, the random selection of anchor points and choice of distance threshold are used to control the rating coverage. Due to the randomness of the anchor points, the coverage of all submatrices will be 1 when the number of anchor points reaches a large enough value. Further, the larger the threshold is, the larger the coverage of each submatrix. In SLOMA, we select seeds, i.e., connectors, in the social graph and propagate from each seed along different number of hops to construct the social groups. For long-tail users who cannot be covered, we predict their ratings with the mean of all the observed ratings, which is the same approach as in [5]. In Section V-A, we can see that the performance of SLOMA is still better than that of LLORMA and RegSVD, demonstrating the superiority of the submatrix construction method in SLOMA. However, it is still very important to cover as many points in the rating matrix as possible. Thus, we try several connector selection methods, which are described below: • Hub: Select a set of users with the largest number of neighbors. • Random: Randomly select a set of users. • Random-Hub: It is an integration of the above two methods, that is, we first select a larger number of hub users, e.g., 1000, and then randomly select a smaller number of the hub users, e.g., 50. • Greedy: Each time we select a connector, we select from those not yet covered by connectors that are already selected. After obtaining the set of connectors, we construct one submatrix around each connector user as shown in Lines 4 and 5 of Algorithm 2. In Section V-D, we compare the performance of these methods as well as the community-based submatrix construction method. We can see that despite the existence of a small number of uncovered users, Hub and Greedy obtain the best performance comparing to the other methods. F. SLOMA++ In addition to constructing the submatices in SLOMA, social connections can be used to enhance each local model by applying social regularization to the factorization of each submatrix. Social regularization adds constraints that users with direct social connections should be closer in the latent space, which is introduced in Equation (5). For each local model, the social regularization term is defined as: ijut St i − ut j2 2, (7) m(cid:88) (cid:88) i=1 j∈F (i) β 2 where St ∈ RU t×U t is the similarity matrix with St ij representing the similarity between users ui and uj in the tth submatrix, and U t is the number of users in the tth submatrix. Note that we apply Equation (7) when each submatrix is factorized. We term this model SLOMA++. To calculate the similarities between two users ui and uj, we utilize the popular Person Correlation Coefficient (PCC) [29]: (cid:114) (cid:80) (Mt if − ¯M t if − ¯M t i )(Mt i )2(cid:114) (cid:80) jf − ¯M t j ) jf − ¯M t j )2 (Mt , f∈T (i,j) (Mt St ij = f∈T (i,j) f∈T (i,j) (cid:80) where ¯M t i represents the mean value of all of the ratings of user ui in the tth submatrix Mt, and I t(i) represents the set of items rated by user ui in the tth submatrix Mt. T (i, j) = I t(i) ∩ I t(j), representing the set of common items rated by ui and uj in the tth submatrix. Note that when constructing the submatrix for a group of users, we take all of the group's ratings in O. Thus, for the same user pair ui and uj, T (i, j) will be the same across all submatrices. This is why we remove the superscript t from T (i, j). Further, we employ a mapping 2 (x+1) to bound the range of PCC similarity function g(x) = 1 to [0, 1]. For each submatrix, we need to solve the following opti- (cid:0)U2 (cid:1) (Oij − uiv(cid:62) j )2 + λ 2 F + V2 F mization problem: (cid:88) m(cid:88) (i,j)∈Ω β 2 min U,V 1 2 + (cid:88) i=1 j∈F (i) Sijui − uj2 2. (8) Note that we leave out the superscript t for simplicity of notations. From this equation we can see that when β = 0, we recover SLOMA. Equation (8) can be solved by gradient descend methods [2], [3]. IV. EXPERIMENT In this section, we introduce the details of the experiments, including the datasets, the evaluation metrics, and the base- lines. A. Datasets We conduct the experiments on two real-world datasets: Yelp and Douban. Yelp is a location-based website where users can give ratings to and write reviews on items like restaurants, theaters, and businesses. The Yelp dataset is provided by the Yelp Dataset Challenge,5 which has been used in previous research in recommendation [5], [18], [30]. Douban is a Chinese website where users can rate and share their opinions on items such as movies and books. The Douban Dataset is obtained from [15]. In both websites, users can build Facebook-style connections to each other and all ratings are in the range 1 to 5. Therefore, the datasets are ideal for evaluating the effectiveness of our proposel model. 5https://www.yelp.com/dataset challenge We preprocess the two datasets by removing users without any friends or with ratings fewer than 5. These users are called, respectively, "social cold-start" and "cold start" users in RSs. The statistics of the two preprocessed datasets are shown in Table I. Note that in the table, R density represents the density of the rating matrix, and S edges and S density represent the social connections and density of the social matrix, respectively. B. Evaluation Metrics We choose two evaluation metrics, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). They are defined as (cid:88) (cid:88) (i,j)∈ ¯Ω 1 ¯Ω (cid:118)(cid:117)(cid:117)(cid:116) 1 ¯Ω (i,j)∈ ¯Ω Oij − Rij, (Oij − Rij)2, MAE = RMSE = where ¯Ω is the set of all user-item pairs (i, j) in the test set, and Rij is the corresponding rating predicted by the algorithm. These metrics are popular for the task of rating prediction in the literature [2], [3], [15]. C. Experimental Settings We compare our proposed models with the following state- of-the-art methods: • RegSVD [1]: It is the standard matrix factorization method with (cid:96)2 regularization. We implement it according to [1]. • LLORMA [5]: It is based on the local low-rank as- sumption, which SLOMA also assumes. We implement it according to [5]. • SocReg [15]: It is a state-of-the-art method that integrates social connections into MF by employing social connec- tions as regularization terms. We implement it according to [15]. • SLOMA: This is our proposed model. It utilizes social connections among users to form social groups for local low-rank factorization. • SLOMA++: This is the same as SLOMA except that it applies social regularization to the factorization of submatrices, as is shown in Equation (7). Following the experimental settings in [5], we also randomly split each dataset into training and test data with a ratio of 8:2. In the training process, the training data are used to fit the model, and the test data are used to calculate the prediction errors of the models. We repeat each experiment five times by randomly splitting the datasets and report the average results. V. ANALYSIS In this section, we present and compare the experimental results of our proposed SLOMA and SLOMA++ against the baselines. We try to answer the following questions: (1). What is the recommendation performance of our pro- posed models comparing to the state-of-the-art methods? STATISTICS OF DATASETS (R density = Yelp Douban Users 76,220 103,054 Items 79,257 57,908 #Ratings TABLE I #U sers×#Items , S density = #2×S edges Ratings 1,352,762 15,129,113 R density 0.022% 0.254% S density 0.022% 0.028% S edges 647,451 753,358 #U sers×#U sers ). PERFORMANCE OF DIFFERENT METHODS IN YELP AND DOUBAN WITH K = 10, 20. THE BEST TWO RESULTS ARE HIGHLIGHTED. IMPROVEMENTS ARE TABLE II Datasets Yelp Douban K 10 20 10 20 LLORMA MEASURED BY THE REDUCTION OF RMSE COMPARING TO SLOMA++. SLOMA Metrics 0.9362 MAE +0.65% Improve 1.1760 RMSE +0.04% Improve 0.9389 MAE +1.59% Improve 1.1788 RMSE +0.76% Improve 0.5744 MAE +2.45% Improve 0.7255 RMSE +2.07% Improve 0.5715 MAE +2.48% Improve 0.7225 RMSE +2.01% Improve RegSVD 0.9478 +1.87% 1.1908 +1.28% 0.9499 +2.73% 1.1918 +1.85% 0.5828 +3.86% 0.7347 +3.29% 0.5803 +3.96% 0.7320 +3.28% SocReg 0.9228 -0.79% 1.1802 +0.40% 0.9190 -0.54% 1.1754 +0.48% 0.5662 +1.04% 0.7165 +0.84% 0.5638 +1.15% 0.7142 +0.87% 0.9459 +1.67% 1.1843 +0.74% 0.9477 +2.50% 1.1862 +1.38% 0.5811 +3.58% 0.7310 +2.80% 0.5779 +3.56% 0.7278 +2.72% SLOMA++ 0.9301 1.1755 0.9240 1.1698 0.5603 0.7105 0.5573 0.7080 (2). How do the parameters in our models affect the over- all performance? Specifically, how do the number of submatrices and hops, i.e., distance threshold, affect recommendation performance? (3). What is the performance of different methods of con- nector selection? (4). Which submatrix construction approach, heuristic or systematic, is better for SLOMA and SLOMA++? A. Recommending Performance To answer Question (1), we show the recommendation performance of the baselines and our proposed methods in Table II. We can see that SLOMA++ consistently outperforms all the other methods under different Ks in both datasets. This demonstrates the efficacy of integrating social connections with local low-rank framework. When comparing RegSVD, LLORMA, and SLOMA, we observe that SLOMA has better performance than LLORMA and RegSVD, which can be at- tributed to the better submatrices SLOMA builds with the help of social connections. When further studying the results of SocReg, SLOMA and SLOMA++, we can see that SLOMA++ is the best. This demonstrates that SLOMA++ makes the best use of social connections. Taking advantage from both the local low-rank assumption and social recommendation, SLOMA++ beats SocReg with the local low-rank framework and SLOMA with social regularization. Through these two comparisons, we can conclude that both social regularization and local low-rank can help improve the recommending per- formance. When comparing SLOMA and SocReg, we can see that the performance of SocReg is consistently a bit better than that of SLOMA. This result not only indicates the power of social regularization but also points to a limitation of SLOMA, that is, it does not cover the long-tail users very well when constructing the submatrices and hence can only predict the long-tail users' ratings based on the mean of all the existing ratings. The latter also explains the phenomenon that the performance gain of SLOMA against RegSVD is limited. Therefore, to design better submatrix construction methods with larger coverage. We leave this for future research. is very important it When comparing the performance on Yelp and Douban, we can see that the improvement on Douban is more significant than that on Yelp, which may be attributed to Douban's larger density of social connections (see Table I). It means that better performance can be obtained when social networks are dense. This is intuitive because the more friends a user has, the more likely his or her behaviors are affected by friends. B. Impact of the Number of Local Models In this section, we show how performance varies with different numbers of local models in LLORMA, SLOMA, SLOMA++. The results are shown in Figure 2. We also plot the RMSEs of RegSVD and SocReg for comparison. From the figures, we can see that with increasing number of local models, RMSE's of LLORMA, SLOMA, and SLOMA++ all decrease. This means that the better performance we can get. However, when the number is large enough, e.g., around 30, the performance gains become marginal. This trend is consistent with the results reported in [5], which indicates that SLOMA and SLOMA++ show similar performance trends to LLORMA with the number of local models varying. the more local models, When the number of local models is smaller than 10, the performance is not as good as what we can achieve at 30. The reason is that fewer submatrices means less overlaps of the local models, which impairs the ensemble prediction accuracy according to Equation (6). Further, it also reduces the overall coverage of SLOMA and SLOMA++, which means more users' ratings are predicted by the mean of other users instead of by the learned model. Again we can see that SLOMA++ consistently outperforms all the other baselines when the local model is large enough, e.g., greater than 10. In practice, it is good enough to set the number of local models to 50, which provides good prediction ability while avoiding higher costs of space and computation. (a) K = 10@Yelp. (b) K = 20@Yelp. (c) K = 10@Douban. (d) K = 20@Douban. Fig. 2. RMSEs vs different numbers of local models on Yelp and Douban. (a) K = 10@Yelp. (b) K = 20@Yelp. (c) K = 10@Douban. (d) K = 20@Douban. Fig. 3. RMSEs vs different hops on Yelp and Douban. C. Impact of the Number of Hops In SLOMA and SLOMA++, we firstly select a small number of connectors, and then take all of their friends within different hops, i.e., the distance threshold d in Algorithm 2, in the social network to construct the social groups. Therefore, different hops mean different numbers of users included in each social group, thus different coverages of each local model, which results in different prediction capabilities of the proposed models. Due to the small-world phenomenon in social networks [31], [32], which states that two people can be connected by a small number of intermediates in social networks, the so-called Six Degree Separation. Thus, we limit the hop to the range [1, 6] in the experiments. Results are shown in Figure 3. As can be seen, the trend of the performance w.r.t. the hop is very similar to that of the number of local models, which is, the performance becomes better with increasing of hops and then becomes stable when the hop is large enough. Specifically, when hop < 3, the performance of SLOMA and SLOMA++ is worse than those of RegSVD and SocReg. The reason is that the coverage of each local model is not large enough, leading to a decrease in overall coverage of all local models. Therefore, the ensemble ability is impaired. However, when hop ≥ 3, the performance of SLOMA and SLOMA++ becomes better, and stable when hop keeps increasing. Therefore, it means that a larger hop is useful for SLOMA and SLOMA++. However, it does not have to be very large because it will lead to higher space and computation cost but only marginal performance gain. An interesting observation is that on Douban, at hop > 3, the performance of SLOMA and SLOMA++ decreases to a level that is very close to those of RegSVD and SocReg. This means that the benefit of the local low-rank assumption vanishes. We analyze the experimental results and find that SLOMA and SLOMA++ can cover all the users in Douban, thus each local model is reduced to the same one as RegSVD. Specifically, the variance of each local model decreases when the hop is too large, thus impairing the ensemble performance of all local models. Therefore, in the previous experiments comparing SLOMA and SLOMA++ against other baselines, we set d = 3 when constructing submatrices according to Algorithm 2. In summary, Question (2) can be understood clearly based on the above two sections. D. Impact of Connector Selection Methods In this section, we try to answer Question (3) by exploring in depth the performance of different connector selection methods (introduced in Section III-E). When adopting the heuristic approaches to construct submatrices, the selection of connectors is very important. It determines the coverages of all local models. As we have elaborated, even though there exist some long-tail users who are rarely selected into any submatrix, we still want to design better heuristic approaches to include as many of them as we can. In Figure 4, we give the comparisons among different selection methods. Note that we also show the results of LLORMA as well as community- based SLOMA, and the results of RegSVD and SocReg are added for SLOMA and SLOMA++, respectively. We set d = 3 for different connector selection methods. 01020304050Number of Local Models1.171.181.191.201.211.22RMSESLOMASLOMA++LLORMARegSVDSocReg01020304050Number of Local Models1.171.181.191.201.211.22RMSESLOMASLOMA++LLORMARegSVDSocReg01020304050Number of Local Models0.710.720.730.740.750.760.77RMSESLOMASLOMA++LLORMARegSVDSocReg01020304050Number of Local Models0.700.710.720.730.740.750.76RMSESLOMASLOMA++LLORMARegSVDSocReg123456HOP1.171.181.191.201.211.22RMSESLOMASLOMA++RegSVDSocReg123456HOP1.171.181.191.201.211.22RMSESLOMASLOMA++RegSVDSocReg123456HOP0.700.750.800.850.90RMSESLOMASLOMA++RegSVDSocReg123456HOP0.700.750.800.850.90RMSESLOMASLOMA++RegSVDSocReg(a) SLOMA(K = 10@Yelp). (b) SLOMA(K = 20@Yelp). (c) SLOMA++(K = 10@Yelp). (d) SLOMA++(K = 20@Yelp). (e) SLOMA(K = 10@Douban). (f) SLOMA(K = 20@Douban). (g) SLOMA++(K = 10@Douban). (h) SLOMA++(K = 20@Douban). Fig. 4. RMSEs v.s different methods of identifying connectors on Yelp and Douban. thus we can guarantee that all From Figure 4, we can see that Hub and Greedy outperform all the other methods with increasing number of local models. The better performance of Greedy lies in the fact that we choose uncovered users as connectors with the most neighbors each time, the connectors will not be too close to each other in the social graph. It will increase the coverage as well as the variance of all the submatrices, which is useful for the ensemble effect of all the local models. However, considering the clarity and simplicity of Hub, we adopt Hub as our selection method in the experiments reported in Section V-A. In Hub, each connector can be regarded as the most active and influential users of the social groups around them, which leads to a low-rank property of the constructed submatrix. Furhter, Hub is simple to implement. We can see that for all the heuristic approaches, even though Random does not provide good performance, the RMSEs of them decrease with increasing number of local models. The underlying reason is that we can cover more entries in the rat- ing matrix with more local models, leading to the performance gain for overall prediction. Taking all the heuristic approaches and LLORMA into consideration, it further demonstrates the efficacy of local low-rank framework. From Figure 4, an obvious observation is E. Discussion on Systematic Submatrices Construction that the community-based method, i.e., systematic submatrix construc- tion, performs very bad in all settings. We make use of the BIGCLAM [26] to detect overlapping communities, each of which corresponds to one submatrix in SLOMA and SLOMA++. The number of communities is set to 50. However, there are two problems facing the communities detected by BIGCLAM. First, BIGCLAM tends to detect closely con- nected groups, which leads to smaller coverage of each subma- trix and fewer overlaps amongest all the submatrices. As we mentioned above, it impairs the ensemble effects in the overall prediction. Second, there are a small portion of long-tail users (it is around 4% in our experiments) who cannot be covered by all the communities, thus impairing the performance further. Therefore, it solves Question (4) as well as explains why we do not report the results of systematic approaches for SLOMA and SLOMA++ in Section V-A. However, we leave it as future work to detect better overlapping communities for SLOMA and SLOMA++. VI. CONCLUSIONS AND FUTURE WORK it In this paper, we propose the SLOMA framework, which addresses the problem of submatrix construction facing LLORMA. To the best of our knowledge, is the first work to incorporate social connections into the local low- rank framework. Based on the social homophily theory, we exploit users' social connections to construct meaningful as well as better submatrices, which leads to superior predic- tion model compared to LLORMA. Moreover, we integrate social regularization to further strengthen SLOMA. Extensive experiments have been conducted on two real-world datasets, comparing our models against MF as well as LLORMA. The results demonstrate that with the help of social connections, SLOMA outperforms LLORMA and MF. VII. ACKNOWLEDGMENT This project is partially supported by Research Grants Coun- cil HKSAR GRF (No.615113) and Microsoft Research Asia. We also thank the anonymous reviewers for their valuable comments and suggestions that help improve the quality of this paper. 01020304050Number of Local Models1.1751.1801.1851.1901.1951.2001.2051.2101.2151.220RMSEHubRandomRand-HubGreedyCommunityLLORMARegSVD01020304050Number of Local Models1.1751.1801.1851.1901.1951.2001.2051.2101.2151.220RMSEHubRandomRand-HubGreedyCommunityLLORMARegSVD01020304050Number of Local Models1.171.181.191.201.211.22RMSEHubRandomRand-HubGreedyCommunityLLORMASocReg01020304050Number of Local Models1.171.181.191.201.211.22RMSEHubRandomRand-HubGreedyCommunityLLORMASocReg01020304050Number of Local Models0.750.800.850.90RMSEHubRandomRand-HubGreedyCommunityLLORMARegSVD01020304050Number of Local Models0.750.800.850.90RMSEHubRandomRand-HubGreedyCommunityLLORMARegSVD01020304050Number of Local Models0.700.750.800.850.90RMSEHubRandomRand-HubGreedyCommunityLLORMASocReg01020304050Number of Local Models0.700.750.800.850.90RMSEHubRandomRand-HubGreedyCommunityLLORMASocReg[24] B. Rees and K. Gallagher, "Overlapping community detection by collective friendship group inference," in International Conference on Advances in Social Network Analysis and Mining, 2010, pp. 375–379. [25] M. Coscia, G. Rossetti, F. Giannotti, and D. Pedreschi, "Demon: a local- first discovery method for overlapping communities," in International Conference on Knowledge Discovery and Data Mining, 2012, pp. 615– 623. [26] J. Yang and J. Leskovec, "Overlapping community detection at scale: a nonnegative matrix factorization approach," in International Conference on Web Search and Data Mining, 2013, pp. 587–596. [27] J. Xie, S. Kelley, and B. K. Szymanski, "Overlapping community detection in networks: The state-of-the-art and comparative study," ACM Computing Surveys, vol. 45, no. 4, p. 43, 2013. [28] J. Whang, D. F. Gleich, and I. S. Dhillon, "Overlapping community detection using seed set expansion," in International Conference on Information and Knowledge Management, 2013, pp. 2099–2108. [29] J. S. Breese, D. Heckerman, and C. Kadie, "Empirical analysis of predic- tive algorithms for collaborative filtering," in Conference on Uncertainty in Artificial Intelligence, 1998, pp. 43–52. [30] J. McAuley and J. Leskovec, "Hidden factors and hidden topics: understanding rating dimensions with review text," in ACM Conference on Recommender Systems, 2013, pp. 165–172. [31] J. Travers and S. Milgram, "The small world problem," Phychology Today, vol. 1, pp. 61–67, 1967. [32] J. Kleinberg, "The small-world phenomenon: An algorithmic perspec- tive," in ACM Symposium on Theory of Computing, 2000, pp. 163–170. Proof. Let PΩ(A) = Aij if (i, j) ∈ Ω, and 0 otherwise. Then, we can express Equation (2) as (9) APPENDIX A PROOF OF LEMMA III.1 (cid:13)(cid:13)PΩ(UV(cid:62) − O)(cid:13)(cid:13)2 F . min U,V 1 2 If U and V are the optimal solution to Equation (9), then they should satisfy first order optimal condition as U(cid:62)PΩ(UV(cid:62) − O) = 0, PΩ(UV(cid:62) − O)V = 0. (10) As U = UQ and V = VQ−1, we have U(cid:62)PΩ( U V(cid:62) − O) = Q(cid:62)U(cid:62)PΩ( U V(cid:62) − O) = Q(cid:62)U(cid:62)PΩ(UV(cid:62) − O) = 0. where the last equality comes from Equation (10). Similarly, we can get PΩ( U V(cid:62) − O) V = 0. Thus, we can see U and V also satisfy the optimal condition of Equation (2), and then the Lemma holds. REFERENCES [1] A. Paterek, "Improving regularized singular value decomposition for collaborative filtering," in KDDcup Workshop, vol. 2007, 2007, pp. 5–8. [2] A. Mnih and R. Salakhutdinov, "Probabilistic matrix factorization," in Advances in Neural Information Processing Systems, 2007, pp. 1257– 1264. [3] Y. Koren, "Factorization meets the neighborhood: a multifaceted col- laborative filtering model," in International Conference on Knowledge Discovery and Data Mining, 2008, pp. 426–434. [4] Q. Yao and J. Kwok, "Accelerated inexact Soft-Impute for fast large- scale matrix completion," in International Joint Conference on Artificial Intelligence. AAAI Press, 2015, pp. 4002–4008. [5] J. Lee and S. Kim, "LLORMA: Local low-rank matrix approximation," Journal of Machine Learning Research, vol. 17, no. 15, pp. 1–24, 2016. [6] C. Chen, D. Li, Y. Zhao, Q. Lv, and L. Shang, "WEMAREC: Accurate and scalable recommendation through weighted and ensemble matrix approximation," in International Conference on Research and Develop- ment in Information Retrieval, 2015, pp. 303–312. [7] C. Chen, D. Li, Q. Lv, J. Yan, S. Chu, and L. Shang, "MPMA: Mixture probabilistic matrix approximation for collaborative filtering," in International Joint Conference on Artificial Intelligence, 2016, pp. 1382–1388. [8] M. Zhang, B. Hu, C. Shi, and B. Wang, "Local low-rank matrix ap- proximation with preference selection of anchor points," in International Conference on World Wide Web Companion, 2017, pp. 1395–1403. [9] Q. Yao and J. Kwok, "Colorization by patch-based local low-rank matrix completion," in AAAI Conference on Artificial Intelligence, 2015, pp. 1959–1965. [10] X. Li, B. Shen, B. Liu, and Y. Zhang, "A locality sensitive low-rank model for image tag completion," IEEE Transactions on Multimedia, vol. 18, no. 3, pp. 474–483, 2016. [11] K. Bhatia, H. Jain, P. Kar, M. Varma, and P. Jain, "Sparse local em- beddings for extreme multi-label classification," in Advances in Neural Information Processing Systems, 2015, pp. 730–738. [12] S. Sangho, C. Jaegul, J. Lee, and K. R. Chandan, "L-EnsNMF: Boosted local topic discovery via ensemble of nonnegative matrix factorization," in International Conference on Data Mining, 2016, pp. 479–488. [13] H. Ma, H. Yang, M. R. Lyu, and I. King, "SoRec: Social recom- mendation using probabilistic matrix factorization," in International Conference on Information and Knowledge Management, 2008, pp. 931– 940. [14] M. Jamali and M. Ester, "A matrix factorization technique with trust propagation for recommendation in social networks," in ACM Confer- ence on Recommender Systems, 2010, pp. 135–142. [15] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King, "Recommender systems with social regularization," in International Conference on Web Search and Data Mining, 2011, pp. 287–296. [16] J. Tang, X. Hu, H. Gao, and H. Liu, "Exploiting local and global social context for recommendation," in International Joint Conference on Artificial Intelligence, 2013, pp. 264–269. [17] G. Guo, J. Zhang, and N. Yorke-Smith, "TrustSVD: Collaborative filtering with both the explicit and implicit influence of user trust and of item ratings." in AAAI Conference on Artificial Intelligence, 2015, pp. 123–129. [18] H. Li, D. Wu, W. Tang, and N. Mamoulis, "Overlapping community regularization for rating prediction in social recommender systems," in ACM Conference on Recommender Systems, 2015, pp. 27–34. [19] M. McPherson, L. Smith-Lovin, and J. M. Cook, "Birds of a feather: Homophily in social networks," Annual Review of Sociology, vol. 27, no. 1, pp. 415–444, 2001. [20] M. Girvan and M. Newman, "Community structure in social and bio- logical networks," in Proceedings of the National Academy of Sciences, vol. 99, no. 12, 2002, pp. 7821–7826. [21] Y.-Y. Ahn, J. P. Bagrow, and S. Lehmann, "Link communities reveal multiscale complexity in networks," Nature, vol. 466, no. 7307, pp. 761– 764, 2010. [22] G. Palla, I. Der´enyi, I. Farkas, and T. Vicsek, "Uncovering the overlap- ping community structure of complex networks in nature and society," Nature, vol. 435, no. 7043, pp. 814–818, 2005. [23] S. Zhang, R.-S. Wang, and X.-S. Zhang, "Identification of overlapping community structure in complex networks using fuzzy c-means cluster- ing," Physica A: Statistical Mechanics and its Applications, vol. 374, no. 1, pp. 483–490, 2007.
1110.5813	4	1110	2012-07-03T19:30:23	Overlapping Community Detection in Networks: the State of the Art and Comparative Study	[ "cs.SI", "cs.DS", "physics.soc-ph" ]	This paper reviews the state of the art in overlapping community detection algorithms, quality measures, and benchmarks. A thorough comparison of different algorithms (a total of fourteen) is provided. In addition to community level evaluation, we propose a framework for evaluating algorithms' ability to detect overlapping nodes, which helps to assess over-detection and under-detection. After considering community level detection performance measured by Normalized Mutual Information, the Omega index, and node level detection performance measured by F-score, we reached the following conclusions. For low overlapping density networks, SLPA, OSLOM, Game and COPRA offer better performance than the other tested algorithms. For networks with high overlapping density and high overlapping diversity, both SLPA and Game provide relatively stable performance. However, test results also suggest that the detection in such networks is still not yet fully resolved. A common feature observed by various algorithms in real-world networks is the relatively small fraction of overlapping nodes (typically less than 30%), each of which belongs to only 2 or 3 communities.	cs.SI	cs	Overlapping Community Detection in Networks: the State of the Art and Comparative Study1 JIERUI XIE ([email protected]) Network Science and Technology, Rensselaer Polytechnic Institute, New York, USA STEPHEN KELLEY ([email protected]) Oak Ridge National Laboratory, Tennessee, USA and BOLESLAW K. SZYMANSKI ([email protected]) Network Science and Technology, Rensselaer Polytechnic Institute, New York, USA This paper reviews the state of the art in overlapping community detection algorithms, quality measures, and benchmarks. A thorough comparison of different algorithms (a total of fourteen) is provided. In addition to community level evaluation, we propose a framework for evaluating algorithms’ ability to detect overlapping nodes, which helps to assess over-detection and under- detection. After considering community level detection performance measured by Normalized Mutual Information, the Omega index, and node level detection performance measured by F- score, we reached the following conclusions. For low overlapping density networks, SLPA, OSLOM, Game and COPRA offer better performance than the other tested algorithms. For networks with high overlapping density and high overlapping diversity, both SLPA and Game provide relatively stable performance. However, test results also suggest that the detection in such networks is still not yet fully resolved. A common feature observed by various algorithms in real-world networks is the relatively small fraction of overlapping nodes (typically less than 30%), each of which belongs to only 2 or 3 communities. Categories and Subject Descriptors: A.1 [General Literature]: INTRODUCTORY AND SUR- VEY; I.5.3 [Clustering]: Clustering—Algorithms; H.3.3 [Clustering]: Information Search and Retrieval—Clustering; E.1 [Data]: DATA STRUCTURES—Graphs and networks General Terms: Algorithms, Performance Additional Key Words and Phrases: Algorithms, overlapping community detection, social net- works 1. INTRODUCTION Community or modular structure is considered to be a significant property of real- world social networks as it often accounts for the functionality of the system. De- spite the ambiguity in the definition of community, numerous techniques have been developed for both efficient and effective community detection. Random walks, spectral clustering, modularity maximization, differential equations, and statistical mechanics have all been used previously. Much of the focus within community detection has been on identifying disjoint communities. This type of detection as- sumes that the network can be partitioned into dense regions in which nodes have more connections to each other than to the rest of the network. Recent reviews 1Citation: JIERUI XIE, STEPHEN KELLEY and BOLESLAW K. SZYMANSKI, Overlapping Community Detection in Networks: the State of the Art and Comparative Study, ACM Computing Surveys, vol. 45, no. 4, 2013 (In press) , Vol. V, No. N, May 2014, Pages 1–37. 2 · Jierui Xie et al. on disjoint community detection are presented in [Danon et al. 2005; Lancichinetti and Fortunato 2009; Leskovec et al. 2010; Fortunato 2010]. However, it is well understood that people in a social network are naturally char- acterized by multiple community memberships. For example, a person usually has connections to several social groups like family, friends, and colleagues; a researcher may be active in several areas. Further, in online social networks, the number of communities an individual can belong to is essentially unlimited because a person can simultaneously associate with as many groups as he wishes. This also happens in other complex networks such as biological networks, where a node might have multiple functions. In [Kelley et al. 2011; Reid et al. 2011], the authors showed that the overlap is indeed a significant feature of many real-world social networks. For this reason, there is growing interest in overlapping community detection algorithms that identify a set of clusters that are not necessarily disjoint. There could be nodes that belong to more than one cluster. In this paper, we offer a review on the state of the art in this area. 2. PRELIMINARIES In this section, we present basic definitions that will be used throughout the paper. Given a network or graph G = {E, V }, V is a set of n nodes and E is a set of m edges. For dense graphs m = O(n2), but for sparse networks m = O(n). The network structure is determined by the n × n adjacency matrix A for unweighted networks and weight matrix W for weighted networks. Each element Aij of A is equal to 1 if there is an edge connecting nodes i and j; and it is 0 otherwise. Each element wij of W takes a nonnegative real value representing strength of connection between nodes i and j. In the case of overlapping community detection, the set of clusters found is called a cover C = {c1, c2, · · · , ck} [Lancichinetti et al. 2009], in which a node may belong to more than one cluster. Each node i associates with a community according to a belonging factor (i.e., soft assignment or membership) [ai1, ai2, · · · , aik] [Nepusz et al. 2008], in which aic is a measure of the strength of association between node i and cluster c. Without loss of generality, the following constraints are assumed to be satisfied 0 ≤ aic ≤ 1 ∀i ∈ V, ∀c ∈ C (1) and C aic = 1, Xc=1 where C is the number of clusters. However, the belonging factor is often solely a set of artificial weights. It may not have a clear or unambiguous physical meaning [Shen et al. 2009]. In general, algorithms produce results that are composed of one of two types of assignments, crisp (non-fuzzy) assignment or fuzzy assignment [Gregory 2011]. With crisp assignment, the relationship between a node and a cluster is binary. That is, a node i either belongs to cluster c or does not. With fuzzy assignment, each node is associated with communities in proportion to a belonging factor. With , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 3 a threshold, a fuzzy assignment can be easily converted to a crisp assignment. Most detection algorithms output crisp community assignments. 3. ALGORITHMS In this section, algorithms for overlapping community detection are reviewed and categorized into five classes which reflect how communities are identified. 3.1 Clique Percolation The clique percolation method (CPM) is based on the assumption that a community consists of overlapping sets of fully connected subgraphs and detects communities by searching for adjacent cliques. It begins by identifying all cliques of size k in a network. Once these have been identified, a new graph is constructed such that each vertex represents one of these k-cliques. Two nodes are connected if the k- cliques that represent them share k−1 members. Connected components in the new graph identify which cliques compose the communities. Since a vertex can be in multiple k-cliques simultaneously, overlap between communities is possible. CPM is suitable for networks with dense connected parts. Empirically, small values of k (typically between 3 and 6) have been shown to give good results [Palla et al. 2005; Lancichinetti and Fortunato 2009; Gregory 2010]. CFinder2 is the implementation of CPM, whose time complexity is polynomial in many applications [Palla et al. 2005]. However, it also fails to terminate in many large social networks. CPMw [Farkas et al. 2007] introduces a subgraph intensity threshold for weighted networks. Only k-cliques with intensity larger than a fixed threshold are included into a community. Instead of processing all values of k, SCP [Kumpula et al. 2008] finds clique communities of a given size. In the first phase, SCP detects k-cliques by checking all the (k−2)-cliques in the common neighbors of two endpoints when links are inserted to the network sequentially in the order of decreasing weights. In the second phase, the k-community is detected by finding the connected components in the (k − 1)-clique projection of the bipartite representation, in which one type of node represents a k-clique and the other denotes a (k − 1)-clique. Since each k-clique is processed exactly twice, the running time grows linearly as a function of the number of cliques. SCP allows multiple weight thresholds in a single run and is faster than CPM. Despite their conceptual simplicity, one may argue that CPM-like algorithms are more like pattern matching rather than finding communities since they aim to find specific, localized structure in a network. 3.2 Line Graph and Link Partitioning The idea of partitioning links instead of nodes to discover community structure has also been explored. A node in the original graph is called overlapping if links connected to it are put in more than one cluster. In [Ahn et al. 2010]3, links are partitioned via hierarchical clustering of edge similarity. Given a pair of links eik and ejk incident on a node k, a similarity can 2http://www.cfinder.org. 3https://github.com/bagrow/linkcomm. , Vol. V, No. N, May 2014. 4 · Jierui Xie et al. be computed via the Jaccard Index defined as S(eik, ejk) = Ni ∩ Nj Ni ∪ Nj , where Ni is the neighborhood of node i including i. Single-linkage hierarchical clustering is then used to build a link dendrogram. Cutting this dendrogram at some threshold yields link communities. The time complexity is O(nk2 max), where kmax is the maximum node degree in the network. Evans [Evans and Lambiotte 2009; 2010] projected the network into a weighted line graph, whose nodes are the links of the original graph. Then disjoint community detection algorithms can be applied. The node partition of a line graph leads to an edge partition of the original graph. CDAEO [Wu et al. 2010] provides a post- processing procedure to determine the extent of overlapping. Once the preliminary partitioning on the line graph is done, for a node i with Eicmin/Eicmax below some predefined threshold, where Eicmin(cmax) is the set of edges in the community with which i has the minimum (maximum) number of connections, links in Eicmin of the line graph are removed. This essentially reduces node i to a single membership. Kim [Kim and Jeong 2011] extended the map equation method (also known as Infomap [M. Rosvall 2008]) to the line graph, which encodes the path of the random walk on the line network under the Minimum Description Length (MDL) principle. Line graph has been extended to clique graph [Evans 2010], wherein cliques of a given order are represented as nodes in a weighted graph. The membership strength of a node i to community c is given by the fraction of cliques containing i which are assigned to c. Although the link partitioning for overlapping detection seems conceptually nat- ural, there is no guarantee that it provides higher quality detection than node based detection does [Fortunato 2010] because these algorithms also rely on an ambiguous definition of community. Note that a link-based extended modularity is also proposed by Nicosia in [Nicosia et al. 2009]. This measure is built on the belonging coefficients of links. Let a link l(i, j) connecting i to j for community c be βl(i,j),c = F (aic, ajc), then the expected belonging coefficient of any possible link l(i, j) from node i to a node j in community c can be defined as βout expected belonging coefficient of any link l(i, j) pointing to node j in community c is defined as βin used as weights for the probability of an observed link (first term in (2)) and the probability of a link starting from i to j in the null model (second term in (2)), respectively, resulting in the new modularity defined as V Pj∈V F (aic, ajc). Accordingly, the V Pi∈V F (aic, ajc). The above belonging coefficients are l(i,j),c = 1 l(i,j),c = 1 QN i ov = 1 mXc Xi,j∈V "βl(i,j),cAi,j − βout l(i,j),cβin l(i,j),c kout i kin j m # , (2) i where kout(in) tal number of edges. Note that QN i F (aic, ajc), which could be the product, average, or maximum of aic and ajc. is the number of outgoing (incoming) links of i and m is the to- ov depends on the link belonging coefficient , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 5 3.3 Local Expansion and Optimization Algorithms utilizing local expansion and optimization are based on growing a nat- ural community [Lancichinetti et al. 2009] or a partial community. Most of them rely on a local benefit function that characterizes the quality of a densely connected group of nodes. Baumes [Baumes et al. 2005] proposed a two-step process. First, the algorithm RankRemoval is used to rank nodes according to some criterion. Then the pro- cess iteratively removes highly ranked nodes until small, disjoint cluster cores are formed. These cores serve as seed communities for the second step of the process, Iterative Scan (IS), that expands the cores by adding or removing nodes until a local density function cannot be improved. The proposed density function can be formally given as f (c) = wc in wc in + wc out , in and wc where wc out are the total internal and external weight of the community c. The worst-case running time is O(n2). The quality of discovered communities depends on the quality of seeds. Since the algorithm allows vertices to be removed during the expansion, IS has been shown to produce disconnected components in some cases. For this reason, a modified version called CIS was introduced in [Kelley 2009], wherein the connectedness is checked after each iteration. In the case that the community is broken into more than one part, only the one with the largest density is kept. CIS also develops a new fitness function f (c) = wc in wc in + wc out + λep incorporating the edge probability ep. The parameter λ controls how the algorithm behaves in sparse areas of the network. The addition of a node needs to strike a balance between the change in the internal degree density and the change in edge density. LFM [Lancichinetti et al. 2009] expands a community from a random seed node to form a natural community until the fitness function f (c) = kc in in + kc (kc out)α (3) in and kc is locally maximal, where kc out are the total internal and external degree of the community c, and α is the resolution parameter controlling the size of the communities. After finding one community, LFM randomly selects another node not yet assigned to any community to grow a new community. LFM depends significantly on the resolution parameter α. The computational complexity for a fixed α-value is roughly O(ncs2), where nc is the number of communities and s is the average size of communities. The worst-case complexity is O(n2). MONC [Havemann et al. 2011] uses the modified fitness function of LFM f (c) = kc in + 1 in + kc out)α , (kc which allows a single node to be considered a community by itself. This avoids violation of the principle of locality. The proposed fitness function enables MONC , Vol. V, No. N, May 2014. 6 · Jierui Xie et al. to find the range of αs (resolution parameter as in LFM) for which a set of nodes is locally optimal. Rather than numerical exploration of these α values, MONC calculates the next lowest value of α which results in further expansion and contin- ues to expand the community. In the case that the natural community of a node i is a subset of another node, the analysis of i stops. In this way, MONC merges communities during processing and, as a result, uncovers the network faster than LFM. OSLOM4 [Lancichinetti et al. 2011] tests the statistical significance of a cluster [Bianconi et al. 2008] with respect to a global null model (i.e., the random graph generated by the configuration model [Molloy and Reed 1995]) during community expansion. To grow the current community, the r value is computed for each neighbor, which is the cumulative probability of having the number of internal connections equal or larger than the number of connections from a neighbor into this community in the null model. If the cumulative distribution of the smallest r value is smaller than a given tolerance, it is considered to be significant, and the corresponding node is added to the community. Otherwise, the second smallest r is checked and so on. OSLOM usually results in a significant number of outliers or singleton communities. The worst-case complexity in general is O(n2), while the exact complexity depends on the community structure of the underlying network being studied. Rather than considering the original network, UEOC [Jin et al. 2011] unfolds the community of a node based on the l-step transition probability of the random walk on the corresponding annealed network [Newman et al. 2001], which represents an ensemble of networks. After sorting nodes according to the transition probabilities in descending order, the natural community is extracted with some proper cutoff. The dominating time complexity is for calculating the transition matrix, which is O(ln2). OCA [Padrol-Sureda et al. 2010] is based on the idea of mapping each node to a d- dimensional vector. Each subset of nodes S is then defined as the sum of individual vectors in this set. The fitness function is defined as the directed Laplacian on function O, where O is the squared Euclidean length of a subset vector. Like LFM, starting from some initial seeds, OCA tries to remove or add a node that results in the largest increase in the value of the fitness function. OCA requires finding the most negative eigenvalue of the adjacency matrix. Chen [Chen et al. 2010] proposed selecting a node with maximal node strength based on two quantities B(u, c) (called belonging degree) and the modified modu- larity Qov for weighted networks. Qov is defined as QC ov = 1 2mXc Xi,j∈V (cid:20)Aij − kikj 2m (cid:21) βicβjc, (4) where βic = kic/Pc′ kic′ is the strength with which node i belongs to community c, and kic =Pj∈c wij is the total weight of links from i into community c. B(u, c) measures how tightly a node u connects to a given community c compared to the rest of the network. Given two thresholds BU and BL, when expanding a 4http://www.oslom.org. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 7 community c, neighboring nodes with B(u, c)>BU are included in c. For nodes with BL ≤ B(u, c) ≤ BU , if Qov increases after adding such a node, u is added to c. The drawbacks of this algorithm are the rather arbitrary selection of the BU and BL thresholds and the expensive computation of Qov whose complexity is O(kn2), where k is the number of communities. iLCD5 [Cazabet et al. 2010] is capable of detecting both static and temporal communities. Given a set of edges created at some time step, iLCD updates the existing communities by adding a new node if its number of second neighbors and number of robust second neighbors are greater than expected values. New edges are also allowed to create a new community if the minimum pattern is detected. Defining the similarity between two communities as the ratio of nodes in common, a merging procedure is performed to improve the detection quality if the similarity is high. iLCD relies on two parameters for adding a node and merging two com- munities. The complexity of iLCD is O(nk2) in general, whose precise quantity depends on community structures and its parameters. Seeds are very important for many local optimization algorithms. A clique has been shown to be a better alternative over an individual node as a seed, serving as the basis for a wide range of algorithms. EAGLE [Shen et al. 2009] uses the agglomerative framework to produce a dendrogram. First, all maximal cliques are found and made to be the initial communities. Then, the pair of communities with maximum similarity is merged. The optimal cut on the dendrogram is determined by the extended modularity with a weight based on the number of overlapping memberships in [Shen et al. 2009]. Even without taking into account the time required to find all the maximal cliques, EAGLE is still computationally expensive with complexity O(n2 + (h + n)s), where s is the number of maximal cliques whose upper bound is 3n/3 (i.e., theoretically exponential) [Moon and Moser 1965], and h is the number of pairs of maximal cliques which are neighbors. This paper also defines an extended modularity that uses the number of communities to which a node belongs as a weight for Q as QE ov = 1 2mXc Xi,j∈c(cid:20)Aij − kikj 2m (cid:21) 1 OiOj , (5) where Oi is the number of communities to which node i belongs. This measure is in the same form as (8), but with a coefficient defined based on the maximal clique. One may argue that they are identical as in [Gregory 2011]. Similar to EAGLE, GCE6 [Lee et al. 2010] identifies maximum cliques as seed communities. It expands these seeds by greedily optimizing a local fitness function. GCE also removes communities that are similar to previously discovered using distance between communities c1 and c2 defined as 1 − c1 ∩ c2 min(c1, c2) . If this distance is shorter than a parameter ǫ, the communities are similar. The time complexity for greedy expansion is O(mh), where m is the number of edges, 5http://cazabetremy.fr/Cazabet_remy/iLCD.html. 6https://sites.google.com/site/greedycliqueexpansion. , Vol. V, No. N, May 2014. 8 · Jierui Xie et al. and h is the number of cliques. In COCD [Du et al. 2008], cores are a set of independent maximal cliques induced on each vertex. Two maximal cliques are said to be dependent if their closeness function is positive. This function is a product of the differences between the size of internal links between two maximal cliques and the number of links connect- ing nodes appearing only in one of the two maximal cliques. Once the cores are identified, the remaining nodes are attached to cores with which they have maxi- mum connections. COCD runs in O(Cmax · T ri2) in the worst case, where Cmax is the maximum size of the detected communities, and T ri is the number of trian- gles, whose lower bound is 9mn−2n3−2(n2−3m)3/2 [Fisher 1989] or O(n3) for a dense enough graph. 27 3.4 Fuzzy Detection Fuzzy community detection algorithms quantify the strength of association between all pairs of nodes and communities. In these algorithms, a soft membership vector, or belonging factor [Gregory 2010], is calculated for each node. A drawback of such algorithms is the need to determine the dimensionality k of the membership vector. This value can be either provided as a parameter to the algorithm or calculated from the data. Nepusz [Nepusz et al. 2008] modeled the overlapping community detection as a nonlinear constrained optimization problem which can be solved by simulated annealing methods. The objective function to minimize is f = n n Xi=1 Xj=1 wij ( sij − sij)2, (6) where wij denotes the predefined weight, sij is the prior similarity between nodes i and j, and the similarity sij is defined as aicajc, (7) sij =Xc where the variable aic is the fuzzy membership of node i in community c, subject to the total membership degree constraint in (1) and a non-empty community constraint. To determine the number of communities k, the authors increased the value of k until the community structure does not improve as measured by a modified fuzzy modularity, which, by weighting Q with the product of a node’s belonging factor, is defined as QN e ov = 1 2mXc Xi,j∈c(cid:20)Aij − kikj 2m (cid:21) aicajc, (8) where aic is the degree of membership of node i in the community c. Zhang [Zhang et al. 2007a] proposed an algorithm based on the spectral clustering framework [Newman 2006; White and Smyth 2005]. Given an upper bound on the number of communities k, the top k − 1 eigenvectors are computed. The network is then mapped into a d-dimensional Euclidean space, where d ≤ k − 1. Instead of using k-means, fuzzy c-means (FCM) is used to obtain a soft assignment. Both detection accuracy and computation efficiency rely on the user specified value k. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 9 With running time O(mkh + nk2h + k3h) + O(nk2), where m is the number of edges, n is the number of nodes, the first term is for the implicitly restarted Lanczos method, and the second term is for FCM, it is not scalable for large networks. An extended modularity that used the average of the belonging factor is also proposed as c , V ) −(cid:18) A(V ′ A(V, V ) (cid:19)2#, QZ c , V ′ c ) A(V, V ) ov =Xc " A(V ′ c ) = Pi,j∈V ′ c , V ′ c where V ′ ing nodes i and j, A(V ′ c is the set of nodes in a community c, wij is the weight of the link connect- c ) + wij (aic + ajc)/2, A(V ′ c , V ) = A(V ′ c , V ′ c ,j∈V \V ′ c Pi∈V ′ wij(aic + (1 − ajc))/2, and A(V, V ) =Pi,j∈V wij . Due to their probabilistic nature, mixture models provide an appropriate frame- work for overlapping community detection [Newman and Leicht 2007]. In general, the number of mixture models is equal to the number of communities, which needs In SPAEM7 [Ren et al. 2009], the mixture model is to be specified in advance. viewed as a generative model for the links in the network. Suppose that πr is the probability of observing community r and community r selects node i with proba- bility Br,i. For each r, Br,i is a multinomial across elements i = 1, 2, · · · , n, where i=1 Br,i = 1. The edge probability eij generated by such finite mixture model is given by n is the number of nodes. Therefore, Pn Xr=1 p(eijπ, B) = k πrBr,iBr,j. The total probability over all the edges present in the network is maximized by the Expectation-Maximization (EM) algorithm. As in [Kim and Jeong 2011], the optimal number of communities k is identified based on the minimum description length. There is another algorithm called FOG8 [Davis and Carley 2008] also trying to infer groups based on link evidence. Similar mixture models can also be constructed as a generative model for nodes [Fu and Banerjee 2008]. In SSDE9 [Magdon-ismail and Purnell 2011], the network is first mapped into a d-dimensional space using the spectral clustering method. A Gaussian Mixture Model (GMM) is then trained via Expectation-Maximization algorithm. The number of communities is determined when the increase in log- likelihood of adding a cluster is not significantly higher than that of adding a cluster to random data which is uniform over the same space. Stochastic block model (SBM) [Nowicki and Snijders 2001] is another type of generative model for groups in the network. Fitting an empirical network to a SBM requires inferring model parameters similar to GMM. In OSBM [Latouche et al. 2011], each node i is associated with a latent vector (i.e., community assignment) Zi with K independent Boolean variables Zik ∈ {0, 1}, where K is the number of communities, and Zik is drawn from a multivariate Bernoulli distribution. Z is inferred by maximizing the posterior probability conditioned on the present of 7http://www.code.google.com/p/spaem. 8http://www.casos.cs.cmu.edu/projects/ora. 9http://www.cs.rpi.edu/~purnej/code.php. , Vol. V, No. N, May 2014. 10 · Jierui Xie et al. edges as in [Ren et al. 2009]. OSBM requires more efforts than mixture models because the factorization in the observed condition distribution for edges given Z is in general intractable. MOSES10 [McDaid and Hurley 2010] combines OSBM with the local optimization scheme, in which the fitness function is defined based on the observed condition distribution. MOSES greedily expands a community from edges. Unlike OSBM, no connection probability parameters are required as input. The worst-case time complexity is O(en2), where e is the number of edges to be expanded. Non-negative Matrix Factorization (NMF) is a feature extraction and dimension- ality reduction technique in machine learning that has been adapted to community detection. NMF approximately factorizes the feature matrix V into two matrices with the non-negativity constraint as V ≈ W H, where V is n × m, W is n × k , H is k × m, and k is the number of communities provided by users. W represents the data in the reduced feature space. Each element wi,j in the normalized W quanti- fies the dependence of node i with respect to community j. In [Zhang et al. 2007b], V is replaced with the diffusion kernel, which is a function of the Laplacian of the network. In [Zarei et al. 2009], V is defined as the correlation matrix of the columns of the Laplacian. This results in better performance than [Zhang et al. 2007b]. In [Zhao et al. 2010], redundant constraints in the approximation are removed, reduc- ing NMF to a problem of symmetrical non-negative matrix factorization (s-NMF). Psorakis [Psorakis et al. 2011] proposed a hybrid algorithm called Bayesian NMF11. The matrix V , where each element vij denotes a count of the interactions that took place between two nodes i and j, is decomposed via NMF as part of the param- eter inference for a generative model similar to OSBM and GMM. Traditionally, NMF is inefficient with respect to both time and memory constraints due to the matrix multiplication. In the version of [Psorakis et al. 2011], the worst-case time complexity is O(kn2), where k denotes the number of communities. Wang et al. [Wang et al. 2009] combined disjoint detection methods with local optimization algorithms. First, a partition is obtained from any algorithm for disjoint community detection. Communities attempt to add or remove nodes. The difference, called variance, of two fitness scores on a community, either including a node i or removing node i, is computed. The normalized variances form a fuzzy membership vector of node i. Ding [Ding et al. 2010] employed the affinity propagation clustering algorithm [Frey and Dueck 2007] for overlapping detection, in which clusters are identified by representative exemplars. First, nodes are mapped as data points in the Euclidean space via the commute time kernel (a function of the inverse Laplacian). The simi- larity between nodes is then measured by the cosine distance. Affinity propagation reinforces two types of messages associated with each node, the responsibility r(i, k) and the availability a(i, k). The probability for assigning node i into the cluster represented by exemplar node k is computed by equation p(i, k) = er(i,k), where r is the normalized responsibility as in [Geweniger et al. 2009]. 10http://sites.google.com/site/aaronmcdaid/moses. 11http://www.robots.ox.ac.uk/~parg/software.html. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 11 3.5 Agent-Based and Dynamical Algorithms The label propagation algorithm [Raghavan et al. 2007; Xie and Szymanski 2011], in which nodes with same label form a community, has been extended to overlap- ping community detection by allowing a node to have multiple labels. In COPRA12 [Gregory 2010], each node updates its belonging coefficients by averaging the co- efficients from all its neighbors at each time step in a synchronous fashion. The parameter v is used to control the maximum number of communities with which a node can associate. The time complexity is O(vm log(vm/n)) per iteration. SLPA13 [Xie et al. 2011; Xie and Szymanski 2012] is a general speaker-listener based information propagation process. It spreads labels between nodes according to pairwise interaction rules. Unlike [Raghavan et al. 2007; Gregory 2010], where a node forgets knowledge gained in the previous iterations, SLPA provides each node with a memory to store received information (i.e., labels). The probability of observing a label in a node’s memory is interpreted as the membership strength. SLPA does not require any knowledge about the number of communities, which is determined by the clustering of labels in the network. The time complexity is O(tm), linear in the number of edges m, where t is a predefined maximum number of iterations (e.g., t ≥ 20). SLPA can also be adapted for weighted and directed networks by generalizing the interaction rules, known as SLPAw. A game-theoretic framework is proposed in [Chen et al. 2010], in which a commu- nity is associated with a Nash local equilibrium. A gain function and a loss function are associated with each agent. The game assumes that each agent is selfish and selects to join, leave and switch communities based on its own utility. An agent is allowed to joint multiple communities to handle overlapping, so long as it results in increased utility. The time complexity to find the best local operation for an agent i is O(Li · L(Ni) · ki), where Li is the communities that agent i wants to joint, L(Ni) is the set of communities that i’s neighbors want to joint, and ki is the node degree. The time takes to reach a local equilibrium is bounded by O(m2), where m is the number of edges. A process in which particles walk and compete with each other to occupy nodes is presented in [Breve et al. 2009]. Particles represent different communities. Each node has an instantaneous ownership vector (similar to belonging factor) and a long term ownership vector. At each iteration, each particle takes either a random walk or a deterministic walk to one of its neighbors with some probability. If the random walk is performed, the visited neighbor updates its instantaneous ownership vector; otherwise, the long term ownership vector is updated. At the end of the process, the long term ownership vector is normalized to produce a soft assignment. Different from SLPA and COPRA, this algorithm takes a semi-supervised approach. It requires at least one labeled node per class. Multi-state spin models [Reichardt and Bornholdt 2004; Lu et al. 2009], in which a spin is assigned to each node, can also be applied to community detection. One of such models is q-state Potts model [Blatt et al. 1996; Reichardt and Bornholdt 2004], where q is the number of states that a spin may take, indicating the maximum 12http://www.cs.bris.ac.uk/~steve/networks/software/copra.html. 13https://sites.google.com/site/communitydetectionslpa. , Vol. V, No. N, May 2014. 12 · Jierui Xie et al. number of communities. The community detection problem is equivalent to the problem of minimizing the Hamiltonian of the model. In the ground states (i.e., local minima of the Hamiltonian), the set of nodes with the same spin state form a community. The overlap of communities is linked to the degeneracy of the minima of the Hamiltonian [Reichardt. and Bornholdt 2006]. Although a co-appearance matrix keeps track of how frequently nodes i and j have been grouped together over multiple runs, it is not clear how to aggregate this information into overlapping communities when analyzing large networks. Another Potts model-like approach was proposed in [Ronhovde and Nussinov 2009] to evaluate the hierarchical or multiresolution structure of a graph via information-based replica correlations. Synchronization of a system that consists of coupled phase oscillators is able to uncover community structures. In such a model (e.g., the Kuramoto model) the phase of each unit evolves in time according to the predefined dynamics. The set of nodes with the same phase or frequency can be viewed as a community [Arenas et al. 2006] while nodes that do not match any observed dynamic behaviors can be considered overlapping nodes [Li et al. 2008]. Like methods utilizing a Potts model, such algorithms are parameter dependent. 3.6 Others CONGA14 [Gregory 2007] extends Girvan and Newman’s divisive clustering algo- rithm (GN) [Girvan and Newman 2002] by allowing a node to split into multiple copies. Both splitting betweenness, defined by the number of shortest paths on the imaginary edge, and the conventional edge betweenness are considered. CONGA in- herits the high computational complexity of GN. In a more refined version, CONGO [Gregory 2008] uses local betweenness to optimize the speed. Gregory [Gregory 2009] also proposed to perform disjoint detection algorithms on the network pro- duced by splitting the node into multiple copies using the split betweenness. Zhang15 [Zhang et al. 2009] proposed an iterative process that reinforces the network topology and propinquity that is interpreted as the probability of a pair of nodes belonging to the same community. The propinquity between two vertices is defined as the sum of the number of direct links, number of common neighbors and the number of links within the common neighborhood. Given the topology, propinquity is computed. Propinquity above a certain threshold is then used to redistribute links, updating the topology. If the propinquity is large, a link is added to the network; otherwise, the link is removed. The propinquity can be used to perform micro clustering on each vertex to allow overlap. Kov´acs et al. [Kov´acs et al. 2010] proposed an approach focusing on centrality- based influence functions. Community structures are interpreted as hills of the influence landscape. For each node i, the influence over each link fi(j, k) is com- puted. Links within a community should have higher influence than those linking distant areas of the network. The influence on a given link c(j, k) is the sum of fi(j, k) over all nodes. The function c(j, k) over each link defines the community landscape, wherein the communities are determined by local maxima and their surrounding regions. 14http://www.cs.bris.ac.uk/~steve/networks/software/conga.html. 15http://dbgroup.cs.tsinghua.edu.cn/zhangyz/kdd09. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 13 Rees [Rees and Gallagher 2010] proposed an algorithm to extract the overlapping communities from the egonet, which is a subgraph including a center node, its neighbors, and the links around them. When all egonets are induced, each center is removed, creating small connected components among neighbors. Then, the center node is added back to each of these components to form so-called friendship group. Clearly, each center node can be in multiple friendship groups. The overlapping communities are determined by merging all friendship groups. Inspired by OPTIC [Ankerst et al. 1999], an algorithm based on techniques from visualization was proposed in [Chen et al. 2009]. Nodes are ordered according to the reachability score (RS) with respect to a starting node. The reachability is based on the probability of the existence of a link between two nodes. By scanning through the obtained sequence of nodes, a community containing consecutive nodes with RS larger than a community threshold is found. Clearly, this algorithm is hard to apply to large networks and requires the introduction of a community threshold. 4. EVALUATION CRITERIA Evaluating the quality of a detected partitioning or cover is nontrivial, and ex- tending evaluation measures from disjoint to overlapping communities is rarely straightforward. Unlike disjoint community detection, where a number of measures have been proposed for comparing identified partitions with the known partitions [Danon et al. 2005; Leskovec et al. 2010], only a few measures are suitable for a set of overlapping communities. Two most widely used measures are the Normalized Mutual Information (NMI) and Omega Index. 4.1 Normalized Mutual Information Lancichinetti [Lancichinetti et al. 2009] has extended the notion of normalized mu- tual information to account for overlap between communities. For each node i in cover C′, its community membership can be expressed as a binary vector of length C′ (i.e., the number of clusters in C′). (xi)k = 1 if node i belongs to the kth cluster C′ k; (xi)k = 0 otherwise. The kth entry of this vector can be viewed as a random variable Xk, whose probability distribution is given by P (Xk = 1) = nk/n, P (Xk = 0) = 1 − P (Xk = 1), where nk = C′ k is the number of nodes in the cluster C′ k and n is the total number of nodes. The same holds for the random variable Yl associated with the lth cluster in cover C′′. Both the empirical marginal probability distribution P (Xk) and the joint probability distribution P (Xk, Yl) are used to further define entropy H(X) and H(Xk, Yl). The conditional entropy of a cluster Xk given Yl is defined as H(XkYl) = H(Xk, Yl) − H(Yl). The entropy of Xk with respect to the entire vector Y is based on the best matching between Xk and any component of Y given by H(XkY ) = minl∈{1,2,··· ,C ′′}H(XkYl). The normalized conditional entropy of a cover X with respect to Y is H(XY ) = H(XkY ) H(Xk) . 1 C′Xk In the same way, one can define H(Y X). Finally the NMI for two covers C′ and , Vol. V, No. N, May 2014. 14 · Jierui Xie et al. C′′ is given by N M I(XY ) = 1 − [H(XY ) + H(Y X)] /2. (9) The extended NMI is between 0 and 1, with 1 corresponding to a perfect matching. Note that this modified NMI does not reduce to the standard formulation of NMI when there is no overlap. 4.2 Omega Index Omega Index [Collins and Dent 1988] is the overlapping version of the Adjusted Rand Index (ARI) [Hubert and Arabie 1985]. It is based on pairs of nodes in agreement in two covers. Here, a pair of nodes is considered to be in agreement if they are clustered in exactly the same number of communities (possibly none). That is, the omega index considers how many pairs of nodes belong together in no clusters, how many are placed together in exactly one cluster, how many are placed in exactly two clusters, and so on. Let K1 and K2 be the number of communities in covers C1 and C2, respectively, the omega index is defined as [Gregory 2011; Havemann et al. 2011] ω(C1, C2) = ωu(C1, C2) − ωe(C1, C2) 1 − ωe(C1, C2) . (10) The unadjusted omega index ωu is defined as ωu(C1, C2) = 1 M max(K1,K2) Xj=0 tj(C1) ∩ tj(C2), where M equal to n(n − 1)/2 represents the number of node pairs and tj(C) is the set of pairs that appear exactly j times in a cover C. The expected omega index in the null model ωe is given by ωe(C1, C2) = 1 M 2 max(K1,K2) Xj=0 tj(C1) · tj(C2). The subtraction of the expected value in (10) takes into account agreements resulting from chance alone. The larger the omega index is, the better the matching is between two covers. A value of 1 indicates perfect matching. When there is no overlap, the omega index reduces to the ARI. In addition to NMI and Omega, some other measures have been proposed, such as the generalized external indexes [Campello 2007; Campello. 2010] and the fuzzy rand index [Hullermeier and Rifqi 2009]. 5. BENCHMARKS It is necessary to have good benchmarks to both study the behavior of a proposed community detection algorithm and to compare the performance across various al- gorithms. In order to accurately perform these two analyses, networks in which the ground truth is known are needed. This requirement implies that real-world networks, which are often collected from online or observed interactions, do not paint a clear enough picture due to their lack of “ground truth”. In light of this re- quirement, we begin our discussion with synthetic networks. In the GN benchmark , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 15 [Girvan and Newman 2002], equal size communities are embedded into a network for a given expected degree and a given mixing parameter µ that measures the ratio of internal connections to outgoing connections. One drawback of this benchmark is that it fails to account for the heterogeneity in complex networks. Another is that it does not allow embedded communities to overlap. A few benchmarks have been proposed for testing overlapping community detection algorithms, all of which are special cases of the planted l -partition model [Condon and Karp 2001] just like GN. Sawardecker [Sawardecker et al. 2009] proposed an extension of GN, in which the probability pij of an edge being present in the network is a non-decreasing function based solely on the set of co-memberships of nodes i and j. With the definition pij = pk, parameter pk is the connection probability of nodes i and j that co-occur k times, subject to p0 < p1 ≤ p2 ≤ · · · . The LFR16 benchmark proposed in [Lancichinetti et al. 2008] introduces hetero- geneity into degree and community size distributions of a network. These distri- butions are governed by power laws with exponents τ1 and τ2, respectively. To generate overlapping communities, On, the fraction of overlapping nodes is speci- fied and each node is assigned to Om ≥ 1 communities. The generating procedure is equivalent to generating a bipartite network where the two classes are the com- munities and nodes subject to the requirement that the sum of community sizes equals the sum of node memberships. LFR also provides a rich set of parameters to control the network topology, including the mixing parameter µ, the average degree k, the maximum degree kmax, the maximum community size cmax, and the minimum community size cmin. The LFR model brings benchmarks closer to the features observed in real-world networks. However, requiring that overlapping nodes interact with the same number of embedded communities is unrealistic in practice. A simple generalization, where each overlapping node may belong to different number of communities has been considered in [McDaid and Hurley 2010]. In [Gregory 2011], crisp communities from LFR are converted to fuzzy associa- tions by adding a belonging coefficient to the occurrence of nodes. This coefficient can be defined as pij = sij p1 + (1 − sij)p0, where pk is the same as in Sawardecker’s model and sij = Pc∈C αicαjc is the similarity of node i and j as defined in (6). In other words, the probability of an edge being present depends not only on the number of communities in which nodes i and j appear together but also on their degree of belonging to these communities. 6. TESTS ON SYNTHETIC NETWORKS In this section, we empirically compare the performance of different algorithms on LFR networks. We focus on algorithms which produce a crisp assignment of vertices to communities. In total, 14 algorithms were collected and tested. They are listed in Table I. Note that the time complexity given is for the worst case. 16http://sites.google.com/site/andrealancichinetti/files. , Vol. V, No. N, May 2014. 16 · Jierui Xie et al. Algorithm Table I. Algorithms included in the experiments. Complexity Reference CFinder LFM EAGLE CIS GCE COPRA Game NMF MOSES Link iLCD UEOC OSLOM SLPA [Palla et al. 2005] [Lancichinetti et al. 2009] [Shen et al. 2009] [Kelley 2009] [Lee et al. 2010] [Gregory 2010] [Chen et al. 2010] [Psorakis et al. 2011] [McDaid and Hurley 2010] [Ahn et al. 2010] [Cazabet et al. 2010] [Jin et al. 2011] [Lancichinetti et al. 2011] [Xie et al. 2011; Xie and Szymanski 2012] - O(n2) O(n2 + (h + n)s) O(n2) O(mh) O(vm log(vm/n)) O(m2) O(kn2) O(en2) O(nk2 max) O(nk2) O(ln2) O(n2) O(tm) Imp C++ C++ C++ C++ C++ Java C++ Matlab C++ C++ Java Matlab C++ C++ For algorithms with tunable parameters, the results with the best setting are re- ported. For LFM, we varied α from 0.8 to 1.6 with an interval 0.1, within which good results have previously been reported [Lancichinetti et al. 2009; Lee et al. 2010]. For iLCD, f Ratio is from {0.75, 0.5, 0.35} and bT hreshold is from {0.5, 0.3, 0.2} as suggested by the authors. For GCE, the minimum clique size k ranges from 3 to 8. For CFinder, k ranges from 3 to 8. For OSLOM, we considered the first two levels. For Link, the threshold varies from 0.1 to 0.9 with an interval 0.1. For COPRA, parameter v is taken from the range [1,10]. For SLPA, parameter r varies from 0.05 to 0.5 with an interval 0.05. Since COPRA and SLPA are non-deterministic, we repeated each of them 10 times on each network instantiation. For NMF, which returns a fuzzy assignment, we applied the same threshold as SLPA to convert it to a crisp assignment. For each parameter set generated via LFR, we generated 10 instantiations. We used networks with sizes n ∈ {1000, 5000}. The average degree is kept at k = 10, which is of the same order as most large real-world social networks17. The rest of the parameters of LFR generator are set similar to those in [Lancichinetti and Fortunato 2009]: node degrees and community sizes are governed by power law distributions with exponents τ1 = 2 and τ2 = 1 respectively, the maximum degree is kmax = 50, and community sizes vary in both small range s = (10, 50) and large range b = (20, 100). The mixing parameter µ is from {0.1, 0.3}, which is the expected fraction of links through which a node connects to other nodes in the same community. The degree of overlap is determined by two parameters. On is the number of overlapping nodes, and Om is the number of communities to which each overlapping node belongs. On is set to 10% and 50% of the total number of nodes, indicating low and high overlapping density respectively. Instead of fixing Om [Lancichinetti and Fortunato 2009; Gregory 2010], we also allow it to vary from 2 to 8 indicating the overlapping diversity of overlapping nodes. By increasing the value of Om, we create harder detection tasks. This also allow us to look in greater detail at the detection accuracy at the node level. 17snap.stanford.edu/data. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 17 Two previously discussed measures, Omega and NMI, are used to quantify the quality of the cover discovered by an algorithm. 6.1 Effects of µ, n and Om We first examined how the performance, measured by NMI, changes as the number of memberships Om varies from small to large values (i.e., 2 to 8) for different network sizes (n) and intra community strength (µ) in Figure 1. In general, changes in the network topology, especially the mixing value µ, have a similar impact for disjoint community detection. That is, the larger the value of µ, the poorer the results produced by detection algorithms (i.e., red curve < blue curve in Figure 1) due to the fact that the connection inside communities is weak for larger µ . This is true for the majority of algorithms with the only exception NMF (see the 5000b case). On the other hand, increasing network size from 1000 to 5000 typically results in slightly better performance (i.e., square > dot in Figure 1), with prominent exceptions for EAGLE, NMF and UEOC. Slight fluctuations in performance are observed for iLCD and Link. Detection performance typically decays at a moderate rate as the diversity of overlapping increases (i.e., Om getting larger), except for OSLOM and UEOC. 6.2 Effects of Community Size Range and Overlapping Density On We evaluated the effects of On and community size ranges on each individual al- gorithm on networks with n = 5000 and µ = 0.3. Results for NMI are shown in Figure 2. As expected, the performance of detection consistently and significantly drops in the case there are many overlapping nodes for all algorithms (i.e., red curves (On = 50%) < blue curves (On = 10%)). However, the difference in performance between small and large community size ranges (gaps between two curves with the same color) is more prominent in the case of low overlapping density. Interestingly, the NMI’s for networks with small community size s = (10, 50) are typically higher than those for networks with large community size range b = (20, 100) (i.e., dot > square in Figure 2). It appears that the well known resolution limit does not play a role here since all the tested algorithms are neither based on a modularity nor an extended modularity. This is evidenced by algorithms including CFinder, LFM, Link, MOSES, Game, iLCD, CIS and OSLOM that have a significant performance gap between small and large community size ranges. Given only small variances in performance in two tested ranges, we also conclude that the community size range has limited impact on algorithms including SLPA, COPRA, EAGLE and GCE. 6.3 Ranking for Community Detection Extensive comparisons have been conducted over different overlapping densities and community size ranges. Performance measured by NMI and Omega for n = 5000 and µ = 0.3 is shown in Figures 3 and 4. To provide an intuitive way for both comparing two measures and also summa- rizing the vast volume of experiment results, we propose RSM (i), the averaged , Vol. V, No. N, May 2014. 18 (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:27)(cid:30)(cid:14)(cid:16) (cid:27)(cid:30)(cid:14)(cid:16) (cid:30)(cid:21)(cid:19) (cid:30)(cid:21)(cid:19) (cid:29)(cid:31)(cid:17)(cid:25) (cid:29)(cid:31)(cid:17)(cid:25) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:30)(cid:22)(cid:23)(cid:33) (cid:30)(cid:22)(cid:23)(cid:33) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:18)(cid:19)(cid:21) (cid:18)(cid:19)(cid:21) · Jierui Xie et al. (cid:18)(cid:35)(cid:10)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:37)(cid:37)(cid:37)(cid:35)(cid:8)(cid:9)(cid:10) (cid:18)(cid:35)(cid:10)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:37)(cid:37)(cid:37)(cid:35)(cid:8)(cid:9)(cid:2) (cid:18)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:37)(cid:37)(cid:37)(cid:35)(cid:8)(cid:9)(cid:10) (cid:18)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:37)(cid:37)(cid:37)(cid:35)(cid:8)(cid:9)(cid:2) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:29)(cid:12)(cid:28) (cid:29)(cid:12)(cid:28) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:12)(cid:20)(cid:27) (cid:12)(cid:20)(cid:27) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:22)(cid:30)(cid:12)(cid:32) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:34)(cid:28)(cid:13)(cid:12) (cid:34)(cid:28)(cid:13)(cid:12) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:12)(cid:21)(cid:22)(cid:23)(cid:24)(cid:25)(cid:26) (cid:12)(cid:21)(cid:22)(cid:23)(cid:24)(cid:25)(cid:26) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:12)(cid:13)(cid:14)(cid:15)(cid:16) (cid:12)(cid:13)(cid:14)(cid:15)(cid:16) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:19)(cid:13)(cid:27)(cid:28)(cid:27) (cid:19)(cid:13)(cid:27)(cid:28)(cid:27) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:20) (cid:19) (cid:18) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:13)(cid:27)(cid:30)(cid:13)(cid:19) (cid:13)(cid:27)(cid:30)(cid:13)(cid:19) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:28)(cid:16)(cid:29)(cid:30)(cid:28) (cid:28)(cid:16)(cid:29)(cid:30)(cid:28) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:13)(cid:17) (cid:5) (cid:6) (cid:7) Fig. 1. The effects of network size n and mixing parameter µ on LFR networks. Plots show NMI’s for networks with large community size range and On = 10%. The order of subplots is from general behaviors to exceptions (see text). , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 19 (cid:21)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:15)(cid:35)(cid:10)(cid:8)(cid:38) (cid:21)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:36) (cid:37)(cid:15)(cid:35)(cid:4)(cid:8)(cid:38) (cid:21)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:39) (cid:37)(cid:15)(cid:35)(cid:10)(cid:8)(cid:38) (cid:21)(cid:35)(cid:4)(cid:8)(cid:8)(cid:8)(cid:39) (cid:37)(cid:15)(cid:35)(cid:4)(cid:8)(cid:38) (cid:12)(cid:13)(cid:14)(cid:15)(cid:16)(cid:17)(cid:18) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:22)(cid:19)(cid:24)(cid:28)(cid:24) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:30)(cid:13)(cid:22) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:29)(cid:31)(cid:20)(cid:17) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:30)(cid:14)(cid:15)(cid:33) (cid:14)(cid:30)(cid:12)(cid:32) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:12)(cid:19)(cid:25)(cid:26)(cid:27) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:12)(cid:23)(cid:24) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:24)(cid:30)(cid:25)(cid:27) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:29)(cid:12)(cid:28) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:10) (cid:8)(cid:9)(cid:11) (cid:8)(cid:9)(cid:7) (cid:8)(cid:9)(cid:6) (cid:8)(cid:9)(cid:5) (cid:19)(cid:24)(cid:30)(cid:19)(cid:22) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:28)(cid:27)(cid:29)(cid:30)(cid:28) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:23) (cid:22) (cid:21) (cid:8)(cid:9)(cid:4) (cid:8)(cid:9)(cid:3) (cid:8)(cid:9)(cid:2) (cid:8)(cid:9)(cid:1) (cid:8)(cid:9)(cid:10) (cid:8) (cid:1) (cid:21)(cid:22)(cid:13) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) (cid:34)(cid:28)(cid:19)(cid:12) (cid:1) (cid:8) (cid:2) (cid:3) (cid:4) (cid:19)(cid:20) (cid:5) (cid:6) (cid:7) Fig. 2. The effects of community size range and overlapping density On on LFR networks. Plots show NMI’s for networks with n = 5000 and µ = 0.3. The order of subplots is from general behaviors to exceptions (see text). , Vol. V, No. N, May 2014. 20 · Jierui Xie et al. (cid:24)(cid:17)(cid:25) (cid:35) (cid:28)(cid:33)(cid:44) (cid:28)(cid:33)(cid:43) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:20) (cid:16) (cid:19) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:35) (cid:28) (cid:39) (cid:35) (cid:24)(cid:38)(cid:25) (cid:28)(cid:33)(cid:44) (cid:12)(cid:26)(cid:27)(cid:28)(cid:28)(cid:28)(cid:29)(cid:30)(cid:31)(cid:32)(cid:26)(cid:28)(cid:33)(cid:34)(cid:30)(cid:31)(cid:6)(cid:12)(cid:26)(cid:35)(cid:28)(cid:36) (cid:35) (cid:24)(cid:37)(cid:25) (cid:28)(cid:33)(cid:44) (cid:17) (cid:45) (cid:14) (cid:18) (cid:6) (cid:28)(cid:33)(cid:43) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:35) (cid:34) (cid:40) (cid:27) (cid:6)(cid:18) (cid:41) (cid:42) (cid:43) (cid:28) (cid:39) (cid:34) (cid:40) (cid:27) (cid:6)(cid:18) (cid:41) (cid:42) (cid:43) (cid:12)(cid:26)(cid:27)(cid:28)(cid:28)(cid:28)(cid:38)(cid:30)(cid:31)(cid:32)(cid:26)(cid:28)(cid:33)(cid:34)(cid:30)(cid:31)(cid:6)(cid:12)(cid:26)(cid:35)(cid:28)(cid:36) (cid:35) (cid:24)(cid:13)(cid:25) (cid:28)(cid:33)(cid:44) (cid:28)(cid:33)(cid:43) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:20) (cid:16) (cid:19) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:35) (cid:28) (cid:39) (cid:17) (cid:45) (cid:14) (cid:18) (cid:6) (cid:28)(cid:33)(cid:43) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:35) (cid:34) (cid:40) (cid:27) (cid:6)(cid:18) (cid:41) (cid:42) (cid:43) (cid:28) (cid:39) (cid:34) (cid:40) (cid:27) (cid:6)(cid:18) (cid:41) (cid:42) (cid:43) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) Fig. 3. Evaluations of overlapping community detection on LFR networks with low overlap density On = 10%. Left column: NMI as a function of the number of memberships Om; Right column: Omega as a function of the number of memberships Om. Results for small community size range are shown in top row (i.e., (a) and (c)), and results for large community size range are shown in bottom row (i.e., (b) and (d)). All resutls are from networks with n = 5000 and µ = 0.3. ranking score for a given algorithm i with respect to some measure M as follows: RSM (i) =Xj=1 rank(i, Oj m), (11) where Oj m is the number of memberships (diversity) in {2, 3, · · · , 8}, and function rank returns the ranking of algorithm i for the given Om. Sorting RSM in increasing order gives the final ranking among algorithms. Whenever it is clear from context, we use the term ranking to refer to the final rank without the actual score value. The results for low overlapping density case in Figure 3 are summarized as four Omega in Table II, where rankings including RSs N MI , RSb N MI , RSs Omega and RSb , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 21 (cid:24)(cid:17)(cid:25) (cid:43) (cid:28)(cid:33)(cid:44) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:40) (cid:20) (cid:16) (cid:19) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:38) (cid:28)(cid:33)(cid:43) (cid:28) (cid:38) (cid:43) (cid:24)(cid:37)(cid:25) (cid:28)(cid:33)(cid:44) (cid:12)(cid:26)(cid:27)(cid:28)(cid:28)(cid:28)(cid:29)(cid:30)(cid:31)(cid:32)(cid:26)(cid:28)(cid:33)(cid:34)(cid:30)(cid:31)(cid:6)(cid:12)(cid:26)(cid:27)(cid:28)(cid:35) (cid:43) (cid:24)(cid:36)(cid:25) (cid:28)(cid:33)(cid:44) (cid:17) (cid:45) (cid:14) (cid:18) (cid:6) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:38) (cid:28)(cid:33)(cid:43) (cid:34) (cid:39) (cid:27) (cid:6)(cid:18) (cid:40) (cid:41) (cid:42) (cid:28) (cid:38) (cid:34) (cid:39) (cid:27) (cid:6)(cid:18) (cid:40) (cid:41) (cid:42) (cid:12)(cid:26)(cid:27)(cid:28)(cid:28)(cid:28)(cid:37)(cid:30)(cid:31)(cid:32)(cid:26)(cid:28)(cid:33)(cid:34)(cid:30)(cid:31)(cid:6)(cid:12)(cid:26)(cid:27)(cid:28)(cid:35) (cid:43) (cid:24)(cid:13)(cid:25) (cid:28)(cid:33)(cid:44) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:40) (cid:20) (cid:16) (cid:19) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:38) (cid:28)(cid:33)(cid:43) (cid:28) (cid:38) (cid:17) (cid:45) (cid:14) (cid:18) (cid:6) (cid:28)(cid:33)(cid:42) (cid:28)(cid:33)(cid:41) (cid:28)(cid:33)(cid:40) (cid:28)(cid:33)(cid:27) (cid:28)(cid:33)(cid:39) (cid:28)(cid:33)(cid:34) (cid:28)(cid:33)(cid:38) (cid:28)(cid:33)(cid:43) (cid:34) (cid:39) (cid:27) (cid:6)(cid:18) (cid:40) (cid:41) (cid:42) (cid:28) (cid:38) (cid:34) (cid:39) (cid:27) (cid:6)(cid:18) (cid:40) (cid:41) (cid:42) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) Fig. 4. Evaluations of overlapping community detection on LFR networks with high overlap density On = 50%. Left column: NMI as a function of the number of memberships Om; Right column: Omega as a function of the number of memberships Om. Results for small community size range are shown in top row (i.e., (a) and (c)), and results for large community size range are shown in bottom row (i.e., (b) and (d)). All resutls are from networks with n = 5000 and µ = 0.3. RSs(b) N MI(Omega) denotes the ranking based on NMI (or Omega) for networks with small (or large) community size range. Results for high overlapping density case in Figure 4 are summarized in Table III. We first compared pairwise similarities of (RSs N MI , RSb Omega). As shown, among the top seven (half of the total fourteen) algorithms in two rankings, there are 4 pairs of matches (ignoring the exact order) for On = 10% and 5 pairs of matches for On = 50% for (RSs Omega). Even more, there are 6 pairs of matches for both On = 10% and On = 50% for (RSb Omega). This suggests that NMI and Omega provide similar overall evaluation to some extent. Omega) and (RSb N MI , RSs N MI , RSs N MI , RSb , Vol. V, No. N, May 2014. 22 · Jierui Xie et al. Table II. The community detection ranking for n = 5000, µ = 0.3 and low overlap- ping density On = 10%. Rank RSs N MI RSs Omega RSb SLPA N MI RSb Omega RS∗ SLPA MOSES MOSES COPRA CFinder COPRA OSLOM SLPA GCE CIS LFM OSLOM Game GCE Game OSLOM COPRA NMF Link iLCD Link CIS LFM NMF iLCD EAGLE EAGLE UEOC UEOC CFinder MOSES CFinder MOSES CFinder Link Link EAGLE EAGLE UEOC iLCD iLCD UEOC SLPA GCE NMF CIS LFM Game OSLOM COPRA Game GCE CIS NMF LFM N MI,Omega SLPA GCE OSLOM CIS Game COPRA LFM MOSES NMF CFinder Link EAGLE iLCD UEOC RS∗ F SLPA CFinder Game OSLOM MOSES COPRA iLCD Link LFM UEOC EAGLE GCE CIS NMF 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Table III. The community detection ranking for n = 5000, µ = 0.3 and high overlapping density On = 50%. Rank RSs N MI RSs N MI RSb MOSES COPRA CFinder Omega RSb SLPA Link Game SLPA GCE COPRA Omega RS∗ SLPA Game N MI,Omega SLPA MOSES Game OSLOM Link Game MOSES MOSES SLPA GCE iLCD CFinder CFinder MOSES OSLOM Game CFinder COPRA OSLOM COPRA OSLOM CIS LFM NMF Link GCE iLCD LFM CIS NMF EAGLE EAGLE UEOC UEOC LFM CIS NMF UEOC iLCD Link EAGLE LFM NMF EAGLE CIS iLCD UEOC GCE COPRA CFinder OSLOM Link GCE LFM CIS iLCD NMF EAGLE UEOC RS∗ F Link UEOC SLPA Game LFM CFinder CIS MOSES OSLOM iLCD GCE COPRA EAGLE NMF Based on these four rankings, we further derive the average ranking RS∗ N MI,Omega as the overall community detection performance. In this final ranking, the top seven algorithms are exclusively agent-based algorithms (SLPA, Game and COPRA) and local expansion based algorithms (GCE, OSLOM, CIS, and LFM), which signifi- cantly outperform the others for networks with low overlapping density (see Figure 3). For high overlapping density, agent-based algorithms remain in the top seven, together with MOSES representing fuzzy algorithms, CFinder representing clique algorithms, Link representing link partitioning. However, given the fact that the performance is actually fairly low (most of them are less than 0.5 for Om > 2) shown in Figure 4, it is fair to conclude that all the algorithms do not yet achieve satisfy- ing performance for networks with high overlapping density and high overlapping diversity (e.g., for On = 50% and Om > 2). , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 23 t n u o C 1000 800 600 400 200 0 500 True SLPA GCE NMF LFM MOSES 20 40 60 80 100 120 Community Size (CS) t n u o C 400 300 200 100 0 20 True OSLOM Game COPRA CFinder Link CIS 6000 4000 2000 0 10 20 40 60 80 100 120 Community Size (CS) Fig. 5. Histogram of the detected community sizes for SLPA, GCE, NMF, LFM and MOSES created from the results for LFR networks with n = 5000, µ = 0.3 and On = 10%. Fig. 6. Histogram of the detected community sizes for OSLOM, Game, COPRA, CFinder, Link and CIS cre- ated from the results for LFR net- works with n = 5000, µ = 0.3 and On = 10%. True EAGLE UEOC iLCD 2000 1500 t n u o C 1000 500 10000 8000 6000 4000 2000 0 0 10 20 0 20 40 60 80 100 120 Community Size (CS) Fig. 7. Histogram of the detected community sizes for EAGLE, UEOC and iLCD crated from the results for LFR networks with n = 5000, µ = 0.3 and On = 10%. 6.4 Comparing Detected Community Size Distribution in LFR In order to provide insight into the behaviors of different algorithms and verify the ranking, we examined the discovered distribution (histograms) of community sizes (CS) and compared it with the known ground truth. Here we only provide analysis for n = 5000, µ = 0.3, On = 10% (the corresponding ranking is RSb N MI in Table II). In this case, we expect the community size distribution to follow the power law with exponent τ2 = 1, a minimum of 20, and a maximum size of 100. Note that the histograms are created from communities over different O′ ms. As shown in Figure 5, SLPA, GCE and NMF find communities whose sizes are distributed in a unimodal distribution with a single peak at CS = 20 in aggreement with the ground truth distribution. This explains why they perform well with respect to ranking RSb N MI . LFM and MOSES have a peak around CS = 5, which lowers their performance. , Vol. V, No. N, May 2014. 24 · Jierui Xie et al. The prominent feature of Figure 6 (see the inset) is a bimodal distribution that has a peak at CS = 1 ∼ 5. This means that algorithms like OSLOM, Game, COPRA, CIS and CFinder find significant numbers of small communities. In Figure 7, the distribution is shifted mostly outside the predefined range 20∼100. Algorithms with such a distribution create relatively small communities and perform poorly with respect to this analysis. Here, we conclude that observations on the community size distribution can be used to explain the performance and ranking. It is worth noticing that in Figure 6, excluding the range that contains the un- desired peak (CS = 1 ∼ 20), the distributions seem to agree well with the ground truth, especially for COPRA. The performances of OSLOM, Game, CIS and CO- PRA with respect to NMI are still fairly stable. This demonstrates that NMI is, to some degree, not sensitive to small size communities (including outliers or singleton communities). 6.5 Identifying Overlapping Nodes in LFR Community overlap manifests itself as the existence of the nodes with membership in multiple communities. Thus, we will refer to nodes with multiple membership as overlapping nodes. In real-world social networks, such nodes are important because they usually represent bridges (or messengers) between communities. For this reason, the ability to identify overlapping nodes, although often neglected, is essential for assessing the accuracy of community detection algorithms. Measures like NMI and Omega focus only on providing an overall measure of algorithmic accuracy. As we see in section 6.4, these measures might not be sensitive enough to provide an accurate picture of what is happening at the node level. In this section, we evaluate an algorithm’s ability to identify overlapping nodes. n and detected memberships Od Similar to the definitions of On and Om, we define the number of detected overlap- ping nodes Od m. Note that the number of overlapping nodes Od n alone is insufficient to accurately quantify the detection performance, be- cause it contains both true and false positive. Ideally, an algorithm should report as many true overlapping nodes as possible (i.e., a balance between quality and quantity). To provide more precise analysis, we consider the identification of over- lapping nodes as a binary classification problem. A node is labeled as overlapping as long as Om>1 or Od m>1 and labeled as non-overlapping otherwise. Within this framework, one can use Jaccard index as in [Ball et al. 2011] or F-score as a measure of detection accuracy. In this paper, we use the later that is defined as F = 2 · precision · recall precision + recall , (12) where recall is the number of correctly detected overlapping nodes divided by the true number of overlapping nodes On, and precision is the number of correctly detected overlapping nodes divided by the total number of detected overlapping nodes Od n. F-score accounts for the balance between detection quantity and quality, and reaches its best and worst value at 1 and 0, respectively. Figures 8 and 9 show the F-score, precision, and recall for different settings of the LFR benchmark. In general, an algorithm achieves better F-score on benchmarks with small community sizes for both low and high overlapping density. However, the behaviors are quite different for the various algorithms. For example, the gain , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 25 in F-score on communities in the small size range for EAGLE is due to the increase in recall (i.e., detect more overlapping nodes shown in (c) and (f) in Figure 8), while the gain for iLCD is mainly due to the increase in precision (see (b) and (e) in Figure 8). Moreover, the F-score (performance) typically decays moderately as overlapping diversity Om increases. It is evident that Om has great impacts on OSLOM with an rapid drop for large Om. SLPA is the only exception that has a positive correlation with Om for low overlapping density case. In terms of the precision, half of the algorithms including SLPA, CFinder, Game, OSLOM, MOSES, EAGLE and iLCD consistently outperform the expected random performance, 10% and 50% for low and high overlapping density respectively (see (b) and (e) in both Figures 8 and 9). The high precision of EAGLE (also CFinder and GCE for Om = 2) shows that clique-like assumption of communities may help to identify overlapping nodes in low overlapping density case. Link performs merely as well as the random classifier. Experiments also reveal an imbalance in precision and recall for some algorithms, which is partially due to either over-detection where more overlapping nodes than there exists are claimed or under-detection where only very few overlapping nodes are identified. Extreme examples are EAGLE and Link. Although EAGLE achieves very high detection precision (e.g., (b) and (e) in Figure 8), it suffers from under- detection (verified in Figure 10), which results in a low recall score ((c) and (f) in Figure 8). As a result, we observe a low F-score ((a) and (b) in Figure 8). For Link, the algorithm does not perform well in terms of F-score even though it has very high recall ((c) and (f) in Figure 8). This is due to the fact that Link claims way more overlapping nodes than excepted (verified in Figure 10). 6.6 Ranking for Overlapping Node Detection The rankings with respect to F-scores for different community size ranges, RSs RSb the average ranking over two community size ranges. F are shown in Tables IV and V for different overlapping density cases. RS∗ F and F is To facilitate comparison, we copy RS∗ F into Tables II and III. It is clearly shown, for example in Table II, that the community quality ranking RS∗ N MI,Omega and node quality ranking RS∗ F might provide quite different pictures of the perfor- mance. For the low overlapping density case (as in Table II), algorithms with a low rank in detecting communities could actually have good performances when it comes to identifying overlapping nodes (e.g., CFinder, iLCD and MOSES), while high-ranking algorithms, including GCE and CIS, might perform badly due to under-detection and over-detection respectively. SLPA has very stable and good performance for the low overlapping density case. These observations suggest the need for a careful treatment of the algorithms with a high NMI or Omega score if the application of these algorithms is aimed at identifying nodes with multiple community memberships. Similar conclusions can be drawn for high overlapping density case. 6.7 Final Ranking Since two types of rankings provide complementary information, we conclude, by considering algorithms that are consistently ranked in the top seven in both RS∗ F and RS∗ N MI,Omega: (a) For low overlapping density networks, SLPA, OSLOM, , Vol. V, No. N, May 2014. 26 · Jierui Xie et al. (cid:44)(cid:32)(cid:45) (cid:5)(cid:6)(cid:4) (cid:29)(cid:38)(cid:9)(cid:5)(cid:5)(cid:5)(cid:15)(cid:40)(cid:41)(cid:42)(cid:38)(cid:5)(cid:6)(cid:8)(cid:40)(cid:41)(cid:11)(cid:29)(cid:38)(cid:7)(cid:5)(cid:43) (cid:44)(cid:30)(cid:45) (cid:5)(cid:6)(cid:4) (cid:29)(cid:38)(cid:9)(cid:5)(cid:5)(cid:5)(cid:39)(cid:40)(cid:41)(cid:42)(cid:38)(cid:5)(cid:6)(cid:8)(cid:40)(cid:41)(cid:11)(cid:29)(cid:38)(cid:7)(cid:5)(cid:43) (cid:19) (cid:18) (cid:17) (cid:16) (cid:15) (cid:14) (cid:13) (cid:5)(cid:6)(cid:10) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:9) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:8) (cid:5)(cid:6)(cid:1) (cid:5)(cid:6)(cid:7) (cid:5) (cid:1) (cid:7) (cid:44)(cid:39)(cid:45) (cid:19) (cid:18) (cid:17) (cid:16) (cid:15) (cid:14) (cid:13) (cid:5)(cid:6)(cid:10) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:9) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:8) (cid:5)(cid:6)(cid:1) (cid:5)(cid:6)(cid:7) (cid:5) (cid:1) (cid:7) (cid:44)(cid:19)(cid:45) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:28) (cid:28) (cid:29) (cid:17) (cid:15) (cid:16) (cid:19) (cid:18) (cid:22) (cid:5)(cid:6)(cid:4) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:1) (cid:28) (cid:28) (cid:29) (cid:17) (cid:15) (cid:16) (cid:19) (cid:18) (cid:22) (cid:5) (cid:1) (cid:7) (cid:44)(cid:16)(cid:45) (cid:46) (cid:46) (cid:32) (cid:16) (cid:19) (cid:25) (cid:5)(cid:6)(cid:4) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:1) (cid:5) (cid:1) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:44)(cid:47)(cid:45) (cid:46) (cid:46) (cid:32) (cid:16) (cid:19) (cid:25) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:5)(cid:6)(cid:4) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:1) (cid:5) (cid:1) (cid:7) (cid:5)(cid:6)(cid:4) (cid:5)(cid:6)(cid:3) (cid:5)(cid:6)(cid:2) (cid:5)(cid:6)(cid:1) (cid:5) (cid:1) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:2) (cid:11)(cid:12) (cid:3) (cid:4) (cid:20)(cid:21)(cid:22)(cid:23) (cid:24)(cid:11)(cid:22)(cid:25)(cid:23) (cid:26)(cid:24)(cid:27) (cid:24)(cid:13)(cid:28)(cid:29)(cid:30)(cid:19)(cid:18) (cid:21)(cid:13)(cid:31) (cid:26)(cid:32)(cid:12)(cid:19) (cid:11)(cid:20)(cid:21)(cid:11)(cid:31) (cid:33)(cid:31)(cid:13) (cid:31)(cid:11)(cid:20)(cid:27)(cid:20) (cid:24)(cid:34)(cid:20) (cid:21)(cid:28)(cid:29)(cid:35) (cid:27)(cid:23)(cid:26)(cid:21)(cid:27) (cid:28)(cid:21)(cid:24)(cid:36) (cid:37)(cid:27)(cid:11)(cid:24) (cid:20)(cid:21)(cid:22)(cid:23) (cid:24)(cid:11)(cid:22)(cid:25)(cid:23) (cid:26)(cid:24)(cid:27) (cid:24)(cid:13)(cid:28)(cid:29)(cid:30)(cid:19)(cid:18) (cid:21)(cid:13)(cid:31) (cid:26)(cid:32)(cid:12)(cid:19) (cid:11)(cid:20)(cid:21)(cid:11)(cid:31) (cid:33)(cid:31)(cid:13) (cid:31)(cid:11)(cid:20)(cid:27)(cid:20) (cid:24)(cid:34)(cid:20) (cid:21)(cid:28)(cid:29)(cid:35) (cid:27)(cid:23)(cid:26)(cid:21)(cid:27) (cid:28)(cid:21)(cid:24)(cid:36) (cid:37)(cid:27)(cid:11)(cid:24) (cid:20)(cid:21)(cid:22)(cid:23) (cid:24)(cid:11)(cid:22)(cid:25)(cid:23) (cid:26)(cid:24)(cid:27) (cid:24)(cid:13)(cid:28)(cid:29)(cid:30)(cid:19)(cid:18) (cid:21)(cid:13)(cid:31) (cid:26)(cid:32)(cid:12)(cid:19) (cid:11)(cid:20)(cid:21)(cid:11)(cid:31) (cid:33)(cid:31)(cid:13) (cid:31)(cid:11)(cid:20)(cid:27)(cid:20) (cid:24)(cid:34)(cid:20) (cid:21)(cid:28)(cid:29)(cid:35) (cid:27)(cid:23)(cid:26)(cid:21)(cid:27) (cid:28)(cid:21)(cid:24)(cid:36) (cid:37)(cid:27)(cid:11)(cid:24) Fig. 8. Evaluations of overlapping node detection on LFR networks with low overlap density On = 10%. Plots show F-score (together with precision and recall) as a function of the number of memberships for n = 5000 and µ = 0.3. Results for small community size range are shown in the left column, and results for large community size range are shown in the righ column. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 27 (cid:34)(cid:17)(cid:35) (cid:26)(cid:31)(cid:43) (cid:12)(cid:24)(cid:25)(cid:26)(cid:26)(cid:26)(cid:45)(cid:28)(cid:29)(cid:30)(cid:24)(cid:26)(cid:31)(cid:32)(cid:28)(cid:29)(cid:6)(cid:12)(cid:24)(cid:25)(cid:26)(cid:33) (cid:34)(cid:13)(cid:35) (cid:12)(cid:24)(cid:25)(cid:26)(cid:26)(cid:26)(cid:27)(cid:28)(cid:29)(cid:30)(cid:24)(cid:26)(cid:31)(cid:32)(cid:28)(cid:29)(cid:6)(cid:12)(cid:24)(cid:25)(cid:26)(cid:33) (cid:26)(cid:31)(cid:43) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:25) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:32) (cid:26)(cid:31)(cid:38) (cid:26)(cid:31)(cid:42) (cid:14) (cid:15) (cid:46) (cid:37) (cid:45) (cid:44) (cid:10) (cid:39) (cid:6)(cid:18) (cid:40) (cid:41) (cid:26) (cid:38) (cid:42) (cid:34)(cid:14)(cid:35) (cid:39) (cid:6)(cid:18) (cid:40) (cid:41) (cid:11) (cid:11) (cid:12) (cid:46) (cid:45) (cid:37) (cid:14) (cid:15) (cid:3) (cid:26)(cid:31)(cid:41) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:38) (cid:14) (cid:15) (cid:46) (cid:37) (cid:45) (cid:44) (cid:10) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:25) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:32) (cid:26)(cid:31)(cid:38) (cid:26)(cid:31)(cid:42) (cid:26) (cid:38) (cid:42) (cid:34)(cid:27)(cid:35) (cid:11) (cid:11) (cid:12) (cid:46) (cid:45) (cid:37) (cid:14) (cid:15) (cid:3) (cid:26)(cid:31)(cid:41) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:38) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) (cid:26) (cid:38) (cid:42) (cid:34)(cid:37)(cid:35) (cid:47) (cid:47) (cid:17) (cid:37) (cid:14) (cid:7) (cid:26)(cid:31)(cid:41) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:38) (cid:26) (cid:38) (cid:39) (cid:6)(cid:18) (cid:40) (cid:26) (cid:38) (cid:42) (cid:41) (cid:34)(cid:36)(cid:35) (cid:39) (cid:6)(cid:18) (cid:40) (cid:41) (cid:47) (cid:47) (cid:17) (cid:37) (cid:14) (cid:7) (cid:26)(cid:31)(cid:41) (cid:26)(cid:31)(cid:40) (cid:26)(cid:31)(cid:39) (cid:26)(cid:31)(cid:38) (cid:26) (cid:38) (cid:39) (cid:6)(cid:18) (cid:40) (cid:41) (cid:39) (cid:6)(cid:18) (cid:40) (cid:41) Fig. 9. Evaluations of overlapping node detection on LFR networks with high overlap density On = 50%. Plots show F-score (together with precision and recall) as a function of the number of memberships for n = 5000 and µ = 0.3. Results for small community size range are shown in the left column, and results for large community size range are shown in the righ column. , Vol. V, No. N, May 2014. 28 · Jierui Xie et al. (cid:6)(cid:12)(cid:37)(cid:28)(cid:33)(cid:38) (cid:28)(cid:33) (cid:32) (cid:27) (cid:31) (cid:26) (cid:30) (cid:25) (cid:29) (cid:24) (cid:28) (cid:12) (cid:6) (cid:35) (cid:36) (cid:13) (cid:12) (cid:6) (cid:34) (cid:6)(cid:12)(cid:37)(cid:30)(cid:33)(cid:38) (cid:24) (cid:28) (cid:12) (cid:6) (cid:13) (cid:36) (cid:12) (cid:6) (cid:24) (cid:25) (cid:6)(cid:18) (cid:26) (cid:27) (cid:24) (cid:25) (cid:26) (cid:27) (cid:6)(cid:18) (cid:1)(cid:2)(cid:3)(cid:4) (cid:5)(cid:6)(cid:3)(cid:7)(cid:4) (cid:8)(cid:5)(cid:9) (cid:5)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:15) (cid:2)(cid:10)(cid:16) (cid:8)(cid:17)(cid:18)(cid:14) (cid:6)(cid:1)(cid:2)(cid:6)(cid:16) (cid:19)(cid:16)(cid:10) (cid:16)(cid:6)(cid:1)(cid:9)(cid:1) (cid:5)(cid:20)(cid:1) (cid:2)(cid:11)(cid:12)(cid:21) (cid:9)(cid:4)(cid:8)(cid:2)(cid:9) (cid:11)(cid:2)(cid:5)(cid:22) (cid:23)(cid:9)(cid:6)(cid:5) Fig. 10. The number of detected overlapping nodes (normalized by On) based on the results for LFR networks with n = 5000b and µ = 0.3. A value larger than 1 is possible. Table IV. The overlapping node de- tection ranking for n = 5000, µ = 0.3 and low overlap density On = 10%. Table V. The overlapping node detec- tion ranking for n = 5000, µ = 0.3 and high overlap density On = 50%. Rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 RSs F SLPA RSb F SLPA RS∗ F SLPA CFinder CFinder CFinder OSLOM Game Game Game OSLOM OSLOM MOSES COPRA MOSES iLCD MOSES COPRA COPRA Link EAGLE Link iLCD LFM UEOC CIS GCE UEOC LFM EAGLE CIS NMF GCE NMF iLCD Link LFM UEOC EAGLE GCE CIS NMF Rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 RSb F Link UEOC SLPA Game LFM CIS RS∗ F Link UEOC SLPA Game LFM CFinder RSs F Link UEOC Game SLPA CFinder LFM CIS GCE CFinder MOSES MOSES MOSES OSLOM OSLOM CIS iLCD iLCD OSLOM COPRA COPRA EAGLE COPRA EAGLE EAGLE NMF NMF GCE iLCD GCE NMF Game and COPRA offer better performance than the other tested algorithms; (b) For high overlapping density networks, both SLPA and Game provide better per- formance. (Note that we do not include Link and UEOC because their high ranks are mainly due to the over-detection.) 7. TESTS ON REAL-WORLD SOCIAL NETWORKS We first examined algorithm performance on a high school friendship network18 where the ground truth is known. This social network from a high school is based on self-reporting from students. It is known that the true partitioning of the network roughly corresponds to the grade (from 7 to 12) of students listed in the survey. The ground truth is a total of 6 communities (see Figure 11) together with two subgroups 18A project funded by the National Institute of Child Health and Human Development. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 29 grade 10 65 27 28 60 47 64 29 48 26 32 57 33 58 63 68 grade 9 43 53 54 44 62 15 56 61 59 31 67 1 0 grade 12 18 17 16 19 13 12 30 36 38 34 37 39 35 grade 11 55 14 45 6 42 41 40 4 52 66 3 grade 7 11 7 2 9 10 5 8 21 23 46 24 25 51 20 49 50 grade 8 22 Fig. 11. High school network (n = 69, k = 6.4). Colors represent known communities correspond- ing to grades ranging from 7 to 12. Grade 9 is separated into two subgroups that correspond to white (upper) and black (lower) students respectively. Numbers are the node id’s. within grade 9 corresponding to a group of white and black students. Even though there are no overlapping nodes reported by the students, each algorithm reports some by its own. Results are shown in Table VI19. Discovered overlapping nodes are listed in the third column. For algorithms that discover more than 10 overlapping nodes, only the total number is shown. We also include NMI and the number of communities for reference. It is easy to verify that all the overlapping nodes in Table VI are connected to at least two different groups. Some of them lie between different grades with strong connections to each individual one, for example, nodes 45, 46, 61, 26, 32 and 33. Some are boundary nodes between subgroups within a grade such as nodes 59, 12 and 18. Node 42 serves as a bridge between groups without having particular coherence to any group. However, it is still not clear whether these nodes are really meaningful or necessary to be considered as “overlapping”. This is one factor that makes the detection (and verification) even more challenging in real-life applications. Next, we tested on a wider range of social networks listed in Table VII. More information about these networks can be found here20. Given that the ground truth is not available for most of these networks, we selected two overlapping modularities QE ov in (2) as quality measures. The former is based on the node ov in (5) and QN i 19For each algorithm, we show results with parameters that output the best NMI score. 20CA-GrQc: a co-authorship network based on papers in General Relativity publishing in Arxiv [Leskovec et al. 2007]. PGP: a network of users of the Pretty-Good-Privacy algorithm [Boguna et al. 2004]. Email: a communication network in Enron via emails [Leskovec et al. 2009]. Epinions: a who-trust-whom on-line social network of a consumer review site Epinions.com [Richardson et al. 2003]. P2P: the Gnutella peer-to-peer file sharing network from August 2002 [Ripeanu et al. 2002]. Amazon: a co-purchase network of the Amazon website [Leskovec et al. 2007]. Data are available at http://www-personal.umich.edu/~mejn/netdata and http://snap. stanford.edu/data. , Vol. V, No. N, May 2014. 30 · Jierui Xie et al. Table VI. Test on a high school friendship network. Algorithm Num. of communities CFinder CIS COPRA EAGLE Game GCE iLCD LFM Link MOSES NMF OSLOM SLPA UEOC 2 9 6 4 10 6 7 7 20 10 7 11 6 7 {18} total 14 {0, 21, 45, 46, 61} Overlapping nodes {12, 18} total 34 total 14 NMI 0.1679 0.7495 0.7966 0.4962 0.4673 0.8333 {5, 21, 26, 29, 31, 32, 33, 46, 61} 0.3713 0.8134 0.3155 0.5037 0.643 0.4315 0.6788 0.8148 {0, 45} total 31 total 18 {0, 12, 18, 26, 29, 45} {0, 12, 18, 45} {45, 46} {1, 42, 45, 59} Table VII. Social networks in the tests Network karate (KR) football (FB) lesmis (LS) dolphins (DP) CA-GrQc (CA) n 34 115 77 62 4730 k 4.5 10.6 6.6 5.1 5.6 Network PGP Email (EM) P2P Epinions (EP) Amazon (AM) n 10680 33696 62561 75877 262111 k 4.5 10.7 2.4 10.6 6.8 belonging factor, and the later is based on the link belonging factor. For the arbitrary function in QN i ov , we adopted the one used in [Gregory 2010], f (x) = 60x − 30. In Figures 12 ∼ 17, networks are shown in the order of increasing number of edges along the x-axis. Lines connecting points are meant merely to aid the reader in differentiating points from the same algorithm. We removed CFinder, EAGLE and NMF from the test due to either their memory or computation inefficiency in large networks. As a reference, we also performed disjoint community detection with the Infomap algorithm [M. Rosvall 2008], which has been shown to be quite accurate in [Lancichinetti and Fortunato 2009]. Figures 12 and 13 show a positive correlation between the two quality measures. Typically, the disjoint partitioning achieves higher QE ov than overlapping cluster- ings, which empirically serves as a bound of the quality of detected overlapping communities. This also holds for QN i ov in general. In general, Link and iLCD achieve lower QN i ov compared to others, while SLPA, LFM, COPRA, OSLOM and GCE achieve higher performance on larger net- works (e.g., last five networks). Moreover, an algorithms may not perform equally well on different types of network structures. Some of them are sensitive to specific structures. For example, only SLPA, LFM, CIS and Game have satisfying perfor- mances in networks with highly sparse structure such as P 2P , for which COPRA finds merely one single giant community and GCE also fails. Another issue is that some algorithms tend to over-detect the overlap, as was the case for LFR networks. CIS and Link fail in the test because they find too many overlapping nodes or memberships relative to the consensus shown by the other algorithms as seen in ov or QE , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 31 E v o Q 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC Infomap KR FB LS DP CA PGP EM EP P2P AM i N v o Q 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC Infomap KR FB LS DP CA PGP EM EP P2P AM Fig. 12. Overlapping modularity QE ov for social networks. Fig. 13. Overlapping modularity QN i ov for social networks. N > / d n O < 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 KR FB LS DP CA PGP EM EP P2P AM SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC N > / d n O < 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC KR FB LS DP CA PGP EM EP P2P AM Fig. 14. The normalized number of de- tected overlapping nodes for social net- works based on the clustering with the best QE ov. Fig. 15. The normalized number of de- tected overlapping nodes for social net- works based on the clustering with the best QN i ov . Figures 14 ∼ 17. Such over-detection happens to other algorithms, including CO- PRA, GCE and UEOC on specific networks, resulting in low performance for these algorithms. Some interesting common features are observed from our tests. As shown in Fig- ures 14 and 15, the fraction of overlapping nodes found by most of the algorithms is typically less than 30%. Results from SLPA, OSLOM and COPRA, which offer good performances in the LFR benchmarks, show an even smaller fraction of over- lapping nodes, less than 20%, in most real-world networks examined in this paper. Moreover, Figures 16 and 17 confirm that the diversity (i.e., membership) of over- lapping nodes in the tested social networks is relatively small as well, typically 2 or 3 . 8. CONCLUSIONS AND DISCUSSIONS In this paper, we review a wide rang of overlapping community detection algo- rithms along with quality measures and several existing benchmarks. A number of tests are performed on the LFR benchmarks, incorporating different network struc- tures and various degree of overlapping. Quality evaluation is performed on both community and node levels to provide complementary information. Results show that the detection in networks with high overlapping density and high overlapping diversity has still space for improvements. The node level evaluation reveals the , Vol. V, No. N, May 2014. 32 · Jierui Xie et al. > d m O < 20 18 16 14 12 10 8 6 4 2 0 SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC > d m O < 20 18 16 14 12 10 8 6 4 2 0 SLPA COPRA GCE LFM Game OSLOM MOSES CIS Link iLCD UEOC KR FB LS DP CA PGP EM EP P2P AM KR FB LS DP CA PGP EM EP P2P AM Fig. 16. The number of detected mem- berships for social networks based on the clustering with the best QE ov. Fig. 17. The number of detected mem- berships for social networks based on the clustering with the best QN i ov . problems of over-detection and under-detection which needs to be considered when designing or evaluating detection algorithms. The results discovered in real-world social networks suggest the sensitivity of some algorithms to sparse networks. A common feature of social networks in view of agreement of different algorithms is relatively small number of overlapping nodes, most of which belong just to a few communities. Moreover, the ambiguity in the definition of overlapping nodes imposes challenges in real-life applications as well. Here, we review work that has been done mostly for unweighted networks. How- ever, there are number of applications where a weight bears significant information (e.g., the correlation network in biological studies [Langfelder and Horvath 2008]). Algorithms that explicitly take weights into account and allow overlapping, such as CPMw and SLPAw, expect to have advantages over others. Despite the large amount of work devoted to developing detection algorithms, there are a number of fundamental questions that have yet to be fully addressed. Two of the most prominent are when to apply overlapping methods and how signif- icant the overlapping is. It is natural to ask whether or not an application of the overlapping detection algorithms captures any additional information that a disjoint algorithm would necessarily miss. Unfortunately, measures like NMI and Omega do not offer a satisfying answer. The discussion on the necessity of overlap has largely been left unexplored. In [Kelley 2009], the authors empirically examined attributes of the vertices in a network representing commenting activity. The authors suggest that, for a pair of communities A and B, the trait similarity between A ∩ B and the sets A − B and B − A be higher than the similarity between A − B and B − A. Such a relationship might offer a way to estimate the validation of the overlap. The significance of community structures has been previously explored only within the context of disjoint community detection and based on the notion of modularity [Reichardt and Bornholdt 2006], [Guimer`a et al. 2004], [Massen and Doye 2005]. The robustness and uniqueness of a discovered partitioning is also examined in [Gfeller et al. 2005; Karrer et al. 2008; Massen and Doye 2007]. Many of these techniques can be extend to assess the overlapping community structure. Interestingly, statistical significance has begun to be included in detection method- ologies such as OSLOM [Lancichinetti et al. 2011]. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 33 ACKNOWLEDGMENTS The work of J.X and B.K.S was supported in part by the Army Research Labora- tory under Cooperative Agreement Number W911NF-09-2-0053 and by the Office of Naval Research Grant No. N00014-09-1-0607. The views and conclusions con- tained in this document are those of the authors and should not be interpreted as representing the official policies either expressed or implied of the Army Research Laboratory, the Office of Naval Research, or the U.S. Government. The submitted manuscript has been co-authored by a contractor of the U.S. Government (S.K.) under contract DE-AC05-00OR22725. Accordingly, the U.S. Government retains a nonexclusive, royalty-free license to publish or reproduce the published form of this contribution, or allows others to do so, for U.S. Government purposes. The authors would like to thank researchers who generously provided their soft- ware and helped setting up our experiments. We especially thank Steve Gregory for his helpful codes and discussions. REFERENCES Ahn, Y.-Y., Bagrow, J. P., and Lehmann, S. 2010. Link communities reveal multiscale com- plexity in networks. Nature 466, 761–764. Ankerst, M., Breunig, M. M., Kriegel, H.-P., and Sander, J. 1999. Optics: ordering points to identify the clustering structure. In Proc. SIGKDD Conf. 49–60. Arenas, A., D´ıaz-Guilera, A., and P´erez-Vicente, C. J. 2006. Synchronization reveals topo- logical scales in complex networks. Phys. Rev. Lett. 96, 11, 114102. Ball, B., Karrer, B., and Newman, M. E. J. 2011. Efficient and principled method for detecting communities in networks. Phys. Rev. E 84, 036103. Baumes, J., Goldberg, M., Krishnamoorthy, M., Magdon-Ismail, M., and Preston, N. 2005. Finding communities by clustering a graph into overlapping subgraphs. In Proc. IADIS. 97–104. Bianconi, G., Pin, P., and Marsili, M. 2008. Assessing the relevance of node features for network structure. Proc. Natl. Acad. Sci. USA 28, 7. Blatt, M., Wiseman, S., and Domany, E. 1996. Superparamagnetic clustering of data. Phys. Rev. Lett. 76, 3251–3254. Boguna, M., Pastor-Satorras, R., Diaz-Guilera, A., and Arenas, A. 2004. Models of social networks based on social distance attachment. Phys. Rev. E 70, 056122. Breve, F., Zhao, L., and Quiles, M. 2009. Uncovering overlap community structure in complex networks using particle competition. In Proc. ICAI. 619–628. Campello, R. J. G. B. 2007. A fuzzy extension of the rand index and other related indexes for clustering and classification assessment. Pattern Recogn. Lett. 28, 833–841. Campello., R. J. G. B. 2010. Generalized external indexes for comparing data partitions with overlapping categories. Pattern Recogn. Lett. 31, 966–975. Cazabet, R., Amblard, F., and Hanachi, C. 2010. Detection of overlapping communities in dynamical social networks. In Proc. SOCIALCOM. 309–314. Chen, D., Shang, M., Lv, Z., and Fu, Y. 2010. Detecting overlapping communities of weighted networks via a local algorithm. Physica A 389, 19, 4177–4187. Chen, J., Zaıane, O. R., and Goebel, R. 2009. A visual data mining approach to find overlapping communities in networks. In Proc. ASONAM Conf. 338–343. Chen, W., Liu, Z., Sun, X., and Wang, Y. 2010. A game-theoretic framework to identify overlapping communities in social networks. Data Min. Knowl. Discov. 21, 224–240. Collins, L. M. and Dent, C. W. 1988. Omega: A general formulation of the rand index of cluster recovery suitable for non-disjoint solutions. Multivar. Behav. Res. 23, 2 (Feb.), 231–242. Condon, A. and Karp, R. M. 2001. Algorithms for graph partitioning on the planted bisection model. Rand. Struct. Algo. 18, 116–140. , Vol. V, No. N, May 2014. 34 · Jierui Xie et al. Danon, L., Duch, J., Arenas, A., and Diaz-guilera, A. 2005. Comparing community structure identification. J. Stat. Mech., 09008. Davis, G. B. and Carley, K. 2008. Clearing the fog: Fuzzy, overlapping groups for social networks. Social Networks 30, 3, 201–212. Ding, F., Luo, Z., Shi, J., and Fang, X. 2010. Overlapping community detection by kernel-based fuzzy affinity propagation. In Proc. ISA Workshop. 1–4. Du, N., Wang, B., and Wu, B. 2008. Overlapping community structure detection in networks. In Proc. CIKM. 1371–1372. Evans, T. 2010. Clique graphs and overlapping communities (arxiv: 1009.0638). Evans, T. and Lambiotte, R. 2010. Line graphs of weighted networks for overlapping commu- nities. Eur. Phys. J. B 77, 265. Evans, T. S. and Lambiotte, R. 2009. Line graphs, link partitions and overlapping communities. Phys. Rev. E 80, 016105. Farkas, I., ´Abel, D., Palla, G., and Vicsek, T. 2007. Weighted network modules. New J. Phys. 9, 6, 180. Fisher, D. C. 1989. Lower bounds on the number of triangles in a graph. Journal of Graph Theory 13, 4, 505–512. Fortunato, S. 2010. Community detection in graphs. Phys. Rep. 486, 75–174. Frey, B. J. and Dueck, D. 2007. Clustering by passing messages between data points. Sci- ence 315, 972–976. Fu, Q. and Banerjee, A. 2008. Multiplicative mixture models for overlapping clustering. In Proc. ICDM. 791–796. Geweniger, T., Zuhlke, D., Hammer, B., and Villmann, T. 2009. Fuzzy variant of affinity propagation in comparison to median fuzzy c-means. In Proc. WSOM. 72–79. Gfeller, D., Chappelier, J.-C., and De Los Rios, P. 2005. Finding instabilities in the com- munity structure of complex networks. Phys. Rev. E 72, 056135. Girvan, M. and Newman, M. E. J. 2002. Community structure in social and biological networks. Proc. Natl. Acad. Sci. USA 99, 12, 7821–7826. Gregory, S. 2007. An algorithm to find overlapping community structure in networks. In Proc. PKDD Conf. 91–102. Gregory, S. 2008. A fast algorithm to find overlapping communities in networks. Lect. Notes Comput. Sci. 5211, 408. Gregory, S. 2009. Finding overlapping communities using disjoint community detection algo- rithms. CompleNet 207, 47–61. Gregory, S. 2010. Finding overlapping communities in networks by label propagation. New J. Phys. 12, 10301. Gregory, S. 2011. Fuzzy overlapping communities in networks. J. Stat. Mech. 2011, 02, P02017. Guimer`a, R., Sales-Pardo, M., and Amaral, L. A. N. 2004. Modularity from fluctuations in random graphs and complex networks. Phys. Rev. E 70, 025101. Havemann, F., Heinz, M., Struck, A., and Glaser, J. 2011. Identification of overlapping communities and their hierarchy by locally calculating community-changing resolution levels. J. Stat. Mech. 2011, 01, P01023. Hubert, L. and Arabie, P. 1985. Comparing partitions. Journal of Classification 2, 193–218. Hullermeier, E. and Rifqi, M. 2009. A fuzzy variant of the rand index for comparing clustering structures. In Proc. IFSA/EUSFLAT Conf. 1294–1298. Jin, D., Yang, B., Baquero, C., Liu, D., He, D., and Liu, J. 2011. A markov random walk under constraint for discovering overlapping communities in complex networks. J. Stat. Mech. 2011, 05, P05031. Karrer, B., Levina, E., and Newman, M. E. J. 2008. Robustness of community structure in networks. Phys. Rev. E 77, 046119. Kelley, S. 2009. The existence and discovery of overlapping communities in large-scale networks. Ph.D. thesis, Rensselaer Polytechnic Institute, Troy, NY. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 35 Kelley, S., Goldberg, M., Magdon-Ismail, M., Mertsalov, K., and Wallace, A. 2011. Hand- book of Optimization in Complex Networks. Springer, Chapter 6. Kim, Y. and Jeong, H. 2011. The map equation for link community (unpublished). Kov´acs, I. A., Palotai, R., Szalay, M., and Csermely, P. 2010. Community landscapes: An integrative approach to determine overlapping network module hierarchy, identify key nodes and predict network dynamics. PLoS ONE 5, 9, e12528. Kumpula, J. M., Kivela, M., Kaski, K., and Saramaki, J. 2008. Sequential algorithm for fast clique percolation. Phys. Rev. E 78, 2, 026109. Lancichinetti, A. and Fortunato, S. 2009. Community detection algorithms: a comparative analysis. Phys. Rev. E 80, 056117. Lancichinetti, A., Fortunato, S., and Kert´esz, J. 2009. Detecting the overlapping and hier- archical community structure of complex networks. New J. Phys. 11, 033015. Lancichinetti, A., Fortunato, S., and Radicchi, F. 2008. Benchmark graphs for testing community detection algorithms. Phys. Rev. E 78, 046110. Lancichinetti, A., Radicchi, F., Ramasco, J. J., and Fortunato, S. 2011. Finding statistically significant communities in networks. PLoS ONE 6, 4, e18961. Langfelder, P. and Horvath, S. 2008. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 1, 559. Latouche, P., Birmele, E., and Ambroise, C. 2011. Overlapping stochastic block models with application to the french political blogosphere. The Annals of Applied Statistics 5, 309–336. Lee, C., Reid, F., McDaid, A., and Hurley, N. 2010. Detecting highly overlapping community structure by greedy clique expansion. In Proc. SNAKDD Workshop. 33–42. Leskovec, J., Adamic, L. A., and Huberman, B. A. 2007. The dynamics of viral marketing. ACM Trans. Web 1, 5. Leskovec, J., Kleinberg, J., and Faloutsos, C. 2007. Graph evolution: Densification and shrinking diameters. ACM TKDD 1, 2. Leskovec, J., Lang, K. J., and andMichael W. Mahoney, A. D. 2009. Community structure in large networks: Natural cluster sizes and the absence of large well-defined clusters. Internet Mathematics 6, 29–123. Leskovec, J., Lang, K. J., and Mahoney, M. 2010. Empirical comparison of algorithms for network community detection. In Proc. WWW Conf. 631–640. Li, D., Leyva, I., Almendral, J., Sendina-Nadal, I., Buldu, J., Havlin, S., and Boccaletti, S. 2008. Synchronization interfaces and overlapping communities in complex networks. Phys. Rev. Lett. 101, 168701. Lu, Q., Korniss, G., and Szymanski, B. K. 2009. The naming game in social networks: com- munity formation and consensus engineering. J. Econ. Interact. Coord. 4, 221–235. M. Rosvall, C. B. 2008. Maps of random walks on complex networks reveal community structure. Proc. Natl. Acad. Sci. 105, 1118–1123. Magdon-ismail, M. and Purnell, J. 2011. Fast overlapping clustering of networks using sampled spectral distance embedding and gmms. Tech. rep., Rensselaer Polytechnic Institute. Massen, C. and Doye, J. 2005. Identifying communities within energy landscapes. Phys. Rev. E 71, 046101. Massen, C. and Doye, J. 2007. Thermodynamics of community structure. Preprint arXiv:cond- mat/0610077v1 . McDaid, A. and Hurley, N. 2010. Detecting highly overlapping communities with model-based overlapping seed expansion. In Proc. ASONAM Conf. 112–119. Molloy, M. and Reed, B. 1995. A critical point for random graphs with a given degree sequence. Rand. Struct. Algo. 6, 161–179. Moon, J. and Moser, L. 1965. On cliques in graphs. Israel Journal of Mathematics 3, 23–28. Nepusz, T., Petr´oczi, A., N´egyessy, L., and Bazs´o, F. 2008. Fuzzy communities and the concept of bridgeness in complex networks. Phys. Rev. E 77, 016107. Newman, M. E. J. 2006. Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E 74, 036104. , Vol. V, No. N, May 2014. 36 · Jierui Xie et al. Newman, M. E. J. and Leicht, E. A. 2007. Mixture models and exploratory analysis in networks. Proc. Natl. Acad. Sci. USA 104, 9564–9569. Newman, M. E. J., Strogatz, S. H., and Watts, D. J. 2001. Random graphs with arbitrary degree distributions and their applications. Phys. Rev. E 64, 2, 026118. Nicosia, V., Mangioni, G., Carchiolo, V., and Malgeri, M. 2009. Extending the definition of modularity to directed graphs with overlapping communities. J. Stat. Mech., 03024. Nowicki, K. and Snijders, T. A. B. 2001. Estimation and prediction for stochastic blockstruc- tures. JASA 96, 455, 1077–1087. Padrol-Sureda, A., Perarnau-Llobet, G., Pfeifle, J., and Munts-Mulero, V. 2010. Over- lapping community search for social networks. In Proc. ICDE. 992–995. Palla, G., Der´enyi, I., Farkas, I., and Vicsek, T. 2005. Uncovering the overlapping community structure of complex networks in nature and society. Nature 435, 814–818. Psorakis, I., Roberts, S., Ebden, M., and Sheldon, B. 2011. Overlapping community detection using bayesian non-negative matrix factorization. Phys. Rev. E 83, 6, 066114. Raghavan, U. N., Albert, R., and Kumara, S. 2007. Near linear time algorithm to detect community structures in large-scale networks. Phys. Rev. E 76, 036106. Rees, B. and Gallagher, K. 2010. Overlapping community detection by collective friendship group inference. In Proc. ASONAM Conf. 375–379. Reichardt, J. and Bornholdt, S. 2004. Detecting fuzzy community structures in complex networks with a potts model. Phys. Rev. Lett. 93, 218701. Reichardt., J. and Bornholdt, S. 2006. Statistical mechanics of community detection. Phys. Rev. E 74, 1, 016110. Reichardt, J. and Bornholdt, S. 2006. When are networks truly modular? Physica D 224, 20–26. Reid, F., McDaid, A. F., and Hurley, N. J. 2011. Partitioning breaks communities. In Proc. ASONAM Conf. 102–109. Ren, W., Yan, G., Liao, X., and Xiao, L. 2009. Simple probabilistic algorithm for detecting community structure. Phys. Rev. E 79, 3, 036111. Richardson, M., Agrawal, R., and Domingos, P. 2003. Trust management for the semantic web. In Proc. ISWC. Vol. 2870. 351–368. Ripeanu, M., Foster, I., and Iamnitchi, A. 2002. Mapping the gnutella network: Properties of large-scale peer-to-peer systems and implications for system design. IEEE Internet Computing Journal 6, 1. Ronhovde, P. and Nussinov, Z. 2009. Multiresolution community detection for megascale net- works by information-based replica correlations. Phys. Rev. E 80, 016109. Sawardecker, E., Sales-Pardo, M., and Amaral, L. 2009. Detection of node group membership in networks with group overlap. Eur. Phys. J. B 67, 277. Shen, H., Cheng, X., Cai, K., and Hu, M.-B. 2009. Detect overlapping and hierarchical com- munity structure. Physica A 388, 1706. Shen, H., Cheng, X., and Guo, J. 2009. Quantifying and identifying the overlapping community structure in networks. J. Stat. Mech. 07, 9. Wang, X., Jiao, L., and Wu, J. 2009. Adjusting from disjoint to overlapping community detection of complex networks. Physica A 388, 5045–5056. White, S. and Smyth, P. 2005. A spectral clustering approach to finding communities in graphs. In Proc. SIAM International Conference on Data Mining. 76–84. Wu, Z., Lin, Y., Wan, H., and Tian, S. 2010. A fast and reasonable method for community detection with adjustable extent of overlapping. In Proc. ISKE Conf. 376–379. Xie, J. and Szymanski, B. K. 2011. Community detection using a neighborhood strength driven label propagation algorithm. In Proc. NSW. 188–195. Xie, J. and Szymanski, B. K. 2012. Towards linear time overlapping community detection in social networks. In Proc. PAKDD Conf. 25–36. , Vol. V, No. N, May 2014. Overlapping Community Detection in Networks: the State of the Art and Comparative Study · 37 Xie, J., Szymanski, B. K., and Liu, X. 2011. SLPA: Uncovering overlapping communities in social networks via a speaker-listener interaction dynamic process. In Proc. ICDM Workshop. 344–349. Zarei, M., Izadi, D., and Samani, K. A. 2009. Detecting overlapping community structure of networks based on vertex-vertex correlations. J. Stat. Mech. 2009, 11, P11013. Zhang, S., Wang, R.-S., and Zhang, X.-S. 2007a. Identification of overlapping community structure in complex networks using fuzzy c-means clustering. Physica A 374, 483–490. Zhang, S., Wang, R.-S., and Zhang, X.-S. 2007b. Uncovering fuzzy community structure in complex networks. Phys. Rev. E 76, 4, 046103. Zhang, Y., Wang, J., Wang, Y., and Zhou, L. 2009. Parallel community detection on large networks with propinquity dynamics. In Proc. SIGKDD Conf. 997–1006. Zhao, K., Zhang, S.-W., and Pan, Q. 2010. Fuzzy analysis for overlapping community structure of complex network. In Proc. CCDC. 3976–3981. Received Month Year; revised Month Year; accepted Month Year , Vol. V, No. N, May 2014.
1610.04170	2	1610	2018-01-17T17:44:51	Network segregation in a model of misinformation and fact checking	[ "cs.SI", "cs.CY", "physics.soc-ph" ]	Misinformation under the form of rumor, hoaxes, and conspiracy theories spreads on social media at alarming rates. One hypothesis is that, since social media are shaped by homophily, belief in misinformation may be more likely to thrive on those social circles that are segregated from the rest of the network. One possible antidote is fact checking which, in some cases, is known to stop rumors from spreading further. However, fact checking may also backfire and reinforce the belief in a hoax. Here we take into account the combination of network segregation, finite memory and attention, and fact-checking efforts. We consider a compartmental model of two interacting epidemic processes over a network that is segregated between gullible and skeptic users. Extensive simulation and mean-field analysis show that a more segregated network facilitates the spread of a hoax only at low forgetting rates, but has no effect when agents forget at faster rates. This finding may inform the development of mitigation techniques and overall inform on the risks of uncontrolled misinformation online.	cs.SI	cs	Network Segregation in a Model of Misinformation and Fact-checking Marcella Tambuscio,1, ∗ Diego F.M. Oliveira,2, 3, 4 Giovanni Luca Ciampaglia,5 and Giancarlo Ruffo1 1Computer Science Department, University of Turin, Italy 2School of Informatics, Computing, and Engineering, Indiana University, Bloomington, IN, USA 3U.S Army Research Laboratory, 2800 Powder Mill Rd., Adelphi, MD 20783 USA. 4Network Science and Technology Center, Rensselaer Polytechnic Institute, 335 Materials Research Center 110 8th St. Troy, NY 12180 USA 5Network Science Institute, Indiana University, IN, USA (Dated: January 18, 2018) Abstract Misinformation under the form of rumor, hoaxes, and conspiracy theories spreads on social media at alarming rates. One hypothesis is that, since social media are shaped by homophily, belief in misinformation may be more likely to thrive on those social circles that are segregated from the rest of the network. One possible antidote to misinformation is fact checking which, however, does not always stop rumors from spreading further, owing to selective exposure and our limited attention. What are the conditions under which factual verification are effective at containing the spreading of misinformation? Here we take into account the combination of selective exposure due to network segregation, forgetting (i.e., finite memory), and fact-checking. We consider a compartmental model of two interacting epidemic processes over a network that is segregated between gullible and skeptic users. Extensive simulation and mean-field analysis show that a more segregated network facilitates the spread of a hoax only at low forgetting rates, but has no effect when agents forget at faster rates. This finding may inform the development of mitigation techniques and raise awareness on the risks of uncontrolled misinformation online. 8 1 0 2 n a J 7 1 ] I S . s c [ 2 v 0 7 1 4 0 . 0 1 6 1 : v i X r a ∗ Corresponding author: [email protected] 1 I. INTRODUCTION Social media are rife with inaccurate information of all sorts [6, 18, 21]. This is in part due to their egalitarian, bottom-up model of information consumption and production [9], according to which users can broadcast to their peers information vetted by neither experts nor journalists, and thus potentially inaccurate or misleading [28]. Examples of social media misinformation include rumors [21], hoaxes [36], and conspiracy theories [3, 24]. In journalism, corrections, verification, and fact-checking are simple yet powerful anti- dotes to misinformation [11], and several newsrooms employ these techniques to vet the information they publish. Moreover, in recent years, several independent fact-checking or- ganizations have emerged with the goal of debunking widely circulating claims online. From now on, we refer to all these practices collectively as fact-checking. Among the leading US-based fact-checking organizations we can cite Snopes [43], FactCheck.org [20], and Poli- tifact [46]. Several more are joining their ranks worldwide [34]. In many cases these organi- zations cannot cope with the sheer volume of misinformation circulating online, and some are exploring alternatives to scale their verification efforts, including automated techniques [16], and collaboration with technology platforms such as Facebook [38] and Google [19]. These trends thus beg a rather fundamental question - is the dissemination of fact- checking information effective at stopping misinformation from spreading on social media? In particular cases, timely corrections are enough to limit a rumor from spreading further [4, 21, 35]. However, administering fact-checking information may also have adverse effects. For example, in some instances it has been observed that correcting an inaccurate or misleading claim can have counterproductive effects, increasing - and not decreasing - belief in it. This is a phenomenon called the backfire effect [36]. Recent work has however failed to replicate this form of backfiring in independent trials, suggesting that it is a rather elusive phenomenon [48]. Fact-checking could also lead to a hypercorrection effect, meaning that providing accurate information to people who have been exposed to misinformation may cause them, on the long term, to forget the former, and remember the latter [12]. Thus, given the growing emphasis put into fact-checking, as well as its unintended side effects, it is clear that, for a better understanding of how to fight social media misinformation, it would be useful to explore the relation between fact-checking and the misinformation it is intended to quell. 2 Recent work has also revealed that, when it comes to misinformation, online conver- sations tend to be highly polarized [10, 18]. This suggests the importance of homophily and segregation in the spread of misinformation. Since social networks are shaped by ho- mophily [29], one hypothesis is that misinformation may be more likely to thrive in those social circles that are segregated from the rest of the network. Social media may be partic- ularly susceptible to this aspect due to the fact that exposure to information is mediated in part by algorithms, whose goal is to filter and recommend stories that have a high potential for engagement. This may create filter bubbles and echo chambers, information spaces that favor confirmation bias and repetition [39, 44]. Recent work has started to measure the extent to which editorial decisions performed automatically by algorithms affect selective exposure, and thus segregation of the information space [7, 33]. Therefore, in modeling the interplay between misinformation and fact-checking, our second goal is to shed light on the role of the underlying social network structure in the spreading process, in particular the presence of communities of users with different attitude toward unvetted and unconfirmed information - which could potentially constitute misinformation. Besides segregation, in the literature there is also disagreement about whether weak ties - the links that connect different communities together - play a role in the diffusion of information. Some studies suggest that weak ties play an important role [8]; others that they do not [37]. In their seminal work on complex social contagion, Centola and Macy argue that the spread of collective actions benefits from bridges, i.e. ties that are "wide enough to transmit strong social reinforcement" [13]. It is well known that misinformation can be propagated thanks to repetition [2, 27], which in some ways can be obtained through social reinforcement, and thus it would be useful to investigate this additional aspect as well. In terms of modeling, there has hitherto been little work on characterizing the epidemic spreading of different types of information, with most efforts devoted to describing mu- tually independent processes [23, 32]. Instead, the presence of the rich cognitive effects just described suggests that misinformation and fact-checking interact and compete for the attention of individuals on social media, and this could lead to non-trivial diffusion dynam- ics. Among the work specifically devoted to competition in the diffusion of information, or memes, the literature has focused on the role of limited attention [25, 47], as well as that of information quality [31, 41]. Several models have been proposed in prior work to describe the propagation of rumors 3 1 − fi(t) − g(t) S fi(t) gi(t) (1 − pv) (1 − pf ) B pf pf pv (1 − pf ) F (1 − pf ) FIG. 1. The transitions states for the generic i-th agent of our hoax epidemic model. To simplify the model, here we set pv = 1 − α. in a complex social networks [1, 14, 17, 30]. Most are based on the epidemic compartmental models like the SIR (Susceptible–Infected–Recovered) [40], contemplating the fact-checking only as a remedy after the hoax infection. Another class of models uses branching processes on signed networks to take into account user polarization [18]. Neither type, however, takes into account in the same model the three aforementioned mechanisms - competition between hoax and fact-checking, forgetting mechanisms and segregation. To consider all these features, here we introduce a simple agent-based model in which individuals are endowed with finite memory and a fixed predisposition toward factual veri- fication. In this model, hoax and fact checks compete on a network formed by two groups, the gullible and the skeptic, marked by a different tendency to believe in the hoax. Varying the level of segregation in the network, as well as the relative credibility of the hoax among the two groups, we look at whether the hoax becomes endemic or instead is eradicated from the whole population. II. MODEL Here we describe a model of the spread of the belief in a hoax and the related fact checking within a social network of agents with finite memory. An agent can be in any of the following three states: 'Susceptible' (denoted by S), if they have not heard about neither the hoax nor the fact checking, or if they have forgotten about it; 'Believer' (B), if they believe in the hoax and choose to spread it; and 'Fact checker' (F ) if they know the hoax is false - for 4 example after having consulted an accurate news source - and choose to spread the fact checking. Let us consider the i-th agent at time step t and let us denote with nX i (t) the number of its neighbors in state X ∈ {S, B, F}. We assume that an agent 'decides' to believe in either the hoax or the fact checking as a result of interaction over interpersonal ties. This could be due to social conformity [5] or because agents accept information from their neighbors [42]. Second, we assume that the hoax displays an intrinsic credibility α ∈ [0, 1], which, all else being equal, makes it more believable than the fact checking. We will discuss later how this parameter can be also related to the users: by now, we consider it as a feature of the hoax. Thus, the probability of transitioning from S to either B or F are given by functions fi, and gi, respectively: fi(t) = β gi(t) = β nB i (1 + α) i (1 + α) + nF nB i (1 − α) nF i (1 + α) + nF nB i (1 − α) i (1 − α) (1) (2) where β ∈ [0, 1] is the overall spreading rate. Furthermore, agents possess finite memory, which means that at each time step, any believer of fact checker can 'forget' about the hoax with fixed probability pf or the fact check and become susceptible again. Finally, any believer who has not forgotten the hoax yet can decide to check the news and stop believing in the hoax, becoming a fact checker. This happens with probability pv. In any other case, an agent remains in its current state. The full model with the transition states are shown in Fig. 1. Observe that, since fi(t) + gi(t) = β, which is equivalent to the infection rate of the SIS model. Indeed, if one considers the two states B and F as single 'Infected' state (I), then our model is reduced to an SIS model, with the only difference that the probability of recovery µ is denoted by pf . Let us denote by si(t) the state of the i-th agent at time t, and let us define, for X ∈ {B, F, S}, the state indicator function sX i (t)(cid:3) describes the probability that a node i is in any of the three states i (t) = δ(si(t), X). The triple pi(t) = (cid:2)pB i (t), pF i (t), pS at time t. The dynamics of the system at time t + 1 will be then given by a random 5 realization of pi at t + 1. Thus, pi(t + 1) can be described as: i (t + 1) = fi(t)sS pB pF i (t + 1) = gi(t)sS pS i (t + 1) = pf (cid:2)sB i (t) + (1 − pf )(1 − pv)sB i (t) + pv(1 − pf )sB i (t) + sF i (t)(cid:3) + [1 − fi(t) − gi(t)] sS i (t) i (t) + (1 − pf )sF i (t) i (t) (3) (4) (5) In previous work [45] we analyzed the behavior of the model at equilibrium. Starting from a well-mixed topology of N agents, in which a few agents have been initially seeded as believers, we derived the expressions for the density of believers, fact checkers, and suscep- tible agents in the infinite-time limit denoted by B∞, F∞, and S∞, respectively. We found that, independent of the network topology (Barab´asi-Albert and Erdos-R´enyi), the value of pv, and of α, S∞ stabilizes around the same values in all simulations. We confirmed such a result using both mean-field equations and simulations. Once the system reaches equilibrium, the relative ratio between believers and fact checkers is determined by α and pv: such as the higher α, the more believers, and conversely for pv. In particular, we showed that there always exists a critical value of pv above which the hoax is completely eradicated from the network (i.e., B∞ = 0). This value depends on α and pf , but not on the spreading rate β. As one can see, the model has several parameters, namely, spreading rate β, credibility of the hoax α, probability of verification pv and probability of forgetting pf . Since, in the present work, we want to consider the role of communities of people with different attitudes to believe to an hoax, the number of parameters is going to increase. As an attempt to reduce the number of parameters, we set pv = 1 − α. (6) This simplification can be motivated by assuming that the more credible a piece of news is, the lower are the chances anybody will try to check its veracity. This means that we restrict the parameters space pv × α to a line. This constrain can be easily observed in Fig. 3 (left), where the curve represents the analytic threshold on the verifying probability. Above it the hoax becomes endemic, over it the hoax is completely removed. We note that even with this additional constrain, this new, simplified model exhibits the same behaviors that our original model can produce (i.e., believers survive or not). 6 FIG. 2. Epidemic threshold for the simplified version of the model given by Eq. 6. The grey area indicates the region of the parameter space where the hoax is completely removed from the network. The white part indicates the region of the parameter space where the hoax can become endemic. Recomputing the mean-field equations with Eq. 6, we obtain now a critical value for pf , a sufficient condition that guarantees the removal of the hoax from the network: pf ≤ (1 − α)2 1 + α2 =⇒ pB(∞) = 0. (7) The behaviour of pf versus α is shown in Fig. 2. For any combination of pf and α below the curve, the hoax is completely removed from the network. For combinations above the curve, the infection is instead endemic. The forgetting probability can be considered as a measure of the attention toward a specific topic, meaning that if there is a large discussion around the subject, then exposed people tend to have a stable opinion about it, otherwise the probability to forget the belief and the news will be higher. The presence of this threshold in Eq. 7 could suggest that the level of attention plays an important role in fake news global spread and persistence. III. RESULTS Two parameters govern the spreading dynamics in our model. These are the credibility α and forgetting probability pf . To address our research question about the role of network 7 FIG. 3. Simplification of the model setting pv = 1 − α, here fixing pf = 0.1. On the left we can observe the phase diagram of the entire parameter space considered for the model. In this new work we are restricting it to the dashed line, but we are keeping all the possible configurations of the model: believers can survive (dark line) or not (pale line), as we can see comparing the dots on the left to the curves on the right that show evolution of believers numbers in the network over time. structure and communities, we consider a simple generative model of a segregated network. Let us consider N agents divided into two groups, one comprised by t < N agents whose beliefs conform more to the hoax than the other one, which is comprised by the rest of the population. We call the former the gullible group, while the latter the skeptic group. To represent this in our framework, we set different values of α for each agent group (either αgu or αsk, αgu > αsk). This is not a contradiction with what we said before: the credibility is a parameter describing the hoax, but of course is also related to the user attitude and personal worldviews, then it is reasonable to think to different values for different groups. To generate the network, we assign M edges at random. Let s ∈(cid:2) 1 2, 1(cid:1) denote the fraction of intra-group edges, regardless of the group. For each edge we first decide, with probability s, whether two individuals from the same group (intra-group tie) or different groups (inter- group tie) should be connected. In the case of an intra-group tie, we select a group with probability proportional to the relative ratio of the total number of possible inter-group ties 8 B(t)=0.4=0.9FIG. 4. Network structure under different segregation regimes between two groups (in this case of equal size). In the figure, three different values of s were used. (a): s = 0.6, (b): s = 0.8, and (c): s = 0.95. Node layout was computed using a force-directed algorithm [22]. (of that group) to that of the whole network; then, we pick uniformly at random two agents from that group and connect them. In the second case, two agents are chosen at random, one per group, and connected. Fig. 4 shows three examples of network with different values of s. To understand the behavior of the model in this segregated network, we performed ex- tensive numerical simulations. We set fixed values for αsk and we considered a wide range of values of αgu, pf , s, and t. Fig. 5 reports the results of the first of these exercises, showing the overall number of believers B∞ in the whole population at equilibrium. Increasing either γ or αgu we see an increase of B∞, all else being equal. However, when we changed the segregation s we observed two different situations: for small pf , an increase 9 FIG. 5. Believers at equilibrium in the phase space of s × αgu. We considered two forgetting low forgetting, pf = 0.1, and (b): high forgetting, pf = 0.8. Other parameters: regimes: (a): αsk = 0.4. Each point was averaged over 50 simulations. of s resulted in an increase of B∞. Conversely - and perhaps a bit surprisingly - under high values of pf increasing s does not change B∞. Trying to better understand the role of pf , we further explore the behavior of the model by varying the size of the gullible group γ and its level of segregation s. In Fig. 6 we report the relevant phase diagrams, breaking down the number of believers at equilibrium by group, i.e., B∞ = Bgu∞ + Bsk∞. If pf is low (Fig 6, left column), the overall number of believers depends heavily on Bgu∞ , whereas Bsk∞ ≈ 0, and the segregation is unimportant,see Fig. 6(a) and Fig. 6(c). Instead, with an high rate of forgetting (right column), B∞ (Fig. 6(f)) depends on both Bsk∞ and Bgu∞ . But in this case we have a different role of the segregation: while in the skeptic group Bsk∞ decreases when s increases, see Fig. 6(d), in the gullible group s has fewer influence, see Fig. 6(b). To give an analytical support to our findings, we obtain mean-field approximation for the model (details in the Appendix) and we perform both numerical integration of the mean field equations and agent-based simulations, which give very similar results. Fig. 7 shows 10 FIG. 6. Believers at equilibrium under low (pf = 0.1) and high forgetting (pf = 0.8) rate. The number of believers at equilibrium is broken down as B∞ = Bgu∞ +Bsk∞. Phase diagrams in the space s × γ for (a) Bgu∞ , low forgetting, (b) Bgu∞ , high forgetting, (c) Bsk∞ low forgetting, (d) Bsk∞ high forgetting, (e) B∞ low forgetting, and (f ) B∞ high forgetting. We fixed αgu = 0.9 and αsk = 0.05. the phase diagrams obtained by numerical simulations of the mean-field equations. Summarizing, segregation can have a very different role on the final configuration of the hoax spreading and this depends on the forgetting rate. Why the number of links among communities with different behaviors is so important? It should be noted that any 'network effect' present in our model will only appear in the infection phase, that is for transitions S → B and S → F . To better understand what happens in both groups, we computed the rate at which these transitions happen, that is, the conditional probability of, being susceptible, becoming either believer or fact checker. Let us consider a susceptible agent in the gullible group. At low forgetting rates, in the gullible group more intra-group ties (i.e. an higher s) increase the chances of becoming a 11 FIG. 7. Mean-field approximation for different values of pv: these phase diagrams represent the density of Believers at equilibrium varying γ and s, exactly as in Figure 6. believer and reduce those of becoming fact checker; see Fig. 8 (top left). In the skeptic group, the segregation effect is almost negligible (top right). This happens because inter-group ties expose the susceptible agents, among the gullible, to more members of the skeptic group, who are largely fact checkers. At high forgetting rates, instead, we observe the opposite behavior: more inter-group ties translate into more exposure, for susceptible users in the skeptic group, to gullible agents who are, by and large, believers. In the gullible group (bottom left of Fig. 8), segregation is not very important while, in the skeptic group, more connections with the gullible mean more believers (bottom right of Fig. 8). In other words, the role of segregation, being related to the abundance of inter-group ties, can have both a positive and negative role in stopping the spread of misinformation: for low forgetting rates, these links can help the spread of the debunking in the gullible group, while for high forgetting rates, they have the opposite effect, helping the hoax spread in the skeptic group. 12 Gullible BelieversSkeptic BelieversTotal Believers0.50.60.70.80.91.00.250.500.751.000.250.500.751.000.250.500.751.00segregationsize0.000.250.500.751.00densityForgetting probability=0.1Gullible BelieversSkeptic BelieversTotal Believers0.50.60.70.80.91.00.250.500.751.000.250.500.751.000.250.500.751.00segregationsize0.000.250.500.751.00densityForgetting probability=0.8FIG. 8. Rate of transitions of type S → B and S → F at equilibrium. We run the simulation until the system has reached the steady state and then compute the average number of transitions per susceptible. The plot shows averages over 50 simulations. IV. DISCUSSION Using agent-based simulations, here we have analyzed the role of the underlying structure of the network on which the diffusion of a piece of misinformation takes place. In particular we consider a network formed by two groups - gullible and skeptic - characterized by different values of the credibility parameter α. In order to study how the social structure shapes information exposure, we introduce a parameter s that regulates the abundance of ties between these two groups. We observe that s has an important role in the diffusion of misinformation. If the probability of forgetting pf is small then the fraction of the population affected by the hoax will be large or small depending on whether the network is respectively segregated or not. However, if the rate of forgetting is large, segregation has no effect on the spread of the hoax. 13 SUSCEPTIBLE TO BELIEVERSUSCEPTIBLE TOFACT-CHECKERtransitions rateThe probability of forgetting could be also interpreted as a measure of how much a given topic is discussed. A low value of pf could perhaps fit well with the scenario of a ideas whose belief tends to be more persistent over time, for example conspiracy theories. A high value pf could fit better with situations where beliefs are short lived, either because the claims are easy to debunk or are no more interesting than mere gossip, whose information value is transient. Hoaxes about the alleged death of celebrities, for instance, could fall within this latter category. On the basis of the findings presented in this paper, further research should be devoted to understanding the role of segregation in the spread of misinformation on social media. For conspiracy theories, it could be useful to analyze what happens if the communication among different groups increases. Moreover, it could be also interesting to consider more realistic situations in which credibility is distributed over the population according to socio-economic features and other individual-level attributes. Our results are also important from a purely theoretical point of view. Indeed, the model we had introduced in prior work, and on which we have build upon here, was an example of an epidemic process that is not affected by the network topology, meaning that the structure does not influence the final configuration of the network - indeed, it could be proved that there are no significant differences in the behavior of the spreading dynamics in random or scale-free networks [45]. In the present work, however, we show that actually network structure can become very important if we add a layer of complexity, characterizing groups of nodes with slightly different behaviors (here different values of the credibility parameter). This can suggest a lot of research, experiments and simulations in order to understand which parameters are sensitive to the segregation level, or some other topology measure, even in models that have usually topology-independent dynamics. In conclusion, understanding the production and consumption of misinformation is a crit- ical issue [15]. As several episodes are showing, there are obvious consequences connected to the uncontrolled production and consumption of inaccurate information [26]. A more thor- ough understanding of rumor propagation and the structural properties of the information exchange networks on which this happens may help mitigate these risks. 14 ACKNOWLEDGMENTS The authors would like to acknowledge Filippo Menczer and Alessandro Flammini for feedback and insightful conversations. DFMO acknowledges the support from James S. McDonnell Foundation. GLC acknowledges support from the Indiana University Net- work Science Institute (iuni.iu.edu) and from the Swiss National Science Foundation (PBTIP2 142353). Appendix A: Appendix 1. Mean-Field computations In previous work we showed mean-field analysis for our model on a homogeneous net- work [45]. Similarly, here we perform a similar analysis for our model on a network segre- gated in two groups, skeptic and gullible: for each group we have three equations (see Eq. 5) representing the spreading process. ig (t) + (1 − pf )(1 − pv)sB pB ig (t + 1) = fig (t)sS ig (t) + pv(1 − pf )sB ig (t + 1) = gig (t)sS pF ig (t) + sF sB pS ig (t + 1) = pf +(cid:2)1 − fig (t) − gig (t)(cid:3) sS ig (t) + (1 − pf )sF ig (t) ig (t) (cid:104) (cid:105) ig (t) ig (t) (t) + (1 − pf )(1 − pv)sB (t) + pv(1 − pf )sB (t)(cid:3) + [1 − fisk(t) − gisk(t)] sS (t) + (1 − pf )sF (t) isk isk isk (t) (t) isk (t + 1) = pf (t) + sF isk isk (t + 1) = fisk(t)sS isk (t + 1) = gisk(t)sS isk (cid:2)sB pB isk pF isk pS isk In these equations we can substitute si(t) with pi(t) and when t → ∞ we can assume g (∞) = pB pi(t) = pi(t+1) = pi(∞) for all i ∈ N . Hereafter we simplify the notation using pB (and analogously for the other cases). Now, let us consider the spreading functions for the g gullible agents. Similar equations can be written for the case of skeptic agents. The spreading functions are: 15 fig (t) = β gig (t) = β nB ig (1 + α) ig (1 + α) + nig iF (1 − α) nB ig (1 − α) nF nB ig (1 + α) + nF ig (1 − α) bors are chosen randomly, we can write nB Assuming that all vertices have the same number of neighbors (cid:104)k(cid:105), and that these neigh- g and i , we can obtain an expression that is not dependent on i. This simplifies the equations and lets us to simulate the process iterating the application i = s·nBig +(1−s)·nB = (1−γ)·(cid:104)k(cid:105)pB sk. Similarly, for nF ig = γ·(cid:104)k(cid:105)pB , where nB nB isk isk of them until the values of pS sk, pB sk, pF sk, pS g , pB g , pF g have reached stability. [1] D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi. Spread of (mis) information in social networks. Games and Economic Behavior, 70(2):194–227, 2010. [2] G. W. Allport and L. Postman. The psychology of rumor. Oxford, England: Henry Holt, 1947. [3] A. Anagnostopoulos, A. Bessi, G. Caldarelli, M. Del Vicario, F. Petroni, A. Scala, F. Zollo, and W. Quattrociocchi. Viral misinformation: the role of homophily and polarization. arXiv preprint arXiv:1411.2893, 2014. [4] C. Andrews, E. Fichet, Y. Ding, E. S. Spiro, and K. Starbird. Keeping up with the tweet- dashians: The impact of 'official' accounts on online rumoring. In Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work & Social Computing, CSCW '16, pages 452–465, New York, NY, USA, 2016. ACM. [5] S. E. Asch. Effects of group pressure upon the modification and distortion of judgements. In M. Henle, editor, Documents of Gestalt Psychology, pages 222–236. University of California Press, Oakland, California, USA, 1961. [6] E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts. Everyone's an influencer: Quan- tifying influence on twitter. In Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, WSDM '11, pages 65–74, New York, NY, USA, 2011. ACM. [7] E. Bakshy, S. Messing, and L. A. Adamic. Exposure to ideologically diverse news and opinion on facebook. Science, 348(6239):1130–1132, 2015. 16 [8] E. Bakshy, I. Rosenn, C. Marlow, and L. Adamic. The role of social networks in information diffusion. In Proceedings of the 21st international conference on World Wide Web, pages 519–528. ACM, 2012. [9] Y. Benkler. The wealth of networks: How social production transforms markets and freedom. Yale University Press, 2006. [10] A. Bessi, M. Coletto, G. A. Davidescu, A. Scala, G. Caldarelli, and W. Quattrociocchi. Science vs conspiracy: Collective narratives in the age of misinformation. PLoS ONE, 10(2):e0118093, Feb. 2015. [11] B. Borel. The Chicago Guide to Fact-Checking. The University of Chicago Press, Chicago, IL, USA, 2016. [12] A. C. Butler, L. K. Fazio, and E. J. Marsh. The hypercorrection effect persists over a week, but high-confidence errors return. Psychonomic bulletin & review, 18(6):1238–1244, 2011. [13] D. Centola and M. Macy. Complex contagions and the weakness of long ties. American journal of Sociology, 113(3):702–734, 2007. [14] F. Chierichetti, S. Lattanzi, and A. Panconesi. Rumor spreading in social networks. In Automata, Languages and Programming, pages 375–386. Springer, 2009. [15] G. L. Ciampaglia, A. Flammini, and F. Menczer. The production of information in the attention economy. Scientific Reports, 5:9452, 2015. [16] G. L. Ciampaglia, P. Shiralkar, L. M. Rocha, J. Bollen, F. Menczer, and A. Flammini. Com- putational fact checking from knowledge networks. PLoS ONE, 10(6):e0128193, June 2015. [17] D. J. Daley and D. G. Kendall. Epidemics and rumours. Nature, 204:1118, 1964. [18] M. Del Vicario, A. Bessi, F. Zollo, F. Petroni, A. Scala, G. Caldarelli, H. E. Stanley, and W. Quattrociocchi. The spreading of misinformation online. Proceedings of the National Academy of Sciences, 2016. [19] X. L. Dong, E. Gabrilovich, K. Murphy, V. Dang, W. Horn, C. Lugaresi, S. Sun, and W. Zhang. Knowledge-based trust: Estimating the trustworthiness of web sources. Proceedings of the VLDB Endowment, 8(9):938–949, 2015. [20] Factcheck.org. A project of the annenberg public policy center, Oct. 2017. [Online; Last accessed: 28-Oct-2017]. [21] A. Friggeri, L. A. Adamic, D. Eckles, and J. Cheng. Rumor cascades. In Proc. Eighth Intl. AAAI Conf. on Weblogs and Social Media (ICWSM), pages 101–110, 2014. 17 [22] T. M. J. Fruchterman and E. M. Reingold. Graph drawing by force-directed placement. Software: Practice & Experience, 21(11):1129–1164, Nov. 1991. [23] S. Funk and V. A. A. Jansen. Interacting epidemics on overlay networks. Phys. Rev. E, 81:036118, Mar. 2010. [24] S. Galam. Modelling rumors: the no plane pentagon french hoax case. Physica A: Statistical Mechanics and Its Applications, 320:571–580, 2003. [25] J. P. Gleeson, J. A. Ward, K. P. O'Sullivan, and W. T. Lee. Competition-induced criticality in a model of meme popularity. Phys. Rev. Lett., 112:048701, Jan. 2014. [26] L. Howell et al. Digital wildfires in a hyperconnected world. In Global Risks. World Economic Forum, 2013. [27] R. H. Knapp. A psychology of rumor. Public opinion quarterly, 8(1):22–37, 1944. [28] H. Kwak, C. Lee, H. Park, and S. Moon. What is twitter, a social network or a news media? In Proceedings of the 19th International Conference on World Wide Web, WWW '10, pages 591–600, New York, NY, USA, 2010. ACM. [29] M. McPherson, L. Smith-Lovin, and J. M. Cook. Birds of a feather: Homophily in social networks. Annual Review of Sociology, 27(1):415–444, 2001. [30] Y. Moreno, M. Nekovee, and A. F. Pacheco. Dynamics of rumor spreading in complex net- works. Physical Review E, 69(6):066130, 2004. [31] A. Nematzadeh, G. L. Ciampaglia, F. Menczer, and A. Flammini. How algorithmic popularity bias hinders or promotes quality. e-print, CoRR, July 2017. [32] M. E. J. Newman and C. R. Ferrario. Interacting epidemics and coinfection on contact net- works. PLOS ONE, 8(8):1–8, Aug. 2013. [33] D. Nikolov, D. F. Oliveira, A. Flammini, and F. Menczer. Measuring online social bubbles. PeerJ Computer Science, 1:e38, 2015. [34] The Duke Reporters' Lab keeps an updated list of global fact-checking sites at https:// reporterslab.org/fact-checking/. [35] B. Nyhan and J. Reifler. The effect of fact-checking on elites: A field experiment on us state legislators. American Journal of Political Science, 59(3):628–640, 2015. [36] B. Nyhan, J. Reifler, and P. A. Ubel. The hazards of correcting myths about health care reform. Medical Care, 51(2):127–132, 2013. [37] J.-P. Onnela, J. Saramaki, J. Hyvonen, G. Szab´o, D. Lazer, K. Kaski, J. Kert´esz, and A.-L. 18 Barab´asi. Structure and tie strengths in mobile communication networks. Proceedings of the national academy of sciences, 104(18):7332–7336, 2007. [38] E. Owens and U. Weinsberg. News feed fyi: Showing fewer hoaxes, 2015. [Online; accessed January 2016]. [39] E. Pariser. The filter bubble: What the Internet is hiding from you. Penguin UK, 2011. [40] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, and A. Vespignani. Epidemic processes in complex networks. Rev. Mod. Phys., 87:925–979, Aug. 2015. [41] X. Qiu, D. F. Oliveira, A. S. Shirazi, A. Flammini, and F. Menczer. Limited individual atten- tion and online virality of low-quality information. Nature Human Behaviour, 1(7):s41562–017, 2017. [42] R. L. Rosnow and G. A. Fine. Rumor and gossip: The social psychology of hearsay. Elsevier, 1976. [43] Snopes.com. The definitive fact-checking site and reference source for urban legends, folklore, myths, rumors, and misinformation., Oct. 2017. [Online; last accessed: 28-Oct-2017. [44] C. Sunstein. Republic.com. Princeton University Press, 2002. [45] M. Tambuscio, G. Ruffo, A. Flammini, and F. Menczer. Fact-checking effect on viral hoaxes: A model of misinformation spread in social networks. In Proceedings of the 24th International Conference on World Wide Web Companion, pages 977–982. International World Wide Web Conferences Steering Committee, 2015. [46] T. B. Times. Fact-checking u.s. politics, Oct. 2017. [Online; last accessed: 28-Oct-2017]. [47] L. Weng, A. Flammini, A. Vespignani, and F. Menczer. Competition among memes in a world with limited attention. Scientific Reports, 2, 2012. [48] T. Wood and E. Porter. The elusive backfire effect: Mass attitudes' steadfast factual adher- ence. e-print, SSRN, Aug. 2016. 19
1709.01773	2	1709	2018-06-25T16:34:10	IAD: Interaction-Aware Diffusion Framework in Social Networks	[ "cs.SI", "physics.soc-ph" ]	In networks, multiple contagions, such as information and purchasing behaviors, may interact with each other as they spread simultaneously. However, most of the existing information diffusion models are built on the assumption that each individual contagion spreads independently, regardless of their interactions. Gaining insights into such interaction is crucial to understand the contagion adoption behaviors, and thus can make better predictions. In this paper, we study the contagion adoption behavior under a set of interactions, specifically, the interactions among users, contagions' contents and sentiments, which are learned from social network structures and texts. We then develop an effective and efficient interaction-aware diffusion (IAD) framework, incorporating these interactions into a unified model. We also present a generative process to distinguish user roles, a co-training method to determine contagions' categories and a new topic model to obtain topic-specific sentiments. Evaluation on large-scale Weibo dataset demonstrates that our proposal can learn how different users, contagion categories and sentiments interact with each other efficiently. With these interactions, we can make a more accurate prediction than the state-of-art baselines. Moreover, we can better understand how the interactions influence the propagation process and thus can suggest useful directions for information promotion or suppression in viral marketing.	cs.SI	cs	JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 IAD: Interaction-Aware Diffusion Framework in Social Networks Xi Zhang, Member, IEEE, Yuan Su, Siyu Qu, Sihong Xie, Member, IEEE, Binxing Fang, and Philip S. Yu, Fellow, IEEE Abstract-In networks, multiple contagions, such as information and purchasing behaviors, may interact with each other as they spread simultaneously. However, most of the existing information diffusion models are built on the assumption that each individual contagion spreads independently, regardless of their interactions. Gaining insights into such interaction is crucial to understand the contagion adoption behaviors, and thus can make better predictions. In this paper, we study the contagion adoption behavior under a set of interactions, specifically, the interactions among users, contagions' contents and sentiments, which are learned from social network structures and texts. We then develop an effective and efficient interaction-aware diffusion (IAD) framework, incorporating these interactions into a unified model. We also present a generative process to distinguish user roles, a co-training method to determine contagions' categories and a new topic model to obtain topic-specific sentiments. Evaluation on the large-scale Weibo dataset demonstrates that our proposal can learn how different users, contagion categories and sentiments interact with each other efficiently. With these interactions, we can make a more accurate prediction than the state-of-art baselines. Moreover, we can better understand how the interactions influence the propagation process and thus can suggest useful directions for information promotion or suppression in viral marketing. Index Terms-Social networks, information diffusion, interaction-aware framework, sentiment analysis, topic model ! 1 INTRODUCTION S Ocial networks are the fundamental medium for the diffusion of information. This diffusion process imitates the spread of infectious disease. Specifically, when a user forwards a contagion (such as a political opinion or product), infection occurs and the contagion gets spread along the edges of the underlying network. The process of a user examining the contagions shared by their neighbours and forwarding some of them results in the information cascades. For the cascade of one contagion, both the users on the cascade and their neighbours are exposed to the contagion, and their behaviors and decisions will be influenced not only by the contents of the contagion but also by their social contacts. In order to effectively employ information diffusion for viral marketing, it is essential to understand the users' adoption decisions under such influences. Much work has been done to understand the information dynamics in social networks, including theoretical models and empirical studies. However, most of the existing studies assume that each piece of information spreads independently [1], [2], [3], [4], [5], [6], [7], regardless of the interactions between contagions. In the real world, multiple contagions may compete or cooperate with each other when they spread at the same time. • • X. Zhang, Y. Su, S. Qu and B. Fang are with the Key Laboratory of Trustworthy Distributed Computing and Service (Beijing University of Posts and Telecommunications), Ministry of Education, Beijing 100876, China. E-mail: {zhangx, timsu, qusiyu, fangbx}@bupt.edu.cn. S. Xie is with Computer Science and Engineering Department, Lehigh University, Bethlehem, PA 18015, USA. E-mail: [email protected]. • P.S. Yu is with the Department of Computer Science, University of Illinois at Chicago, Chicago, IL 60607, USA, and the Institute for Data Science, Tsinghua University, Beijing 100084, China. E-mail: [email protected]. Manuscript received April 19, 2005; revised August 26, 2015. For example, the news about the banning of Samsung Galaxy S7 in airports may promote the spreading of the battery explosion events of S7, but suppress the news that Samsung is releasing other exciting products. Thus, in this example, the contagion- contagion interaction can be seen as a "competition" between the popularity of two pieces of information. Taking the interactions into account is crucial to address the question of how much a user would like to adopt a contagion. Recent diffusion models have started to consider interactions between contagions [8], [9], [10], [11], [12], [13], [14], [15], [16], however, in most cases, the interactions they learned are latent factors and thus are difficult to understand. For example, in [9], the interactions it considered are between latent topics, making the promotion or suppression effects difficult to interpret. Specifically, given two contagions that are of unrelated content or subject matter, it is difficult to infer whether they will interact with each other when they spread si- multaneously. Actually, what interests us is the interactions among explicit categories, namely whether contagions belonging to one category (say food) would have some positive/negative effects on the spreading of contagions belonging to another category (say health). These interactions can be used to design viral marketing strategies to promote or suppress some products or news. For example, if it can be inferred that contagions belonging to sports usually have positive effects on the adoption of energy drinks, advertisements on energy drinks can be exhibited alongside with sports news in a user's input stream of posts to promote the sales. However, it is challenging to assign each contagion to its category by human due to the large volume in social networks. How to find an efficient way to classify contagions with only minimum supervision is one of the key challenges in this work. Even if a methodology is proposed to obtain explicit categories, the starting point is always a set of latent topics. In addition to the categories that each contagion belongs to, JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2 among the contagions and users. To this end, this paper focuses on the discovery of the interactions of users and contagions in the information diffusion process, as well as the impact of these interactions on the user's adoption behavior. The scenario we study is that when a user is exposed to a set of contagions posted or forwarded by her neighbours, whether she would like to forward a specific one of them. The task we conduct is hence to predict the infection probability by exploiting the inherent popularity of the contagion as well as the interactions among contagions and users. We approach this problem through an interaction-aware diffusion model (IAD), which considers three kinds of interac- tions: (1) User-Contagion Interaction, (2) User-User Interaction, (3) Contagion-Contagion Interaction. Fig. 1 describes an example of the three kinds of interactions when user ua is exposed to a set of contagions posted by her neighbours. When ua is examining whether to adopt the contagion mi posted by ud, there exists User- User interactions between ua and ud, User-Contagion Interaction between ua and mi, as well as Contagion-Contagion Interaction between m1 and mi. We then describe each contagion from two views, say category and sentiment, and refine the contagion- related interactions and capture Category-Category Interaction, Sentiment-Sentiment Interaction, User-Category Interaction, and User-Sentiment Interaction. As the interactions are rather complex and often exhibit strong coupling relations, rather than exploiting each kind of the interactions separately, we incorporate all the interactions into a unified framework. Due to the large volume of contagions and users in the online social networks, learning interactions for each pair of contagions and users is impractical. To address this challenge, we use the interactions among the categories of users and contagions instead. Specifically, we first apply a mixture of Gaussians model to ex- plain the generation process of user network features, and use EM algorithm to extract the social role distribution for each user. We then propose a new topic model to capture topics and sentiments simultaneously for each contagion. To get the category of each contagion, a classification approach based on co-training [20] is then developed, with only a small number of labeled data. Overall, our model can statistically learn the interactions effectively, aiming to better understand the information diffusion process and provide a more predictive diffusion model. Our contributions are summarized as follows: 1) We propose an IAD framework to model the user's forwarding behavior by jointly incorporating various kinds of interactions among contagions, interaction among users as well as interactions among users and contagions. This framework provides new insights into how forwarding decisions are made. interactions: 2) Due to the large volume of interactions to learn, we devise a co-training based method to classify the contagions to categories, and apply a generative process to obtain the social roles for users, to significantly reduce the complexity of the fitting process. 3) To obtain the sentiment-related from contagions, we develop a new topic model called Latent Dirichlet Allocation with Sentiment (LDA-S), a variation of LDA model [21], to obtain the sentiment and topic simultaneously in short texts. interactions Fig. 1: An example of the interacting scenario. User ua is exposed to contagions {m1, ..., mK} (here K = 2) and mi (forwarded by ua's neighbour ub), and is examining whether to adopt mi. ua 's decision is influenced by the interaction between ua and mi, the interaction between ua and ud, and interactions among mi and other exposing contagions (m1 and m2). recent studies show that the sentiments exhibited in the contagions may also affect the dynamics of information diffusion [17], [18]. Suppose that some user has just been infected by negative news (e.g, disasters or losing games) and thus in bad mood, and when she is exposed to a funny story, it is less likely for her to forward it. On the contrary, in happy days (e.g., holidays), users are more likely to forward positive news than negative news. Thus, sentiment analysis is crucial for information diffusion as well and can provide a new dimension to understand the users' adopting behaviours. However, how to effectively uncover the sentiments from contagions remains an open problem. Furthermore, as the sentiments and topics are both extracted from the contagions, they are hence not independent of each other and their coupling relations should also be taken into account in the information diffusion process, which has not been studied in previous works. Apart from the categories and sentiments extracted from the contagions, social roles have also been proved to play important roles in the information diffusion process [19]. Besides, a user may play multiple roles with respect to different communities, and each social role may present different influences on their neighbours. To study how social roles affect is necessary to distinguish the users' social roles and then capture their interactions when modeling the user's adoption decisions. the diffusion process, it Since there are interactions among contagions and the inter- actions among users in the information diffusion process, it is natural to ask whether there are interactions between users and contagions. The answer is obvious since most of the users have their own preferences on some messages, and they would like to forward those to their tastes and skip the others. Although a few methods have been proposed to estimate a user's taste on a specific topic, it remains an open question from the perspective of information diffusion, especially when complex interactions are involved. Motivated by the above examples and prospects, there is a clear opportunity to extend the present understanding of in- formation diffusion by jointly studying the interactions existed 𝑢𝑢𝑎𝑎 𝑢𝑢𝑏𝑏 14:15 𝑢𝑢𝑐𝑐 14:03 𝑢𝑢𝑑𝑑 13:42 Contagion-Contagion Interaction User-User Interaction User-Contagion Interaction 𝑚𝑚1: In order to keep healthy, make sure to eat plenty of vitamins every day. 𝑚𝑚2:The food in this Chinese restaurant is delicious and healthy. 𝑚𝑚𝑖𝑖: Greek yogurt is on sale at Whole Foods for 10/$10. Hurry in today, sale ends Thursday! read Examining (adopt 𝑚𝑚𝑖𝑖?) Category: Food Sentiment: Positive Category: Health Sentiment: Neutral Category: Advertisement Sentiment: Positive JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 3 on a 4) Experimental evaluation large-scale Weibo dataset [22] shows that IAD outperforms state-of-art works in terms of F1-score, accuracy and fitting time. Besides, the estimated interactions reveal some interesting findings that can provide compelling principles to govern content interaction in information diffusion. The remainder of this paper is organized as follows. Section 2 introduces the related works. In Section 3, we describe the proposed IAD framework. Section 4 describes how to classify contagions to explicit categories and Section 5 illustrates the pro- cess to infer sentiment-related interactions. In Section 6, extensive experiments have been conducted to show the effectiveness of the proposed approach. Section 7 concludes this paper. 2 RELATED WORK 2.1 Information Diffusion In recent years, researchers have extensively studied the informa- tion diffusion in social networks [23]. A collection of models are proposed to explain the diffusion process from various perspec- tives [24], [25], [26], [27], [28], [29], [30], [31], while some other models are proposed to predict whether a piece of information will diffuse [9], [32]. However, most of the prior models assume the spreading of each piece of information is independent of others, e.g., the Linear Threshold Model [33], [34], the Independent Cascade Model [1], [2], SIR and SIS Model [3], [4]. The diffusion of multiple contagions has been covered in several recent works [12], [13], [14], [15], [16]. The scenario discussed by these works is that one contagion is mutually exclusive to others, i.e., only involving the competition of contagions. In [8], an agent-based model is employed to study whether the competition of information for user's finite attention may affect the popularity of different contagions, but this model does not quantify the interactions between them. Our research is not limited to the mutual exclusivity condition, but instead a user can adopt multiple contagions. We offer a comprehensive consideration for the inter-relations of the contagions in online social networks. The most related work to ours is the IMM model proposed in [9], which statistically learns how different contagions interact with each other through the Twitter dataset. It models the probabil- ity of a user's adoption of information as a function of the exposure sequence, together with the membership of each contagion to a cluster. However, this model doesn't consider the user roles, which has been proved to play an important role in information diffusion in our work. In addition, the clusters in this model are latent variables without real-world meanings. In contrast, our proposal can infer interactions among explicit categories, which are easy to interpret. Nonetheless, IMM model is implemented as a baseline to compare with. Other studies [10], [11], [35] also neglect the influence of user roles' interactions, and do not discover the interactions of actual categories of contagions, which are different from our work. 2.2 Sentiment Analysis Although recent works suggest the sentiments in the contents can play important roles in various applications such as product and restaurant reviews [36], stock market prediction [37], few existing studies quantify the effects of contents sentiment on the dynamics of information diffusion. Empirical analysis on German political blogosphere indicates that people tend to participate more in emotionally-charged (either positive or negative) discussions [38]. A recent study on Twitter exhibits the effect of sentiment on information diffusion [18], and reveals different diffusion patterns for positive and negative messages respectively. However, different from our IAD model, these works still treat each contagion in isolation and thus do not take the interactions into account. Besides, most of the previous studies try to extract only the sentiments. However, sentiments polarities are often dependent on topics or aspects. Therefore, detecting on which topics of the users are expressing their opinions is very important. Several models have been proposed to infer the topic and sentiment simultaneously. Mei et al. [39] propose the TSM model which can reveal the latent topical facets in a Weblog collection, the subtopics in the results of an ad hoc query, and their associated sentiments. Lin et al. [36], [40] propose a novel probabilistic modeling framework based on LDA, called joint sentiment/topic model (JST), which detects sentiment and topic simultaneously from a text. This model assumes that each word is generated from a joint topic and sentiment distribution, and hence doesn't distinguish the topic word and opinion word distributions. Liu et al. [41] propose a topic-adaptive sentiment classification model which extracts text and non-text features from twitters as two views for co-training. Tan et al. [42] propose a LDA based model, Foreground and Background LDA (FB-LDA), to distill foreground topics and filter out longstanding background topics, which can give potential interpretations of the sentiment variations. There are some other topic models considering aspect-specific opinion words [43], [44], [45], [46]. A recently proposed TSLDA model can estimate different opinion word distributions for individual sentiment categories for each topic [37], and has been successfully applied to stock prediction. One weakness of TSLDA is that it divides a document into several sentences and sample the topic and sentiment of each sentence. Therefore, its performance is limited when it is applied to Weibo where most of the messages have only one or two sentences. The other weakness is that it lacks prior information, making it difficult to achieve good results for short texts. To address the aforementioned problems, we propose a variation of TSLDA model, namely LDA-S, to make it work for short texts such as Weibo and Twitter. 3 INTERACTION-AWARE DIFFUSION FRAMEWORK In this section, we first state and formulate the problem, and then propose our framework and the corresponding learning process. Before going into details of IAD framework, we define some important notations shown in Table 1. 3.1 Problem Statement In a social network, when some new contagion is originated from a user, we assume that the user's neighbours would see this contagion, or would be exposed to this contagion. This assumption is consistent with [9]. The exposed contagion is called an exposure. Since users have limited attention [8], we make the assumption as [9] that at a given time, a user can read through all the contagions her neighbours have forwarded, but only the most recent K exposures that she can keep in mind. In social networks like Weibo and Twitter, tweets in a user's reading screen are arranged in time descending order, i.e., users will first read JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 TABLE 1: Description of Symbols in IAD Framework Symbol Description Users u Contagions m User roles r Contagion latent topics t Contagion sentiments s Contagion categories c ∆ ∈ Ru×u User-user interaction matrix Λ ∈ Rm×m Contagion-contagion interaction matrix Ω ∈ Ru×m User-contagion interaction matrix ∆role ∈ Rr×r User role-role interaction matrix Λtopic ∈ Rt×t Contagion topic-topic interaction matrix Λt−s ∈ Rts×ts Contagion (topic-sentiment) - (topic-sentiment) interaction matrix Λs ∈ Rs×s Contagion sentiment - sentiment interaction matrix topic ∈ Rr×t User role - contagion topic Ωrole interaction matrix t−s ∈ Rr×ts User role - contagion Ωrole (topic-sentiment) interaction matrix ∈ Rr×s User role - contagion sentiment Ωrole interaction matrix Λcateg. ∈ Rc×c Contagion category-category interaction matrix categ. ∈ Rr×c User role - contagion category Ωrole s interaction matrix the most recent contagions and then go backward. 1 Therefore, there is a sliding attention window going back K contagions, and the K contagions in the window may affect a user's adoption behavior. Then the problem we focus is that when a user reads a contagion that has been forwarded by one of her neighbours, given the sequence of contagions the user has previously read, what's the probability of the user adopting this contagion. Figure 1 describes the interaction scenario studied in this pa- per, where the set {m1,m2,...mK} is a sequence of K contagions user ua has read and kept in mind, and mi (i (cid:54)= 1, 2, ..., K) is the contagion which is previously forwarded by a user ud and now examined by ua. ua will determine whether to adopt (i.e., forward) mi. In this scenario, the forwarding decision made by ua is not only decided by the inherent characteristics of mi, but also by external interactions described as follows: • User-Contagion Interaction: The interaction between the examining user and the examined contagion. As shown in Fig. 1, it is ua's preference over mi. • User-User Interaction: The interaction between the examining user and the neighbour who has forwarded the examined contagion previously. In Fig. 1, it is the effect ub has on ua. • Contagion-Contagion Interaction: The interaction among the examined contagion and other contagions the user has read recently. In Fig. 1, it is the effect contagions m1 and m2 (K = 2) has on mi. 1. Please note that the dataset that we analyze was collected in 2012, and at that time Weibo and Twitter still showed tweets in the reversed chronological order. Though they have stopped showing contagions in this simple order at present, our model can still work if we could identify which set of contagions are read simultaneously by a user. 4 Given a collection of the interacting scenarios, our task is to model the users' adoption behaviour by incorporating the aforementioned interactions, and fitting the model to infer the interactions. Meanwhile, we can make more accurate predictions on users' adoption behaviors. The problem will be formulated in the next subsection. 3.2 Formulation According to the interacting scenario, given {m1,m2,...mK} and ub, the probability of infection by mi to ua is P (Imi(ua)Emi(ub), E{m1,m2,...,mK}), (1) where Imi(ua) is the infection of ua by mi, Emi(ub) is the exposure of mi which is forwarded by ub, and E{m1,m2,...,mK} is the exposure set {m1, m2, ..., mK}. We make the same as- sumption as [9] that for any k and l, Emk is independent of Eml. Applying Bayes' rule, we model Eq. (1) by P (Imi(ua)Emi(ub), E{m1,m2,...,mK}) P (Imi(ua))P (Emi(ub), E{m1,m2,...,mK}Imi(ua)) = P (Imi(ua))P (Emi(ub)Imi(ua))(cid:81)K P (Emi(ub), E{m1,m2,...,mK}) k=1 P (EmkImi(ua)) P (Emi(ub))(cid:81)K k=1 P (Emk ) P (Imi(ua )Emi(ub ))P (Emi(ub )) (2) = = = P (Imi(ua)) × K(cid:89) k=1 P (Imi(ua )) k=1 P (Emk ) P (Imi(ua)Emk )P (Emk ) P (Emi(ub))(cid:81)K K(cid:89) P (Imi(ua)) P (Imi(ua)Emi(ub)) P (Imi(ua))K k=1 P (Imi(ua)Emk ). Here we need to model P (Imi(ua)), P (Imi(ua)Emi(ub)) and P (Imi(ua)Emk ) for each k ∈ {1, ..., K}, which are enforced between 0 and 1. Since each contagion has its inherent infec- tiousness, P (Imi) is defined as the prior infection probability of mi, which can be obtained through dividing the number of its infections by the number of its exposures. We define Ω(ua, mi) as the effect user ua has on contagion mi (User-Contagion Interaction), ∆(ua, ub) as the effect user ub has on user ua (User-User Interaction), and Λ(mi, mk) as the effect contagion mk has on contagion mi (Contagion-Contagion Interaction). Then we model P (Imi(ua)), P (Imi(ua)Emi(ub)) and P (Imi(ua)Emk ) as P (Imi(ua)) ≈ P (Imi ) + Ω(ua, mi) P (Imi(ua)Emi(ub)) ≈ P (Imi(ua)) + ∆(ua, ub) ≈ P (Imi ) + Ω(ua, mi) + ∆(ua, ub) P (Imi(ua)Emk ) ≈ P (ImiEmk ) + Ω(ua, mi) ≈ P (Imi ) + Λ(mi, mk) + Ω(ua, mi) (3) (4) (5) Example 1. In Fig. 1, assuming that P (Imi) = 0.5, Ω(ua, mi) = 0.02, ∆(ua, ub) = −0.03, Λ(mi, mk1) = −0.04 and Λ(mi, mk2) = −0.05, then we can derive that P (Imi(ua)) ≈ 0.52 according to Eq. (3), P (Imi(ua)Emi(ub)) ≈ 0.49 according to Eq. (4) and P (Imi(ua)Emk1 ) ≈ 0.43 accord- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 ing to Eq. (5). Integrating these into Eq. (2), we can obtain that the probability of ui adopting mi is 0.327. Please note that, as shown in Eq. (3), Eq. (4), and Eq. (5), the proposed model adopts summations to combine the interaction matrices to the prior infection probability P (Imi). In addition, we also develop a model which adopts multiplications to combine them, however, the experimental results show that the model with summations performs better. The possible reason is that, the form of Eq. (2) will be like a linear function when adopting multipli- cations, and its expressive power would be weaker compared to that with the additive schema. Thus we apply the additive model in this paper. Though we have connected the infection probability with three interaction matrices: (1) Ω ∈ Ru×m, (2) ∆ ∈ Ru×u, and (3) Λ ∈ Rm×m, where u is the number of users and m is the number of contagions, these matrices are impractical to learn because u and m are extremely large in social networks. Thus, to decrease the parameters to fit, we model User Role - Contagion Topic Interaction, User Role-Role Interaction and Contagion Topic-Topic Interaction instead. Moreover, we also involve sentiments into IAD framework and model the User Role - Contagion Sentiment Interaction. It will be described in detail in the next subsection. 3.3 The Proposed Approach Overview To decrease the fitted parameters in the interaction matrices, user roles and contagion categories are introduced, and then the interactions between user roles and categories can be learned efficiently. To this end, we utilize the network structures to infer users' social roles, and use the contagion texts to extract the contagions' topics. To obtain the sentiments and topics from contagions, we propose an extension of LDA model called LDA- S to extract sentiments. The whole process of IAD framework is shown in Fig. 2, comprising of the following five components: • User roles generation (C1): A generative process for user roles is proposed to distinguish different kinds of users. • Contagion latent topics extraction or Contagion Latent Topic-Sentiment Extraction (C2): We have two ways to deal with the contagion texts. One way is to apply LDA model to extract latent topics, and the other is to extract both topics and sentiments with our proposed LDA-S. The output of this component will be used as features for statistical model learning (C4) and contagion classification (C3). • Contagion classification (C3): Based on the latent topics from C2, a co-training method of contagion classification is proposed to assign the contagions with explicit categories. • • Statistical model learning (C4): Based on the outputs of C1 and C2, a statistical model is learned. Interactions inference (C5): Given the results of contagion classification (C3) and the statistical model (C4), interac- tions among contagions and users can be inferred. Next we will introduce the process of user roles generation and contagion latent topic extraction in detail, and then describe 5 statistical model learning. C3 will be illustrated in Section 4. Note that in C2 and C5, the dashed boxes are newly added parts over the earlier version of this work [48], in order to involve sentiments into the IAD framework, which will be illustrated in Section 5. User Role-Role Interaction. User roles are categorized into au- thority users, hub users and ordinary users in our work. Intuitively, an authority user commonly has a large number of followers, indicating that the node of authority user has a large in-degree but its out-degree is small, while a hub user has lots of followees, which means the node of hub user has a small in-degree but large out-degree. And an ordinary user usually does not have a lot of followers or followees, i.e., the in-degree and out-degree of the node are both small. A user may play multiple roles, for instance, an authority user may also be a hub user to some extent, and therefore we adopt a probability distribution over social roles for each user. Then we infer the interactions among different social roles. The results can demonstrate how a user, with a specific roles distribution, influence other user's probability of adopting a contagion. We use PageRank score [49], HITS authority and hub val- ues [50], in-degree and out-degree scores as features of users. A mixture of Gaussians model is proposed to explain the features generation process. Specifically, we assume the features of each user is sampled as a multivariate Gaussian distribution. Intuitively, users with the same roles have similar features and share the same multivariate Gaussian distribution, e.g., two authority users are both likely to have a large number of followers. Define r = [r1, r2, r3](cid:62) as a user role vector, representing a probability distribution over social roles for each user, e.g., [0.5, 0.2, 0.3](cid:62). Then for each role rj, we generate multivariate Gaussian distri- bution urj ∼ N (µj, Σj). EM algorithm is used to extract the role distribution for each user. After that, we determine each rj as the most relevant one of the three roles, according to the fact that authority users commonly have lots of followers and hub users have lots of followees. Rather than modeling the User-User Interaction denoted by ∆ ∈ Ru×u, we would model User Role-Role Interaction instead, which is denoted by ∆role ∈ Rr×r. ∆role(ri, rj) is the effect role rj has on role ri. Define ϑa,i as the probability i ϑa,i = 1, and then of user ua belonging to role ri, and (cid:80) ∆(ua, ub) in Eq. (4) can be updated by ∆(ua, ub) = ϑa,i∆role(ri, rj)ϑb,j (6) (cid:88) (cid:88) i j Contagion Topic-Topic Interaction. Each contagion is as- sumed to have a distribution on several topics, and t denotes the set of latent topics. LDA [21] is used to extract the latent topic distribution of each contagion. Then, instead of modeling Λ ∈ Rm×m, we would model a matrix Λtopic ∈ Rt×t, which denotes the Contagion Topic-Topic Interaction. We define θi,a as the probability of contagion mi belonging to topic ta, a θi,a = 1. Let Λtopic(ta, tb) denote the impact of topic tb has on topic ta. Now, Λ(mi, mk) in Eq. (5) can be updated by and therefore(cid:80) Λ(mi, mk) = θi,aΛtopic(ta, tb)θk,b (7) a b Besides the Contagion Topic-Topic Interaction, we also pro- pose a topic-sentiment model to get Contagion (Topic-Sentiment)- (Topic-Sentiment) Interaction to update Eq. (5). User Role - Contagion Topic Interaction. Instead of learning (cid:88) (cid:88) JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 6 Fig. 2: IAD framework. topic ∈ Rr×t to denote the User Role - Ω, we build a matrix Ωrole Contagion Topic Interaction. Then Ω(ua, mi) in Eq. (3), Eq. (4) (cid:88) and Eq. (5) can be updated by (cid:88) Ω(ua, mi) = ϑa,jΩrole topic(rj, tb)θi,b. (8) j b 3.4 Model Learning The input of our model is a set of interacting scenarios. An exam- ple of the interacting scenario is shown in Fig. 1, which consists of the examining user ua, the examined contagion mi, user ua's neighbour ub who has forwarded the examined contagion, and the exposing contagion set {m1, m2, ..., mK} (i (cid:54)= 1, 2, ..., K). All the interacting scenarios comprise a set {x1, x2, ..., xn}, where xi is the ith interacting scenario and n is the total number of scenarios. For each interacting scenario, we can observe whether the examining user has adopted the examined contagion or not, and denote it as yi ∈ {0, 1} (1 for adoption and 0 for not). Then the training set {(x1, y1), (x2, y2), ..., (xn, yn)} will be obtained. Let π(xi) denote Eq. (1) for simplicity. Now π(xi) can be updated by Ωrole topic, ∆role and Λtopic, and the log-likelihood function is L(Ωrole n(cid:88) topic, ∆role, Λtopic) (yilogπ(xi) + (1 − yi)log(1 − π(xi))) = (9) Our goal i=1 is to estimate the parameters in Ωrole topic, ∆role and Λtopic to maximize the log-likelihood function. Stochastic gradient ascent is adopted to fit the model. In each iteration in the parameters updating process, if it makes any variable with probability meaning smaller than 0 or larger than 1, we don't conduct any updating in this iteration, and go directly to the next iteration. topics, which are difficult 4 CLASSIFICATION OF CONTAGIONS The interaction matrix Λtopic and Ωrole topic learned through our model are comprised of latent to interpret. In this section, we illustrate how to obtain interactions among explicit categories. We define c = 15 categories based on the Weibo dataset, involving advertisement, constellation, culture, economy, food, health, history, life, movie, music, news, politics, sports, technology and traffic. One contagion on Weibo is always a short text with only a few words, and is mostly affiliated to only one specific class. Thus, the multi-label classification techniques such as [47] are not applied in this paper. To discover interactions among categories, contagions should be classified into categories first. However, contagions spreading in Weibo [22] are not labeled to intrinsic categories. Labeled contagions are extremely expensive to obtain because large hu- man efforts are required. Thus, only a few labeled contagions are available for learning. A classification approach based on co-training [20] is proposed. Co-training is a semi-supervised learning technique that assumes each example is described us- ing two different feature sets (or views) that provide different, complementary information about the instance. It first learns a separate classifier for each view using an initial small set of labeled examples, and then use these classifiers on the unlabeled examples. The most confident predictions are iteratively added to the labeled training data. Network DataUser Network Features ExtractingMultivariate Gaussian DistributionRole Distribution of UsersContagion Context DataLDALatent Topics of ContagionsContagion Context DataLDA-SLatent Topic -Sentiment of ContagionsorModel LearningStatistical ModelUser Role-Contagion Sentiment InteractionContagion Sentiment-Sentiment InteractionUser Role-Role InteractionContagion Topic-Topic InteractionContagion Category-Category InteractionUser Role-Contagion Topic InteractionUser Role-Contagion Category InteractionLabeled and Unlabeled Contagion Category DataContagion Infection DataCo-trainingContagion CategoryTopic Distribution of CategoriesC1User Roles GenerationC4Statistical Model LearningC2Contagion Latent Topics Extraction or Contagion Latent Topic-Sentiment ExtractionC3Contagion ClassificationC5Interactions InferenceJOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 7 Specifically, for our task, each contagion in the dataset is described in two distinct views. One is the contagion itself, and the other is a set of the other contagions posted by the same user. The intuition here is that contagions created from the same user are prone to have the similar category. Then we build two classifiers for two views, and choose the latent topics as the features for each classifier. As described in Section 3.3, contagion mi's latent topic distribution, denoted by θi,a(a ∈ 1, ..t, where t denotes the number of latent topics), can be extracted using LDA. We define a contagion set Mi = {m1, m2, ..., mk} to contain the other contagions created by the same user. The latent topic distribution Θi,a(a ∈ 1, ..t) of Mi is obtained by . Now, the two classifiers are listed as follows, and LIBSVM [51] is used for multi-class classification. (cid:80)k j=1 θj,a k • Classifier 1: θi,a(a ∈ 1, ..t) as features for each contagion mi. • Classifier 2: Θi,a(a ∈ 1, ..t) as features for each conta- gion set Mi. We labeled a minimum number of contagions for each category by hand for training in the beginning. The number of manually annotated contagions for each category is 100, and the total number of annotated data is 1500. After the initial training process, two classifiers go through the unlabeled contagions to make predictions. If the results from the two classifiers are the same for a contagion, this contagion is added to the labeled set and removed from the unlabeled set. Then a new set for training is obtained, and another iteration starts. In each iteration, there are some contagions moved from the unlabeled set to the labeled set. After enough contagions being labeled, we can derive the following two interactions. Contagion Category-Category Interaction. If the set of conta- gions {m1, m2, ..., mk} belongs to category ci, the latent topic distribution ϕi,a(a ∈ 1, ..t) of category ci can be obtained . We define Λcate. ∈ Rc×c to denote Con- through tagion Category-Category Interaction, where Λcate.(ci, ck) is the impact of category ck on ci, that is (cid:80)k j=1 θj,a k Λcateg.(ci, ck) = ϕi,aΛtopic(ta, tb)ϕk,b (10) (cid:88) (cid:88) a b (cid:88) User Role - Contagion Category Interaction. Similarly, define cate. ∈ Rr×c to denote User Role - Contagion Category Ωrole Interaction, where Ωrole cate.(ri, cj) is the interaction from user role ri to category cj, that is Ωrole categ.(ri, ck) = Ωrole topic(ri, tb)ϕk,b (11) b 5 SENTIMENT-RELATED INTERACTIONS In this section, we first introduce the LDA-S model, and then discuss how to use this model to obtain the sentiment-related interactions. 5.1 LDA-S Model 5.1.1 Motivation and Basic Idea We consider how to incorporate sentiments into the IAD frame- work since a user's forwarding behavior may be affected by the sentiments expressed in the contagions. A natural way is to extract the sentiments from each contagion, and then model their interactions if they are simultaneously exposed to a user. However, extract only sentiments may not be enough as sentiment polarities are usually dependent on topics or domains [43]. In other words, the exact same word may express different sentiment polarities for different topics. For example, the opinion word "low" in the phrase "low speed" may have negative orientation in a traffic- related topic. However, if it is in the phrase "low fat" in a food- related topic, the word "low" usually belongs to the positive sentiment polarity. Therefore, it is necessary to incorporate the topic information for sentiment analysis, in order to get topic- specific sentiments, namely topic-sentiments in this work. This can have two benefits. On one hand, extracting the sentiments corresponding to different topics can improve the sentiment clas- sification accuracy. On the other hand, particularly for this work, such method can refine the inferred interactions between topics. Specifically, we use the interactions between topic-sentiments instead of the interactions between topics in the IAD framework, which would improve the accuracy of forwarding prediction. Inspired by Topic Sentiment Latent Dirichlet Allocation (TSLDA) [37], we propose LDA-S, an extension of LDA [21] model, to infer sentiment distribution and topic distribution simul- taneously for short texts. The difference of LDA-S and TSLDA will be illustrated in Section 5.1.4. LDA-S model consists of two steps. The first step aims to obtain the topic distribution of each document (or contagion in our work), and then set the contagion's topic as the one that has the largest probability. The second step gets the sentiment distribution of each document. The opinion words are usually adjectives or adverbs, whereas the topics words are usually nouns. The words in a document are classified into three categories, the topic words (c = 1), the sentiment words (c = 2) and the background words (c = 0). We adopt a sentiment word list called NTUSD [52], which contains 4370 negative words and 4566 positive words. If a word is an adjective but not in the sentiment word list, the sentiment label of this word is set as neutral. If a word is a noun, it is considered as a topic word. Otherwise, it is considered as a background word. In our model, different topics have different opinion word distributions. For each topic, we distinguish opinion word distributions for different sentiment meanings such as positive, negative or neutral. We will show the graphical model and the generation process in the following part. 5.1.2 Generation Process . Figure 3 shows the graphical model of LDA-S. Shaded circles indicate observed variables, and clear circles indicate hidden vari- ables. A word is an item from a vocabulary indexed by {1, ..., V }. We represent words using standard basis vectors that have a single component equal to one and all other components equal to zero. Thus, using superscripts to denote components, the vth word in the vocabulary is represented by a V -vector w such that wv = 1 and wu = 0 for u (cid:54)= v. A document d is a sequence of N words denoted by w = (wd,1, wd,2, ..., wd,N ), where wd,n is the nth word in the sequence. A corpus is a collection of M documents denoted by D = {w1, w2, ..., wM}. Let zt d denote the topic assignment of document d and other notations is introduced in Table 2. The generation process of LDA-S is shown as Algorithm 1. To generate a document, we will generate every word in order. Based on Fig. 3 and the above generation process, we can get JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 8 Algorithm 1 LDA-S Input: documents 1: Set the prior distributions α, β, λ and γ 2: Set the number of iterations n 3: for iter = 1 to n do 4: Choose a distribution of background words Φb ∼ 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: Dirichlet(α) for each topic k do Choose a distribution of topic words Φt for each sentiment s of topic k do Choose a distribution of sentiment words Φo k ∼ Dirichlet(α) k,s ∼ Dirichlet(λ) for each document d do Choose a topic distribution θt Choose a sentiment distribution θo for each word wd,n do Determine the category of wd,n if the category is 0 then d ∼ Dirichlet(β) d ∼ Dirichlet(γ) choose a word wd,n ∼ M ultinomial(Φb) if the category is 1 then topic choose a choose d), M ultinomial(θt M ultinomial(Φt zt d,n ) assignment word zt d,n wd,n a ∼ ∼ if the category is 2 then choose a d,n ∼ M ultinomial(θo d), zo M ultinomial(Φo d,zo zt d,n ) sentiment assignment a word wd,n ∼ choose 20: return θt d,θo d when the document is d and the topic is k. nd,s is the number of times that the document d is endowed with sentiment s while γd,s specifies the value of γ when the document is d and the sentiment is s. When c = 0, we only consider the background words. Specifically, we use αwv to denote the value of α when the word wv is a background word, and the background word has no topic assignment in this work. Then the generation probability of wv is (cid:80)V0 p(wv) ∝ Bwv + αwv v=1(Bwv + αwv ) (13) When c = 1, we only consider the topic words. The generation probability of wv is p(wv, kd) = p(wvk) · p(kd) ∝ (cid:80)V1 (cid:80)K · nk,wv + αk,wv v=1(nk,wv + αk,v) nd,k + βd,k k=1(nd,k + βd,k) (14) When c = 2, we only consider the sentiment words. When the topic of document d is zt d, the generation probability of wv is p(wv, sd, zt ∝ (cid:80)V2 d) = p(wvs, zt d) · p(sd, zt d) nzt d,s,wv + λzt d,s d,s,wv + λzt d,s,wv ) nd,s + γd,s s=1(nd,s + γd,s) v=1(nzt · (cid:80)S (15) Then Eq. (14) and Gibbs Sampling are implemented for infer- Fig. 3: Graphical Model Representation of LDA-S. TABLE 2: Description of Symbols in LDA-S Description Symbol α, β, γ, λ Dirichlet prior vectors c K S B V V0 V1 V2 Φb Φt Φo θt d θo d θt−o d zt zo Category of words # of topics # of sentiments # of stop words # of words in corpus # of words with category 0 in corpus # of words with category 1 in corpus # of words with category 2 in corpus Distribution over background words Distribution over topic words Distribution over sentiment words Topic distribution for document d Sentiment distribution for document d Topic-Sentiment distribution for document d Topic assignment Sentiment assignment  the generation probability of each word wd,n in document d depending on the category of wd,n as shown in Eq. (12). p(wd,n, zt d,nd, zt d,nd) = p(wd,nzt d,n, zt d) = p(wd,nzo p(wd,n) = p(wd,n) d,nd) d,nd, zt d) d,n) · p(zt d) · p(zo c = 0 c = 1 c = 2 p(wd,n, zo (12) Then we get the approximate expression of Eq. (12) in Eq. (13), Eq. (14) and Eq. (15). We will define some notations to explain the equations. Note that we represent wd,n in document d as wv in the corpus. Let Bwv denote the number of times that the background word wv appears in all documents. nk,wv is the number of times that the topic word wv with the topic k appears in all documents (i.e., corpus) and αk,wv specifies the value of α when the topic is k and the word is wv. Let nzt d,s,wv be the number of times that the sentiment word wv with the sentiment s under the topic zt d,s,wv specifies the value of λ when the topic is zt d, the sentiment is s and the word is wv. nd,k denotes the number of times that the document d is endowed with topic k while βd,k specifies the value of β d appears in all documents and λzt DKSαλVwβγc𝛷𝑏 𝛷𝑡 𝛷𝑜 𝑧𝑡 𝑧𝑜 𝜃𝑡 𝜃𝑜 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 ence in LDA-S in which we only consider words with category 1. It will sequentially sample the hidden variables zt d,n (the topic assignment of the nth word in document d) given a set of observed variables w and a set of hidden variables zt−(d,n). A bold-front variable denotes the list of the variables. For example, w denotes all the words in all documents. zt−(d,n) denotes all d,n. −(wv) stands for all topic assignment variables zt except zt the words except wv. nd,k,−wv denotes the number of times that words are endowed with the topic k except the word wv. w−wv denotes all the words in all documents except the word wv. The equation of topic sampling is shown in Eq. (16). d,n = kzt−(d,n), w) d,n = k, wd,n = wvzt−(d,n), w−wv ) p(zt ∝ p(zt (cid:80)K (cid:80)V1 ∝ · nd,k,−n + βd,k k=1(nd,k,−n + βd,k) nk,−wv + αk,wv v=1(nk,−wv + αk,wv ) (16) After we get the topic distribution of document d, the topic that has the largest probability will be set as document d's topic. For example, if θt d = (0.1, 0.2, 0.3, 0.4), we will choose the last one as document d's topic. If there are more than one topics with the same probability, we choose one of them randomly. Then Gibbs Sampling will be adopted again together with Eq. (15), in which we only consider words with category 2. We use the observed d as well as the hidden variable zo−(d,n) to sample variables w, zt the hidden variable zo d,n which denotes the sentiment assignment of nth word in document d. The equation of sentiment sampling is shown in Eq. (17). d,n = szo−(d,n), w, zt d) d,n = s, wn = wvzo−(d,n), w−wv , zt d) p(zo ∝ p(zo (cid:80)S ∝ nd,s,−n + γd,s s=1(nd,s,−n + γd,s) nzt d,s,−wv + λzt d,s,wv v=1(nzt d,s,−wv + λzt d,s,wv ) (cid:80)V2 · (17) After that, we can approximate the multinomial parameter sets with the samples from Eq. (16) and Eq. (17). The distributions of topics and sentiments in document d are shown in Eq. (18) and Eq. (19) respectively, (cid:80)K (cid:80)S θt d,a = θo d,b = nd,a + βd,a k=1(nd,k + βd,k) nd,b + γd,b s=1(nd,s + γd,s) (18) (19) where a denotes one of the topics, and b denotes one of the sentiments. Then the joint distribution of topic a and sentiment b is obtained in Eq. (20). θt−o d,ab = θt d,a · θo d,b (20) words in the corpus. (cid:80)V0 (cid:80)V1 (cid:80)V2 Φb r = Φt k,r = Φo r,k,s = Br + αr v=1(Bv + αv) nk,r + αk,r v=1(nk,v + αk,v) nk,s,r + λk,s,r v=1(nk,s,v + λk,s,v) 9 (21) (22) (23) 5.1.3 With Prior Information For the hyperparameters of this LDA-S, namely, α, β, λ and γ, we set α=0.1 and β=0.01 as TSLDA [37]. But for prior λ and γ, we adopt a sentiment word list to obtain them. The intuition is that incorporating prior information or subjective lexicon (i.e., words bearing positive and negative polarity) are able to improve the sentiment detection accuracy [36]. The prior parameters of Dirichlet distribution can be regarded as the "pseudo-counts" from "pseudo-data". To get the priors, we match the documents to the sentiment word list. Then the number of the words with a specific sentiment s appearing in a document d can be obtained, which can be set as the value of γd,s. Similarly, the value of λk,v,s can be set as the count of the word v with sentiment s under topic k. With the prior information, our experimental results show a significant improvement in the sentiment classification accuracy, but the results are omitted due to the limit of space. 5.1.4 Comparision with TSLDA TSLDA is designed for long documents which consist of a set of sentences, and is not suitable to short texts with only one or two sentences. Specifically, the limit in length makes it difficult to achieve good learning results. Compared to TSLDA, LDA- S adopts different assumptions and different sampling methods. Moreover, LDA-S incorporates prior information at the beginning of the generation process, which is expected to be more suitable to short texts. The differences are listed as follows. • Sampling Method. With TSLDA, each sentence is supposed to express only one topic and one opinion on that topic, and we need to sample topics and sentiments for each sentence in the generation process. But in LDA-S, each contagion (which usually consists of one or two sentences) bears a topic distribution and a sentiment distribution, and we don't involve sentences in the generation process. • Prior Information. In the initialization step of TSLDA, each document and word are assigned with topics and sentiments randomly, limiting its performance. In LDA- S, we use a paradigm word list consists of a set of positive and negative words, and compare each word token in the contagions against the words in the sentiment word list in the initialization of the generation process. The background word distribution is shown in Eq. (21) in which we only consider the background words, the topic word distribution of topic k is shown in Eq. (22) in which we only consider the topic words and the sentiment word distribution of sentiment s under the topic k is shown in Eq. (23) in which we only consider the sentiment words, where r denotes one of the 5.2 Interactions with Sentiment After the topic-sentiment distribution (as shown in Eq. (20)) is obtained, the Contagion (Topic-Sentiment) - (Topic-Sentiment) In- teraction and User Role - Contagion (Topic-Sentiment) Interaction will be introduced in our framework. Now, π(xi) (i.e., Eq. (1)) can be updated by Ωrole t−s , ∆role and Λt−s. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 10 Specifically, each contagion is assumed to have a distribution not only on topics but also on sentiments, and ts denotes the set of tuples that contain a latent topic and a kind of senti- ment. LDA-S is used to extract the joint distribution of latent topic and sentiment on each contagion. Therefore, the value of ts is actually the number of topics (20 in our setting) times the number of sentiment polarities (3 in our setting). Then, instead of modeling Λtopic ∈ Rt×t, we would model a matrix Λt−s ∈ Rts×ts, which denotes the Contagion (Topic- Sentiment) - (Topic-Sentiment) Interaction. We define θi,ax as the probability of contagion mi belonging to topic ta and sentiment ax θi,ax = 1. Let Λt−s(tax, tby) denote the impact of topic-sentiment tby has on topic-sentiment tax. Then Λ(mi, mk) in Eq. (5) can be updated by sx, and therefore(cid:80) Λ(mi, mk) = θi,axΛt−s(tax, tby)θk,by (24) Instead of learning Ω, we build a matrix Ωrole t−s ∈ Rr×ts to denote the User Role - Contagion (Topic-Sentiment) Interactions. Then Ω(ua, mi) in Eq. (3), Eq. (4) and Eq. (5) can be updated by Ω(ua, mi) = ϑa,jΩrole t−s (rj, tby)θi,by (25) j by Now, π(xi) can be updated by Ωrole t−s , ∆role and Λt−s, and the log-likelihood function is (cid:88) (cid:88) ax by (cid:88) (cid:88) L(Ωrole t−s , ∆role, Λt−s) (yilogπ(xi) + (1 − yi)log(1 − π(xi))) n(cid:88) i=1 = Our goal is to estimate the parameters in Ωrole t−s , ∆role and Λt−s to maximize the log-likelihood function. Stochastic gradient (rj, ty) and ascent is adopted again to fit the model. Then Ωrole Λs(tx, ty) can be obtained by s (cid:88) (cid:88) b 1 K 1 K (cid:88) a b Ωrole s (rj, ty) = Ωrole t−s (rj, tby) Λs(tx, ty) = 1 K Λt−s(tax, tby) (27) (28) 6 EVALUATION In this section, we conduct experiments based on a public Weibo dataset to evaluate IAD framework, and then discuss various qualitative insights. 6.1 Experimental Settings 6.1.1 Dataset Weibo is a Twitter-like social network which provides microblog- ging service. The Weibo dataset [22] provides a list of users who have forwarded contagions, as well as the forwarding timestamp. Users' friendship links are also recorded. Because of the crawling strategy, the distribution of retweet counts in different months is highly imbalanced. Thus, we select the diffusion data from July 2012 to December 2012, in which the retweet count per month is large enough and the distribution is more balanced. Consequently, we get 19,388,727 retweets on 140,400 popular microblogs. We delete the inactive users without any retweets in this period and obtain 1,077,021 distinct users for the experiment. Then we do statistical analysis to extract interacting scenarios from the dataset as illustrated in Sec. 3.1. Please note that we remove the interacting scenarios where all microblogs are neutral, and obtain a set of scenarios that have contagions with positive or negative sentiments. As LDA-S is proposed to study how the sentiments can affect the diffusion process, thus the contagions that we focus on should be those whose sentiments are evident. Based on this intuition, we refine the dataset and only retain the contagions whose positive portion or negative portion in the sentiment distribution is above 0.7, and thus obtain a new dataset consisting of contagions with greater sentiment intensities. Thus, the obtained instances after filtering is actually a subset of the instances used in the earlier version of this work [48]. If the examined contagion is adopted, the interacting scenario is a positive instance, otherwise it is a negative instance. We observe that the positive and negative instances are highly unbalanced in the dataset, so we sample a balanced dataset with equal numbers of positive and negative instances. In total, we get 64,456 microblogs and 14,537,835 interacting scenarios when K = 1, while 75,945 microblogs and 17,462,853 interacting scenarios when K = 2. We use 5-fold cross-validation, and set the number of latent topics as t = 20. 6.1.2 Baselines We compare the performance of our proposed model with several existing models. These methods are: (26) • IP Model. Infection Probability Model is based on the Independent Cascade Model [1], [2], which assigns the infection probability of a contagion to be simply the prior infection probability. IP model doesn't consider any interactions among users and contagions. • • UI Model. User Interaction (UI) Model is actually one component of our IAD framework. UI model only consid- ers the user-user interactions, more specifically, the user role-role interactions, and ignores the other interactions. IMM Model [9]. IMM incorporates the interactions among contagions into its model. IMM models the inter- actions between clusters (i.e., latent topics) of contagions. At the same time, it learns to which cluster each contagion belongs to. However, the clusters (i.e., latent topics) cannot be interpreted and thus prevent us getting deeper insights. IAD w/o S. IAD w/o S refers to IAD framework without sentiment, which is also the proposal in our previous work [48]. IAD w/ TSLDA. In our proposed IAD framework, we replace LDA-S with TSLDA algorithm [37], which intends to compare the performance difference between LDA-S and TSLDA. • • the predicting result To make a fair comparison, we use the same set of instances and the same setting of parameters. In our proposal and the baselines, we set to 0 if the predicting infection probability is less than 0.5, otherwise we set the pre- dicting result to 1. Our model and the baselines are evaluated in terms of Precision, Recall, F1-score, as well as Accuracy. All experiments are performed on a dual-core Xeon E5-2620 v3 pro- cessor. The code of our proposal has been made publicly available via https://www.dropbox.com/s/lp296h2omhhfw6e/IAD.zip?dl=0. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 TABLE 3: Performance of IAD Methods Compared to Baselines When K = 1 11 TABLE 5: Performance of IAD Methods with Varying Sentiment Intensities When K = 1 Model Name IP UI IMM IAD w/o S (t = 20) IAD w/ TSLDA (t = 20) IAD w/ LDA-S (t = 20) Precision 0.7042 0.7534 0.8005 0.8112 0.8156 0.8243 Recall Accuracy 0.6338 0.6789 0.6923 0.6134 0.7234 0.6376 0.7332 0.6505 0.7391 0.6579 0.6608 0.7439 F1-score 0.6671 0.6762 0.7098 0.7220 0.7293 0.7365 TABLE 4: Performance of IAD Methods Compared to Baselines When K = 2 Model Name IP UI IMM IAD w/o S (t = 20) IAD w/ TSLDA (t = 20) IAD w/ LDA-S (t = 20) Precision 0.7124 0.7221 0.7345 0.7616 0.7757 0.7775 Recall Accuracy 0.6799 0.6031 0.6846 0.6089 0.6954 0.6275 0.7178 0.6559 0.6601 0.7267 0.7298 0.6664 F1-score 0.6532 0.6607 0.6768 0.7048 0.7132 0.7177 6.2 Performance Results 6.2.1 Effectiveness Table 3 and Table 4 show the performance of our proposal and the baselines when K = 1 and K = 2 respectively. IAD w/ LDA-S denotes the IAD framework with our LDA-S method for extracting sentiments. It can be observed that UI and IMM models perform better than IP model, indicating that interactions among users and among contagions are vital to user's forwarding decisions. Moreover, since IMM outperforms UI, the interaction among contagions plays more important roles than the interaction among users. It can be observed that our proposed IAD meth- ods (no matter with or without sentiments) almost consistently outperform IP, UI and IMM, indicating that solely involving the interactions between contagions (e.g., IMM model) or between user interactions (e.g., UI) is not sufficient. Besides, we need to involve both of the aforementioned interactions as well as the interactions between users and contagions to get a more accurate prediction. Then we examine the effectiveness of the sentiment factors in the forwarding decision process, and observe that IAD w/ LDA-S and IAD w/ TSLDA both perform better than IAD w/o S in terms of F1-score when K = 1 and K = 2. Thus, it indicates that involving sentiments into the the framework is beneficial to the information diffusion prediction. When comparing the performance of IAD w/ LDA-S and IAD w/ TSLDA, we can find that IAD w/ LDA-S outperforms IAD w/ TSLDA all the time in terms of both F1-score and Accuracy, indicating that our proposed LDA-S can better capture the effects of the sentiments on the information diffusion than TSLDA. We also conduct experiments with a larger K, and the results are not better than those when K = 1 and K = 2. The possible reason is that individual users have limited attention, and can only jointly consider a very limited number of contagions. Moreover, to demonstrate how the sentiment intensity affects the prediction performance, we construct three datasets whose contagions are with different sentiment intensities. Specifically, the positive or negative portions in the sentiment distribution in the three datasets are above τ=0.6, τ=0.7 and τ=0.8, and we denote them as the dataset (a), (b) and (c) respectively for simplicity. It is obvious that the sentiment intensities of contagions in (c) are generally greater than that in (a) and (b). We perform Model Name IAD w/o S (t=10, τ=0.6) IAD w/o S (t=10, τ=0.7) IAD w/o S (t=10, τ=0.8) IAD w/o S (t=20, τ=0.6) IAD w/o S (t=20, τ=0.7) IAD w/o S (t=20, τ=0.8) IAD w/ LDA-S (t=10, τ=0.6) IAD w/ LDA-S (t=10, τ=0.7) IAD w/ LDA-S (t=10, τ=0.8) IAD w/ LDA-S (t=20, τ=0.6) IAD w/ LDA-S (t=20, τ=0.7) IAD w/ LDA-S (t=20, τ=0.8) Precision 0.7759 0.8156 0.7734 0.8187 0.7707 0.8205 0.8167 0.8208 0.8112 0.8243 0.8087 0.8259 Recall Accuracy 0.6589 0.7223 0.7250 0.6411 0.7243 0.6657 0.7343 0.6442 0.7198 0.6512 0.7379 0.6483 0.6564 0.7386 0.7413 0.6592 0.7332 0.6505 0.7439 0.6608 0.7348 0.6495 0.6657 0.7475 F1-score 0.7126 0.7179 0.7155 0.7210 0.7059 0.7243 0.7278 0.7312 0.7220 0.7365 0.7204 0.7372 TABLE 6: Performance of IAD Methods with Varying Sentiment Intensities When K = 2 Model Name IAD w/o S (t=10, τ=0.6) IAD w/o S (t=10, τ=0.7) IAD w/o S (t=10, τ=0.8) IAD w/o S (t=20, τ=0.6) IAD w/o S (t=20, τ=0.7) IAD w/o S (t=20, τ=0.8) IAD w/ LDA-S (t=10, τ=0.6) IAD w/ LDA-S (t=10, τ=0.7) IAD w/ LDA-S (t=10, τ=0.8) IAD w/ LDA-S (t=20, τ=0.6) IAD w/ LDA-S (t=20, τ=0.7) IAD w/ LDA-S (t=20, τ=0.8) Precision 0.7459 0.7608 0.7423 0.7621 0.7395 0.7658 0.7651 0.7746 0.7616 0.7775 0.7579 0.7789 Recall Accuracy 0.7137 0.6502 0.7186 0.6538 0.7103 0.6478 0.7212 0.6576 0.7076 0.6443 0.6595 0.7237 0.7221 0.6590 0.7263 0.6632 0.7178 0.6559 0.7298 0.6664 0.7139 0.6510 0.6678 0.7306 F1-score 0.6948 0.7033 0.6918 0.7060 0.6886 0.7087 0.7081 0.7146 0.7048 0.7177 0.7004 0.7191 the experiments on the three datasets, and can observe that our proposal performs better on the dataset whose sentiment intensities is greater. For example, in terms of accuracy, when t=10, the performance gaps between IAD with LDA-S and without LDA- S are 0.27%, 1% and 1.81% using dataset (a) and (b) and (c) respectively; when t=20, the corresponding performance gaps are 0.27%, 1.07% and 1.27%. It is intuitive as LDA-S is proposed to study how the sentiments can affect the diffusion process, and thus applying LDA-S on the contagions whose sentiments are evident is supposed to be more beneficial to the prediction than the contagions whose sentiments are mostly neutral. The detailed results are shown in Table 5 and Table 6. These results indicate that sentiment is an important factor for information diffusion. Please note that in addition to improving the predictive capability, taking sentiments into consideration can also provide a new perspective to understand the information diffusion process, and the analysis is described in Sec. 6.3. 6.2.2 Efficiency Taking the model complexity into consideration, IAD w/o S is much more efficient than IMM. The number of parameters to learn Fig. 4: Efficiency of IAD Compared to IMM 020406080100120140K=1K=2Time of Learning Process (in hours)IAD w/o SIAD w/ TSLDAIAD w/ LDA-SIMMJOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 in IAD w/o S is 469 (when r = 3, and t = 20), whereas the corresponding number in IMM model is 2,070,520 and 2,141,740 respectively when K = 1 and K = 2 (obtained by T × T × K + W × T , where T = 20, W = 103, 506 and 107, 047 when K = 1 and K = 2). For IAD w/ LDA-S, the number of parameters to infer is 3789. Fig. 4 compares the time cost in the learning process, and the results confirm the efficiency of the IAD models, especially the IAD w/o S model. In particular, IAD w/o S is faster than IMM by an order of magnitude. IAD w/ LDA-S and IAD w/ TSLDA take almost the same time in learning, and they are both slower than IAD w/o S but much faster than IMM. When K = 2, similar trends can be observed. 6.3 Analysis of Interactions Throughout this section, we provide qualitative insights into the extent to which the interactions influence the adoption of con- tagions. Due to limits of space, we only show the interaction results under IAD w/ LDA-S. Please note that interaction results under IAD w/ LDA-S and IAD w/o S should be identical except sentiment-related interactions are further derived with LDA-S. After fitting the IAD w/ LDA-S model, ∆role, Ωrole t−s and Λt−s are obtained. The results can be simply processed to obtain Ωrole topic categ. can be derived by the and Λtopic, and then Λcateg. and Ωrole category classification method. In addition, we can also derive the and Λs. Fig. 5 shows the sentiment-related interactions, i.e., Ωrole inferred interactions when t = 20 and K = 2. s • User Role-Role Interaction. In Fig. 5(a), it can be observed that authority users are more likely to adopt contagions forwarded by other authority users, rather than those from hub users or ordinary users, which indicates a status gradient on social roles seniority. Hub users would like to adopt contagions from ordinary users, rather than from hub users. The ordinary users prefer to adopt contagions from other ordinary users or hub users, while don't forward the contagions from authority users with a larger probability. The reason is that, although the number of adoptions from authoritative users is commonly large, the forwarding ratio, that is, (the number of forwarded contagions) / (the number of exposed contagions), from authoritative users is not necessarily high, due to the large exposure count in the denominator. • User Role-Contagion Sentiment Interaction. Fig. 5(b) shows that no matter what roles the users play, they are more likely to forward neutral and positive contagions than the negative contagions. It is intuitive that the users commonly prefer positive contagions to the negative ones. One possible reason why users prefer the neutral ones may be because the neutral contagions are not aggressive and less likely to conflict with other contagions, and thus easy to get accepted. From the perspective of the user roles, hub users are more likely to forward contagions than authority users and ordinary users. Particularly, the hub users like to forward the positive contagions very much. This finding is consistent with Fig. 5(d), where the hub users like to forward advertisements, as most of the advertisements are supposed to be positive. • Contagion Sentiment-Sentiment Interaction. Fig. 5(c) shows that when a negative contagion meets another neg- ative contagion, it can promote each other's propagation. This phenomenon is reasonable as they probably express 12 similar opinions and cooperate with each other to get spread. The neutral sentiment seldom suppresses itself, neither does the positive sentiment. The mutual suppres- sion between positive contagions and negative contagions is very strong, and stronger than the suppressions between positive/negative and neutral contagions. It is intuitive as positive and negative contagions commonly express oppo- site opinions, and it is more likely for them to compete rather than cooperate with each other. • User Role-Contagion Category Interaction. Fig. 5(d) shows that authority users are more likely to adopt con- tagions on economy, news, and politics, and don't like to forward advertisements. One possible reason is that the authority users such as the news media give more attention to big events such as politics rather than small events in common life. On the contrary, ordinary users prefer contagions such as movies and food, and they don't like to forward advertisements and political contagions. Hub users tend to adopt contagions about advertisement and movies, and one possible reason is that they may be spam users. • Contagion Category-Category Interaction. Fig. 5(e) reveals how different categories of contagions compete or cooperate with each other to get propagated. It can be observed that generally the entries in the matrix are negative, which indicates that relationships between dif- ferent categories are mainly competition. It validates the existing conclusion that attention is limited for individual users to adopt contagions [8]. It is obvious that the colors of the entries on the diagonal line are lighter, indicating that contagions affiliated to the same category are less likely to suppress each other. It is intuitive that when similar topics meet each other, they tend to cooperate to become hot topics and thus attract more attention, making themselves easier to get spread. It also shows that contagions belonging to food category are more likely to get adopted when simultaneously propagating with conta- gions belonging to other categories, i.e., the propagation of contagions on food are more likely to suppress the propagation of other contagions. In addition, contagions about constellation and life also attract a lot of attention. On the contrary, contagions belonging to categories such as the advertisement and news are less likely to suppress other contagions' propagation, revealing that commonly users don't like to forward them. 7 CONCLUSION In this paper, a new information diffusion framework called IAD is proposed to analyze the users' behaviors on adopting a contagion, in consideration of the interactions involving users and contagions as a whole. With this framework, we can quantitatively study how these interactions would influence the propagation process. To efficiently learn the interactions, we use a generative process to infer user roles and a co-training method to classify the contagions into explicit categories. To involve sentiment factors into the user's forwarding behavior, we also propose a LDA-S model which are able to extract the sentiment distribution and topic distribution simultaneously from contagions. Experimental results on the large-scale Weibo dataset demonstrate that IAD methods can outperform the state-of-art baselines in terms of F1-score, JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 13 (a) User Role-Role Interactions (b) User Role - Contagion Sentiment Interactions (c) Contagion Sentiment-Sentiment Interactions (d) User Role - Contagion Category Interactions (e) Contagion Category-Category Interactions Fig. 5: The inferred interactions when K = 2 and t = 20. (a) User Role-Role Interaction ∆role(ri, rj), with ri as the ordinate and rj as the abscissa, denoting the willingness of users in role ri adopting contagions posted or forwarded by users in role rj; (ri, sx), denoting the willingness of users in role ri adopting contagions of (b) User Role - Contagion Sentiment Interactions Ωrole sentiment sx; (c) Contagion Sentiment-Sentiment Interactions Λ(sx, sy), with sx as the ordinate and sy as the abscissa, denoting the influence of contagions in sentiment sy has on the decision of whether to adopt contagions in sentiment sx; (d) User Role - Contagion categ.(ri, ck), denoting the willingness of users in role ri adopting contagions of category ck; (e) Contagion Category Interactions Ωrole Category-Category Interactions Λcateg.(ci, ck), with ci as the ordinate and ck as the abscissa, denoting the influence of a contagion in category ck has on the decision of whether to adopt a contagion in category ci. s accuracy and runtime. Moreover, IAD with sentiment can further improve the prediction performance of IAD without sentiments. Last but not least, various kinds of interactions can be obtained and interesting findings can be observed, which are useful to various domains such as viral marketing. ACKNOWLEDGMENTS This work was supported in part by State Key Development Program of Basic Research of China (No. 2013CB329605), the Natural Science Foundation of China (No. 61300014, 61672313), and NSF through grants IIS-1526499, IIS-1763325, and CNS- 1626432, and DongGuan Innovative Research Team Program (No.201636000100038). REFERENCES [1] D. Kempe, J. Kleinberg, and E. Tardos, "Maximizing the spread of influence through a social network," in Proc. 9th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2003. [2] J. Goldenberg, B. Libai, and E. Muller, "Talk of the network: A complex systems look at the underlying process of word-of-mouth," Marketing letters, vol. 12, no. 3, pp. 211–223, 2001. [3] H. W. Hethcote, "The mathematics of infectious diseases," SIAM review, vol. 42, no. 4, pp. 599–653, 2000. [4] M. E. J. Newman, "The structure and function of complex networks," SIAM review, vol. 45, no. 2, pp. 167–256, 2003. [5] S. Wang, X. Hu, P. S. Yu, and Z. Li, "MMRate: inferring multi-aspect dif- fusion networks with multi-pattern cascades," in Proc. 20th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2014, pp. 1246–1255. Scalable influence estimation in continuous-time diffusion networks [6] E. Cohen, D. Delling, T. Pajor, and R. F. Werneck, "Sketch-based influence maximization and computation: Scaling up with guarantees," in Proc. 23st ACM Int. Conf. Inf. Knowl. Manage., 2014. [7] Y. Tang, Y. Shi, and X. Xiao, "Influence maximization in near-linear time: a martingale approach," in Proc. ACM Int. Conf. Special Interest Group Manage. Data, 2015. [8] L. Weng, A. Flammini, A. Vespignani, and F. Menczer, "Competition among memes in a world with limited attention," Scientific Reports, vol. 2, 2012. [9] S. A. Myers and J. Leskovec, "Clash of the contagions: Cooperation and competition in information diffusion," in Proc. 12th IEEE Int. Conf. Data Mining, 2012. [10] X. Rong and Q. Mei, "Diffusion of innovations revisited: from social network to innovation network," in Proc. 22th ACM Int. Conf. Inf. Knowl. Authority User Hub User Ordinary User Ordinary User Hub User Authority User -0.06-0.04-0.0200.020.040.060.08 Negative Positive Neutral Ordinary User Hub User Authority User 00.0050.010.0150.020.0250.030.035 Negative Positive Neutral Neutral Positive Negative -0.2-0.18-0.16-0.14-0.12-0.1-0.08-0.06-0.04-0.020 Ordinary User Hub User Authority User Traffic Technology Sports Politics News Music Movie Life History Health Food Economy Culture ConstellationAdvertisement00.010.020.030.040.050.060.07 Traffic Technology Sports Politics News Music Movie Life History Health Food Economy Culture ConstellationAdvertisementAdvertisementConstellation Culture Economy Food Health History Life Movie Music News Politics Sports Technology Traffic-0.018-0.016-0.014-0.012-0.01-0.008-0.006-0.004-0.0020JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 14 Manage., 2013. [11] Y. Bi, W. Wu, and Y. Zhu, "Csi: Charged system influence model for human behavior prediction," in Proc. 13th IEEE Int. Conf. Data Mining, 2013. [12] M. Coscia, "Competition and success in the meme pool: a case study on quickmeme. com," in Proc. Int. AAAI Conf. Weblogs Soc. Media, 2013. [13] I. Valera and M. Gomez-Rodriguez, "Modeling adoption and usage of competing products," in Proc. 15th IEEE Int. Conf. Data Mining, 2015. [14] N. Pathak, A. Banerjee, and J. Srivastava, "A generalized linear threshold model for multiple cascades," in Proc. 10th IEEE Int. Conf. Data Mining, 2010. [15] B. A. Prakash, A. Beutel, R. Rosenfeld, and C. Faloutsos, "Winner takes all: competing viruses or ideas on fair-play networks," in Proc. 21rd Int. Conf. World Wide Web, 2012. [16] B. Karrer and M. E. J. Newman, "Competing epidemics on complex networks," Physical Review E, vol. 84(3): 036106, 2011. [17] R. Fan, J. Zhao, Y. Chen, and K. Xu, "Anger is more influential than joy: Sentiment correlation in weibo," PloS one, vol. 9, no. 10, p. e110184, 2014. [18] E. Ferrara and Z. Yang, "Quantifying the effect of sentiment on infor- mation diffusion in social media," PeerJ Computer Science, vol. 1, p. e26, 2015. [19] Y. Yang, J. Tang, C. W. Leung, Y. Sun, Q. Chen, J. Li, and Q. Yang, "Rain: Social role-aware information diffusion," in Proc. AAAI Conf. Artificial Intell., 2015. [20] A. Blum and T. Mitchell, "Combining labeled and unlabeled data with co-training," in Proc. 11th Annual Conf. Comput. Learn. Theory, 1998. [21] D. Blei, A. Ng, and M. Jordan, "Latent dirichlet allocation," J. Machine Learn. Research, vol. 3: 993-1022, 2003. [22] J. Zhang, B. Liu, J. Tang, T. Chen, and J. Li, "Social influence locality for modeling retweeting behaviors," in Proc. 23th Int. Joint Conf. Artificial Intell., 2013. [23] J. Zhang, Z. Fang, W. Chen, and J. Tang, "Diffusion of following links in microblogging networks," IEEE Trans. Knowl. Data Eng., vol. 27, no. 8, pp. 2093–2106, Aug 2015. [24] M. G. Rodriguez, J. Leskovec, and A. Krause, "Diffusive logistic model towards predicting information diffusion in online social networks," in Proc. 13th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2010, pp. 1019–1028. [25] J. Leskovec, M. Mcglohon, C. Faloutsos, N. Glance, and M. Hurst, "Cascading behavior in large blog graphs," in Proc. SAIM Int. Conf. Data Minining, 2007, pp. 551–556. [26] S. Wu, J. Hofman, W. Mason, and D. Watts, "Who says what to whom on twitter," in Proc. 20rd Int. Conf. World Wide Web, 2011, pp. 705–714. [27] E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts, "Everyone's an influencer: quantifying influence on twitter," in Proc. 4th ACM Int. Conf. Web Search Data Mining, 2011, pp. 65–74. [28] K. Sneppen, A. Trusina, M. H. Jensen, and S. Bornholdt, "A minimal model for multiple epidemics and immunity spreading," PloS one, vol. 5(10): e13326, 2010. [29] J. Yang and J. Leskovec, "Modeling information diffusion in implicit networks," in Proc. 10th IEEE Int. Conf. Data Mining, 2010. [30] Y. Zhang, A. Adiga, S. Saha, A. Vullikanti, and B. A. Prakash, "Near- optimal algorithms for controlling propagation at group scale on networks," IEEE Trans. Knowl. Data Eng, vol. 28, no. 12, pp. 3339–3352, 2016. [31] D. Li, S. Zhang, X. Sun, H. Zhou, S. Li, and X. Li, "Modeling informa- tion diffusion over social networks for temporal dynamic prediction," IEEE Trans. Knowl. Data Eng, vol. PP, no. 99, pp. 1–1, 2017. [32] S. Petrovic, M. Osborne, and V. Lavrenko, "Rt to win! predicting message propagation in twitter," in Proc. Int. AAAI Conf. Weblogs Soc. Media, 2011. [33] M. Granovetter, "Threshold models of collective behavior," American J. [34] T. Schelling, Micromotives and Macrobehavior. New York: Norton, Soc., vol. 1420-1443, 1978. 2006. [35] Y. Su, X. Zhang, L. Liu, S. Song, and B. Fang, "Understanding informa- tion interactions in diffusion: an evolutionary game-theoretic perspective," Frontiers of Computer Science, vol. 10, no. 3, pp. 518–531, 2016. [36] C. Lin and Y. He, "Joint sentiment/topic model for sentiment analysis," in Proc. 18th ACM Conf. Inf. Knowl. Manage., 2009, pp. 375–384. [37] T. Nguyen and K. Shirai, "Topic modeling based sentiment analysis on social media for stock market prediction," in Meeting Assoc. Comput. Lingui. and Int. Joint Conf. Natural Language Proc. Asian Federation Natural Language Proc., 2015. [38] L. Dang-Xuan and S. Stieglitz, "Impact and diffusion of sentiment in political communication-an empirical analysis of political weblogs." in Proc. Int. AAAI Conf. Weblogs Soc. Media, 2012. [39] Q. Mei, X. Ling, M. Wondra, H. Su, and C. X. Zhai, "Topic sentiment mixture: modeling facets and opinions in weblogs." in Proc. 16rd Int. Conf. World Wide Web, 2007, pp. 171–180. [40] C. Lin, Y. He, R. Everson, and S. Ruger, "Weakly supervised joint sentiment-topic detection from text," IEEE Trans. Knowl. Data Eng, vol. 24, no. 6, pp. 1134–1145, June 2012. [41] S. Liu, X. Cheng, F. Li, and F. Li, "TASC:topic-adaptive sentiment classification on dynamic tweets," IEEE Trans. Knowl. Data Eng., vol. 27, no. 6, pp. 1696–1709, June 2015. [42] S. Tan, Y. Li, H. Sun, Z. Guan, X. Yan, J. Bu, C. Chen, and X. He, "Interpreting the public sentiment variations on twitter," IEEE Trans. Knowl. Data Eng, vol. 26, no. 5, pp. 1158–1170, 2014. [43] W. X. Zhao, J. Jiang, H. Yan, and X. Li, "Jointly modeling aspects and opinions with a maxent-lda hybrid," in Proc. Conf. Emp. Methods Natural Language Proc. Association for Computational Linguistics, 2010, pp. 56–65. [44] C. P. Ho, "Exploiting coherence for the simultaneous discovery of latent facets and associated sentiments," in Proc. SIAM Int. Conf. Data Mining. SIAM, 2011. [45] S. Wang, Z. Chen, and B. Liu, "Mining aspect-specific opinion using a holistic lifelong topic model," in Proc. 25th Int. Conf. World Wide Web. International World Wide Web Conferences Steering Committee, 2016, pp. 167–176. [46] K. Schouten and F. Frasincar, "Survey on aspect-level sentiment analy- sis," IEEE Trans. Knowl. Data Eng, vol. 28, no. 3, pp. 813–830, 2016. [47] T. N. Rubin, A. Chambers, P. Smyth, M. Steyvers, "Statistical topic models for multi-label document classification," Machine Learning, vol. 88(1-2), pp.157–208, 2012. [48] Y. Su, X. Zhang, P. S. Yu, W. Hua, X. Zhou, and B. Fang, "Understanding information diffusion under interactions," in Proc. 26th Int. Joint Conf. Artificial Intell., 2016. [49] L. Page, S. Brin, R. Motwani, and T. Winograd, "The pagerank citation ranking: bringing order to the web," Stanford InfoLab, Technical Report, 1999. [50] J. Kleinberg, "Authoritative sources in a hyperlinked environment," J. ACM (JACM), vol. 46(5): 604-632, 1999. [51] C. Chang and C. Lin, "Libsvm: A library for support vector machines," ACM Trans. Intell. Syst. and Technol., vol. 2(3): 27, 2011. [52] L.-W. Ku and H.-H. Chen, "Mining opinions from the web: Beyond relevance retrieval," J. American Soc. Inf. Science Tech., vol. 58, no. 12, pp. 1838–1850, 2007. Xi Zhang received the PhD degree in Com- puter Science from Tsinghua University. He is an associate professor in Beijing University of Posts and Telecommunications, and is also the vice director of Key Laboratory of Trustworthy Distributed Computing and Service, Ministry of Education, China. He was a visiting scholar at the University of Illinois at Chicago. His research interests include data mining and computer ar- chitecture. He is a member of IEEE. Yuan Su received the master degree in Com- puter Science from Beijing University of Posts and Telecommunications in 2012. He is now a PhD student in Key Laboratory of Trustworthy Distributed Computing and Service (Beijing Uni- versity of Posts and Telecommunications), Min- istry of Education, China. He was a visiting stu- dent at the University of Queensland from May 2015 to May 2016. His research interests include data mining and social network analysis. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15 Siyu Qu received the bachelor degree in Com- puter Science from Xidian University in 2012. She is now a master student in Key Laboratory of Trustworthy Distributed Computing and Ser- vice (Beijing University of Posts and Telecom- munications), Ministry of Education, China. Her research interests include data mining and ma- chine learning. Sihong Xie received the PhD Degree from the Department of Computer Science at University of Illinois at Chicago. He is an assistant profes- sor in Department of Computer Science and En- gineering, Lehigh University. His research inter- ests include big data, data mining, and machine learning. Binxing Fang received his PhD from Harbin Institute of Technology, China in 1989. He is a member of the Chinese Academy of Engineering and a professor in School of Cyberspace Secu- rity at Beijing University of Posts and Telecom- munications. He is currently the chief scientist of State Key Development Program of Basic Re- search of China. His current interests include big data and information security. Philip S. Yu received the PhD degree in elec- trical engineering from Stanford University. He is a distinguished professor in computer sci- ence at the University of Illinois, Chicago, and is the Wexler chair in Information Technology. His research interests include big data, data mining, data stream, database, and privacy. He was the editor-in-chief of the IEEE Transac- tions on Knowledge and Data Engineering and the ACM Transactions on Knowledge Discovery from Data. He received the ACM SIGKDD 2016 Innovation Award, a Research Contributions Award from the IEEE Inter- national Conference on Data Mining (2003), and a Technical Achieve- ment Award from the IEEE Computer Society (2013). He is a fellow of the IEEE and the ACM.
1711.05462	1	1711	2017-11-15T09:10:02	A Machine Learning Approach to Modeling Human Migration	[ "cs.SI", "cs.LG", "physics.soc-ph" ]	Human migration is a type of human mobility, where a trip involves a person moving with the intention of changing their home location. Predicting human migration as accurately as possible is important in city planning applications, international trade, spread of infectious diseases, conservation planning, and public policy development. Traditional human mobility models, such as gravity models or the more recent radiation model, predict human mobility flows based on population and distance features only. These models have been validated on commuting flows, a different type of human mobility, and are mainly used in modeling scenarios where large amounts of prior ground truth mobility data are not available. One downside of these models is that they have a fixed form and are therefore not able to capture more complicated migration dynamics. We propose machine learning models that are able to incorporate any number of exogenous features, to predict origin/destination human migration flows. Our machine learning models outperform traditional human mobility models on a variety of evaluation metrics, both in the task of predicting migrations between US counties as well as international migrations. In general, predictive machine learning models of human migration will provide a flexible base with which to model human migration under different what-if conditions, such as potential sea level rise or population growth scenarios.	cs.SI	cs	A Machine Learning Approach to Modeling Human Migration Caleb Robinson and Bistra Dilkina Georgia Institute of Technology [email protected], [email protected] School of Computational Science and Engineering Atlanta, Georgia. 30332 7 1 0 2 v o N 5 1 ] I S . s c [ 1 v 2 6 4 5 0 . 1 1 7 1 : v i X r a Abstract Human migration is a type of human mobility, where a trip involves a person moving with the intention of changing their home location. Predicting human migration as accurately as possible is important in city planning applications, interna- tional trade, spread of infectious diseases, conservation plan- ning, and public policy development. Traditional human mo- bility models, such as gravity models or the more recent ra- diation model, predict human mobility flows based on pop- ulation and distance features only. These models have been validated on commuting flows, a different type of human mo- bility, and are mainly used in modeling scenarios where large amounts of prior ground truth mobility data are not available. One downside of these models is that they have a fixed form and are therefore not able to capture more complicated mi- gration dynamics. We propose machine learning models that are able to incorporate any number of exogenous features, to predict origin/destination human migration flows. Our ma- chine learning models outperform traditional human mobility models on a variety of evaluation metrics, both in the task of predicting migrations between US counties as well as inter- national migrations. In general, predictive machine learning models of human migration will provide a flexible base with which to model human migration under different what-if con- ditions, such as potential sea level rise or population growth scenarios. 1 Introduction Models of human mobility in their different forms are im- portant for many reasons. Models of human commuting can help reduce traffic congestion and pollution, and can be used to drive land use policy and development choices (De Mon- tis et al. 2010). Models of human migration are equally im- portant to policy makers as they can give broader estimates of how the population of an area will change in upcoming years, how labor markets might be affected (Dinkelman and Mariotti 2016), how infectious diseases spread (Balcan et al. 2009; Sorichetta et al. 2016), and how international trade will change (Fagiolo and Mastrorillo 2013). Much recent re- search focuses on modeling human commuting flows (Ma- succi et al. 2013; Lenormand, Bassolas, and Ramasco 2016); however little has focused on explicitly modeling human mi- gration. Human mobility has been traditionally modeled with the so-called gravity model, which posits that the probability of a trip between two locations decays directly as a function of the distance between them. This model was introduced in its modern form in 1946 (Zipf 1946) and has been used in many applications since (Schneider 1959; Lee 1966; Clark and Ballard 1980; Greenwood 1985; Letouz´e et al. 2009; Noulas et al. 2012). More recently, the radiation model (Si- mini et al. 2012) has been shown to capture long range trips better than gravity based models, and is described as 'a uni- versal model for mobility and migration patterns'. The ra- diation model posits that the probability of a trip will de- cay indirectly with distance and directly with the amount of intervening opportunities, a notion first proposed by Stouf- fer (Stouffer 1940). The radiation model has been extended several times since being proposed in 2012 (Masucci et al. 2013; Yang et al. 2014; Ren et al. 2014). In general, grav- ity models have been shown to be more capable of repro- ducing commuting flows, i.e. human mobility at small spa- tial scales (Lenormand, Bassolas, and Ramasco 2016), while radiation models have been shown to be better at repro- ducing migration flows, i.e. human mobility at larger spa- tial scales (Simini et al. 2012). Additionally, human mi- grations have been estimated by fitting generalized linear models derived from the gravity model (Cohen et al. 2008; Dennett and Wilson 2013). Both gravity and radiation mod- els are analytical models with crafted functional forms and limited input data requirements. These models are focused on explaining human migration, rely on linear relationships between independent variables, and use hand crafted fea- tures for each zone. These approaches, while useful for ex- plaining human migration, trade predictive power for inter- pretability. Data sources such as the World Bank and US Census provide many zone-based features that can be algo- rithmically combined in a non-linear manner by tree or neu- ral network based models to best predict human migration. Our key contributions are as follows: 1. We develop the first general machine learning formulation of the human migration prediction problem. 2. We develop a pipeline for training machine learning mod- els to tackle this problem that includes procedures to deal with dataset imbalance, hyperparameter tuning, and per- formance evaluation. 3. We develop a custom loss function for training artificial neural networks that is more suitable for the migration prediction task. 4. We compare the performance of machine learning models to traditional models of human migration on two datasets, and show that the machine learning models outperform the traditional models in all cases. 2 Traditional Migration Models ing to all other destinations sum to 1, i.e.(cid:80)n In human mobility modeling, we are usually given n zones with the goal of predicting the number of people, Tij, that move from every zone i to every other zone j. Traditional mobility models, such as the radiation model (Simini et al. 2012) and the gravity model (Lenormand, Bassolas, and Ra- masco 2016), break the problem of estimating Tij into two pieces: estimating the total number of people Gi that leave zone i (also referred to as a production function), and es- timating the probability Pij of a move occurring from i to j. The predicted number of migrants from i to j would be Tij = GiPij. As a convention, the probabilities are normal- ized such that, given an origin i, the probabilities of travel- j=1 Pij = 1. If prior information about the number of incoming and outgo- ing travelers per zone is known, then a constrained frame- work, such as the one described in Lenormand et al (Lenor- mand, Bassolas, and Ramasco 2016), can be used. If this information is not known, which is the case in predicting migrations, then a production function must be used to es- timate the number of outgoing travelers per zone instead. A simple production function for a dataset can be found by expressing the number of outgoing migrants of a zone as a constant fraction of the population of that zone (for counties in the US, this percentage is ≈ 0.03 (Simini et al. 2012)). Given prior timestep's data on the population mi and the corresponding number of outgoing migrants Oi in the cur- rent timestep for every zone i, the production function is ex- pressed as Gi = M (mi) = αmi, where α is the slope of the line of best fit through the pairs (mi, Oi). The traditional models of human mobility that we include in this study are: the radiation model (Simini et al. 2012), ex- tended radiation model (Yang et al. 2014), and gravity mod- els with both power and exponential distance decay func- tions (Lenormand, Bassolas, and Ramasco 2016). The only information used by these models is: mi, population of a zone i for both the origin and destination zones; dij, the dis- tance between two zones i and j; and sij, a metric of in- tervening opportunities measured as the total population of all intervening zones between i and j, defined as all zones whose centroid falls in the circle centered at i with radius dij (not including zones i or j). See Table 1 for a description of each model. All of these models have two parameters that must be tuned using historical data: the production function param- eter, α, that determines what fraction of the population of a zone will migrate away in a given year, and a model pa- rameter, β. From a socio-economic point of view, traditional models have the advantage of being interpretable, however we will show that this interpretability comes at a cost of pre- dictive accuracy, as machine learning models can use similar historic data to achieve better results. 3 Evaluation Methods To evaluate how well alternative models perform, we use four metrics that compare how well a predicted migration matrix, T, recreates the ground truth values, T. The first two of these (CP C, CP Cd) have been used in previous litera- ture to evaluate human mobility models (Lenormand et al. 2012; Lenormand, Bassolas, and Ramasco 2016), and the other two are standard regression metrics: Common Part of Commuters (CP C) This metric di- travelers between the rectly compares numbers of predicted and ground truth matrices. It will be 0 when the two matrices have no entries in common, and 1 when they are identical. We note that this metric, as used in previous studies of commuting flows, is identical to the Bray- Curtis similarity score used to compare abundance data in ecological studies (Faith, Minchin, and Belbin 1987; Legendre and Legendre 2012). 2(cid:80)n i,j=1 Tij +(cid:80)n (cid:80)n i,j=1 min(Tij, Tij) Tij i,j=1 CP C(T, T) = Common Part of Commuters Distance Variant (CP Cd) This metric measures how well a predicted migration matrix recreates trips at the same distances as the ground truth data. In this definition, N is a histogram where a bin Nk contains the number of migrants that travel between 2k − 2 and 2k kilometers. It will be 0 when the two matrices do not have any migrations at the same distance, and 1 when all fall within the same distances. 2(cid:80)∞ (cid:80)∞ k=1 Nk +(cid:80)∞ k=1 min(Nk, Nk) Nk k=1 CP Cd(T, T) = Root mean squared error (RM SE) This is a standard re- gression measure that will "punish" larger errors more than small errors. This score ranges from 0 in a perfect match, to arbitrarily large values as the predictions be- come worse. (1) (2) (cid:118)(cid:117)(cid:117)(cid:116) 1 n n(cid:88) i,j=1 RM SE(T, T) = (Tij − Tij)2 (3) Coefficient of determination (r2) This score measures the goodness of fit between a set of predictions and the ground truth values. This score ranges from 1, in a perfect fit, to arbitrarily negative values as a fit becomes worse, and is 0 when the predictions are equivalent to the expec- tation of the ground truth values. r2(T, T) = 1 − (4) (cid:80)n (cid:80)n i,j=1(Tij − Tij)2 i,j=1(Tij − ¯T )2 ing migrants for a zone, i, is calculated as vi = (cid:80)n In addition to the previous four metrics, we compare the ground truth number of incoming migrants and the predicted number of incoming migrants per zone using mean abso- lute error (M AE) and r2. The predicted number of incom- j=1 Tji. We argue that it is important to explicitly measure how well each model performs at estimating the number of incoming migrants, because the number of incoming migrants to a lo- cation will be the most important measure for policy makers in that area. Incoming migrant predictions can inform pop- ulation growth estimates and hence infrastructure planning and job analysis. Model Radiation Extended Radiation Equation Tij = M (mi) Tij = M (mi) [(mi+mj +sij )β−(mi+sij )β ](mβ (mi+sij )(mi+mj +sij ) mimj [(mi+sij )β +1][(mi+mj +sij )β +1] Gravity with Power Law −β mimj d ij k=1 mkd Gravity with Exponential Law Tij = M (mi) mimj e Tij = M (mi) (cid:80)n (cid:80)n k=1 mke−βdik −β ik −βdij i +1) Table 1: Traditional migration models 4 Learning Migration Models Formally, the problem of modeling human migration is as follows: given n zones, d1 features describing each zone, F ∈ Rn×d1, and d2 joint features describing features be- tween a pair of zones, J ∈ Rn×n×d2, at some timestep t, the objective is to predict the origin/destination migration ma- trix T ∈ Nn×n at the next timestep, t + 1, where an entry Tij represents the estimated number of migrants relocating from zone i to zone j. Our goal is to estimate a function f (Fi:, Fj:, Jij) = Tij, which takes the features of zone i and j, as well as the joint features between them, as input, and di- rectly outputs the estimated number of migrants that travel from i to j. This approach is different from how the tradi- tional migration models work as it does not require a produc- tion function, but instead directly predicts Tij. This formula- tion contains the simplifying assumption that migrant flows are static in time, meaning that they can be entirely deter- mined by the features from the previous timestep. In reality migrant flows will be dependent on temporal features, such as long term developmental trends, however many places will not have enough data to take advantage of these pat- terns. With this formulation, our models can be applied more broadly to predict future migration patterns in locations that have only collected a single year of data. Hyperparameter optimization To fit f for a given dataset, we will train two machine learn- ing models, "extreme" gradient boosting regression (XG- Boost model) (Chen and Guestrin 2016), and a deep learn- ing based artificial neural network model (ANN model) (Le- Cun, Bengio, and Hinton 2015). Each of these models con- tains several hyperparameters that must be tuned to obtain good performance on a given learning task. Our first model, the XGBoost model, is a standard machine learning model based on gradient boosting trees (Mason et al. 1999) that of- ten performs very well on many regression and classification tasks. One benefit of this model is that it gives a ranking of the relative feature importances (Friedman 2001). The pa- rameters of the XGBoost model that we consider for hyper- parameter tuning are the maximum tree depth, number of es- timators, and learning rate. Our ANN model is composed of densely connected layers with ReLU activation layers1. We tune the following ANN parameters: loss function, number 1Our model is implemented in Python with the Keras library: https://keras.io/ of layers, layer width, number of training epochs, and train- ing mini-batch size. Dealing with zero-inflated data We observe that migration data is heavily zero-inflated, where in any given year, most pairs of zones do not have any migrants traveling between them, i.e. Tij = 0 for most (i, j) pairs. Considering migrations between US counties (Pierce 2015), less than 1% of the possible pairs of counties have migrations between them. This imbalance will cause prob- lems for machine learning models. To address this problem, when creating a training dataset we undersample "negative" samples between pairs of zones for which there are no ob- served migrations. This is a naıve technique that will neces- sarily throw out available information (Guo et al. 2008). To offset this, we introduce a hyperparameter k that determines the number of "negative" examples of migrations to train with. If there are nt pairs of zones where there are observed migrations, "positive examples", we include all nt, and an additional ntk randomly chosen zone pairs where there are no observed migrants. This hyperparameter is included in both the XGBoost and ANN model parameter searches. We give further details on the hyperparameter tuning process in Section 5. Custom ANN loss function Previously we mentioned that our ANN model will consider different loss functions as a hyperparameter. Common loss functions for regression tasks include "mean squared error" (MSE), "mean absolute error" (MAE), and "mean absolute percentage error" (MAPE). Our preliminary experimental results show that MAE and MAPE loss functions perform poorly, partially due to their inability to punish large errors and deal with many zeros respectively. To contrast with the aforementioned zero-inflation problem, we observe that the distribution of migrant counts has a long tail, whereby few pairs of zones consistently experience large volumes of mi- grations. Considering these observations, and because the CPC metric is one of the key metrics of interest (described in detail in Section 3), we derive a new loss function based on CPC to train the ANN model with. This loss function is given as: L(y, y) = 1 − 2(cid:80)n (cid:80)n i=1 yi +(cid:80)n i=1 min(yi, yi) i=1 yi Where yi is a migration flow entry from T. The gradient update for this loss function is: (5) 2(cid:80)n ((cid:80)n i=1 yi +(cid:80)n i=1 min(yi, yi) i=1 yi)2− (cid:26)2 i=1 yi +(cid:80)n (cid:80)n yj < yj otherwise 0 i=1 yi ∂L(y, y) ∂ yj = Intuitively, this loss will 'reward' predictions where the number of migrants matches the ground truth, while also enforcing per link absolute error to be minimized. This loss function is not exactly equivalent to the CP C, as during the ANN training it will only consider a single mini-batch worth of samples at a time (in our case y = y = mini batch, where as the CP C metric is a function of the entire migra- tion matrix). Empirically this custom loss function results in better validation performance and faster training times than MSE loss. 5 Experiments Datasets We perform experiments comparing the performance of tra- ditional models to the performance of our machine learning on two datasets, the USA Migration dataset and the Global Migration dataset. The USA Migration dataset consists of yearly intra-county migrations in the USA between 3106 counties from the IRS Tax-Stats data (Pierce 2015) for the 11 years in the range from 2004 to 2014. We supplement the migration data with the following 7 per-county features (taken from the US Census estimates and calculated from the Census TIGER line maps of US county boundaries): population, land area, population density, median household income, county wa- ter area, is a coastal county, and number of neighboring counties. In addition to these 7 per-county features, we add the following between-county features: distance, interven- ing population, intervening land area, intervening number of counties, intervening population density, and interven- ing median income. The intervening features are calculated based on the idea of "intervening opportunities" presented in the radiation model. For any given county-level variable, x, e.g. population, the intervening amount of that variable between counties i and j is defined as sx i,j, the sum of all x in the intervening counties that fall within the circle centered at county i with a radius to county j (excluding xi and xj). The Global Migration dataset consists of decadal inter- country migration data between 207 countries from the World Bank Global Bilateral Migration Database ( Ozden et al. 2011). The Global Migration dataset contains 5 timesteps, one every 10 years from 1960 to 2000. In the Global Migration dataset we add the following 5 per- country features (taken from World Bank World Develop- ment Indicators data (World Bank 2017)): population, pop- ulation density, population growth, agricultural emissions, and land area. Additionally, we include 3 between-country features: distance, intervening population, and intervening land area. For each year of data in the USA Migration and Global Migration datasets we create an 'observation' for each pair of zones, an origin zone and destination zone (counties in the USA Migration dataset and countries in the Global Mi- gration dataset). Each observation consists of the per-zone (6) features for both the origin zone and destination zone (pop- ulation of origin, population of destination, etc.) and the between-zone features of the origin and destination. This corresponds exactly to the Fi:, Fj:, Jij of the function f (de- scribed in Section 4) that we want to learn. An observation is associated with the target number of migrants, Tij, travel- ing from the origin to the destination. Formally, for a given year, t, number of zones, n, number of per-zone features, d1, and number of between-zone features, d2, we create a matrix of observations Xt ∈ Rn2×(2d1+d2) and vector of targets Yt ∈ Rn2, capturing the migration flows observed in year t + 1. as for three the models "years" of data sets. Specifically, the hyperparameters of 2 and Experimental Setup that To select 4, we consider we described in Sections training, validation, triplets of of testing and years {(Xt−2, Yt−2), (Xt−1, Yt−1), (Xt, Yt)}, we data call (Xt−2, Yt−2) the training set, (Xt−1, Yt−1) the validation set, and (Xt, Yt) the test set. We tune the hyperparameters of the models using a randomized grid search with 50 sampled hyperparameters using the training and validation sets. We select the best set of hyperparameters according to the CP C score, then use those parameters to train a model on the validation set and record its performance on the test set. We repeat this process for each (t − 2, t − 1, t) triplet of years in each dataset. Our final results are reported as averages over the test set results. A hyperparameter present in both XGBoost and ANN models is the downsampling rate, k. As a preprocessing step for a given year of training data (Xt, Yt), we include all ob- servations, Xi where Yi > 0 (let this number of samples be m), and choose k ∗ m random samples with replacement from the remaining observation (where Yi = 0). This sam- pling process only takes place when training a model. When testing a model on the validation or test sets, the full datasets are always used. For experiments with the USA Migration dataset we consider values of k in a uniform distribution of integers from 5 to 100, while for experiments with the Global Migration dataset we consider the uniform distribu- tion of integers from 1 to 5, because the average percentage of non-zeros is < 1% and 20% respectively. For the XGBoost models, we sample the following pa- rameters: maximum tree depth from U{2, 7}2, number of estimators from U{25, 275}, and learning rate uniformly in the range from 0 to 0.5. For the ANN models we sam- ple the following parameters: network loss function uni- formly from {'CPC Loss', 'MSE'}, number of layers uni- formly from U{1, 5}, layer width from U{16, 128}, number of training epochs from U{10, 50}, and training mini-batch size uniformly from {29, 210, 211, 212, 213, 214}. We calibrate the parameters of the traditional models in a similar manner. Every traditional model, except for the ra- diation model, has two parameters, α and β, that must be calibrated to give useful results (the radiation model only uses α). For each {(Xt−1, Yt−1), (Xt, Yt)} pair of data (that we refer to as training and testing sets respectively), we find 2U{a, b} is the discrete uniform distribution. USA Migrations Production Function Gravity Model Exponential Decay Gravity Model Power Law Decay Radiation Model Extended Radiation Model XGBoost model - traditional features ANN model - traditional features XGBoost model - extended features ANN model - extended features No Production Function XGBoost model - traditional features ANN model - traditional features XGBoost model - extended features ANN model - extended features CP C 0.53 +/- 0.01 0.56 +/- 0.01 0.53 +/- 0.01 0.58 +/- 0.01 0.51 +/- 0.08 0.63 +/- 0.01 0.58 +/- 0.03 0.68 +/- 0.01 CP C 0.54 +/- 0.11 0.63 +/- 0.02 0.62 +/- 0.04 0.69 +/- 0.01 Metrics on full matrix CP Cd 0.66 +/- 0.02 0.78 +/- 0.02 0.76 +/- 0.02 0.83 +/- 0.01 0.74 +/- 0.07 0.86 +/- 0.02 0.78 +/- 0.02 0.89 +/- 0.02 CP Cd 0.99 +/- 0.02 0.88 +/- 0.06 0.99 +/- 0.02 0.93 +/- 0.05 RM SE 87.4 +/- 9.0 75.7 +/- 8.0 47.6 +/- 5.0 35.6 +/- 3.0 28.6 +/- 5.4 35.1 +/- 3.2 24.2 +/- 1.4 29.8 +/- 2.7 RM SE 18.5 +/- 6.1 35.3 +/- 3.5 13.0 +/- 1.5 28.0 +/- 3.6 Metrics on incoming migrants vector (Average Incoming Migrants = 3,196) r2 -1.48 +/- 0.28 -0.86 +/- 0.26 0.26 +/- 0.09 0.59 +/- 0.03 0.72 +/- 0.10 0.60 +/- 0.03 0.81 +/- 0.02 0.71 +/- 0.02 M AE 1,216 +/- 128 1,129 +/- 129 1,346 +/- 148 1,123 +/- 117 1,151 +/- 249 911 +/- 107 968 +/- 56 935 +/- 98 r2 M AE 0.88 +/- 0.08 0.60 +/- 0.04 0.94 +/- 0.02 0.75 +/- 0.03 3,091 +/- 1,740 1,188 +/- 259 2,060 +/- 622 909 +/- 48 r2 0.67 +/- 0.03 0.72 +/- 0.04 0.80 +/- 0.02 0.86 +/- 0.02 0.86 +/- 0.04 0.91 +/- 0.01 0.89 +/- 0.02 0.91 +/- 0.02 r2 0.41 +/- 0.85 0.84 +/- 0.16 0.76 +/- 0.28 0.92 +/- 0.04 Table 2: USA Migration results. Comparison of the ANN and XGBoost models with and without a production function to tradi- tional migration models. The values shown in the table are the average and standard deviations of the models' test performance on 2006 through 2014 data. Bold values indicate the best values per column. Global Migrations Metrics on full matrix Metrics on incoming migrants vector (Average Incoming Migrants = 674,858) Production Function Gravity Model Exponential Decay Gravity Model Power Law Decay Radiation Model Extended Radiation Model XGBoost model - traditional features ANN model - traditional features XGBoost model - extended features ANN model - extended features No Production Function XGBoost model traditional features ANN model traditional features XGBoost model extended features ANN model extended features CP C 0.16 +/- 0.00 0.16 +/- 0.00 0.16 +/- 0.00 0.16 +/- 0.00 0.18 +/- 0.01 0.19 +/- 0.01 0.21 +/- 0.01 0.22 +/- 0.02 CP C 0.33 +/- 0.02 0.33 +/- 0.01 0.43 +/- 0.03 0.40 +/- 0.02 CP Cd 0.16 +/- 0.00 0.15 +/- 0.00 0.14 +/- 0.00 0.14 +/- 0.00 0.14 +/- 0.01 0.15 +/- 0.01 0.16 +/- 0.01 0.17 +/- 0.01 CP Cd 0.59 +/- 0.03 0.37 +/- 0.04 0.64 +/- 0.02 0.43 +/- 0.02 RM SE 62,218 +/- 5,341 61,523 +/- 5,278 62,173 +/- 5,277 62,108 +/- 5,299 58,377 +/- 5,141 60,272 +/- 4,610 57,909 +/- 5,409 58,887 +/- 4,477 r2 0.02 +/- 0.03 0.05 +/- 0.00 0.02 +/- 0.00 0.03 +/- 0.00 0.14 +/- 0.01 0.08 +/- 0.02 0.16 +/- 0.02 0.12 +/- 0.02 M AE 651,194 +/- 80,220 628,678 +/- 79,474 614,483 +/- 79,378 618,576 +/- 76,150 597,478 +/- 79,178 589,789 +/- 79,542 573,090 +/- 56,987 563,259 +/- 74,127 RM SE r2 M AE 52,729 +/- 5,455 56,005 +/- 882 47,329 +/- 5,073 50,921 +/- 3,556 0.26 +/- 0.26 0.20 +/- 0.11 0.42 +/- 0.13 0.33 +/- 0.13 938,905 +/- 172,834 537,351 +/- 44,034 577,473 +/- 77,315 459,841 +/- 55,479 r2 0.00 +/- 0.02 0.03 +/- 0.03 0.04 +/- 0.02 0.03 +/- 0.02 0.10 +/- 0.01 0.26 +/- 0.03 0.15 +/- 0.02 0.24 +/- 0.06 r2 0.18 +/- 0.28 0.53 +/- 0.16 0.48 +/- 0.34 0.52 +/- 0.30 Table 3: Global Migration results. Comparison of the ANN and XGBoost models with and without a production function to traditional migration models. The values shown in the table are the average and standard deviations of the models' test performance on 2006 through 2014 data. Bold values indicate the best values per column. USA Features Intervening number of counties Population of origin (trad) Population of destination (trad) Intervening population (trad) Is destination coastal Distance between counties (trad) Intervening area Area of destination Number of neighbors destination Water area of origin Global Features Importance 25.3% +/- 2.4% Population growth of origin 15.7% +/- 1.7% Intervening population (trad) 14.2% +/- 0.9% Agricultural emissions of destination 6.1% +/- 1.2% Intervening land area 4.3% +/- 4.6% Population growth of destination 3.7% +/- 0.9% Population of destination (trad) 3.6% +/- 0.9% Distance between counties (trad) 3.5% +/- 0.5% Population of origin (trad) 3% +/- 1.6% Population density of destination 3% +/- 1.3% Land area of origin Importance 19.5% +/- 16.0% 12.3% +/- 3.7% 10.6% +/- 5.8% 8.7% +/- 5.6% 7.9% +/- 6.2% 6.9% +/- 0.8% 6.6% +/- 1.9% 6.1% +/- 1.6% 5.7% +/- 4.5% 5.2% +/- 3.1% Table 4: Top 10 most important (extended) features in both the USA Migration and Global Migration datasets. The values in the table show the average and standard deviations of the information gain feature importances from an XGBoost model trained on the extended feature set for each timestep of data. the value of α that gives the best production function on the training set. Similarly, we find the value of β that maximizes the CP C score of each traditional models on the training set. We then use these α and β values to run each model on the testing set, and report the results as averages over all test set results. To directly compare how the ML models and traditional models perform under the same conditions, we perform ex- periments where the ML models are used with the same pro- duction functions as the traditional models. This imposes an artificial constraint on the ML models, as these models are able to directly estimate the number of migrations between two zones, without supplemental information on the number of outgoing migrants from each zone. To apply a produc- tion function, M, to the predictions made by a ML model, T, we create a new set of predictions, T(cid:48), where an entry T (cid:48) ij = M (mi) Tij(cid:80)n . Tik k=1 Results Tables 2 and 3 show the average results over all years of data of the ML models and traditional models in the USA Migra- tion and Global Migration datasets respectively. From these tables we observe that the best traditional model for the USA Migration dataset is the Extended Radiation model, beating the other traditional models in all metrics. None of the tradi- tional models are able to capture the migration dynamics in the Global Migration dataset; they all have an r2 score near 0, meaning that a model which predicts the average num- ber of migrants for every link would perform just as well. The ML models perform much better. In the case where the ML models are constrained to the same conditions as the tra- ditional models, using traditional features and a production function, the ANN model beats all of the traditional models in 5 out of the 6 measures in the USA Migration datasets, and the XGBoost model beats all the traditional models in 4 out of the 6 metrics. Similarly, the ML models considerably outperform the traditional models in the Global Migration, outperforming them in all metrics. Considering the extended feature set results, the ML models perform even better. The ANN and XGBoost models without a production function outperform the same models with a production function in 5 out of the 6 metrics. The XGBoost model outperforms the ANN model in 3 out of the 4 metrics that evaluate the mod- els' per link predictions, however the ANN model performs better on the two metrics that evaluate the aggregate incom- ing migrant prediction performance. These results suggest that more features than those which are used by the traditional models, are needed to accurately predict human migrations. The ability of ML models to in- corporate any number of additional features is one of the key motivations for using them to obtain more accurate re- sults. Considering this, it will be insightful to understand, which of the features are most informative to the ML mod- els. Since feature importance analysis for ANNs is quite challenging, we report in Table 4 the top 10 most important features for both datasets (based on information gain) in the XGBoost model trained on the extended feature set, aver- aged over all years of data. In both datasets, the intervening population feature is in the top 4 important features which validates the intuition that intervening opportunities are im- portant in predicting migrations. The most important feature in the USA Migration dataset is the number of intervening counties between two locations, a simpler form of the in- tervening population idea. In the Global Migration dataset, the population growth of the origin is the most important feature on average, with a large standard deviation. In some years this feature is very important, however in other years it is less so. Intuitively, population growth will be correlated with the amount of incoming migration. During relatively stable years, with small population growth, other features will be more predictive of migration. In Figure 1 we show the difference between the actual and predicted numbers of incoming migrants per county for the two best traditional models, and all of the ML models with- out production functions. From these maps we can see that between ML models, those trained with the extended feature set perform better than those trained with only the traditional features. Specifically, without the extended features, the ML models underpredict the number of migrations to the west- ern portion of the United States. When the extended features are taken into account, the models are able to correct for this spatial bias. Holding with the experimental results, we can see that the ANN model with extended features best captures the incoming migrant distributions per county. The ANN model is able to more accurately match the number of mi- grants that travel to rural areas (e.g. to the midwestern US), compared to the traditional models that consistently over es- timate the numbers of migrants to rural areas. In general, these maps agree with our empirical results, that the ANN model (with the lowest average incoming migrants M AE) is able to best predict migrations. 6 Conclusion With the increasing availability of high resolution socio- economic data in countries that also record human migrant flows, it is possible to use machine learning models of hu- man migration rather than traditional gravity or radiation models. Machine learning models offer greater levels of modeling flexibility, as they can combine many input fea- tures in non-linear ways that can not be captured by static equations. Furthermore, machine learning models can be easily customized to the problem or country at hand. We develop two machine learning based models for the task of predicting human migration flows, for both between counties in the US and between countries across the world. We compare these models to traditional human migration models using two sets of features and show that our models outperform the traditional models in most of the evaluation metrics. We would like to extend this work to better explain hu- man migration through a more complete analysis of features included in the model, and through incorporating different models. While the XGBoost model can provide a ranking of feature importances, this does not fully explain the dynam- ics that drive human migration. Additionally we would like to study how these migration models could be specialized to predict migrations under extreme weather events. Hurri- canes and other natural disasters can displace large popula- tions, and determining where these populations will resettle would provide an unique planning tool for policy makers. Figure 1: USA Migrations modeling error. These maps show the difference between the ground truth number of incoming migrants and predicted number of incoming migrants per county for 6 models in 2014. Blue corresponds to overestimation by the model, red to underestimation by the model, and white if the model accurately predicts the correct number of incoming migrants. Top row shows the results for the Extended Radiation model and Gravity model with power law distance decay. Middle row shows the results for ANN models trained with the extended and common feature sets. Bottom row shows the results for XGBoost models trained with the extended and common feature sets. To achieve these goals, higher resolution migration data, on both spatial and temporal scales, will need to be obtained. Extreme weather events, by definition, are short lived and their effects will be better estimated and predicted at local scales. References [Balcan et al. 2009] Balcan, D.; Colizza, V.; Gonc¸alves, B.; Hu, H.; Ramasco, J. J.; and Vespignani, A. 2009. Multi- scale mobility networks and the spatial spreading of infec- tious diseases. Proceedings of the National Academy of Sci- ences 106(51):21484–21489. [Chen and Guestrin 2016] Chen, T., and Guestrin, C. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 785–794. ACM. [Clark and Ballard 1980] Clark, G. L., and Ballard, K. P. 1980. Modeling out-migration from depressed regions: the Frean, M. R. 1999. Boosting algorithms as gradient descent. In NIPS, 512–518. [Masucci et al. 2013] Masucci, A. P.; Serras, J.; Johansson, A.; and Batty, M. 2013. Gravity versus radiation models: On the importance of scale and heterogeneity in commuting flows. Physical Review E 88(2):022812. [Noulas et al. 2012] Noulas, A.; Scellato, S.; Lambiotte, R.; Pontil, M.; and Mascolo, C. 2012. A tale of many cities: PloS one universal patterns in human urban mobility. 7(5):e37027. [ Ozden et al. 2011] Ozden, C¸ .; Parsons, C. R.; Schiff, M.; and Walmsley, T. L. 2011. Where on earth is everybody? the evolution of global bilateral migration 1960–2000. The World Bank Economic Review 25(1):12–56. [Pierce 2015] Pierce, K. 2015. SOI migration data. A new approach: Methodological improvements for SOICs United States population migration data, calendar years 2011-2012. Technical report, Statistics of Income, Internal Revenue Ser- vice. [Ren et al. 2014] Ren, Y.; Ercsey-Ravasz, M.; Wang, P.; Gonz´alez, M. C.; and Toroczkai, Z. 2014. Predicting com- muter flows in spatial networks using a radiation model based on temporal ranges. Nature communications 5. [Schneider 1959] Schneider, M. 1959. Gravity models and trip distribution theory. Papers in Regional Science 5(1):51– 56. [Simini et al. 2012] Simini, F.; Gonz´alez, M. C.; Maritan, A.; and Barab´asi, A.-L. 2012. A universal model for mobility and migration patterns. Nature 484(7392):96–100. [Sorichetta et al. 2016] Sorichetta, A.; Bird, T. J.; Ruktanon- chai, N. W.; zu Erbach-Schoenberg, E.; Pezzulo, C.; Tejedor, N.; Waldock, I. C.; Sadler, J. D.; Garcia, A. J.; Sedda, L.; et al. 2016. Mapping internal connectivity through human migration in malaria endemic countries. Scientific Data 3. [Stouffer 1940] Stouffer, S. A. 1940. Intervening opportu- nities: a theory relating mobility and distance. American sociological review 5(6):845–867. [World Bank 2017] World Bank. //databank.worldbank.org/data/reports. aspx. [Yang et al. 2014] Yang, Y.; Herrera, C.; Eagle, N.; and Gonz´alez, M. C. 2014. Limits of predictability in commut- ing flows in the absence of data for calibration. Scientific reports 4. [Zipf 1946] Zipf, G. K. 1946. The p 1 p 2/d hypothesis: on the intercity movement of persons. American sociological review 11(6):677–686. http: 2017. significance of origin and destination characteristics. Envi- ronment and Planning A 12(7):799–812. [Cohen et al. 2008] Cohen, J. E.; Roig, M.; Reuman, D. C.; and GoGwilt, C. 2008. International migration beyond gravity: A statistical model for use in population projec- tions. Proceedings of the National Academy of Sciences 105(40):15269–15274. [De Montis et al. 2010] De Montis, A.; Chessa, A.; Cam- pagna, M.; Caschili, S.; and Deplano, G. 2010. Modeling commuting systems through a complex network analysis: A study of the italian islands of sardinia and sicily. Journal of Transport and Land Use 2(3). [Dennett and Wilson 2013] Dennett, A., and Wilson, A. 2013. A multilevel spatial interaction modelling framework for estimating interregional migration in europe. Environ- ment and Planning A 45(6):1491–1507. [Dinkelman and Mariotti 2016] Dinkelman, T., and Mariotti, M. 2016. The long-run effects of labor migration on hu- man capital formation in communities of origin. American Economic Journal: Applied Economics 8(4):1–35. [Fagiolo and Mastrorillo 2013] Fagiolo, G., and Mastrorillo, M. 2013. International migration network: Topology and modeling. Physical Review E 88(1):012812. [Faith, Minchin, and Belbin 1987] Faith, D. P.; Minchin, P. R.; and Belbin, L. 1987. Compositional dissimilarity as a robust measure of ecological distance. Vegetatio 69(1-3):57– 68. [Friedman 2001] Friedman, J. H. 2001. Greedy function ap- proximation: a gradient boosting machine. Annals of statis- tics 1189–1232. [Greenwood 1985] Greenwood, M. J. 1985. Human migra- tion: Theory, models, and empirical studies*. Journal of re- gional Science 25(4):521–544. [Guo et al. 2008] Guo, X.; Yin, Y.; Dong, C.; Yang, G.; and Zhou, G. 2008. On the class imbalance problem. In Natural Computation, 2008. ICNC'08. Fourth International Confer- ence on, volume 4, 192–201. IEEE. [LeCun, Bengio, and Hinton 2015] LeCun, Y.; Bengio, Y.; Nature and Hinton, G. 521(7553):436–444. [Lee 1966] Lee, E. S. 1966. A theory of migration. Demog- raphy 3(1):47–57. [Legendre and Legendre 2012] Legendre, P., and Legendre, L. F. 2012. Numerical ecology, volume 24. Elsevier. [Lenormand, Bassolas, and Ramasco 2016] Lenormand, M.; Bassolas, A.; and Ramasco, J. J. 2016. Systematic compar- ison of trip distribution laws and models. Journal of Trans- port Geography 51:158–169. [Lenormand et al. 2012] Lenormand, M.; Huet, S.; Gargiulo, F.; and Deffuant, G. 2012. A universal model of commuting networks. PloS one 7(10):e45985. [Letouz´e et al. 2009] Letouz´e, E.; Purser, M.; Rodr´ıguez, F.; Cummins, M.; et al. 2009. Revisiting the migration- development nexus: a gravity model approach. Human De- velopment Research Paper 44. [Mason et al. 1999] Mason, L.; Baxter, J.; Bartlett, P. L.; and Deep learning. 2015.
1604.00175	2	1604	2017-09-04T11:23:10	Quantifying Location Sociality	[ "cs.SI", "physics.soc-ph" ]	The emergence of location-based social networks provides an unprecedented chance to study the interaction between human mobility and social relations. This work is a step towards quantifying whether a location is suitable for conducting social activities, and the notion is named location sociality. Being able to quantify location sociality creates practical opportunities such as urban planning and location recommendation. To quantify a location's sociality, we propose a mixture model of HITS and PageRank on a heterogeneous network linking users and locations. By exploiting millions of check-in data generated by Instagram users in New York and Los Angeles, we investigate the relation between location sociality and several location properties, including location categories, rating and popularity. We further perform two case studies, i.e., friendship prediction and location recommendation, experimental results demonstrate the usefulness of our quantification.	cs.SI	cs	Quantifying Location Sociality Jun Pang FSTC & SnT University of Luxembourg [email protected] Yang Zhang CISPA, Saarland University Saarland Informatics Campus [email protected] 7 1 0 2 p e S 4 ] I S . s c [ 2 v 5 7 1 0 0 . 4 0 6 1 : v i X r a ABSTRACT The emergence of location-based social networks provides an un- precedented chance to study the interaction between human mo- bility and social relations. This work is a step towards quantifying whether a location is suitable for conducting social activities, and the notion is named location sociality. Being able to quantify loca- tion sociality creates practical opportunities such as urban planning and location recommendation. To quantify a location's sociality, we propose a mixture model of HITS and PageRank on a heteroge- neous network linking users and locations. By exploiting millions of check-in data generated by Instagram users in New York and Los Angeles, we investigate the relation between location sociality and several location properties, including location categories, rating and popularity. We further perform two case studies, i.e., friend- ship prediction and location recommendation, experimental results demonstrate the usefulness of our quantification. CCS CONCEPTS • Human-centered computing → Social networking sites; Ubiq- uitous computing; KEYWORDS Online social networks; location-based social networks; data min- ing; friendship prediction; location recommendation 1 INTRODUCTION Online social networks (OSNs) have been the most successful web applications during the past decade. Leading companies, including Facebook,1 Twitter2 and Instagram,3 have gained a large number of users. More recently, with the development of positioning technol- ogy on mobile devices, OSNs have been extended to geographical space. Nowadays, it is quite common for OSN users to share their geographical locations, i.e., check-ins. Moreover, a special type of OSNs dedicated to location sharing are created, namely location- based social networks (LBSNs). Foursquare and Yelp are two repre- sentative companies. With the emergence of LBSNs, a large quantity of data concerning human mobility become available. This gives us 1https://www.facebook.com/ 2https://twitter.com/ 3https://www.instagram.com/ Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. HT'17, July 4-7, 2017, Prague, Czech Republic. © 2017 ACM. 978-1-4503-4708-2/17/07...$15.00 DOI: http://dx.doi.org/10.1145/3078714.3078729 an unprecedented opportunity to understand human mobility and moreover to study the interaction between social relations and mo- bility. Some previous works have been focused on inferring social relationships from mobility, such as [13, 28, 32, 37, 41], others ex- ploit users' social information to predict their future locations, such as [1, 3]. More recently, researchers propose new understandings of locations by using user generated data such as happiness [30] and walkability [29]. Location has been recognized as an important factor for social activities back in 1950s. In his seminal work [10], Erving Goffman described social interactions as a series of performance given by social actors, and physical setting, i.e., location, is an important aspect of a social actor's performance. In [10], Goffman stated that "A setting tends to stay put, geographically speaking, so that those who would use a particular setting as a part of their performance cannot begin their act until they have brought themselves to the appropriate place". Based on Goffmann's study, Milligan [21] further proposed that "physical sites (however defined by the participants) become the stages for social interaction, stages that are both physically and socially constructed". She explained that not only being physically constructed (by architects, facility managers, property owners and others), a location will also be socially constructed by people who conduct social interactions there. Following this theory, we argue that social construction will make some locations more suitable for social activities than others. In the current work, we aim to quantify whether a location is a suitable for conducting social activities. The notion we quantify is named location sociality. We define a location's sociality as the degree to which individuals tend to conduct social activities at that location. A location is considered social if friends frequently visit, especially for the purpose of socializing or recreation, and vice versa. Studying location sociality could advance the boundary of our understanding on the interaction between social relations and mobility. It can also help us to solve challenging problems such as urban planning and traffic control. In practice, location sociality can be also used to build appealing applications such as location recommendation. Contributions. In the current work, we make the following con- tributions: • We propose a framework to quantify location sociality (Sec- tion 2). Our framework is based on the assumption that a location's sociality and its visitors' social influence are mu- tually reinforced. To model this assumption, we construct a heterogeneous network consisting of users (in a social network) and locations (a user-location network). Then, we propose a mixture model of HITS [16] and PageRank to quantify location sociality on this heterogeneous network. • Following our solution, we exploit millions of check-in data from Instagram in New York and Los Angeles to quantify location sociality (Section 3). We then study the relation between location sociality and several location proper- ties including location categories, rating and popularity. Our discoveries include: certain types of locations (music venues and nightclubs) are more social than others; loca- tion sociality shares a positive relation with location rating given by users; social locations distribute more uniformly w.r.t. geographical space than popular locations. • To demonstrate the usefulness of our quantification of loca- tion sociality, we perform a case study on friendship predic- tion in Section 4.1. We extract two users' common locations and define features based on these common locations' so- ciality for machine learning classification. Experimental results show that with very simple location sociality fea- tures, we are able to achieve a strong prediction. Moreover, adding location sociality into a state-of-the-art prediction model achieves a 5% performance gain. • We perform another case study on using location sociality for location recommendation in Section 4.2. We integrate our quantification into a random walk with restart frame- work. Experimental results show that the recommender based on location sociality achieves a better recommen- dation performance (at least 5%) than the baseline recom- mender that does not consider location sociality. We discuss some implications and limitations of the current work in Section 5. Related works are discussed in Section 6 and Section 7 concludes the paper. 2 PROPOSED SOLUTION In this section, we first discuss the intuition of our solution on quan- tifying location sociality (Section 2.1), then we formally describe the solution (Section 2.2). 2.1 Intuition Our intuition on quantifying location sociality in this paper is based on the assumption that a location's sociality and its visitors' social influence are mutually reinforced. To explain this intuition, we start by addressing socially influential users. In the society, if a person is considered socially influential, he must visit different social places frequently to organize or participate in different social activities and events. On the other hand, if a location is frequently visited by influential users, then it must be suitable for conducting social activities, i.e., it is a social place. Following this, we establish a mutual reinforcement relation between user influence and location sociality, i.e., more social a location is, more socially influential users visit it, and vice versa. In addition to visiting many social places, an influential user should also occupy an important position in the social network, e.g., he should have many friends who are also socially influential. Following the above discussion, our intuition on quantifying location sociality can be summarized as the following two assumptions. Assumption 1. Location sociality and users' social influence are mutually reinforced. Figure 1: A model of the network. Assumption 2. Users' social influence can be quantified from the social network. The first intuition can be naturally formulated into a HITS-style framework [16]. For the second intuition, we apply PageRank on the social graph to quantify each user's influence. In the rest of the paper, we use location sociality and sociality interchangeably. 2.2 Our framework We start by modeling users, locations and their relationships into two types of networks including social network and user-location network. Social network. A social network, denoted as G𝒰 = (𝒰,ℰ𝒰), is an unweighted graph with nodes in set 𝒰 representing all users. ℰ𝒰 ⊆ 𝒰 × 𝒰 is a symmetric relation containing the edges in G𝒰. If ui and uj are friends, then (ui , uj) ∈ ℰ𝒰 and (uj , ui) ∈ ℰ𝒰. We use matrix X to represent G𝒰 where Xi, j = 1 if (ui , uj) ∈ ℰ𝒰 and Xi, j = 0 otherwise. It is easy to see that X is symmetric. We X i, j = Xi, j further use X to denote the column stochastic matrix of X where User-location network. A user-location network, denoted as G𝒰,ℒ = (𝒰,ℒ,ℰ𝒰,ℒ), is a weighted bipartite graph. ℰ𝒰,ℒ ⊆ 𝒰×ℒ consists of the edges in G𝒰,ℒ. Each edge (ui , ℓj) ∈ ℰ𝒰,ℒ, also written as 𝒰,ℒ defined as the number e i, j of times that the user ui has visited (checked in) the location ℓj (denoted by ci(ui , ℓj)). We use matrix Y to represent G𝒰,ℒ with . The transpose of Y is further denoted by YT . In the Yi, j = w end, we use Y and YT to denote the column stochastic matrices of Y and YT , respectively. , is associated with a weight w Figure 1 shows an example of the heterogeneous graph. Within our framework two sets of values, locations' sociality and users' social influence, can be obtained. Each location ℓ's sociality is de- fined as κ(ℓ) and η(u) for each user's social influence. Following the 𝒰,ℒ i, j . k Xk, j 𝒰,ℒ i, j `1`2`3`4u1u2u3u4η(ui) = η(ui) = κ(ℓj) = j j i X i, j · η(uj) Y i, j · κ(ℓj) YT j,i · η(ui) (1) (2) (3) intuition in Section 2.1, our model is formulated into the following equations: Equations 1 is the PageRank implementation for quantifying users' social influence from G𝒰. Equations 2 and 3 are an instance of the HITS framework which establishes the mutual reinforcement relationship between locations and users. We then linearly combine the above equations as η(ui) = α · X i, j ·η(uj) +(1−α)· κ(ℓj) = j YT j,i · η(ui) j Y i, j ·κ(ℓj) (4) (5) i where α specifies the contributions of each component to users' social influence. In our experiments, α is set to 0.5 which indicates the social network structure and user mobility are equally important on quantifying users' social influence. Note that α = 0.5 is a typical setting in many fields such as [36] where the authors aim to discover salient sentences for document summarization. We further use two vectors η and κ to denote users' social in- fluence and locations' sociality. Then the above equations can be written into the following matrix form. η = α · X · η + (1 − α) · Y · κ κ = YT · η (6) (7) Equations 6 and 7 can be computed through an iterative updating process. We set all locations' (users') initial sociality (social influ- ence) to be 1ℒ ( 1𝒰 ). According to our experiments, the computa- tion stops after around 10 iterations, when the maximal difference between κs of two consecutive iterations is less than 0.00001. 3 EXPERIMENTS In this section, we first introduce the dataset used for our experi- ments. Then, we present the results of our quantification: we start by discussing the top social locations and location categories; then we focus on the relation between location sociality and location rating; in the end, the correlation between location sociality and popularity is discussed. 3.1 Dataset description Instagram is a photo-sharing social network with a fast growing user number. By now, it has 400M monthly active users and with 75M photos published everyday. Similar to other social network services such as Facebook and Twitter, Instagram allows users to share their locations when publishing photos. Moreover, unlike Twitter where only a small amount of tweets are geo-tagged, the authors of [19] have shown that Instagram users are much more willing to share their locations (31 times more than Twitter users), which makes Instagram a suitable platform to study the interaction between mobility and social relations. Figure 2: Check-ins in New York. # check-ins # active users # edges (active users) # locations New York 6,181,169 12,280 74,230 8,683 Los Angeles 4,705,079 8,643 44,994 6,908 Table 1: Dataset summary. We collect the geo-tagged photos, i.e., check-ins, in New York and Los Angeles from Instagram through its public API4. Since locations' category information is an important aspect of our anal- ysis, and fortunately the API of Instagram is linked with the API of Foursquare, a leading location-based social network with resource- ful information about each place, thus we exploit the following methodology to collect our data. We first resort to Foursquare to extract all location ids within each city, meanwhile we collect each location's category information together with its rating (number of tips and number of likes). Then for each Foursquare's location id, we query Instagram's API to get its corresponding location id in Instagram. After this, we query each location's recent check- ins in Instagram several times a day from August 1st, 2015 until March 15th, 2016. In the end, more than 6M check-ins have been collected in New York and 4.7M in Los Angeles5. To resolve the data sparseness issue, we focus on users with at least 20 check-ins (considered as active users) and locations with at least 10 check-ins. Figure 2 depicts a sample check-in distribution in New York. Since Foursquare organizes location categories into a tree structure6, we take its second level categories to label each location. To obtain users' social networks, we exploit Instagram's API to query each active user's follower/followee list7. We consider two users as friends if they mutually follow each other in Instagram. To further guarantee that users we have collected are not celebrities or business accounts, we filter out the top 5% of users with most followers. Also, only the relations among active users (users with 4https://www.instagram.com/developer/ 5It is worth noticing that the authors of [20] has applied a similar methodology. 6https://developer.foursquare.com/categorytree 7Since Instagram's API only provides one page with 50 follower/followees per query, we perform multiple queries until all follower/followees of each user are obtained. New York Top Unsocial Staples Top Social Webster Hall Madison Square Park Rockwood Music Hall Washington Square Park Manhattan Theatre Club 17 Frost Gallery China Institute Baby's All Right El Rey Del Taco II (a) New York (b) Los Angeles Los Angeles Top Social The Fonda Theatre Avalon Hollywood The Echo Hermosa Beach Pier Exchange LA Top Unsocial Panda Express Gap Ebar Palms Super Market 7-Eleven Table 2: The most and least social locations in New York and Los Angeles. New York Social Categories Unsocial Categories Music Venue Nightclub Harbor Museum Park Laundry Service Convenience Store Post Office Pharmacy Fast Food Restaurant Figure 3: Distributions of log-transformed location sociality. Los Angeles at least 20 check-ins) are kept. In the end, the social network con- tains 74,230 edges for New York and 44,994 edges for Los Angeles. Table 1 summarizes the dataset. For the sake of experimental result reproducibility, our dataset is available upon request. 3.2 Location sociality vs. location category Figure 3 depicts the log transformed distributions of location social- ity, both of which indicate that most locations have a middle value of sociality while only a few locations are very social or unsocial. This is different from other location measurement, for instance, the number of mobility transitions from or to each location follows a power law distribution [24]. Table 2 presents the top five locations with highest and lowest sociality. For New York, Webster Hall (music venue) is the most social place followed by Madison Square Park (park). On the other hand, the least social place is one Staples store (convenience store) in midtown. For Los Angeles, The Fonda Theatre (concert hall) has the highest sociality. Meanwhile, one Panda Express (fast food restaurant) is the least social place. From Table 2, we can see a clear distinction between social and unsocial places w.r.t. their categories. Next, we take a deeper look at the relation between sociality and location category. Table 3 lists the top five location categories with the highest and lowest average location sociality. Nightclub and music venue are Social Categories Concert Hall Nightclub Music Venue Mall Beach Unsocial Categories Convenience Store Vintage Store Fast Food Restaurant Pet Service Automotive Shop Table 3: Top 5 location categories with highest and lowest average location sociality in New York and Los Angeles. in the top 3 in both cities. On the other hand, convenience store seems to be less attractive to friends. Besides, we also observe some interesting difference between the two cities. For example, beach is the No.5 social choice for people living in Los Angeles while it is not New Yorkers' choice since there are no beaches in Manhattan. As music venue and nightclub have high rankings in both cities, we further list the top 5 music venues and nightclubs in Table 4. Although the ranking of music venues and nightclubs are rather subjective, we have checked several blogs and articles (listed in the additional material) and most of our top-social nightclubs and music venues have received positive reviews and been recommended by these blogs and articles. We conclude that a location's sociality is related to its category. Normally, location category itself is not sufficient to judge whether a location is social or not. Next, we study location socialtiy (log)-10-8-6probability00.010.020.030.040.050.06location socialtiy (log)-10-9-8-7-6probability00.020.040.060.08New York Social Music Venues Social Nightclubs Webster Hall Rockwood Music Hall Baby's All Right Bowery Ballroom Music Hall of Williamsburg Stage 48 Marquee Pacha NYC 1 OAK VIP Room NYC Los Angeles Social Music Venus Avalon Hollywood The Echo The Roxy The Troubadour The Hollywood Bowl Social Nightclubs Exchange LA OHM Nightclub Sound Nightclub Club Los Globos Create Nightclubs (a) New York. Table 4: Top 5 music venues and nightclubs with highest lo- cation sociality in New York and Los Angeles. other properties of locations and their relationship with location sociality. 3.3 Location sociality vs. rating, tips and likes For each location, Foursquare provides us with not only its category information, but also other properties including rating8, number of tips and number of likes generated by Foursquare users. Next, we study whether it is possible to use these properties to explain location sociality. To proceed, we build a linear regression model with rating, number of tips and number of likes as explanatory variables while location sociality as the dependent variable. By fitting the model with ordinary least square method, we obtain a coefficient of determination (R2) of 0.192 in New York and 0.280 in Los Angeles, meaning that 19.2% (28.0%) of the variability of location sociality in New York (Los Angeles) can be explained by these properties. By checking the parameters of our linear model, we discover that the major predictive power is driven by location rating. We further plot the average location sociality as a function of rating in Figure 4: the two variables share a positive relation. Espe- cially when location rating is high (≥ 8), location sociality increases sharply for both cities. This indicates that social places are assigned with high ratings by users. 3.4 Location sociality vs. location popularity A social location is often popular in the sense that it attracts many people. On the other hand, to conduct social activities, everyone has his own preference on choosing locations. For example, one may prefer to go to a bar near his home, which might not be well-known at the city level. The relationship between a location's sociality and its popularity is worth investigation: a location's sociality should be correlated with its popularity, while the two notions should exhibit difference. 8In Foursquare, rating is in the range from 1 to 10. (b) Los Angeles. Figure 4: Location sociality vs. location rating. le(ℓ) = − ci(u, ℓ) ci(ℓ) log ci(u, ℓ) ci(ℓ) , By far, the most common notion for quantifying a location's popularity is location entropy [4], it is formally defined as where ci(u, ℓ) is user u's number of check-ins at location ℓ (Sec- tion 2) and ci(ℓ) is the total number of check-ins of location ℓ. More popular a location is, higher location entropy it has. To check the difference between the two measurements, we plot the heatmaps w.r.t. location entropy and sociality of New York. As expected, midtown and downtown New York are "hot" areas in both maps. On the other hand, we observe that location sociality is more uniformly distributed than location entropy. For example, the areas marked by green circles in Figure 5a are obviously lighter than those in Figure 5b. After having a close look, we discover that bars and restaurants are the "hot" locations inside these areas. Data in Los Angeles exhibits a similar result and is not shown. We further extract the top 20 popular locations in both cities (listed in the additional material). In New York, the most popu- lar locations are parks and museums (e.g., The MET, MoMA and Guggenheim). On the other hand, in Los Angeles, the most popular locations concentrate on malls followed by museums. Moreover, in both cities, famous landmarks have high location entropy, such as Rockefeller Center in New York and Hollywood Walk of Fame in Los Angeles. This is quite different from the ranking in Table 2 and Table 3: social locations are mainly music venues and nightclubs rating46810location sociality×10-40123rating46810location sociality×10-4012345(a) Location entropy (a) New York (b) Location sociality Figure 5: Heatmaps in New York. while popular locations are mainly tourist attractions. In the end, we conclude that there exists a large difference between social and popular locations. 4 CASE STUDIES In this section, we perform two case studies to show how location sociality can be used to build real world applications. The appli- cations we focus on are friendship prediction and new location recommendation, both of which are essential for social network services. 4.1 Friendship prediction Following the seminal work of Liben-Nowell and Kleinberg, friend- ship prediction has been extensively studied [17], resulting in ap- pealing applications such as friendship recommendation which is essential for OSNs to increase user engagement. During the past five years, with the development of LBSNs, many researchers start to exploit users' location data as a new source of information for friendship prediction. Model. In our case study, we consider friendship prediction as a binary classification problem. Each pair of friends is treated positive if they are friends (mutually following each other in Instagram) and negative otherwise. We extract the common locations of two users and construct the feature space based on these common locations. Here, two users' common locations are the intersection of the places they have checked in, regardless of time. For two users ui and uj, (b) Los Angeles Figure 6: Evaluation results on our model w.r.t. four classifi- cation algorithms. we find their common locations' sociality and utilize the average, maximal, minimal and standard deviation of theses sociality as features for classification. Experiment setup. To resolve the data sparseness issue, we filter out pairs of users who have only one or zero common location. This leaves us 7,525 pairs of friends, i.e., positive cases, in New York and 3,961 pairs of friends in Los Angeles. For negative case, we randomly sample the same number of non-friend pairs -- each pair of them has at least two common places as well. It is worth noticing that this way of sampling negative cases increases the hardness of classification since a non-friend pair also has at least two common locations. Therefore, we can further evaluate the usefulness of location sociality. We have adopted four classification algorithms in our experiments including logistic regression, gradient boosting, AdaBoost and random forest. Accuracy, F1score and AUC (area under the ROC curve) are used as our metrics. We randomly split the dataset with 70% for training and 30% for testing, this random split is repeated for 10 times and we report the average result. Results. Figure 6 depicts the performance of our classifications. Among all the classifiers, random forest performs the best with AUC = 0.82 and Accuracy = 0.77 in the two cities. Meanwhile, we have F1score = 0.82 in New York and F1score = 0.83 in Los Angeles. AUCAccuracyF1score00.20.40.60.81Logitstic RegressionGradient BoostingAdaBoostRandom ForestLogistic RegressionAUCAccuracyF1score00.20.40.60.81Logitstic RegressionGradient BoostingAdaBoostRandom ForestLogistic RegressionNew York [32] sociality+[32] Los Angeles [32] sociality+[32] AUC Accuracy 0.77 0.83 0.87 0.81 F1score 0.82 0.85 AUC Accuracy 0.77 0.82 0.86 0.80 F1score 0.82 0.85 and location 5: Evaluation result on [32] Table sociality+[32] . (a) New York Indeed, there exist other solutions for friendship prediction such as considering two users' meeting events [13, 28, 37]. However, the main issue for this method is that meeting events (reflected in OSNs) are rare, for example, with 6M check-ins in New York, we only observe around 100 meeting events. Moreover, we want to emphasize that friendship prediction is not the focus of the current work. Therefore, we choose the most straightforward features for training our classifiers. Nevertheless, our prediction still achieves a strong performance showing that location sociality is a good indica- tor for recommending friends. Further investigation on integrating location sociality into the state-of-the-art friendship prediction models is worth studying and we leave it as a future work. (b) Los Angeles Figure 7: ROC curves for our classification and the number of common locations. AdaBoost and gradient boosting have a comparable performance. On the other hand, logistic regression performs the worst. The number of common locations is further adopted as a naive baseline model for comparison, i.e., we tune the threshold (the number of common locations) for classification to obtain the ROC curve. Figure 7 plots the results. As we can see, all our classifiers based on location sociality outperform this naive baseline. Next, we check whether adding location sociality into a state- of-the-art model as proposed in [32] 9 can increase prediction per- formance. The model in [32] (location feature setting) extracts two users' common locations and design features mainly with these common locations' entropies (see Section 3.4 for location entropy), such as the minimal location entropy. In our experiments, we com- bine our four location sociality features with the features in [32] and fit them into our best performing classifier random forest. The results in Table 5 show that the classification with location social- ity improves [32] by around 5% among all three metrics in both New York and Los Angeles. This further demonstrates that location sociality is useful for friendship prediction. 9Most of the recent solutions focus on two users' meeting events i.e., two users at the same location at roughly the same time, while [32] and our model only exploit common locations. 4.2 Location recommendation The second case study we perform is recommending new locations for users to visit. Location recommendation has a great potential to build appealing applications. During the past five years, it has attracted academia a lot of attention (e.g. [2, 6, 7, 42]). Our goal here is to demonstrate the usefulness of location sociality in recom- mending new locations. In order to integrate location sociality into a location recommender, we adopt a classical approach, namely random walk with restart [34]. Model. In a typical setting of random walk with restart for recom- mendation, in the beginning we define a matrix Q as (cid:18) (cid:19) Q = 0 YT Y 0 where Y and YT represent user-location network (location-user network) (Section 2). Meanwhile, ¯Q denotes the column stochastic version of Q [39]. Then to recommend locations to a user ui, we modify ¯Q to allow every node in the graph having a certain proba- bility (15% in the experiments) to jump to the node representing ui. Formally, for every Qa,b ∈ Q, ¯Qa,b is defined as + c · 1 if the ath row represents ui otherwise  (1 − c) · Qa,b (1 − c) · Qa,b j Qj,b j Qj,b ¯Qa,b = where c = 0.15. By applying the same method of solving PageRank, e.g., power method, we can obtain the steady state distribution over ¯Q, which is the relevance score of all nodes (both locations and users) to ui. Locations with high relevance scores are recommended to ui. Noulas et al. [23] have exploited this approach for location FPR00.20.40.60.81TPR00.20.40.60.81Common Location CountLogitstic RegressionGradient BoostingAdaBoostRandom ForestLogistic RegressionFPR00.20.40.60.81TPR00.20.40.60.81Common Location CountLogitstic RegressionGradient BoostingAdaBoostRandom ForestLogistic Regression(cid:18) 0 𝒯 (cid:19) T 0 recommendation10, where the weight on an edge between a user ui and a location ℓj is simply the user's number of visits to that location, i.e., Yi, j = w = ci(ui , ℓj) in Section 2. To integrate location sociality into the edge weight for location 𝒰,ℒ i, j recommendation, we change Y to T , i.e., Q is modified to: Q = where Ti, j is defined as 1 Ti, j = ci(ui , ℓj) · − log(κ(ℓj)) . (8) Here, κ(ℓj) is the location sociality of ℓj, meanwhile 𝒯 is the trans- pose of T . Under this formulation, Equation 8 assigns higher weight to locations with high sociality, which will bias the recommended locations to be more social. In the end, by performing power method on the column stochastic version of the modified Q, we obtain the recommended locations for each user. Experiment setup. The check-in dataset is partitioned temporally with each one covers consecutively 60 days [23]. For each partition, we use the data of the first 30 days to train the model while the left 30 days for testing. Since our aim is to perform new location recommendation, for each user we further filter out his locations in the testing set that he has already been to in the training set. In the end, we perform random walk with restart with location sociality (rwr-ls) to recommend locations for each user, and exploit rwr without location sociality (rwr), i.e., the one in [23], as the baseline model. Two metrics including precision@10 and recall@10 are adopted for evaluation. Results. Table 6 presents the results for location recommendation in both cities. As we can see, rwr-ls outperforms rwr in all months. For precision@10, rwr-ls outperforms rwr by 10%, while for re- call@10, even in the worst case in New York, rwr-ls still has 3.4% improvement on rwr. Even though the absolute precision and re- call of our recommendation is not high, it is worth noticing that the similar performances of location recommendation have been obtained by [7, 18, 38], thus our results are reasonable. Similar to Gao et al. [7], we emphasize that the focus here is to compare the relative performance, in order to demonstrate the usefulness of our quantification. Many state-of-the-art algorithms exploit other factors for lo- cation recommendation such as geographical distance and users' published contents, one of our future works is to integrate location sociality into these algorithms to further improve recommendation. 5 IMPLICATIONS AND LIMITATIONS We discuss implications and limitations of the current work. Implications. Location sociality as a measurement can character- ize a social map for a city. The applications based on it, including the friendship prediction and location recommendation addressed in Section 4, can benefit several parties. For city administrator, understanding where people prefer to socialize can help them make better city plans. For example, during conventional social time such as Friday nights, the city government 10They [23] also consider social network in Q, here we ignore it for better demonstrat- ing location sociality's usefulness. could make specific transportation plans, such as more buses and taxis, or security plans such as deploying more policemen, for areas with high sociality. Moreover, location sociality can be a good reference for the government to plan future city development. For city residents, location sociality provides a good reference for them to find or discover new places to socialize with their friends. In addition, location sociality can be used as an important factor for location recommendation services for social network services, such as Yelp and Foursquare. For visitors, visiting high sociality places is a good way to engage local people's social life. This can help visitors understand the city's culture in a better way. For business owners, knowing where people like to go to conduct social activities is an important factor for them to determine where to open new business. Limitations. We point out the following limitations of the current work. First, we only focus on the data from Instagram which cannot reflect the general population both socially and geographically. Socially, the authors of [22, 33] have shown that most of Instagram users are center around a younger age (25 yeas old); geographically, most of the check-ins concentrate on the city center (Figure 2). Second, our quantification does not take into account the tem- poral factor which could also be important for understanding loca- tions. One approach for considering time would be concentrating on users' meeting events, i.e., users check in at the same location at the same time, however, as we have discussed previously, meeting events are rare even in large datasets as ours. In the future, we plan to incorporate data with rich temporal information from other sources to further improve our quantification. Third, we quantify a location's sociality based on its visitors' information. In some cases, a location's own property can also contribute to its sociality. For example, it is pointed out that the decoration, the space allocation and even the bicycle parking design can result in different number of visitors to a cafe located in a university campus [21]. However, we argue that data of this kind is hard and expensive to obtain at a large scale. 6 RELATED WORK The emergence of LBSNs has brought us an unprecedented op- portunity to study human mobility and its interaction with social relationships. Many works have been done on understanding hu- man mobility and its interaction with social relations. There mainly exist two research directions: one is to exploit users' location infor- mation to understand social relations, e.g., see [13, 28, 32, 37, 41]; the other is to use social relations to understand mobility and locations, including the current work. Backstrom et al. [1] present one of the pioneer works on friendship- based location prediction. They analyze Facebook users' home lo- cation and discover that friends tend to live closer to each other than strangers. Then they build a maximal likelihood estimator to predict a user's home location. They have shown that their model outperforms significantly the method based on IP addresses. Cho et al. [3] study a general problem: instead of predicting home lo- cation, they aim to predict where a user is at a certain time. They construct a dynamic Gaussian mixture model with the assumption that each user's mobility is centered around two states, such as home and work. The experimental results show that their model 15.8-15.10 rwr rwr-ls 15.9-2015.11 rwr rwr-ls 15.10-15.12 rwr rwr-ls 15.8-15.10 rwr rwr-ls 15.9-2015.11 rwr rwr-ls 15.10-15.12 rwr rwr-ls Precision@10 0.009 0.010 Precision@10 0.010 0.011 Precision@10 0.009 0.010 Precision@10 0.010 0.011 Precision@10 0.013 0.015 Precision@10 0.010 0.012 New York Recall@10 0.021 0.024 Recall@10 0.032 0.034 Recall@10 0.028 0.029 15.11-16.1 rwr rwr-ls 15.12-16.2 rwr rwr-ls 16.1-16.3 rwr rwr-ls Los Angeles Recall@10 0.028 0.030 Recall@10 0.038 0.042 Recall@10 0.025 0.029 15.11-16.1 rwr rwr-ls 15.12-16.2 rwr rwr-ls 16.1-16.3 rwr rwr-ls Precision@10 0.009 0.010 Precision@10 0.009 0.010 Precision@10 0.008 0.009 Precision@10 0.008 0.009 Precision@10 0.012 0.013 Precision@10 0.007 0.009 Recall@10 0.028 0.031 Recall@10 0.026 0.028 Recall@10 0.027 0.029 Recall@10 0.024 0.026 Recall@10 0.047 0.048 Recall@10 0.035 0.044 Table 6: Precision@10 and recall@10 for location recommendation. achieves a promising accuracy. Other works include [8, 14, 26, 27]. More recently, Jurgens et al. [15] perform a comprehensive study on most of the existing works in the field and points out some future directions. Besides predicting a user's location, researchers begin to advance our understandings of locations based on the data from social net- works. In [30], the authors focus on recommending pleasant paths between two locations in a city. Unlike the traditional shortest path recommendation, they assign three values to describe whether a street is quiet, beautiful and happy, respectively. Then they adjust the path recommendation algorithm with these factors and recom- mend the most pleasant path for users. In [29], the authors quantify whether a street is suitable for walk, namely walkability. To as- sess their results, they propose to use concurrent validity. Their discoveries, to mention a few, include walkable streets tend to be tagged with walk-related words on Flickr and can be identified by location types on those streets. The authors of [5] exploit the data from Foursquare to analyze different neighborhoods in a city. They extract some signature features to profile each neighborhood and propose an algorithm to match similar neighborhoods across dif- ferent cities. Experimental results show that they are able to match tourists areas across Paris and Barcelona, and expensive residential areas in Washington D.C. and New York. More recently, Hristova et al. [11] propose four location measurements, including brokerage, serendipity, entropy and homogeneity, under a heterogenous social and location network model. Their experiments are conducted with a Foursquare dataset collected in London, and the authors show that their proposed measurements can be used to describe dynam- ics that is hard to capture including gentrification and deprivation. Other recent works include [9, 12, 25, 31]. The current work also falls into the field of urban informat- ics [43], a newly emerging field where researchers tend to use the ubiquitous data to understand and improve the city where we live. Besides the research literature, several open projects have been established as well. To mention a few examples, Yuan et al. [40] focus on discovering the function of each region in a city; Venerandi et al. [35] measure the socio-economic deprivation of a city. An- other excellent example is the goodcitylife project11, where the team members try to imitate human beings' five senses on food to understand cities. 7 CONCLUSION In this paper, we have proposed a new notion namely location sociality to describe whether a location is suitable for conducting social activities. We constructed a heterogenous network linking locations and users and proposed a mixture model of HITS and PageRank to quantify location sociality. Experimental results on millions of Instagram check-in data validate location sociality with some in-depth discoveries. Two case studies including friendship prediction and location recommendation demonstrate the useful- ness of our quantification. 11http://goodcitylife.org/ Location data do not only come from LBSNs, but many other sources, such as GPS traces and WIFI points. In the future, we are interested in establishing more connections between LBSN data and other sources to gain a deep understanding of cities. [2] REFERENCES [1] Lars Backstrom, Eric Sun, and Cameron Marlow. 2010. Find me if you can: improving geographical prediction with social and spatial proximity. In Proc. 19th International Conference on World Wide Web (WWW). ACM, 61 -- 70. Jie Bao, Yu Zheng, David Wilkie, and Mohamed Mokbel. 2015. Recommendations in location-based social networks: a survey. GeoInformatica 19, 3 (2015), 525 -- 565. [3] Eunjoon Cho, Seth A. Myers, and Jure Leskovec. 2011. Friendship and mobility: user movement in location-based social networks. In Proc. 17th ACM Conference on Knowledge Discovery and Data Mining (KDD). ACM, 1082 -- 1090. Justin Cranshaw, Eran Toch, Jason Hone, Ankiet Kittur, and Norma Sadeh. 2010. Bridging the gap between physical location and online social networks. In Proc. 12th ACM International Conference on Ubiquitous Computing (UbiComp). ACM, 119 -- 128. [5] Geraud Le Falher, Aristides Gionis, and Michael Mathioudakis. 2015. Where Is the soho of Rome? Measures and algorithms for finding similar neighborhoods in cities. In Proc. 9th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 228 -- 237. [4] [6] Huiji Gao, Jiliang Tang, Xia Hu, and Huan Liu. 2013. Exploring temporal effects for location recommendation on location-based social networks. In Proc. 7th ACM Conference on Recommender Systems (RecSys). ACM, 93 -- 100. [7] Huiji Gao, Jiliang Tang, Xia Hu, and Huan Liu. 2015. Content-Aware Point of Interest Recommendation on Location-Based Social Networks. In Proc. 29th AAAI Conference on Artificial Intelligence (AAAI). The AAAI Press, 1721 -- 1727. [8] Huiji Gao, Jiliang Tang, and Huan Liu. 2012. Exploring social-historical ties on location-based social networks. In Proc. 6th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 114 -- 121. [9] Petko Georgiev, Anastasios Noulas, and Cecilia Mascolo. 2014. Where businesses thrive: predicting the impact of the Olympic games on local retailers through location-based services data. In Proc. 8th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 151 -- 160. [10] Erving Goffman. 1959. The Presentation of Self in Everyday Life. Random House. [11] Desislava Hristova, Matthew J. Williams, Mirco Musolesi, Pietro Panzarasa, and Cecilia Mascolo. 2016. Measuring Urban Social Diversity Using Interconnected Geo-Social Networks. In Proc. 25th International Conference on World Wide Web (WWW). ACM, 21 -- 30. [12] Hsun-Ping Hsieh, Cheng-Te Li, and Shou-De Lin. 2015. Estimating Potential Customers Anywhere and Anytime Based on Location-Based Social Networks. In Proc. 2015 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD). Springer, 576 -- 592. [13] Hsun-Ping Hsieh, Rui Yan, and Cheng-Te Li. 2014. Where You Go Reveals Who You Know: Analyzing Social Ties from Millions of Footprints. In Proc. 24th ACM International Conference on Information and Knowledge Management (CIKM). ACM, 1839 -- 1862. [14] David Jurgens. 2013. That's what friends are for: Inferring location in online social media platforms based on social relationships. In Proc. 7th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press. [15] David Jurgens, Tyler Finethy, James McCorriston, Yi Tian Xu, and Derek Ruths. 2015. Geolocation prediction in Twitter using social networks: a critical analysis and review of current practice. In Proc. 9th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 188 -- 197. Jon Kleinberg. 1999. Authoritative sources in a hyperlinked environment. J. ACM 46, 5 (1999), 604 -- 632. [17] David Liben-Nowell and Jon Kleinberg. 2007. The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology 58, 7 (2007), 1019 -- 1031. [18] Bin Liu and Hui Xiong. 2013. Point-of-interest recommendation in location based social networks with topic and location awareness. In Proc. 13th SIAM International Conference on Data Mining (SDM). SIAM, 396 -- 404. [19] Lydia Manikonda, Yuheng Hu, and Subbarao Kambhampati. 2014. Analyzing user activities, demographics, social network structure and user-generated content on Instagram. CoRR abs/1410.8099 (2014). [20] Yelena Mejova, Hamed Haddadi, Anastasios Noulas, and Ingmar Weber. 2015. #FoodPorn: obesity patterns in culinary interactions. In Proc. the 5th International Conference on Digital Health (DH). ACM, 51 -- 58. [21] Melinda J Milligan. 1998. Interactional past and potential: The social construction [16] of place attachment. Symbolic Interaction 21, 1 (1998), 1 -- 33. [22] Minyue Ni, Yang Zhang, Weili Han, and Jun Pang. 2016. An Empirical Study on User Access Control in Online Social Networks. In Proc. 21st ACM Symposium on Access Control Models and Technologies (SACMAT). ACM, 13 -- 23. [26] [27] [23] Anastasios Noulas, Salvatore Scellato, Neal Lathia, and Cecilia Mascolo. 2012. A Random Walk around the City: New Venue Recommendation in Location-Based Social Networks. In Proc. 2012 International Confernece on Social Computing (SocialCom). IEEE, 144 -- 153. [24] Anastasios Noulas, Blake Shaw, Renaud Lambiotte, and Cecilia Mascolo. 2015. Topological properties and temporal dynamics of place networks in urban en- vironments. In Proc. 24th International Conference on World Wide Web (WWW Companion). ACM, 431 -- 441. [25] Alexandra-Mihaela Olteanu, Kévin Huguenin, Reza Shokri, Mathias Humbert, and Jean-Pierre Hubaux. 2017. Quantifying interdependent privacy risks with location data. IEEE Transactions on Mobile Computing 3 (2017), 829 -- 842. Jun Pang and Yang Zhang. 2015. Exploring communities for effective location prediction (poster paper). In Proc. 24th International Conference on World Wide Web Conference (Companion Volume) (WWW). ACM, 87 -- 88. Jun Pang and Yang Zhang. 2015. Location prediction: Communities speak louder than friends. In Proc. 3rd ACM Conference on Online Social Networks (COSN). ACM, 161 -- 171. [28] Huy Pham, Cyrus Shahabi, and Yan Liu. 2013. EBM: an entropy-based model to infer social strength from spatiotemporal data. In Proc. 2013 ACM International Conference on Management of Data (SIGMOD). ACM, 265 -- 276. [29] Daniele Quercia, Luca Maria Aiello, Rossano Schifanella, and Adam Davies. 2015. The digital life of walkable streets. In Proc. 24th International Conference on World Wide Web (WWW). ACM, 875 -- 884. [30] Daniele Quercia, Rossano Schifanella, and Luca Maria Aiello. 2014. The shortest path to happiness: recommending beautiful, quiet, and happy routes in the city. In Proc. 25th ACM Conference on Hypertext and Social Media (HT). ACM, 116 -- 125. [31] Daniele Quercia, Rossano Schifanella, Luca Maria Aiello, and Kate McLean. 2015. Smelly maps: the digital life of urban smellscapes. In Proc. 9th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 237 -- 236. [32] Salvatore Scellato, Anastasios Noulas, and Cecilia Mascolo. 2011. Exploiting place features in link prediction on location-based social networks. In Proc. 17th ACM Conference on Knowledge Discovery and Data Mining (KDD). ACM, 1046 -- 1054. [33] Flavio Souza, Diego de Las Casas, Vinicius Flores, SunBum Youn, Meeyoung Cha, Daniele Quercia, and Virgilio Almeida. 2015. Dawn of the selfie era: the whos, wheres, and hows of selfies on Instagram. In Proc. 3rd ACM on Conference on Online Social Networks (COSN). ACM, 221 -- 231. [34] Hanghang Tong, Christos Faloutsos, and Jia yu Pan. 2006. Community detection in networks with node attributes. In Proc. 6th IEEE International Conference on Data Mining (ICDM). IEEE CS, 613 -- 622. [35] Alessandro Venerandi, Giovanni Quattrone, Licia Capra, Daniele Quercia, and Diego Saez-Trumper. 2015. Measuring urban deprivation from user generated content. In Proc. 18th ACM Conference on Computer Supported Cooperative Work & Social Computing (CSCW). ACM, 254 -- 264. [36] Xiaojun Wan, Jianwu Yang, and Jianguo Xiao. 2007. Towards an iterative rein- forcement approach for simultaneous document summarization and keyword extraction. In Proc. 45th Annual Meeting of the Association of Computational Linguistics (ACL). 552 -- 559. [37] Hongjian Wang, Zhenhui Li, and Wang-Chien Lee. 2014. PGT: Measuring mobility relationship using personal, global and temporal factors. In Proc. 14th IEEE International Conference on Data Mining (ICDM). IEEE, 570 -- 579. [38] Mao Ye, Xingjie Liu, and Wang-Chien Lee. 2012. Exploring social influence for recommendation: a generative model approach. In Proc. 35th ACM SIGIR conference on Research and Development in Information Retrieval (SIGIR). ACM, 671 -- 680. [39] Mao Ye, Dong Shou, Wang-Chien Lee, Peifeng Yin, and Krzysztof Janowicz. 2011. On the semantic annotation of places in location-based social networks. In Proc. 17th ACM Conference on Knowledge Discovery and Data Mining (KDD). ACM, 520 -- 528. Jing Yuan, Yu Zheng, and Xing Xie. 2012. Discovering regions of different functions in a city using human mobility and POIs. In Proc. 18th ACM Conference on Knowledge Discovery and Data Mining (KDD). ACM, 186 -- 194. [41] Yang Zhang and Jun Pang. 2015. Distance and friendship: A distance-based model for link prediction in social networks. In Proc. 17th Asia-Pacific Web Conference (APWeb). Springer, 55 -- 66. [42] Vincent W. Zheng, Yu Zheng, Xing Xie, and Qiang Yang. 2010. Collaborative location and activity recommendations with GPS history data. In Proc. 19th International Conference on World Wide Web (WWW). ACM, 1029 -- 1038. [43] Yu Zheng, Licia Capra, Ouri Wolfson, and Hai Yang. 2014. Urban computing: concepts, methodologies, and applications. ACM Transactions on Intelligent Systems and Technology 5, 3 (2014), 1 -- 55. [40]
1611.02388	1	1611	2016-11-08T04:52:24	PNP: Fast Path Ensemble Method for Movie Design	[ "cs.SI" ]	How can we design a product or movie that will attract, for example, the interest of Pennsylvania adolescents or liberal newspaper critics? What should be the genre of that movie and who should be in the cast? In this work, we seek to identify how we can design new movies with features tailored to a specific user population. We formulate the movie design as an optimization problem over the inference of user-feature scores and selection of the features that maximize the number of attracted users. Our approach, PNP, is based on a heterogeneous, tripartite graph of users, movies and features (e.g., actors, directors, genres), where users rate movies and features contribute to movies. We learn the preferences by leveraging user similarities defined through different types of relations, and show that our method outperforms state-of-the-art approaches, including matrix factorization and other heterogeneous graph-based analysis. We evaluate PNP on publicly available real-world data and show that it is highly scalable and effectively provides movie designs oriented towards different groups of users, including men, women, and adolescents.	cs.SI	cs	PNP: Fast Path Ensemble Method for Movie Design ∗ Danai Koutra University of Michigan {dkoutra}@umich.edu † Stratis Ioannidis Northeastern University [email protected] Abhilash Dighe University of Michigan [email protected] Smriti Bhagat†, Udi Weinsberg† Facebook {smr, udi}@fb.com Christos Faloutsos Carnegie Mellon University [email protected] Jean Bolot Technicolor [email protected] 6 1 0 2 v o N 8 ] I S . s c [ 1 v 8 8 3 2 0 . 1 1 6 1 : v i X r a ABSTRACT How can we design a product or movie that will attract, for exam- ple, the interest of Pennsylvania adolescents or liberal newspaper critics? What should be the genre of that movie and who should be in the cast? In this work, we seek to identify how we can design new movies with features tailored to a specific user population. We formulate the movie design as an optimization problem over the inference of user-feature scores and selection of the features that maximize the number of attracted users. Our approach, PNP, is based on a heterogeneous, tripartite graph of users, movies and fea- tures (e.g., actors, directors, genres), where users rate movies and features contribute to movies. We learn the preferences by leverag- ing user similarities defined through different types of relations, and show that our method outperforms state-of-the-art approaches, in- cluding matrix factorization and other heterogeneous graph-based analysis. We evaluate PNP on publicly available real-world data and show that it is highly scalable and effectively provides movie designs oriented towards different groups of users, including men, women, and adolescents. 1. INTRODUCTION Creating products that satisfy the market is critical to companies as it determines their success and revenue [3]. Online services, such as Amazon, eBay, and Netflix, use data-driven approaches to recommend products to their customers. Recently, however, com- panies like Netflix have employed their direct knowledge of user viewing habits to make licensing and original programming deci- sions; this was indeed the case with, e.g., 'House of Cards', one of their most-watched shows [8, 18]. Can we take this decision- making process one step further, and leverage the large amounts of available data (e.g., user ratings or (dis)likes, reviews, product characteristics) to inform the design of new products and services that will likely satisfy the target customers? Market fit for products and services [3], team formation [17, 2], and team replacement [19] are related problems with a similar goal: designing a new product, service, or team that will succeed, given a target market or task. The large number of possible choices, com- binations, and constraints, as well as the challenge of assessing au- dience demand, make this a considerably hard problem. Currently, experts use their judgement to estimate solutions to this problem, but this does not scale or allow leveraging massive datasets. The goal of this paper is to formalize the movie design prob- lem and solve it in a tractable and scalable fashion. Our motivation comes from the movie entertainment industry: We investigate how ∗This work started during an internship at Technicolor. †This work started while at Technicolor. Figure 1: Tripartite graph of the target users U, their rated movies M and the movie features F. to design a successful movie that will attract the interest of an ar- bitrary targeted demographic. Specifically, we tackle the following problem (in short, MD): PROBLEM 1. [MOVIEDESIGN or MD] Given (a) a set of users, target users, movies, and features, (b) some user movie ratings, and (c) the feature-movie memberships; design a movie which will attract most of the target users. Given additional data about the cost of each feature, the prob- lem can be augmented to handle budget constraints as well. The MD problem resembles recommender systems [23, 20, 13, 34, 26] and group recommendations [25, 14]. Its main difference is that it does not aim to find the best existing movie that the target users are likely to enjoy. Instead, given the user preferences, it determines the features for a new movie that is likely to attract the most users. Solving MD poses several challenges. Clearly, identifying an ef- fective design relies on how movie features affect user preferences. As extensively documented in the recommendation literature [16, 5], and also corroborated by our experiments, "collaborative fil- tering" approaches, exploiting similarities between users, tend to significantly outperform regressing individual user preferences in isolation. It is not immediately clear how to solve MD by lever- aging both explicit features and user similarities in a principled, tractable fashion. An additional challenge arises from the fact that the MD problem can be applied to an arbitrary set of users U. In- deed, different strategists may wish to attract varied demographics, and our design should be able to cope with arbitrary, repeated re- quests. As a result, we would like to efficiently determine movie designs targeting any given group. To address these challenges, we propose a graph-theoretic ap- proach for the MD problem and contribute: • Problem Formulation: We formally pose MD as a problem for designing successful new movies for a target audience. More- over, we adopt an approach that separates the problem solution into two phases: user-feature preference inference, and model-based de- sign. This separation allows us to infer user preferences and effi- ciently handle arbitrary "design queries", as defined by targeted user groups. • Path-Based Training: To infer the user-feature scores, we pro- pose a novel model based on predefined walks on a heterogeneous graph consisting of users, movies, and features, which treats "dis- likes" in a natural way. This methodology allows us to leverage both features and user similarity. Although we focus on the setting of movies, our method is generalizable to any setting where user ratings and product features are available. • Model-Based Design: Having inferred user-feature preferences, we formulate the selection of features that compose a movie as an optimization problem with capacity and other constraints, and es- tablish conditions under which the problem solution is tractable. • Experiments: By using real-world data with ∼5 million movie- ratings and 175 000 movie-features, we show that our model of user behavior succeeds in describing user preferences with very high accuracy. In addition, combined with our optimization method, it results in movie design choices that significantly outperform those of the competitors. The paper is organized as follows: We describe our proposed method, PNP, the dataset, and the experimental evaluation in Sec- tions 2, 3, and 4, respectively. Sections 5 and 6 present the related work and our conclusions. 2. PROPOSED METHOD: PNP For the MD problem, we assume that the data consists of: a set of users U, a set of movies M, and a set of movie features F. In addition, the feature set F is partitioned into types (e.g., actors, directors, genres): we denote by T the set of types, and by F(cid:96) ⊂ F the set of features of type (cid:96) ∈ T . The relations between entities consist of: (a) the user-movie ratings, containing tuples of the form (i, j, rij), where i ∈ U and j ∈ M, and are organized in a u×m matrix R, with zeros indicating absent ratings; (b) the movie- feature memberships stored in a m × f binary matrix F, where a non-zero entry (j, k) means that feature k belongs to movie j. Throughout the paper, following observations in the literature [12], we say that user i 'likes' movie j if the rating rij is larger than i's average rating ¯ri among non-zero entries in R. Table 1 presents the most commonly used symbols. The goal of the MD problem (Problem 1) is to design a new movie by selecting its features so that it is liked by as many of the people in the target audience as possible. To make the problem tractable, and handle the challenges that we described in the in- troduction, we reduce it to solving two separate subproblems. We tackle these subproblems in Sections 2.1, and 2.2, respectively. PROBLEM 2. [Inferring User-Feature Scores] Given (a) a set of users U, movies M, and features F, (b) the user movie ratings R, and (c) the movie-feature memberships F; find the user-feature preference scores. The first subproblem amounts to learning user-feature preference scores, capturing the propensity of a given user to like a movie that contains a given feature; the higher the score, the more likely a user is to like this movie. In the second subproblem, we use these scores to formulate the objective of maximizing the number of users liking the movie, among a targeted set. PROBLEM 3. [Designing the Movie] Given (a) a set of target users U(cid:48), features F, and types T , and (b) the user-feature pref- erences; select features for a new movie s.t. the number of users who will probably like it is maximized. If additionally each feature k is associated with a cost ck, we can augment the problem by incorporating budget constraints. These Table 1: Symbols and Definitions. Bold capital: matrices; Low- ercase: vectors; Plain font: scalars. Symbol G U(U(cid:48)),M,F,T u(u(cid:48)), m, f, t R F W Q τ S G(S) x ck B(cid:96) Description tripartite input graph the set of (target) users, movies, features, and types, resp. number of (target) users, movies, features, and types, resp. u × m matrix of user-movie ratings, with elements rij m × f movie-feature membership matrix, with elements fij inferred u × f matrix of user-feature preferences proposed u × m matrix of user-movie ratings, with elmnts. qij threshold in the linear threshold model the set of chosen features user conversion function 1 × f binary characteristic vector of S cost per feature k (e.g., salary) budget per type (cid:96) constraints would express the requirement that the total expense on each feature type (actor, director, etc.) does not exceed a budget B(cid:96) per type. Next we discuss our solutions to these two problems. 2.1 Step 1. Inferring user-feature scores The first constituent problem of MD, Problem 2, can be ap- proached as a classification problem, in which binary 'like' labels of every user are regressed from movie features. Traditional meth- ods (e.g., random forests, matrix factorization) can be used to solve such a problem; as a byproduct, these methods often quantify the effect of each individual feature on a user's decision. However, these methods do not perform well (cf. Section 4) as they either ignore commonalities between users or make use of latent features which cannot be used to identify user-feature preferences. We propose to solve the MD problem by handling both con- stituent problems via a graph-theoretic approach, which cleverly leverages both explicit features and user similarities, and, most im- portantly, its resulting scores also lead to a tractable movie design optimization for an arbitrary set of target users (Section 2.2). We model the input data as an undirected, heterogeneous, tri-partite network (Figure 1), where the nodes are users, movies, and movie features. The relationships or edge types are 'rated-by', and 'belong- to'. We give a pictorial overview of our approach in Figure 2. Specifically, our proposed method, PNP (Positive / Negative Paths), infers the user-feature scores by performing walks of fixed length on the tripartite graph of users, movies and features, thus lever- aging information about user-movie ratings as in 'collaborative' approaches, as well as movie 'content' (features). PNP is based on meta-paths, which was first introduced for similarity search in heterogeneous information networks [29]. Informally, a meta-path consists of a sequence of relations between node types. Formally: DEFINITION. A meta-path or predefined-path P on a het- erogeneous graph G with object types A1, A2, . . . and relations R1, R2, . . . : P = A1 . . . At = A1 R1−−→ A2 R2−−→ A3 R3−−→ . . . Rt−1−−−→ At defines a composite relation R = R1 ◦ R2 ◦ ··· ◦ Rs−1 between the source node type A1 and the target node type At, where ◦ is the composition operator on relations. In our setting, the object types include users U, movies M, and features F , and there exist two types of relations: 'rated-by' and 'belongs-to', as well as their inverses. To capture user-feature preferences, we propose a random-walk- based score restricted over predefined paths starting from F and ending in U. We view the proximity of a user to a feature through such a path as an indication of the user's possible 'preference' to- wards this feature. For example, a 2-step path via movies cap- tures direct user-feature preferences based solely on the movies they have rated. Unlike our approach, the original path similar- Figure 2: MD: Proposed pipeline for designing a new movie based on user ratings and movie features. (a) Direct 2-step Path (b) User-based 4-step Path (c) Feature-based 4-step Path Figure 3: Combination of three types of predefined walks on the heterogeneous tri-partite graph. The red arrows show the predefined direction of each walk. ity [29] is computed as a normalized count of paths between enti- ties. For different types of preferences, we consider three prede- fined paths: • 2-step Path: The path P = F M U in Figure 3(a) finds the preferences of each user based on her ratings, and, thus, does not exploit a collaborative setting. It computes accurate preferences for the features that appear in movies she has rated, but does not infer preferences for other features. • User-based 4-step Path: The path P = F M U M U in Fig- ure 3(b) computes the user preferences based on the user's in- put, as well as that of other users who are similar to her. The similarity between users is defined via the movies they have watched; two users who have rated the same set of movies with comparable scores are similar. • Feature-based 4-step Path: The path P = F M F M U in Figure 3(c) finds the user preferences based on the user- movie ratings and similar movies to the ones she rated. The similarity between movies is content- or feature-based. One way to compute a user-feature score is through the prob- ability that a random walk starting at a given node, restricted to follow only paths of the above three forms, terminates at a given feature. Computing this probability corresponds to matrix multipli- cations involving the transition matrices induced by each bipartite graph (i.e., user-movies and movie-features). PNP improves upon this approach in two ways, by considering 'positive' and 'negative' random walks, and incorporating edge weights, as described below. Positive and Negative Walks. One issue with the above approach for preference inference is that the random walk model cannot han- dle 'dislikes' or 'negative' weights. Constructing the user-movies graph by creating an edge for each rating results in feature scores that only increase in magnitude despite the fact that a rating may be below average, indicating a 'dislike'. However, this is not desired in our MD problem; we want the features that contribute to movies disliked by a user to be penalized by receiving smaller scores than features that contribute to her favorite movies. To handle this case, we introduce the concept of POSITIVE and NEGATIVE walks. Since the movie ratings, which incorporate 'dis- like' information, appear in the bipartite user-movie graph, we fo- cus only on that. To obtain the POSITIVE graph from the original bipartite graph, we derive a graph that captures 'like' relations [12], (a) Original graph (b) POSITIVE graph (c) NEGATIVE graph (d) Numeric Example Figure 4: Toy example of POSITIVE and NEGATIVE user rat- ings graphs for our proposed method PNP. i.e. edges with weight (rating) above the corresponding average user rating. Similarly, we obtain the NEGATIVE graph by capturing only 'dislike' relations. The process of generating the POSITIVE and NEGATIVE graphs is illustrated in Figure 4. Edge Weights. To better leverage the user preferences, we assign weights to edges between users and movies on both the POSITIVE and NEGATIVE graphs, so that the walker performs a walk over a weighted graph. Since user ratings are often biased [16] (e.g., some users give high scores objectively, even if they did not enjoy the movie), we do not directly use the ratings as edge weights, but we center user i's rating for movie j through rij − ¯ri, where ¯ri is the user's average rating. That is, we adjust the ratings according to each user's average movie rating and refer to it as a 'Centered Rat- ing' (CR). Second, to emphasize big deviations from the average user rating, we use a non-linear reweighing schema: qij = 2δ∗(CR) = 2δ·rij−¯ri (1) which puts more weight on very low and very high ratings, where low and high are subjective for each user. We denote the resulting u×m weighted adjacency matrix by Q, and consider random walks over the corresponding weighted graph. We study the effect of the reweighing schema and its non-linearity in Section 4. PNP Score Computation. Putting all this together, PNP pro- ceeds as follows. It computes the probability that a random walk starting at a user reaches a feature, under the following assump- tions: the walk is (a) constrained to occur only over paths of the three proposed predefined types, and (b) occurs on either the POS- ITIVE or the NEGATIVE graph, with edges appropriately weighted − be the probability vec- through Equation (1). Let wp tors over features, for path type p and user i ∈ U. Then, we define − as the weighted user-feature preference vector wp = wp over path type p. It is clear that, as desired, PNP penalizes features in movies that are disliked by a user, by subtracting some value from their preference score. We define the final weighted feature vector of PNP for i, which represents i's preference over features in F as a linear combination of the weighted user-feature vectors over the three predefined path types: + − wp + and wp w = α w2step + β w4step-usr + γ w4step-feat (2) where w = (w+−w−). In matrix form, the feature preferences for all users are: W = α W2step + β W4step-usr + γ W4step-feat, (3) where α,β,γ ∈ R+ are the combination parameters satisfying the equation α+β+γ=1, and W is a u × f matrix. 2.2 Step 2. Designing the movie The second constituent problem of MD, Problem 3, tackles the actual selection of features, S, for a new movie so that the num- ber of users who will likely enjoy it is maximized. We propose to formulate Problem 3 as an optimization problem with cardinal- ity constraints (and, if available, budget constraints). Following the literature on node-specific threshold models for influence max- imization [15, 10], we consider a user i converted with respect to k∈S wik > τi, where wik is user i's preference score for feature k and τi is the user-specific threshold. Treating τi as a random variable, we de- fine the user conversion function, G(S), as the expected number of (cid:88) users in the target set U(cid:48) that are converted: (i.e., 'likes') a movie consisting of features S if(cid:80) (4) where x ∈ {0, 1}f is the characteristic vector of set S (i.e., xk = 1 iff k ∈ S). Based on these assumptions, we frame the MD task as a general optimization problem: wik > τi] = G(S) = (cid:88) (cid:88) F [wixT ] i∈U(cid:48) i∈U(cid:48) k∈S P [ Maximize subject to (cid:80) G(S) = G(x) =(cid:80) k∈F(cid:96) x ∈ {0, 1}f . bk xk ≤ B(cid:96), i∈U(cid:48) F [wixT ] for each (cid:96) ∈ T , (5) (6) In the special case where bk = 1 for all k ∈ F(cid:96), the above prob- lem effectively amounts to maximizing G(S) subject to a cardinal- ity constraint S ∩ F(cid:96) ≤ B(cid:96). Such constraints also make sense as, e.g., no more than one director is needed per movie. If the cost ck per feature is available, then we can set bk = ck, and, thus, cap- ture budget constraints. We note that the optimization problem has a linear constraint, and the objective depends on the inner product of scores with x; as such, it does not model dependencies between features. Next, based on evidence from the real-world data, we explain why we do not model such dependencies. Why not model feature dependencies? We followed a data-driven approach to evaluate whether dependencies between movie features correlate with movie ratings (e.g., two actors should always play to- gether because then they lead to successful movies, while indepen- dently they do not). Using the IMDB dataset described in Section 3, we performed: (1) frequent item mining (a priori algorithm) to find frequent k-itemsets that have appeared in at least ten movies; and then (2) linear and lasso regression with 5-fold cross validation to select the k-itemsets (singletons, pairs or triplets of features) that are most predictive of the movie rating. Frequent item mining showed that only 1% of the features are frequent (with support ≥ 0.5%), half of which contribute to fre- quent pairs, and 20% contribute to frequent triplets. The results of linear and lasso regression are summarized in Table 2. In general, regression on the dataset with frequent singleton features is better or comparable (in terms of error and model compactness) to regres- sion on datasets that capture feature dependencies. Specifically, the MSE (Mean Squared Error) of linear regression is minimum when the input dataset has only singleton features. The MSE of lasso re- gression becomes slightly smaller when we consider frequent pairs of features rather than only frequent single features (1.04 vs. 1.03). However, the majority of selected features are singletons, and the number of selected features are almost the same. Therefore, the feature dependencies are very few and do not affect the predictabil- ity of ratings significantly. Uniform Threshold Model. We propose to use a linear thresh- old model where each user i picks her conversion threshold τi uni- formly at random [15], i.e., we assume no background information about the user conversion thresholds. Since the PNP user-feature preferences computed are in [−1, 1], we set the thresholds to follow the uniform distribution in the same interval, i.e. τi ∼ U[−1, 1]. Under these assumptions, the user conversion function (4) becomes f (S) = 1T W1 + 1 2 U 2 PROOF. Starting from Equation (4) and applying our assump- tions, we obtain: f (S) =(cid:80) i∈U(cid:48) P [(cid:80) =(cid:80) 2 ((cid:80) (cid:80) (cid:80) (cid:80) i∈U(cid:48) 1 k∈S = 1 2 1 2 k∈S wik > τi] k∈S wik + 1) = i∈U(cid:48) wik + U 2 τi∼U [−1,1] = wk=(cid:80) i∈U(cid:48) wik === = k∈S wk + movie features that maximize(cid:80) U(cid:48) 2 k∈S wk. Hence, under the uniform threshold model, we want to pick the In this case, the optimization problem reduces to knapsack un- der several separable constraints, which, though NP-hard, can be solved with an FPTA scheme [32]. In the unit-cost (i.e., cardinal- ity) case, the problem is solvable in polynomial time: sorting fea- tures k ∈ F in decreasing order of costs wk, and picking the top B(cid:96) features, is optimal. Sigmoid Threshold Model. The uniform threshold model has the advantage of resulting in an easy-to-solve optimization problem. To further leverage background information that might be available in the data, data-driven models may be considered for the user- specific thresholds. For instance, if the average rating score per user follows a sigmoid distribution and it is being used as the user- specific threshold, then F in Equation (4) can be replaced with the CDF of the logistic distribution, FS. formulating the problem as the maximization of(cid:80) This results in a sigmoidal programming problem [31], which is NP-hard even in its relaxed form with non-integral solution. By i∈U FS[yi], we introduce an additional set, F, of constraints yi = wixT , which Table 2: Dependencies between movie factors do not seem to correlate with movie success. Features singletons singletons + pairs only pairs singletons + pairs + triplets MSE (std) 1.0395 (0.0533) 1.0303 (0.0328) 1.1103 (0.0327) 1.0452 (0.0526) Lasso Regression (5-fold C.V.) non-0 coeff. (total) 211 (621) 207 (969) 141 (348) 235 (1104) Singletons Pairs Triplets n/a n/a n/a 11 211 161 n/a 179 n/a 46 141 45 MSE (std) Linear Regression (5-fold C.V.) non-0 coeff. 620 899 100 951 1.5448 (0.0467) 3.025 (0.3498) 4.0232 (17.3369) 2.9254 (0.8767) can be solved approximately by using the branch-and-bound based method in [31]. The decimal solutions can then be converted to in- tegral by using pipage rounding [1]. However, the approximation error of the solution depends on the number of constraints, and the method may solve many convex optimization problems exponen- tially, which suggests that the method will be impractical for the MD problem with thousands of constraints. 2.3 Complexity of PNP Naïve approach. Performing the walks over the 2-step and 4-step predefined paths on the positive and negative graphs is computa- tionally expensive and is dominated by sparse matrix and sparse- dense matrix multiplications (e.g., Q F for P = F M U, or Q QT Q F for P = F M U M U, where Q, QT , F are u × m, m × u, and m × f transition matrices). If we assume naïve matrix multiplica- tions, the complexity of performing the walks on the 2-step, user- based 4-step, and feature-based 4-step paths would be: Θ(f mu), Θ(f mu + 2u2m) and Θ(3f mu). We note that these computa- tions can be trivially parallelized, so the complexity corresponds to the maximum of the three. Moreover, the output is a dense u × f matrix W that captures the user-feature preferences. Although the bulk of the computation can happen offline, we can support quick query processing by just summing up the rows in matrix W that correspond to the target users U(cid:48) and performing linear feature se- lection in O(u(cid:48)f ) after sorting the scores in decreasing order in O(f log f ). Next, we show that we can significantly speed up the computation and reduce the storage requirements. Fast approach. Noticing that the uniform threshold model requires only the sum of the target users' feature preferences (and not the whole W matrix) helps us to significantly speed up the computa- tions and reduce the storage requirements. By using an indicator vector x for the users in U(cid:48), all the walks can be re-designed (back- wards) as fast sparse matrix-vector multiplications-for example, FT ( QT ( Q( QT x))) in Θ((3 nnz( Q) +nnz( F))u(cid:48)), where nnz() is the support size (i.e., number of non-zeros) of the corresponding matrix. The transpose of the resulting score matrix is W. The inter- mediate and final computations generate and store vectors of vari- able lengths. The feature selection can be done in O(f log f + f ). As we show in Section 4.1.3, the fast approach is very scalable and up to 100× faster than the naïve approach. 3. DATA We compiled our data from two different and publicly available sources: (i) the Flixster website [35], a social movie site allowing users to watch and rate movies, from which we obtained user-movie ratings, and (ii) the MGO website [30] which has information about movie features. The Flixster and MGO datasets consist of 66,726 and 83,017 movies, respectively. In addition to the user ratings, the Flixster dataset provides some user demographics, including their gender, age, and location (as free text). To merge the two datasets, first we dropped the movie remakes. In order to be able to infer the user-feature or user-movie prefer- ences (Problem 1) using 5-fold cross validation, we iteratively fil- tered the data so that: every movie has been rated by at least 20 users and has at least 2 features; every user has rated at least 20 movies; every feature has participated in at least two movies. The resulting user-movie-feature tripartite graph has 66, 407 nodes and 4, 567, 253 edges (Table 3). The movie features include: 25, 721 actors, 1, 322 directors, 4, 305 producers, 654 studios, and 27 gen- res. The user ratings are between 1 and 5, with 0.5 increments. For our dependency analysis we used an IMDB dataset consist- ing of 1, 893 movies, their features ( 947 directors, 5, 231 produc- ers, 1, 209 studios, 51, 500 actors, 27 genres, and 4 seasons), and average movie ratings. Movies Users Features Table 3: Statistics of our movie dataset. 5,881 User-Movie Ratings 28,482 Movie-Feature Memberships 32,029 4,435,359 131,894 4. EXPERIMENTS In this section, we give the experimental analysis of PNP, and evaluate its performance for designing new movies for specific au- diences. We seek to answer four main questions: Q1. How does PNP compare to the baseline approaches? Q2. Is PNP robust with respect to its parameters? Q3. Is PNP scalable? Q4. How successful are the movies we design? The first three questions are related to the inference of user-feature- score preferences (Problem 2), and the last question refers to the optimization problem (Problem 3). 4.1 Analysis of Step 2: Inferring feature scores 4.1.1 Q1: PNP vs. baseline methods To compare the predictive power of PNP against baseline meth- ods, we cast Problem 2 as a binary classification problem where the goal is to predict whether a given user will like or dislike a new movie. More formally, recall that we consider two classes [12]: User i likes (dislikes) movie j if she rated it higher (lower) than her average user score, i.e., Cij=1 if rij ≥ ¯ri (Cij=0 if rij < ¯ri). Each observation corresponds to a user-movie rating, where the in- dependent variables are the movie features (binary), and the de- pendent variable is the user's preference (like or dislike). We note that the mapping of the user-ratings to binary values results in 55% 'liked' movies, and 45% 'disliked' movies. In all the cases that we describe below, we learn a per-user classifier on their movie rat- ings, and test it on new movies. To predict the user's preference for a new movie using PNP, we multiply the user-based feature- preference vector wi with the new binary movie vector, and com- pute her movie score. This problem can be solved by several traditional methods, among which we consider: Baseline 1.1: Naive Bayes. NB can predict whether or not a user will like a new movie, i.e., it cannot infer the user-feature prefer- ences (Problem 2). NB makes the assumption that the features act independently. To avoid the zero-probabilities issue, we use the Laplace correction. Baseline 1.2: Logistic Regression. LR is a commonly-used gen- eralized linear model for prediction. Similarly to NB, LR also as- sumes feature independence. LR also generates a weight vector of the features' contribution in the prediction. Since the number of features is greater than the number of samples (Table 3), we used L1 regularization to ensure sparsity of the produced weight vector. To find the best value of the regularizer parameter, we performed a cross-validated grid-search over the values {0.5, 1, 5, 10}. Baseline 1.3: Random Forests. RF is an ensemble method for classification and regression, which is robust to the inclusion of irrelevant features and overfitting. For each user, the RF classi- fier generates N random subsets on the training data in terms of features, and constructs a separate decision tree for each subset. To predict a new user-movie preference, the movie's feature vector (new sample) runs through all the generated user-specific decision trees, and the mode of the classes is assigned as the predicted class. A drawback of this approach is its runtime, since it requires gen- erating many decision trees per user. In our experiments, for the maximum depth of the trees, we performed grid-search over the values {25, 100, 500, 1000}, and generated 10 random subsets of the data. Overall, RF was very slow despite the small number of grid-search values. Baseline 1.4: Matrix Factorization. MF [16, 5] with user and item biases predicts the rating of user i for item j: rij = (cid:104)ui, vj(cid:105) + µ + bi + bj where ui and vj are d-length vectors (for some small d) that cap- ture the user and item latent vectors, respectively, µ is the global average rating, bi is the user bias, and bj is the item bias. The main advantage over the other approaches is that we learn a model for each user using a small dense matrix, where the number of rated movies is commonly larger than the number of features. However, MF operates on a latent feature space, thus using it to design a new movie is not straightforward. We performed MF over the ratings matrix R to compute user and feature vectors with a dimension of d = 10. These were computed using 20 iterations of stochastic gradient descent, and (cid:96)2 regularization parameters selected through cross validation. The predicted ratings of user/movie pairs were used in the AUC computations. Baseline 1.5: Content-based Matrix Factorization. We further extended basic MF by incorporating content-dependent biases: be- yond user and movie biases, we include additional bias in our model, one for every item feature. These can be thought of as explicit, rather than latent, features of a movie. We again compute parame- ters for (cid:96)2 regularization penalties through cross validation. Baseline 1.6: Heterogeneous Entity Recommendation. The method in [33] finds the top-k movies to recommend to a user by perform- ing non-negative matrix factorization on a set of user-movie pref- erence diffusion matrices learned via meta-path similarity [29] and user clustering. We note that this approach cannot be used for fea- ture preference inference since the model applies only for entities that are directly linked to and rated by users in the heterogeneous graph (e.g., movies). Thus, in the binary classification task, we use the inferred user-movie scores which are directly learned over the metapaths that correspond to our proposed method's predefined paths. We set k = 20 for the low-rank approximations and pick the method's parameters through cross validation. Table 4: PNP outperforms all the baselines. We report the pre- diction accuracy (avg AUC and its std in parentheses). α, β, γ are the path combination parameters, and δ is the rating- reweighing parameter. Method Naive Bayes Logistic Regression Random Forests Matrix Factorization Content-based Matrix Factorization Heterogeneous Entity Recommendation PNP (α, β, γ, δ) = (0.7, 0.1, 0.2, 1) (α, β, γ, δ) = (0.5, 0.2, 0.3, 1) (α, β, γ, δ) = (0.3, 0.2, 0.5, 1) (α, β, γ, δ) = (0.5, 0.2, 0.3, 0.5) (α, β, γ, δ) = (0.5, 0.2, 0.3, 1.5) Avg AUC over 5 folds 0.5340 (0.0002) 0.6621 (0.1369) 0.6418 (0.1390) 0.6396 (0.0825) 0.7043 (0.1420) 0.5998 (0.0112) 0.9146 (0.0043) 0.9143 (0.0043) 0.9132 (0.0042) 0.9171 (0.0019) 0.9062 (0.0018) Results To evaluate the methods, we perform 5-fold cross validation, com- pute the AUC (Area Under the Curve) per user, and report its aver- age and standard deviation over all users in Table 4. OBSERVATION 1. PNP outperforms all baseline approaches on user-feature preference inference. The main shortcomings of the baselines are: (i) Each per-user model for NB, LR, and RF leverages information specific to that user only-i.e., her movie ratings, and the corresponding movie features-, and thus suffers from the sparsity problem; (ii) MF falls in the collaborative space, but does not leverage information about the movie features, and operates on a latent feature space; content- based MF also partially relies on latent movie features, which again prohibits its use in movie design. The heterogeneous entity recom- mendation approach infers the movie preferences, but not the fea- ture preferences, and, thus, cannot be used for movie design either. In contrast, our proposed method, PNP, solves the MD problem by exploiting the latent similarities between users (collaborative), movies, and features, and operates explicitly on the movie features, thus directly enabling us to use its output to design movies. 4.1.2 Q2: Robustness of PNP Robustness to the reweighing schema. To evaluate our reweigh- ing schema in Eq. 1, we learn a PNP classifier per user for differ- ent values of the parameter δ. In Figure 5(a), we give the average AUC (5-fold cross validation) of PNP for using the centered ratings without exponentiation (simple reweighing), and our proposed ap- proach with δ ranging from 0 to 3, and for combination parameters α = 0.5, β = 0.2 and γ = 0.3. The performance is stable for δ = 0 − 1.5, which corresponds to moderate scaling of the user ratings. OBSERVATION 2. PNP is robust to moderate rescaling of the user-adapted movie ratings. Robustness to the combination parameters. As described in Eq. (3), the proposed method, PNP, has three parameters which de- fine the involvement of each predefined path. Next, we explore how these parameters affect the user preference inference. As before, we learn a PNP classifier per user for different values of {α, β, γ}, and compute the average AUC and its standard deviation (y-axis) by performing 5-fold cross validation (Figure 5(b)). In this experi- ment we set δ = 1. We vary the values of α and γ from 0 to 1 with step 0.1 (β is uniquely determined since α + β + γ = 1). (a) Robustness wrt reweighing schema (b) Robustness wrt combination params (c) Scalability Figure 5: (a) PNP is robust to moderate rescaling of the centered ratings-x-axis: exponent of our proposed reweighing schema in Eq. 1; y-axis: avg AUC over 5-fold cross validation. (b) PNP is robust to the combination parameters of the predefined paths-x- axis: values of parameter α; y-axis: avg AUC. (c) PNP is scalable. The fast approach is faster and needs less space than the naive approach. OBSERVATION 3. parameters (α, β, γ) of the predefined walks. PNP is quite robust to the 'combination' The lowest accuracy is obtained when the direct user preferences (2-step path) are ignored, and the highest accuracy when the user- based 4-step path has small participation, likely because this walk 'blurs' the individual preferences by relying on 'similar' users. In practice, these parameters can be set by performing cross validation on the data at hand. 4.1.3 Q3: Scalability We evaluate the scalability of our method with respect to the number of ratings, which corresponds to the number of non-zeros in the ratings matrix R. For our experiment, we vary the number of ratings from 1000 to 4 435 359, which is the total number of ratings in our dataset. For each number of ratings, we generate five matrices with randomly chosen ratings from the original matrix, we run PNP and report the average and standard deviation of the runtime. For comparison, we ran both the naïve and fast approaches (we give their theoretical complexities in Section 2.3). Figure 5(c) shows the runtime of the two methods on: (a) a standard machine (STM): AMD Opteron Processor 854 @ 2.80GHz, 2 cores, 32GB RAM; (b) a high-performance machine (HPM): AMD Opteron Proces- sor 6282 SE @ 2.60GHz, 16 cores, 264GB RAM. We note that the naïve approach runs out of memory for more than 10 000 ratings on the standard machine, while our fast approach is highly scalable and has comparable runtime on both machines. OBSERVATION 4. The fast approach (sparse matrix-vector mul- tiplications) is up to 100× faster than the naïve approach, and requires less memory. 4.2 Analysis of Step 3: Designing the Movie 4.2.1 Q4: Quantitative Evaluation To evaluate the quality of our movie designs quantitatively, we compare it with two intuitive baselines: • Baseline 2.1: POPULAR. Using the same capacity constraints as in the optimization problem (Eq. 5), we design a movie by choosing the most popular features for the target audience U(cid:48) (the ratings are inherited from the movies to which they be- long). The 1 × f popularity vector is given by: p = v · F, where v is the vector of total ratings per movie, with ele- i=1 1rij >0, 1 is an indicator function, and F ments vj =(cid:80)u(cid:48) is the binary movie-feature membership matrix. • Baseline 2.2: TOP. Applying the same capacity constraints, we design a movie by selecting the most highly rated fea- tures for the target users U(cid:48). We assume that the features inherit their movies' ratings, so the vector of feature ratings can be computed as: t = ¯rj ·(cid:101)F, where ¯rj is the vector with the average movie ratings over all users in U(cid:48), and (cid:101)F is the column-normalized movie-feature membership matrix. We introduce two quantitative measures to compare the PNP movie designs to the POPULAR and TOP designs. • kNN: Given a movie j∗, we find the set N of its k nearest neighbors (via cosine similarity) and infer its score as rj∗ = j∈N ¯rj, where ¯r is the vector with the average movie 1N ratings over all users in U(cid:48). We set k = 20. • W-kNN: This is the weighted version of kNN, where the movie scores are weighted by their cosine similarity to the j∈N sim(j∗, j) · ¯rj, where (cid:80) given movie j∗, i.e., rj∗ = α(cid:80) α = 1/(cid:80) j∈N sim(j∗, j) is a normalization factor. Results. We applied all methods to the movie dataset (Section 3) in order to design movies (with capacity constraints, e.g., six ac- tors, two directors, two genres) for different target audiences: (i) all women; (ii) all men; (iii) 13- to 15- year-old children; (iv) 13- to 18- year-old children; (v) men 18-25 years old; (vi) women 18-25 years old. OBSERVATION 5. As shown in Figures 6(a) and (b), PNP outperforms the competing baselines POPULAR and TOP designs and obtain higher kNN and W-kNN scores. 4.2.2 Anecdotes To highlight how the movie design changes with the age of the target audience, we present the movies designed for men across different age groups. For teenage boys (13-18years), the movie designed has the following composition. Top genres are Martial Arts and Action; top actors chosen are Will Smith (according to IMDB, best known for the Men In Black series) and Adam San- dler (known for Grown Ups); directors are Clint Eastwood (known for Million Dollar Baby) and Spike Lee (known for Malcolm X); producers are Mark Burg (known for Two and a Half Men) and Jack Giarraputo (known for 50 First Dates); studios are Revolu- tion and Warner Brothers. Interestingly, the top chosen genre for a movie for 19-25 year-old men is Romance, with Ned Bellamy (best known for Being John Malkovich) and Alan Rickman (known for the Harry Potter series) as the leading actors, and Ron Howard as the director (known for Cinderella Man). As another example, the movie designed for men of all ages has Martial Arts (the genre for movies like Rush Hour 2) as top-genre, Gary Oldman (known for very popular in the recent years, and efforts have focused on adapt- ing and extending approaches intended for homogeneous graphs, such as similarity search [29] and random walk with restarts [4, 24]. Sun et al. [29] introduced the idea of meta-paths for similar- ity search. In our work, we use meta-paths or predefined paths to perform 'random' walks, with the ultimate goal of inferring user- feature preferences for the MD problem. The heterogeneous entity recommendation approach [33] addresses the problem of top-k en- tity recommendation by performing matrix factorization on user- movie preference matrices obtained by employing the definition of meta-path similarity [29] (normalized path counts). The goal of [33] is to rank existing items by user interest, assuming that all the items that are rated by a user are of interest to her (although she might have disliked some of the items). This is a weaker prob- lem than the one that our work and typical recommender systems tackle. Moreover, [33] cannot directly infer the user-feature pref- erences, making it unsuitable for the MD problem. Unlike these works, we infer the user preferences with respect to item features by defining random-walk scores over meta-paths and effectively in- corporate likes and dislikes, which leads to high accuracy. Influence Maximization. Influence maximization is a funda- mental underlying problem in viral marketing [9], the early adop- tion of products [28], targeted advertising, and more. The goal is to identify which users to target in order to maximize the adoption of an existing product, or service. One of the most influential papers in the area by Kempe, Tardos, and Kleinberg [15] introduced the independent cascades and linear threshold models for user conver- sion; more efficient models have been proposed since then [10]. In our work, we use a variation of the linear threshold model for the MD problem. 6. CONCLUSIONS This paper introduces the MOVIE-DESIGN problem for specific target audiences by leveraging user-movie preferences and movie content. The MD problem that we have proposed is complemen- tary to recommendation systems. In addition to contributing a novel formulation of the MD problem as an optimization problem, we have introduced a new random walk-based algorithm, PNP on a POSITIVE and NEGATIVE graph, which efficiently handles dislikes in the user preferences. We have shown that PNP is superior to baseline methods in terms of movie preference predictions. Finally, we have applied our method on large, real-world datasets to gen- erate movies for specific audiences, and introduced ways for the qualitative and quantitative evaluation of the designs. Although we have tailored our approach to the design of movies, it can be gen- eralized to other products as long as reviews and product features can be identified. Future work includes extending our method with other elements of the creative movie process, notably the plot of the movie, the screenplay, and music. There are additional signals that can be in- corporated, such as, the credits order for the actors in the movie as an indicator of their contribution and weights on a movie's genre, e.g., for a movie that is mostly drama yet has an element of ro- mance, we can weigh those two genres unequally. 7. REFERENCES [1] A. A. Ageev and M. I. Sviridenko. Pipage rounding: a new method of constructing algorithms with proven performance guarantee. Journal of Combinatorial Optimization, 8, 2004. [2] Aris Anagnostopoulos, Luca Becchetti, Carlos Castillo, Aristides Gionis, and Stefano Leonardi. Online Team Formation in Social Networks. In WWW, 2012. (a) kNN scores per target audience. (b) W-kNN scores per target audience. Figure 6: PNP outperforms the TOP and POPULAR movie designs. the Batman series) and Matt Damon (Bourne series) as lead actors, and Steven Spielberg as lead director. Our model leverages the dis- tinct patterns in the movie interests of different demographics to design movies that specifically cater to those audiences. That said, movie making is a highly creative process, in which there are many elements that need to come together. Taking such a data-driven ap- proach to designing movies is a step towards the world of targeted and personalized movies. 5. RELATED WORK Next we review three main related areas: (i) recommendation systems, (ii) analysis of heterogeneous graphs, and (iii) influence maximization. Recommendation systems. Our work is related to recommen- dation methods, and specifically group recommendation, which seeks to recommend existing items that are likely to match the tastes of multiple users at the same time [25, 14]. The area of recommen- dation systems is very active with numerous algorithms that lever- age the user preferences and similarities [13, 7], or item similari- ties [27, 20] in order to provide personalized recommendations to the users. The user-based CF approaches suffer from the sparsity problem, especially for new users (cold-start problem), which leads to very low prediction quality and scalability. To overcome these problems, [21] leverages the social interactions and connections be- tween users, [34] proposes a unified approach that combines CF with friendships and memberships, and content-based methods use side information, such as demographics, and item features (e.g., stylistic visual attributes) [23, 26, 11]. A popular, scalable and ac- curate approach is matrix factorization [16], which relies on latent factors for both users and items. Privacy is always a major con- cern in online systems including recommendation systems, which has led to privacy-preserving systems [22, 6]. Despite the similar- ities with group recommendation, our objective is to design a new movie (or, generally, a new product), so that the expected number of endorsers in the intended audience is maximized. Heterogeneous graphs. Heterogeneous networks have become [3] Marc Andreessen. Part 4: The only thing that matters. http://pmarchive.com/guide_to_startups_part4.html, 2007. [4] Andrey Balmin, Vagelis Hristidis, and Yannis Papakonstantinou. Objectrank: Authority-based keyword search in databases. In VLDB, 2004. [5] James Bennett and Stan Lanning. The Netflix prize. In KDD Cup, 2007. [6] Smriti Bhagat, Udi Weinsberg, Stratis Ioannidis, and Nina Taft. Recommending with an Agenda: Active Learning of Private Attributes Using Matrix Factorization. In RecSys, 2014. [7] John S. Breese, David Heckerman, and Carl Myers Kadie. Empirical Analysis of Predictive Algorithms for Collaborative Filtering. In UAI, 1998. [8] David Carr. Giving viewers what they want. New York Times, 2013. [9] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for prevalent viral marketing in large-scale social networks. In KDD, 2010. [10] Wei Chen, Yajun Wang, and Siyu Yang. Efficient influence maximization in social networks. In KDD, 2009. [11] Yashar Deldjoo, Mehdi Elahi, Paolo Cremonesi, Franca Garzotto, Pietro Piazzolla, and Massimo Quadrana. Content-based video recommendation system based on stylistic visual features. J Data Semant, 2016. [12] Stephan Dernbach, Nina Taft, Jim Kurose, Udi Weinsberg, Christophe Diot, and Azin Ashkan. Cache Content-Selection Policies for Streaming Video Services. 2016. [13] Yifan Hu, Yehuda Koren, and Chris Volinsky. Collaborative Filtering for Implicit Feedback Datasets. In ICDM, 2008. [14] Anthony Jameson and Barry Smyth. In The Adaptive Web, chapter Recommendation to Groups. Springer-Verlag, 2007. [15] David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of influence through a social network. In KDD, 2003. [16] Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender systems. Computer, 42(8), 2009. [17] Theodoros Lappas, Kun Liu, and Evimaria Terzi. Finding a Team of Experts in Social Networks. In KDD, 2009. [18] Jessica Leber. "House of Cards" and our future of algorithmic programming. MIT Technology Review, 2013. [19] Liangyue Li, Hanghang Tong, Nan Cao, Kate Ehrlich, Yu-Ru Lin, and Norbou Buchler. Replacing the Irreplaceable: Fast Algorithms for Team Member Recommendation. In WWW, 2015. [20] G. Linden, B. Smith, and J. York. Amazon.com recommendations: item-to-item collaborative filtering. IEEE Internet Comput., 7(1):76–80, 2003. [21] Hao Ma, Haixuan Yang, Michael R. Lyu, and Irwin King. SoRec: Social Recommendation Using Probabilistic Matrix Factorization. In CIKM, 2008. [22] Frank McSherry and Ilya Mironov. Differentially Private Recommender Systems: Building Privacy into the Netflix Prize Contenders. In KDD, 2009. [23] Prem Melville, Raymod J. Mooney, and Ramadass Nagarajan. Content-boosted Collaborative Filtering for Improved Recommendations. In AAAI, 2002. [24] Zaiqing Nie, Yuanzhi Zhang, Ji-Rong Wen, and Wei-Ying Ma. Object-level ranking: bringing order to web objects. In WWW, 2005. [25] Mark O'Connor, Dan Cosley, Joseph A. Konstan, and John Riedl. Polylens: A recommender system for groups of users. ECSCW, 2001. [26] Royi Ronen, Noam Koenigstein, Elad Ziklik, and Nir Nice. Selecting content-based features for collaborative filtering recommenders. In RecSys, 2013. [27] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. Item-based Collaborative Filtering Recommendation Algorithms. In WWW. ACM, 2001. [28] Yaron Singer. How to win friends and influence people, truthfully: influence maximization mechanisms for social networks. In WSDM, pages 733–742. ACM, 2012. [29] Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S. Yu, and Tianyi Wu. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. PVLDB, 2011. [30] Technicolor. MGO. http://www.mgo.com/, 2014. [31] Madeleine Udell and Stephen Boyd. Maximizing a sum of sigmoids, 2014. Working paper. [32] Vijay V Vazirani. Approximation algorithms. Springer Science & Business Media, 2013. [33] Xiao Yu, Xiang Ren, Yizhou Sun, Quanquan Gu, Bradley Sturt, Urvashi Khandelwal, Brandon Norick, and Jiawei Han. Personalized entity recommendation: A heterogeneous information network approach. In WSDM, 2014. [34] Quan Yuan, Li Chen, and Shiwan Zhao. Factorization vs. Regularization: Fusing Heterogeneous Social Relationships in Top-n Recommendation. In RecSys. ACM, 2011. [35] R. Zafarani and H. Liu. Social computing data repository at ASU, 2009.
1908.02951	2	1908	2019-12-24T13:32:20	Research Leadership Flow Determinants and the Role of Proximity in Research Collaborations Networks	[ "cs.SI", "cs.DL", "physics.soc-ph" ]	Characterizing the leadership in research is important to revealing the interaction pattern and organizational structure through research collaboration. This research defines the leadership role based on the corresponding author's affiliation, and presents, the first quantitative research on the factors and evolution of five proximity dimensions (geographical, cognitive, institutional, social and economic) of research leadership. The data to capture research leadership consists of a set of multi-institution articles in the life sciences and biomedical during 2013-2017 from Web of Science Core Citation Database. Our sample consists of 484,903 articles from 244 Chinese institutions, which have been the primary affiliation of the corresponding author for at least one paper (with multiple institutions) in each year. A Tobit regression-based gravity model indicates that research leadership mass of both the leading and participating institutions and the geographical, cognitive, institutional, social and economic proximity are important factors of the flow of research leadership among Chinese institutions. In general, the effect of these proximity for research leadership flow has been declining recently. The outcome of this research sheds light on the leadership evolution and flow among Chinese institutions, and thus can provide evidence and support for grant allocation policy to facilitate scientific research and collaborations.	cs.SI	cs	Research Leadership Flow Determinants and the Role of Proximity in Research Collaborations Chaocheng Hea,b Jiang Wua Qingpeng Zhangb,1 a. School of Information Management, Wuhan University, Wuhan, Hubei, China. b. School of Data Science, City University of Hong Kong, Kowloon, Hong Kong, China. Abstract Characterizing the leadership in research is important to revealing the interaction pat- tern and organizational structure through research collaboration. This research defines the leadership role based on the corresponding author's affiliation, and presents, the first quantitative research on the factors and evolution of five proximity dimensions (geographical, cognitive, institutional, social and economic) of research leadership. The data to capture research leadership consists of a set of multi-institution articles in the fields of "Life sciences & biomedicine", "Technology", "Physical Sciences", "Social Sciences" and "Humanities & Arts" during 2013-2017 from Web of Science Core Ci- tation Database. A Tobit regression-based gravity model indicates that research leader- ship mass of both the leading and participating institutions and the geographical, cog- nitive, institutional, social and economic proximities are important factors of the flow of research leadership among Chinese institutions. In general, the effect of these prox- imities for research leadership flow has been declining recently. The outcome of this research sheds light on the leadership evolution and flow among Chinese institutions, and thus can provide evidence and support for grant allocation policy to facilitate sci- entific research and collaborations. Keyword Research collaborations, Social network analysis, Collaboration networks, Research leadership, Gravity model, Proximity dimensions 1 Corresponding author: Qingpeng Zhang ([email protected]); all authors have equal contributions. 1 Introduction Research collaboration is highly valued by academic communities for that it com- bines and diffuses complementary knowledge expertise of different background (Padial, Nabout et al. 2010). Research collaborations have been increasingly prevalent in aca- demia (Sun, Wei et al. 2017). Research collaboration is defined as the working together of researchers to achieve a common goal of producing new scientific knowledge (Katz and Martin 1997). It has been recognized that the research collaborations among re- searchers/institutions/countries form a task-oriented complex systems with various in- teraction and organizational patterns (Newman 2004). Characterizing the collaboration patterns have been a focus of scientometrics (Nguyen, Ho-Le et al. 2017). In particular, identifying key proximity factors that pro- mote research collaboration can inform actionable policy making to facilitate effective collaborations that carry out good research, which could eventually enhance the pro- spects of science as a whole (Scherngell and Hu 2011). Existing research mainly fo- cuses on examining the geographic, socioeconomic, and cognitive proximity as poten- tial determinants of research collaboration using a gravity model (Hoekman, Frenken et al. 2010, Plonikova and Rake 2014, Sidone, Haddad et al. 2017). To summarize, the intensity of research collaboration is greater if researchers are closer to each other geo- graphically (physical distance between the two researchers/institutions/countries) (Fer- nandez, Ferrandiz et al. 2016, Jiang, Zhu et al. 2018); socioeconomic factors such as educational R&D are positively correlated with research activities (Wagner, Park et al. 2015); the cognitive factors such as the similarity in research topics also influence the intensity of research collaboration (Gilsing, Nooteboom et al. 2008). Therefore, under- standing the proximity effect is important to characterize the research collaborations. The previous studies on research collaboration and its proximity factors have three limitations. First, the collaboration relationship was homogeneously assigned among authors for the same article, without considering the research leadership. However, the collaboration relationship between two non-corresponding authors is different from the relationship between a co-author and the corresponding author, who is the leader of the research (Jarneving 2010, Wang, Xu et al. 2013) . Typically, the first author is usually an early-career scientist who undertakes the research (e.g. a PhD student or postdoc); the corresponding author is usually the senior person who supervises and shapes the study (e.g. supervisor) (Sekara, Deville et al. 2018). Other co-authors are usually the researchers who assist the first and corresponding authors in the carrying out the work or writing the paper (not necessarily in the same group) (Hu, Rousseau et al. 2010). Therefore, the common practice is to define the first and corresponding author as the research leaders. The relationship between co-authors and research leaders are stronger than those among non-leaders (Wang, Xu et al. 2013). With the recent trend of the in- creasing size of multi-institution authorship, the effect of research leadership is becom- ing even more significant (Wang and Wang 2017). Second, although Boschma (2005) identified five notions of proximity (geographical, cognitive, institutional, organiza- tional, and social), most studies failed to systematically and comprehensively examine the relationships between these key identified factors and research collaboration (Fer- nández, Ferrándiz et al. 2016, Jiang, Zhu et al. 2018). There are, to the best of our knowledge, only two studies involving all five proximity dimensions, but these two studies used static data without revealing the evolution of these five proximities (Plot- nikova and Rake 2014, Fernandez, Ferrandiz et al. 2016). Third, existing studies were mainly based on regional/country-level data, which is with a relatively low resolution of depicting the research collaborations and leadership (Hoekman, Frenken et al. 2010, Gui, Liu et al. 2018). Studying the collaborations at the institutional-level can provide high-resolution of collaboration patterns and inform actionable policy making, and thus is critically needed for characterizing the flow of research leadership. To fill these research gaps, we extend the literature in several ways. First, we intro- duce a new measure, research leadership (RL), where we only account for the collabo- rative relationship between a co-author and the first author or corresponding author. Second, we focus on proximity as one of the main determinants of research collabora- tion (Gui, Liu et al. 2018, Plonikova and Rake 2014). Following Boschma (2005), we examine the effect of all the four proximity dimensions (geographical, cognitive, insti- tutional, and social proximity) on research collaboration using a gravity model. Note that we do not have organizational proximity since we are focusing on the RL among institutions in specific domains. In addition, we also examine the effect of economic proximity, which has been found to be associated with research performance and col- laborations (Hwang 2008, Acosta, Coronado et al. 2011). Third, we examine the re- search collaboration between Chinese institutions. Because the corresponding author- ship is explicitly set to be the primary criterion in the research evaluation system in China (for example, promotion practice and grant review) (Hu, Rousseau et al. 2010). The paper is organized as follows. Section 2 reviews the literature. Section 3 de- scribes the methods and data. Section 4 and Section 5 present the results and robustness check, respectively. The paper is concluded in Section 6 with discussions of the limita- tions and future work. 2 Literature review 2.1 Research collaboration Research collaboration is a form of intense interaction that allows for effective com- munication and sharing of capabilities and other resources (Heffner 1981). Research collaboration may lead to various outcomes, such as co-authored paper, patents, deep- ened personal contact, or nothing at all. Among these outcomes, journal papers have been widely adopted as a reliable proxy to measure the intensity of research collabora- tion (Newman 2004, Acosta, Coronado et al. 2011). Therefore, the volume of co-au- thored journal papers is adopted to reflect the degree of intensity of research collabora- tion. 2.2 Research leadership in research collaboration It has been recognized that the first author and corresponding author often lead the research collaboration in most fields such as biological, engineering, and management (Sekara, Deville et al. 2018). Typically, the person who carries out the research and writes the paper is the first author (Hu, Rousseau et al. 2010). The corresponding author, on the other hand, is responsible for leading the research team, securing grant to sustain the study, shaping the research ideas, designing the roadmap of a project and corre- sponding with editors, coordinating the collaboration among other co-authors (Wang and Wang 2017). The other co-authors are usually with complementary contributions to assist in undertaking the research. Recently, there is a trend of the increasing size of authorship and more frequent cross-institutional collaborations (Osorio 2018), and thus makes the leadership role of the corresponding author more pronounced (Wang, Wu et al. 2014). At institution-level, the corresponding author's affiliation is commonly recognized as the leading institution in the project (Alvarez-Betancourt and Garcia-Silvente 2014). Therefore, the corresponding authorship is widely adopted to be a key measure of re- search contribution and the leadership role (Hu, Rousseau et al. 2010). In many coun- tries/regions (such as China), the corresponding authorship is set to be the primary cri- terion for research evaluations (for example, promotion and grant review) (Hu, Rous- seau et al. 2010). Although the first authorship is also a key indicator of leadership, the first and corresponding authors usually belong to the same institution (for example, the first author is the student/postdoc of the corresponding author) (Wang, Wu et al. 2014). Therefore, we set the corresponding author's affiliation as the leader in cross-institution collaborations. 2.3 Proximity of research collaboration The dynamic interplay of various proximity factors influences the outcome of re- search collaboration (Hoekman, Frenken et al. 2010). Proximity is an underpinning the- oretical framework in investigating the determinants of research collaboration (Berge 2017, Hoekman, Frenken et al. 2010, Gui, Liu et al. 2018, Plonikova and Rake 2014). Boschma (2005) identified five commonly adopted proximities dimensions: geograph- ical, cognitive, institutional, organizational, and social. Recent research has highlighted the significant role that the economic proximity plays during research collaboration. Since we focus on the research leadership between institutions in specific domains, the organizational proximity is not applicable here. The detailed descriptions of selected proximities are as follows. Geographical distance refers to the physical distance between researchers, which can be measured by absolute metrics such as kilometers, or relative metrics such as travel times (Plotnikova and Rake 2014). There is plenty of evidence showing that the geo- graphical distance between researchers is negatively associated with the intensity of collaborations between them, indicating that researchers/institutions that are close to each other are more likely to collaborate (Hoekman, Frenken et al. 2010). Cognitive proximity refers to the similarity between researchers' research interests and background. The cognitive proximity partly determines the researchers' absorptive capacity in collaborative projects. The cognitive proximity has been recognized to be positively associated with the intensity of collaborations (Scherngell and Barber 2011). There is also evidence showing that close cognitive proximity does not always benefit research collaboration because of the lack of complementary backgrounds in the re- search team (Gilsing, Nooteboom et al. 2008, Balland 2012). Institutional proximity is defined as the similarity between institutions, such as cul- tural norms and economic development. High institutional proximity could reduce the uncertainty and costs of interactions, and facilitate the establishment of mutual trust (Boschma 2005, Boschma and Frenken 2009). It is usually modeled as a dummy vari- able to capture whether the collaborators belong to the same administrative division (such as province and state) or the same country. It has been recognized that research collaboration is more frequent among institutions within the same administrative divi- sion/ country/ linguistic area (Hoekman, Frenken et al. 2010). Social proximity refers to the social embeddedness based on the more intangible closeness, such as friendship or prior collaborations between researchers, which is ex- pected to stimulate knowledge interactions and research collaborations owing to the established trust and commitment (Coenen, Moodysson et al. 2004, Boschma 2005). Social proximity is usually measured by the existing/prior collaborations and has been found to be positively associated with future collaborations between individuals, insti- tutions, and countries (Niedergassel and Leker 2011, Hoekman, Scherngell et al. 2013, Plotnikova and Rake 2014) . Economic proximity, also named as the socioeconomic proximity, refers to the degree of closeness of academic-related economic resource between institutions (Fernandez, Ferrandiz et al. 2016). According to the center-periphery hypothesis, the difference in academic economic resources between institutions could influence the research collab- orations (Schubert and Sooryamoorthy 2009). Economic proximity has been found to have a mixed effect on the collaboration intensity at country level (Hwang 2008, Par- reira, Machado et al. 2017, Ling, Zhu et al. 2018). There is few research on examining the influence of economic proximity on the collaboration between institutions. Table 1. Summary of Literature in Collaboration Proximity Proximity Measurement Remarks and Citation The distance between the capitals of the col- laborating countries The distance between collaborating institu- tions It measures the research collaborations at the country level (Gui, Liu et al. 2018, Jiang, Zhu et al. 2018) It measures the research collaboration at the institution level (Fernandez, Ferrandiz et al. 2016) Geographical distance The distances between the biggest cities of It measures the research collaborations at the the two countries country level (Plotnikova and Rake 2014) The distance between the geographic centers of the regions It measures the research collaboration at the regional level (Scherngell and Hu 2011, Berge 2017). The distance between the latitudinal and lon- It measures the research collaboration at the gitudinal centroid of each country country level (Parreira, Machado et al. 2017). The cosine similarity of the publication It measures the research collaboration in mul- counts in 12 disciplines between two institu- tiple research fields at the institution level tions (Fernandez, Ferrandiz et al. 2016). The cosine similarity of the publication counts in 6 disciplines between two countries It measures the research collaboration in mul- tiple research fields at the country level (Hoekman, Frenken et al. 2010). Cognitive The cosine similarity of the publication It measures the research collaboration in phar- proximity counts in 15 sub-areas of therapeutic between maceutical research at the country level (Plot- two countries nikova and Rake 2014). The cosine similarity of the publication It measures the research collaboration in mul- counts in 22 disciplines between two coun- tiple research fields at the country level (Gui, tries Liu et al. 2018). The simultaneous use of two dimensions: It measures the university -- industry collabora- tion (Garcia, Araujo et al. 2018). the cosine similarity of the publication counts in 6 scientific fields and the cosine similarity of the publication counts in 12 industry sectors pairs Dummy variable: It measures cross-country research collabora- 1 if two institutions are in the same country; tion at institutional level (Hoekman, Frenken 0 otherwise. Dummy variable: 1 if two municipalities both have public uni- versity campuses; Institutional 0 otherwise. proximity Dummy variable: et al. 2009, Fernandez, Ferrandiz et al. 2016). It measures research collaboration between two municipalities in Brazil (Sidone, Haddad et al. 2017). 1 if two countries share a common official It measures research collaboration at country language; 0 otherwise. level (Gui, Liu et al. 2018). The cosine similarity of 9 indicators in Insti- tutional Profiles Database between two coun- tries Dummy variable: It measures research at country level (Plot- nikova and Rake 2014). It measures research collaboration at institu- 1 if two institutions have prior collaborated; tional level (Petruzzelli 2011, Fernandez, 0 otherwise. Ferrandiz et al. 2016). The count of joint-publications by the two It measures research collaboration at country countries level (Plotnikova and Rake 2014). The degree of overlap of two countries' prior It measures research collaboration at country level (Gui, Liu et al. 2018). The stochastic actor-based model SIENA to model the network dynamic and to estimate parameters (Balland 2012). It measures cross-country research collabora- tion at institutional level (Fernandez, Ferrandiz et al. 2016). It measures research collaboration among Eu- ropean universities (Acosta, Coronado et al. 2011). It measures research collaboration at country level (Jiang, Zhu et al. 2018) It measures research collaboration at regional level in China (Scherngell and Hu 2011). Social proximity collaborators Dummy variable: 1 if two collaborators have a partner in com- mon; 0 otherwise. Absolute difference in the average Higher ed- ucation R&D expenditures as % of the GDP. Economic proximity Absolute difference in per capita higher edu- cation R&D expenditure Absolute difference in real GDP per capita Absolute difference of gross regional product 2.4 Impact Measures using Gravity model Spatial interaction patterns, such as inter-regional/international trade, population commuting, and migration, can be modeled by the analogy to Newton's law of universal gravitation (Burger, Van Oort et al. 2009). The basic idea of the gravity model stems from the law of universal gravitation, hypothesizing that the gravitational force between two objects is directly related to the mass of the objects, and is inversely proportional to the physical distance between them. Initially, the gravity model has been successfully used in predicting trade flows between countries (Isard 1954, Anderson 1979). The model has been applied to a variety of fields. In particular, because the gravity model can capture the joint-effect of both the mass (for example, the number of publications) and distance/proximity (for example, physical distance) of entities, it has been exten- sively applied to modeling research collaborations (Acosta, Coronado et al. 2011, Parreira, Machado et al. 2017). In this paper, we will integrate the five measures of proximity to measure their impact on RL flows using a gravity model. 3 Methods and Data 3.1 Measurement of research leadership The intensity of research collaboration captures the frequency of the research collab- oration between two corresponding institutions. In the literature, there are mainly two measures of the intensity of research collaborations: the full count and the fractional count (Berge 2017). The full count is 1 if a paper at least one coauthor from institution 𝑎 and at least one coauthor from institution 𝑏, while the fractional count is 2/(𝑁-1)𝑁, where 𝑁 is the number of institutions. Following Jiang, Zhu et al. (2018), we used the fractional count to measure research collaboration intensity as it takes the contribution into consideration, instead of simply measuring the participating frequency. Because we defined the role of leadership, there exists the relationship between leaders and oth- ers being led to conduct the research. Therefore, the RL is directed (from the leading institution to non-leading participating institutions) and weighted (on the frequency of prior collaborations). We define that a paper possesses a total leadership mass of 1, and the RL flows from the leading institution (the affiliation of the corresponding author) to all other participating institutions (the affiliations of co-authors). Since we focus on the flow of RL, we do not consider a complete network of coauthors; instead, we con- sider a star-like network (Figure 1) with the leading institution in the center and partic- ipating institutions connecting to it (including a self-loop from the leading institution to itself). If there are multiple leading institutions (a paper with multiple corresponding authors affiliated with multiple institutions), the RL mass contributed by them is evenly distributed (Figure 1). In our data, papers with multiple leading institutions account for 3%. The RL flow intensity 𝐶𝑎𝑏,𝑖 from the leading institution 𝑎 to institution 𝑏 in the paper 𝑖 is expressed as 𝐶𝑎𝑏,𝑖 = 1 𝐿𝐼𝑁𝑖 × 1 𝑁𝑖 , (1) where 𝐿𝐼𝑁𝑖 is the number of leading institutions in paper 𝑖. And 𝑁𝑖 is the number of institutions in paper 𝑖. Thus, the RL mass contributed by the leading institution is the total RL flow minus the self-loop: 𝑁𝑖−1 𝐿𝑀𝑎,𝑖 = ∑ 𝐶𝑎𝑏,𝑖 = 𝑏=1 1 𝑁𝑖 × 1 𝐿𝐼𝑁𝑖 × (𝑁𝑖 − 1) = 1 𝐿𝐼𝑁𝑖 (1 − 1 𝑁𝑖 ), (2) Therefore, the more institutions involved in paper 𝑖, the more RL mass contributed by the leading institution 𝑎. And the total RL flow intensity 𝐶𝑎𝑏 from institution a to institution 𝑏 is calculated as 𝑀𝑏 𝐶𝑎𝑏 = ∑ 𝐶𝑎𝑏,𝑖 , (3) 𝑖=1 where 𝑀𝑏 is the number of papers where 𝑎 is the leading institution and 𝑏 is a par- ticipating institution. And institution 𝑎's total RL mass, total RL flow intensity to all other institutions is calculated as 𝐵 𝐿𝑀𝑎 = ∑ 𝐶𝑎𝑏 , (4) 𝑏=1 where 𝐵 is the number of institutions that institution 𝑎 has led. Figure 1 The RL flow from the leading institution(s) to participating institutions 3.2 Data We performed a data collection in Web of Science Core Citation Database. More specifically, the data was retrieved using the search term CU = A AND SU = B AND PY = C, where A is "PEOPLES R CHINA", B is research areas in a certain field2, and C is 2013-2017. Eventually, we obtained the complete authorship and collaboration in- formation of 886105 articles published by all Chinese institutions during 2013-2017. Eventually, for "Life Sciences & Biomedicine" field, we sampled 484,903 articles pub- lished by 244 Chinese institutions, which have been the primary affiliation of the cor- responding author for at least one paper (with multiple institutions) in each year. Simi- larly, we sampled 179,840 articles by 209 Chinese institutions for "Technology" field, 2 http://images.webofknowledge.com/WOKRS532FR6/help/WOS/hp_research_areas_easca.html I7I4I5I3I8I6I2I1Paper mAuthor A1, Institution I1Author A2, Institution I2Author A3, Institution I3Author A4, Institution I4Author A5, Institution I5Author A6, Institution I6Author A7, Institution I7Author A8, Institution I8Author A9, Institution I1Author A10, Institution I2Corresponding Author :Author A1, Institution I1Author A2, Institution I2I7I6I2I4I5I3I1Paper nAuthor A1, Institution I1Author A2, Institution I2Author A3, Institution I3Author A4, Institution I4Author A5, Institution I5Author A6, Institution I6Author A7, Institution I7Author A8, Institution I1Corresponding Author :Author A1, Institution I1A paper with one leading institutionA paper with multipal leading institutions204,320 articles in "Physical Sciences" by 221 Chinese institutions, 13,628 articles in "Social Sciences" by 95 Chinese institutions, and 3,414 articles in "Humanities & Art" by 41Chinese institutes. 3.3 Model and variables To analyze the determinants of research leadership among different institutions, we adopt a gravity model. In its elementary form, the gravity model can be expressed as 𝐶𝑖𝑗 = 𝐾 𝑀𝑖 𝛽2 𝛽1 × 𝑀𝑗 𝛽3 𝑑𝑖𝑗 , (5) where 𝐶𝑖𝑗 is the collaboration intensity (i.e., the number of co-publication) between institutions 𝑖 and 𝑗; K is a proportionality constant; 𝑀𝑖 and 𝑀𝑗 are the numbers of publications of institutions 𝑖 and 𝑗; 𝑑𝑖𝑗 is the geographical distance between the two institutions; 𝛽1 and 𝛽2 reflect the potential to collaboration and 𝛽3 is an impedance factor reflecting the distance decaying factor for research collaboration. Taking loga- rithms of both sides and adding a random disturbance term, the multiplicative form of gravity model (1) can be converted into a testable linear stochastic model: ln 𝐶𝑖𝑗 = ln 𝐾 + 𝛽1 ln 𝑀𝑖 + 𝛽2 ln 𝑀𝑗 − 𝛽3 ln 𝑑𝑖𝑗 + 𝜖𝑖𝑗, (6) And there has been a substantial body of literature involving the subject of which regression model are suitable to estimate (6) (Burger, Van Oort et al. 2009). Because RL of an institution reflects the extent to which it leads collaborations, and the afore- mentioned proximity measures quantify the distance between institutions, the gravity model is an appropriate modeling framework for our study. Given the fractional count nature of the data and the existence of a large number of zeros (many institution pairs have no research collaboration), the ordinary least square (OLS) estimates with a censored dependent variable may be biased and inconsistent. Tobit regression, one of the limited dependent variable models (Li 2002), can effec- tively estimate linear relationship among variables when there was either left- or right- censoring in the dependent variables (Wooldridge 2003, Zhang and Lin 2018). In line with previous studies (Plotnikova and Rake 2014, Fernandez, Ferrandiz et al. 2016), we adopt a Tobit regression model where we consider zero leadership as left-censoring of the distribution. To explore the roles and their dynamic evolutions of multiple proximities in shaping RL flows, we first conduct a cross-section estimate by the pooling data of 2013-2017, and then we perform cross-section estimates using two sub-period data. The equation that will be estimated is: 𝐶𝑖𝑗 = 𝛽0 + 𝛽1 𝑙𝑛(𝐿𝑀𝑖) + 𝛽2 𝑙𝑛(𝐿𝑀𝑗) + 𝛽3 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) + 𝛽4 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) + 𝛽5𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 + 𝛽6𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 + 𝛽7𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) + 𝜖𝑖𝑗, (7) The dependent variable 𝐶𝑖𝑗 in our model is the RL flow intensity from institution 𝑖 (the leading institution) to institution 𝑗 during the period 2013-2017. In independent variables, time lags are used to avoid endogeneity and reverse causality (Fernandez, Ferrandiz et al. 2016, Zhang 2016, Gui, Liu et al. 2018). Specifically, the independent variables are lagged, capturing the information for 2008-2012. They are defined as fol- lows, The 𝐿𝑀𝑖 and 𝐿𝑀𝑗 refer to the RL mass of leading institution 𝑖 and participating institution 𝑗 respectively during 2008-2012. Geographical proximity (𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗 ) is the spatial distance between institution 𝑖 and institution 𝑗. It's calculated following the big circle formula with the latitude and longitude of the institution from Google Map (Gui, Liu et al. 2018). Cognitive proximity represents the extent of overlap or the closeness of researchers' knowledge. In this paper, we adopt the Latent Dirichlet Allocation (LDA) model, to categorize the papers into 50 topics based on the keywords of each paper. And then we embed all the institutions into a 50-dimensional feature vector according to their publi- cations during 2008-2012. The cognitive proximity between two institutions is calcu- lated as the cosine similarity of two corresponding feature vectors3. Institutional proximity (𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗) is defined by the similarity on both formal and informal laws and practice. China's research resource and policy is distributed at the provincial level. In this paper, institutional proximity is a dummy variable (Lander 3 Supplementary material provides a details of measurement of cognitive proximity. 2015), which equals to 1 if institutions 𝑖 and 𝑗 are in the same province, and 0 other- wise. Note that in our data, there is little collinearity concern between 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 and 𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗. Social proximity (𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗) draws on the social embeddedness based on the more intangible closeness, e.g., friendship or prior collaboration experience between re- searchers. Here, social proximity is a dummy variable, which equals to 1 if 𝐶𝑖𝑗 >0 or 𝐶𝑗𝑖 >0 during 2008-2012, and 0 otherwise (Fernandez, Ferrandiz et al. 2016). Economic proximity (𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) indicates the difference in academic resources between institutions 𝑖 and 𝑗. Acosta and Coronado used the difference in the total amount of R&D to measure the economic proximity between two entities (Acosta, Coronado et al. 2011). In this paper, we measure the 𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗 as the absolute dif- ference in the number of National Natural Science Foundation of China (NSFC) pro- jects secured by institutions 𝑖 and 𝑗 during 2008-2012.4 4 http://www.nsfc.gov.cn/ Table 2. Description of dependent variables and independent variables Variable 𝐶𝑖𝑗 𝐿𝑀𝑖 𝐿𝑀𝑗 Description Source RL flow intensity from i ( leading institution) to j ( participating institu- Web of Sci- tion) in the period 2013-2017 ence RL mass of institution i (leading institution) in period 2008-2012 (Varia- Web of Sci- ble in logarithms) ence RL mass of institution j (participating institution) in period 2008-2012 Web of Sci- (Variable in logarithms) ence 𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗 Geographical distance between institution i and j, in kilometers (Variable Google Map in logarithms) 𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗 Cosine similarity between institution vector pairs in period 2008-2012 Web of Sci- (Variable in logarithms) ence 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 Dummy variable, which take value 1 when institution i and j are in the Google Map same province 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 Dummy variable, which takes value 1 if institution i and j have collabo- Web of Sci- rated in period 2008-2012 ence 𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗 Absolute difference in the number of Natural Science Foundation Project NSF, China of institution i and j in period 2008-2012 (Variable in logarithms) 4 Results For simplicity, we mainly introduce the results in the field of "Life sciences & biomed- icine". We introduce the difference of the results in other fields in sub-section 4.4. 4.1 Descriptive analysis For contextualization, we present the top 15 institutions by RL mass (𝐿𝑀𝑖) and top 15 institution pairs by RL flow intensity (𝐶𝑎𝑏) in Table 3 and Table 4, respectively. And Figure 2 and 3 show the spatial pattern and the dynamic pattern of RL flows, re- spectively. As shown in Table 3, all these top ranked institutions (with large RL mass) are fa- mous universities or institutions in biological and medical research in China. In partic- ular, the top four institutions are the only four Chinese universities ranked top 200 in the field of Clinical Medicine and Pharmacy in Academic Ranking of World Universi- ties 2016 edition5. These 15 institutions account for 33.41% of total RL among the 244 Chinese institutions. It's worth noting that the rank here only measures the RL in the collaborations with Chinese institutions. The RL in international collaborations is not accounted. Table 3 Top 15 institutions by leadership mass Institution Leadership Mass (𝐿𝑀) 𝐿𝑀/Total (%) Acum (%) Shanghai Jiao Tong Univ Peking Univ Sun Yat Sen Univ Fudan Univ Zhejiang Univ Shandong Univ Capital Med Univ Nanjing Med Univ Chinese Acad Med Sci Sichuan Univ Chinese Acad Agr Sci 7386.08 7077.45 6344.97 6313.76 5904.04 5785.08 4663.21 4475.98 4359.86 3600.02 3450.40 Huazhong Univ Sci & Technol 3377.42 Peoples Liberat Army Gen Hosp 3331.95 Cent S Univ Second Mil Med Univ Others Total 3155.35 3100.68 123267.94 185114.79 3.99 3.82 3.43 3.41 3.19 3.13 2.52 2.42 2.36 1.94 1.86 1.82 1.80 1.70 1.68 66.59 100 3.99 7.81 3.82 7.25 10.66 13.85 16.98 19.50 21.91 24.27 26.21 28.08 29.90 31.70 33.41 100 5 http://www.shanghairanking.com/FieldMED2016.html Table 4 shows the top-15 institution pairs by RL flow intensity from the leading insti- tution to participating institution. A clear reciprocal pattern presents. Except for the RL flow from "China Agr Univ" to "Chinese Acad Agr Sci", each RL flow its reciprocal form among top 15. For example, the RL flow intensity from "Fudan Univ" to "Shang- hai Jiao Tong Univ" is the largest, and the RL flow intensity from "Shanghai Jiao Tong Univ" to "Fudan Univ" is the second largest. Furthermore, all the top 15 institution pairs consist of institutions in the same city. Table 4 Top 15 Institution pairs by leadership intensity Leading institution Leading city Participating institution Participating city RL flow intensity Fudan Univ Shanghai Shanghai Jiao Tong Univ Shanghai 646.47 Shanghai Jiao Tong Univ Shanghai Fudan Univ Shanghai 601.55 Sun Yat Sen Univ Guangzhou Guangzhou Med Univ Guangzhou 377.53 Shanghai Jiao Tong Univ Shanghai Tongji Univ Shanghai 351.73 Peking Univ Beijing Capital Med Univ Beijing 324.11 Nanjing Univ Nanjing Nanjing Med Univ Nanjing 315.64 Tongji Univ Shanghai Shanghai Jiao Tong Univ Shanghai 297.83 Second Mil Med Univ Shanghai Shanghai Jiao Tong Univ Shanghai 274.93 Fudan Univ Shanghai Tongji Univ Shanghai 270.54 Shanghai Jiao Tong Univ Shanghai Second Mil Med Univ Shanghai 261.11 Nanjing Med Univ Nanjing Nanjing Univ Nanjing 259.17 Capital Med Univ Beijing Peking Univ Beijing 257.21 China Agr Univ Beijing Chinese Acad Agr Sci Beijing 237.37 Tongji Univ Shanghai Fudan Univ Shanghai 224.59 Guangzhou Med Univ Guangzhou Sun Yat Sen Univ Guangzhou 209.28 Figure 2 sheds light on the spatial pattern of RL flows on province level. Generally, the RL flows are mainly concentrated on the eastern part of China, consistent with China's economic and population distributions. Specifically, the most prominent RL flows are from/to Beijing and Shanghai, indicating that many institutions have a RL relationship with institutions in major cities. Although the weights on many edges are small, indicating the low frequency of collaborations between the two provinces, most provinces have RL relationships with each other. Figure 2 Spatial pattern of RL flows at the province level Because of the existence of the proximities, the intensity of RL could be heterogene- ously distributed. For example, institutions in the same city usually have higher chance to collaborate (Table 4). We adopt the disparity metric to measure the heterogeneity of RL intensity distribution. The disparity of institution 𝑖 is calculated as follows, 𝐷𝑖𝑠𝑝𝑎𝑟𝑖𝑡𝑦𝑖 = (𝑁𝑖 − 1) ∑ 𝑁𝑖 𝑗=1,𝑗≠𝑖 2 ) ( 𝐶𝑖𝑗 𝐿𝑀𝑖 − 1 𝑁𝑖 − 2 , (8) where 𝑁𝑖 is the number of institutions that institution 𝑖 has led. If institution 𝑖's RL is evenly distributed to all its participating institutions, 𝐷𝑖𝑠𝑝𝑎𝑟𝑖𝑡𝑦𝑖 → 0. The more un- evenly the institution 𝑖's RL is distributed, the larger the 𝐷𝑖𝑠𝑝𝑎𝑟𝑖𝑡𝑦𝑖 is. In the extreme case, when almost all of 𝑖's RL flow to one participating institution, 𝐷𝑖𝑠𝑝𝑎𝑟𝑖𝑡𝑦𝑖 → 1. Figure 3 shows the kernel density distribution of institutions' leadership disparity dur- ing 2013-2017. In general, most institutions' disparity is small, ranging from 0 to 0.2. There exist a few institutions with very high disparity. Over these five years, the kernel density distribution changed steadily. The disparity of many institutions decreased, re- sulting in higher density around 0.1 and lower density over 0.2. This indicates that the RL flow intensity has become more evenly distributed. A possible explanation is that the hindering effect of proximities has declined over time. We will discuss this in Sec- tion 4.3 Estimation for different sub-periods. Figure 3. Kernel density distribution of institutions' RL disparity over time Table 5 presents the descriptive statistics and the correlation matrix between varia- bles. All variables variance inflation factors (VIFs) are lower than 3, indicating that there is no significant multicollinearity in the data. Table 5 Descriptive statistics and correlations Variables Mean SD VIF 4 5 6 7 8 1.30 9.69 - 1 1 2 4.84 1.27 1.61 0.18* 1 3 4.84 1.27 1.57 0.15* -0.00 1 𝐶𝑖𝑗 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 6.59 1.24 2.33 -0.17* -0.04* -0.04* 1 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) -0.07 0.04 1.63 0.15* 0.43* 0.43* 0.03* 1 0.07 0.25 2.34 0.18* 0.02* 0.02* -0.75* -0.04* 1 0.13 0.34 1.31 0.28* 0.35* 0.28* -0.18* 0.31* 0.18* 1 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 4.54 1.70 1.37 0.10* 0.36* 0.36* -0.03* 0.32* -0.01* 0.23*** 1 p<0.10; * p<0.05; *p<0.01 4.2 Estimation result In order to measure the determinants of research leadership, we also used Tobit gravity model to estimate the impacts. Table 6 reports the estimation results of our Tobit re- gression-based gravity model. Model 1 is the base gravity model, which only includes 𝐿𝑀𝑖, 𝐿𝑀𝑗, and 𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗. Models 2 to 5 have incrementally added independent var- iables. Model 5 presents the full model with all variables. The positive and significant coefficient of 𝐿𝑀𝑖 and 𝐿𝑀𝑗 indicates that the RL mass of both the leading and participating institutions are positively associated with the RL flow intensity. Previous studies showed that the counts of publications of two insti- tutions are positive determinants of their co-publication (Hoekman, Frenken et al. 2010, Parreira, Machado et al. 2017). Our results extended the understanding of the collabo- ration patterns between institutions by showing that the institutions with rich leading experience are more likely to lead future research, as well as to participate in others' research. In addition, we found that the coefficients of 𝑙𝑛(𝐿𝑀𝑖) are larger than that of 𝑙𝑛(𝐿𝑀𝑗) in all models, indicating that there is a larger influence of the leading institu- tion's RL mass on RL flow intensity. 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) _cons N Left-censored Uncensored Table 6 Estimation results of Tobit gravity model Model Model Model Model Model (1) 8.440* (0.105) 7.024* (2) 5.963* (0.114) 4.433* (3) 5.999* (0.114) 4.453* (4) 4.551* (0.116) 3.360* (5) 4.417* (0.124) 3.232* (0.101) (0.114) (0.113) (0.114) (0.122) -4.246* -4.606* -1.963* -1.600* -1.590* (0.087) (0.088) (0.143) (0.142) (0.142) 237.725* 241.254* 200.344* 201.403* (5.894) (5.889) (5.76) (5.783) 15.371* 12.547* 12.648* (0.679) (0.671) (0.673) 13.224* 13.210* (0.326) (0.327) 0.277* (0.089) -71.408* -28.875* -47.347* -41.155* -41.178*** (1.027) (1.352) (1.591) (1.568) (1.569) 59292 46,757 12,535 LR chi2 14871.14 16827.87 17331.69 19000.90 19010.52 Log-likelihood -66308.096 -65329.73 -65077.822 -64243.217 -64238.407 Pseudo R2 Prob > chi2 0.1008 0.0000 0.1141 0.0000 0.1175 0.0000 0.1288 0.0000 0.1289 0.0000 p<0.10; * p<0.05; *p<0.01 The negative and significant coefficient of 𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗 suggests that RL flow intensity decays with longer geographical distance. This is in line with previous literature on research collaboration (Berge 2017, Parreira, Machado et al. 2017). Coordination ac- tivities such as seminars, meetings, exchange of personnel, and sharing lab facilities become more difficult and expensive as distance increases. In addition, successful col- laborative research projects involve intensive face-to-face discussions, which are diffi- cult if the two institutions are far away from each other. Cognitive proximity is positively significant, indicating that RL flows are more likely to occur between institutions with similar research experience. This echoes the previous findings that researchers need a shared cognitive base to understand, absorb, and ex- plore the unknown successfully Boschma (2005). The institutional proximity is positively significant, indicating that positive and sig- nificant, showing that factors such as similar policies and culture background could facilitate the RL flows. This result is in line with existing studies (Fernandez, Ferrandiz et al. 2016), and can also explain the phenomenon that the leading institution pairs are mostly located in the same city. Prior collaborations are found to be positively associated with the RL flows. This result indicates that prior collaborations lead to higher trust and confidence in future collaborations, and is in line with the previous studies (Lazzeretti and Capone 2016). The positive and significant coefficient of economic proximity suggests that RL flows are more likely to occur between institutions with a diverse academic economic re- sources. This is consistent with the center-periphery hypothesis (Schott 1998) that re- searchers in peripheral regions are willing to collaborate with those in core regions to gain access to research resources, while core region researchers are also willing to seek complementarities to participate in their research (Hwang 2008). 4.3 Estimation for different sub-periods To analyze the determinants of RL flows from a dynamic perspective, we estimate the full model for two sub-periods (2013-2014 and 2016-2017) and compare their parame- ters over time. We take 2-year lagged independent variables to address the endogeneity and reverse causality concerns (Fernandez, Ferrandiz et al. 2016, Zhang 2016, Gui, Liu et al. 2018). Table 7 presents the estimation results for different sub-periods. To com- pare the fitted model in different sub-years, we adopt the Chow test (Chow 1960) to determine whether the independent variables have significant differences in the sub- periods. The Chow test result rejects no difference specification (p<0.01), indicating a clear difference between the two models for two sub-periods. Table 7 Estimation results of Tobit gravity model for sub-period 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) _cons N Left-censored Uncensored Model (1) Model (2) 2013-2014 2016-2017 2.181* (0.079) 1.615* (0.077) 2.121* (0.062) 1.784* (0.060) -0.857* -0.797* (0.094) (0.078) 83.603* 84.956* (4.382) 4.592* (0.444) (3.649) 4.970* (0.366) 14.668* 12.034* (0.225) 0.069 (0.063 (0.181) 0.102* (0.051) -27.354* -27.994* (0.953) 59,292 52,142 7,150 (0.775) 59,292 50,518 8,774 LR chi2 19219.97 22033.39 Prob > chi2 Pseudo R2 0.0000 0.2320 0.0000 0.2263 Log likelihood -31817.755 -37667.064 Chow test 43.91*** p<0.10; * p<0.05; *p<0.01 Consistent with the model fitted with all data (Table 6), the coefficients for 𝑙𝑛(𝐿𝑀𝑖) and 𝑙𝑛(𝐿𝑀𝑗) are both positively significant in both sub-periods. More specifically, alt- hough the coefficient of the 𝑙𝑛(𝐿𝑀𝑖) have been larger than that of 𝑙𝑛(𝐿𝑀𝑗), the coef- ficient of 𝑙𝑛(𝐿𝑀𝑖) decreased, while the coefficient of 𝑙𝑛(𝐿𝑀𝑗) increased over time. This indicates that the RL of the leading institution has a more prominent effect than that of the participating institution on the future RL flow, but the difference has been becoming smaller over time. The geographical proximity remains negatively significant, while its absolute value of the negative coefficient decreased, indicating that the negative effect of geographical proximity has declined over time. On the other hand, the cognitive proximity, and in- stitutional proximity remain positively significant and their coefficients increased, in- dicating that RL flows have become more likely to occur between institutions that have a similar research background and are located in the same province. The social prox- imity is positively significant in both sub-periods, but its coefficient declines, indicating that although the prior collaborations enhance the chance of future RL flows, its influ- ence is decreasing over time. It's interesting that the economic proximity was statistically insignificant in 2013- 2014, and became positively significant in 2016-2017. This result indicates that the economic proximity has recently become an important determinant of RL flows; insti- tutions have become increasingly likely to collaborate if there is a gap in economic resource (measured by the number of NSFC projects). The center -- periphery theory provides a possible explanation for this: the institutions with fewer economic resource tend to participate in the projects led by those with more resources; institutions with much economic resource are willing to collaborate with those less resource-intensive institutions for the access to non-economic research resources (e.g. patient subjects, the habitat for certain animals/plants, human resources, etc.). 4.4 The role of proximity on RL in multiple research fields To have a comprehensive understanding of RL, we also exam the role of proximity on RL in different research fields. We combined "Humanities & Arts" and "Social Sci- ences" into "Humanities, Arts and Social Sciences" (HASS) because they have been found to have similar collaboration pattern, outcome form and collaborative scale (En- gels, Ossenblok et al. 2012). Existing literature usually combined these two fields to explore the collaboration patterns (Sidone, Haddad et al. 2017). In addition, the number of publications in these two fields are far fewer than that of natural sciences and medi- cine by Chinese institutions. Combining them can avoid unstable results caused by the unbalanced data. Table 8 shows the estimation results of Tobit gravity model in these fields during 2013-2017. Generally, the results are in line with that of "Life Sciences & Biomedicine". Table 9 presents the estimation results of Tobit gravity model for sub-period in other fields. For these fields, the RL mass of both leading and participating institutions remain a positive and significant coefficient. And the gap is narrowing over time. Dif- ferent from the observations in "Life Sciences & Biomedicine", both effect of RL mass of leading and participating institutions are decreasing. The result of geographical prox- imity, cognitive proximity and social proximity are in line with that of "Life Sciences & Biomedicine". As for the institutional proximity and economic proximity, the results of Physical Sciences and Technology remain the same as that of "Life Sciences & Bio- medicine". However, in HASS, these two coefficients are declining, indicating the hin- der effect of disparity in policy, norms, rules, and economics has been becoming less significant in HASS RL flows. Table 8 Estimation results of Tobit gravity model in other fields 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) _cons N Left-censored Uncensored Technology Physical Sciences HASS 2013-2017 2013-2017 2013-2017 1.493* (0.026) 1.348* 1.356* (0.023) 1.267* (0.025) (0.023) 0.497* (0.046) 0.411* (0.046) -0.454* -0.446* -0.289* (0.026) (0.026) 18.452* 17.177* (0.982) 2.210* (0.047) 2.043* (0.116) 0.031* (0.017) (0.982) 2.237* (0.044) 2.157* (0.108) 0. 099* (0.016) -3.851* -4.093* (0.215) (0.204) 43,472 30,646 12,826 48,620 34,040 14,580 (0.055) 5.805* (0.590) 1.981* (0.011) 0.762 (0.244) 0.172* (0.034) -0.488* (0.420) 8,930 7,138 1,792 LR chi2 24880.19 26879.96 2 2451.48 Log-likelihood -40594.615 -46303.218 -6093.6622 Pseudo R2 Prob > chi2 0.2346 0.0000 0.2250 0.0000 0.1675 0.0000 p<0.10; * p<0.05; *p<0.01 Table 9 Estimation results of Tobit gravity model for sub-period in other fields 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) _cons N Left-censored Uncensored Technology Physical Sciences HASS 2013-2014 2016-2017 2013-2014 2016-2017 2013-2014 2016-2017 0.916* 0.668* 0.864* 0.642* 0.570* 0.431* (0.018) (0.015) (0.016) (0.015) (0.037) (0.049) 0.792* 0.630* 0.762* 0.616* 0.474* 0.409* (0.018) (0.015) (0.016) (0.015) (0.036) (0.049) -0.249* -0.223* -0.224* -0.203* -0.178* -0.091* (0.020) (0.018) (0.018) (0.017) (0.044) (0.058) 15.689* 19.765* 13.555* 15.827* 4.946* 6.404* (0.987) (0.705) (0.793) (0.705) (0.666) (0.608) 1.637* 1.656* 1.634* 1.696* 1.555* 1.335* (0.036) (0.031) (0.033) (0.030) (0.109) (0.074) 1.171* 1.149* 1.123* 1.008* 0.947* 0.373* (0.085) 0.002* (0. 012) (0.079) 0.049* (0.011) (0.077) 0.031 (0. 011) (0.075) 0.044* (0.011) (0.257) 0.079 (0. 035) (0.194) 0.051 (0.026) -3.084* -4.923* -3.387* -4.706* -2.483* -2.092* (0.165) (0.152) (0.149) (0.150) (0.455) (0.328) 43,472 36,936 6,536 43,472 33,150 10,322 48,620 40,807 7,813 48,620 37,359 11,261 8,930 8250 680 8,930 7610 1,320 LR chi2 15342.68 22533.58 17854.47 23926.17 1300.24 2175.00 Prob > chi2 Pseudo R2 0.0000 0.2764 0.0000 0.2786 0.0000 0.2747 0.0000 0.2712 0.0000 0.2126 0.0000 0.2035 Log likelihood -20088.156 -29176.044 -23568.254 -32147.945 -2408.0288 -4256.5959 Chow test 285.18* 254.35* 26.69* p<0.10; * p<0.05; *p<0.01 Figure 4.a, 4.b, 4.c show the kernel density distributions of institutions' RL disparity during 2013-2017 in "Technology", "Physical Sciences", and HASS. The results are in line with Figure 3, indicating that the RL flow intensity has become more evenly dis- tributed. It is worth noting that the change of HASS is more pronounced, which is con- sistent with the discussions above. Figure 4 Kernel density distributions of institutions' RL disparity over time Technology (a), Physical Sciences (b), HASS (c) 5 Robustness check To further test the robustness of our results, we apply the "full counting" method (i.e. the count of participating institutions) to measure the RL flow intensity and then esti- mate a negative binomial regression for the data of "Life sciences & biomedicine". For example, in a paper with three institutions, the leading institution's RL flow to each of the two participating institutions is 1. The RL mass 𝐿𝑀𝑖 obtained by the leading insti- tution is, therefore 2. Since the dependent variable is count data with over-dispersion (i.e. its variance is greater than its mean) and there exist a large number of zeros, a zero- inflated negative binomial regression is adopted (Cameron and Trivedi 2013). The rest of the independent variables remain the same as the main model. The results of the zero-inflated negative binomial regression (Table 10) leads to the same conclusions and confirms the robustness of the main model in Table 6. Table 10 Estimation results of zero-inflated negative binomial regression ("Full counting" of RL flows and mass) Model (1) 0.578* (0.010) 0.455* (0.010) Model (2) 0.428* (0.011) 0.322* (0.010) Model (3) 0.439* (0.011) 0.330* (0.010) Model (4) 0.287* (0.011) 0.232* (0.010) Model (5) 0.307* (0.012) 0.249* (0.011) 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) -0.343* -0.353* -0.137* -0.091* -0.092* (0.008) (0.007) (0.012) (0.012) (0.012) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 13.896* 14.804* 12.722* 12.208* (0.564) (0.565) 1.230* (0.058) (0.547) 0.980* (0.057) 1.271* (0.029) (0.555) 0.983* (0.0566) 1.260* (0.028) -0.030* (0.008) _cons -2.925* -0.544* -2.150* -1.660* -1.728* (0.097) (0.127) (0.149) (0.145) (0.145) Inflated (Logit) -0.782* -0.517* -0.520* -0.433* -0.371* (0.018) (0.018) (0.018) (0.021) (0.023) -0.663* -0.381* -0.383* -0.310* -0.252* (0.017) 0.277* (0.015) (0.017) 0.317* (0.015) (0.018) 0.131* (0.023) (0.020) 0.084* (0.026) (0.021) 0.077* (0.026) -23.581* -23.645* -19.337* -20.168* (0.861) (0.881) (0.933) (0.938) 𝑙𝑛(𝐿𝑀𝑖) 𝑙𝑛(𝐿𝑀𝑗) 𝑙𝑛(𝐺𝑒𝑜𝑝𝑟𝑜𝑥𝑖𝑗) 𝑙𝑛(𝐶𝑜𝑔𝑛𝑝𝑟𝑜𝑥𝑖𝑗) 𝐼𝑛𝑠𝑡𝑝𝑟𝑜𝑥𝑖𝑗 𝑆𝑜𝑐𝑝𝑟𝑜𝑥𝑖𝑗 𝑙 𝑛(𝐸𝑐𝑜𝑛𝑝𝑟𝑜𝑥𝑖𝑗) -1.286* -0.979* -0.996* (0.110) (0.123) (0.122) -1.985* -2.017* (0.110) (0.281) 0.387* 1.473 7.9+04 19.24 (0.111) -0.104* (0.015) 2.597* (0.280) 0.384* 1.468 7.9e+04 19.03 _cons 6.021* 1.542* 2.864* 2.725* Lnalpha Alpha (0.156) 0.837* 2.311 (0.208) 0.640* 1.897 (0.253) 0.597* 1.817 Likelihood ratio test 1.1e+05 8.9e+04 8.6e+04 Vuong test alpha LR chi2 Number of obs Nonzero obs 19.26 7960.96 13.84 8744.57 p<0.10; * p<0.05; ***p<0.01 14.84 9221.41 11697.79 11712.96 59,292 12,585 6 Conclusions In this paper, we proposed the concept and measurement of "research leadership" (RL) in research collaborations and examined the determinants of RL flow between Chinese institutions using a comprehensive bibliographic dataset during 2013-2017. We find that research leadership among Chinese institutions is highly concentrated. Top 15 in- stitutions account for over one-third of the total research leadership in the field of bio- medical field. The most prevalent research leadership flows occur between institutions in the same city. In addition, the flows of research leadership from the leading institu- tion are observed to be (increasingly) evenly distributed to multiple participating insti- tutions. Our empirical results drawn from a Tobit regression-based gravity model show that the research leadership mass of both the leading and participating institutions and the geographical, cognitive, institutional, social and economic proximities are important factors of the flow of research leadership among Chinese institutions. These results remain robust to several sensitive checks. In particular, the leading institution's research leadership mass has a higher influence than that of the participat- ing institution, though the gap is narrowing. The constraining effect of geographical and social barriers have become less significant. But cognitive proximity is playing increasingly important roles in science and technology fields. Notably, we also obtain evidence that economic proximity has recently become an important factor of research leadership flows in science and technology fields. Combining the results from both the descriptive statistics and the gravity model, there is clear evidence that although there are a number of significant proximities for research leadership flows, the effects of these proximities have been declining recently. The col- laborations among institutions become more "flat" and convenient, possibly due to the advances in transportation, communication, and the general research capability of Chi- nese institutions. This research leads to the following policy implications. Given the positive signifi- cance of both the leading and participating institutions research leadership mass, insti- tutions should actively lead research projects to obtain a more significant role in both leading and participating important research in the future. Given the hindering effect of geographical proximity social proximity and institutional proximity, the funding bodies should encourage cross-provincial collaborations. In the meanwhile, policymakers in different provinces should facilitate the research leadership flow across multiple prov- inces by unifying their research policy and norms. In addition, policy makers and fund- ing bodies should facilitate the establishment of new collaborations between institutions that have not collaborated before, and between the institutions with rich economic re- source and those with less economic resource. By doing so, the hindering effects of social and economic proximities among institutions are reduced, so that we can (a) take advantage of the network effect brought by additional links in the collaboration network, and (b) integrate the economic and complementary non-economic research resources possessed by different institutions。 Results of this study also shed light on future applications of the proposed research leadership concept in analyzing other scientific collaboration datasets, such as patent, grant, conference organizations, and journal editorial. It is also important to delve into the research leadership roles of individual scholars and the whole country. References Acosta M, Coronado D, Ferrándiz E, León MD. (2011). Factors affecting inter-regional academic scientific collaboration within Europe: The role of economic distance. Scientometrics, 87(1): 63-74. Alvarez-Betancourt Y, Garcia-Silvente M. (2014) An overview of iris recognition: a bibliometric analysis of the period 2000-2012. Scientometrics, 101(3): 2003-2033. Anderson J E. (1979). A theoretical foundation for the gravity equation. The American Economic Review, 69(1): 106-116. Balland P A. (2012). Proximity and the evolution of collaboration networks: evidence from research and development projects within the global navigation satellite system (GNSS) industry. Regional Studies, 46(6): 741-756. Bergé L R. (2017). Network proximity in the geography of research collaboration. Papers in Regional Science, 96(4): 785-815. Boschma R, Frenken K. (2009). Some notes on institutions in evolutionary economic geography. Economic geography, 85(2): 151-158. Boschma R.(2005). Proximity and innovation: a critical assessment. Regional studies, 39(1): 61-74. Burger M, Van Oort F, Linders G J. (2009). On the specification of the gravity model of trade: zeros, excess zeros and zero-inflated estimation. Spatial Economic Analysis, 4(2): 167-190. Cameron A C, Trivedi P K. (2013). Regression analysis of count data. Cambridge university press. Chow G C. (1960). Tests of equality between sets of coefficients in two linear regressions. Econometrica: Journal of the Econometric Society, 6(2): 591-605. Coenen L, Moodysson J, Asheim B T. (2004). Nodes, networks and proximities: on the knowledge dynamics of the Medicon Valley biotech cluster. European Planning Studies , 12(7): 1003-1018. Engels, T. C., Ossenblok, T. L., & Spruyt, E. H. (2012). Changing publication patterns in the social sciences and humanities, 2000 -- 2009. Scientometrics, 93(2), 373-390. Fernández A, Ferrándiz E, León M D. (2016). Proximity dimensions and scientific collaboration among academic institutions in Europe: The closer, the better?. Scientometrics, 106(3): 1073-1092. Garcia, R., Araujo, V., Mascarini, S., Gomes Dos Santos, E. and Costa, A. (2018). Is cognitive proximity a driver of geographical distance of university -- industry collaboration? Area Development and Policy, 3(3): 349-367. Gilsing V, Nooteboom B, Vanhaverbeke W, Duysters G, van den Oord A. (2008). Network embeddedness and the exploration of novel technologies: Technological distance, betweenness centrality and densit. Research policy, 37(10): 1717-1731. Gui Q, Liu C, Du D. (2018). International knowledge flows and the role of proximity. Growth and Change, 49(3): 532-547. Heffner A. (1981). Funded research, multiple authorship, and subauthorship collaboration in four disciplines. Scientometrics, 3(1): 5-12. Hoekman J, Frenken K, Tijssen R J W. (2010). Research collaboration at a distance: Changing spatial patterns of scientific collaboration within Europe. Research policy, 39(5): 662-673. Hoekman, J., K. Frenken and F. Van Oort (2009). The geography of collaborative knowledge production in Europe. The Annals of Regional Science, 43(3): 721-738. Hoekman J, Scherngell T, Frenken K, et al (2013). Acquisition of European research funds and its effect on international scientific collaboration. Journal of economic geography, 13(1): 23-52. Hu X, Rousseau R, Chen J. (2010). In those fields where multiple authorship is the rule, the h-index should be supplemented by role-based h-indices. Journal of Information Science, 36(1): 73-85. Hwang K. (2008). International collaboration in multilayered center-periphery in the globalization of science and technolog. Science, Technology, & Human Values, 33(1): 101-133. Jarneving B. (2009). Regional research and foreign collaboration. Scientometrics, 83(1): 295-320. Jiang LA, Zhu N, Yang Z, Xu S, Jun M. (2018), The relationships between distance factors and international collaborative research outcomes: A bibliometric examination. Journal of Informetrics, 12(3): 618-630. Katz J S, Martin B R. (1997). What is research collaboration?. Research policy, 26(1): 1-18. Lander B. (2015). Proximity at a distance: the role of institutional and geographical proximities in Vancouver's infection and immunity research collaborations. Industry and Innovation, 22(7): 575-596. Lazzeretti L, Capone F. (2016). How proximity matters in innovation networks dynam- ics along the cluster evolution. A study of the high technology applied to cultural goods. Journal of Business Research, 69(12): 5855-5865. Li T. (2002). Econometric Analysis of Cross Section and Panel Data. Journal of eco- nomic literature, 40(4): 1239-1241. Ling, J., N. Zhu, Z. Yang, X. Shen and M. Jun (2018). "The relationships between dis- tance factors and international collaborative research outcomes: A bibliometric exami- nation." Journal of Informetrics, 12(3): 618-630. Newman M E J. (2004). Coauthorship networks and patterns of scientific collaboration. Proceedings of the national academy of sciences, 101(1): 5200-5205. Nguyen T V, Ho-Le T P, Le U V. (2017). International collaboration in scientific research in Vietnam: an analysis of patterns and impact. Scientometrics, 110(2): 1035- 1051. Niedergassel B, Leker J. (2011). Different dimensions of knowledge in cooperative R&D projects of university scientists. Technovation, 31(4): 142-150. Osório A. (2018) On the impossibility of a perfect counting method to allocate the credits of multi-authored publications. Scientometrics, 116(3): 2161-2173. Padial A, Nabout J, Siqueira T, Bini L, Diniz-Filho J. (2010).Weak evidence for determinants of citation frequency in ecological articles. Scientometrics, 85(1): 1-12. Parreira MR, Machado KB, Logares R, Diniz-Filho JA, Nabout JC. (2017). The roles of geographic distance and socioeconomic factors on international collaboration among ecologists. Scientometrics, 113(3): 1539-1550. Petruzzelli, A. M. (2011). The impact of technological relatedness, prior ties, and geo- graphical distance on university -- industry collaborations: A joint-patent analysis. Tech- novation, 31(7): 309-319. Plotnikova T, Rake B. (2014). Collaboration in pharmaceutical research: exploration of country-level determinants. Scientometrics, 98(2): 1173-1202. Scherngell T, Barber M J. (2011). Distinct spatial characteristics of industrial and pub- lic research collaborations: evidence from the fifth EU Framework Programme. The Annals of Regional Science, 46(2): 247-266. Scherngell T, Hu Y. (2011). Collaborative knowledge production in China: Regional evidence from a gravity model approach. Regional Studies, 45(6): 755-772. Schott T. (1998). Ties between center and periphery in the scientific world-system: ac- cumulation of rewards, dominance and self-reliance in the center. Journal of World- Systems Research, 4(2): 112-144. Schubert T, Sooryamoorthy R. (2009). Can the centre -- periphery model explain patterns of international scientific collaboration among threshold and industrialised countries? The case of South Africa and Germany. Scientometrics, 83(1): 181-203. Sekara V, Deville P, Ahnert SE, Barabási AL, Sinatra R, Lehmann S. (2018). The chap- erone effect in scientific publishing. Proceedings of the National Academy of Sciences, 115(50): 12603-12607. Sidone O J G, Haddad E A, Mena‐Chalco J P. (2017). Scholarly publication and col- laboration in B razil: The role of geography. Journal of the Association for Information Science and Technology, 68(1): 243-258. Sun Y, Wei Y, Zhang L. (2017). International academic impact of Chinese tourism research: A review based on the analysis of SSCI tourism articles from 2001 to 2012. Tourism Management, 58: 245-252. Wagner C S, Park H W, Leydesdorff L. (2015). The continuing growth of global coop- eration networks in research: A conundrum for national governments. PLoS One, 10(7): e0131816. Wang L, Wang X. (2017).Who sets up the bridge? Tracking scientific collaborations between China and the European Union. Research Evaluation, 26(2): 124-131. Wang W, Wu Y, Pan Y. (2014). An investigation of collaborations between top Chinese universities: a new quantitative approach. Scientometrics, 98(2): 1535-1545. Wang X, Xu S, Wang Z, Peng L, Wang C. (2013). International scientific collaboration of China: Collaborating countries, institutions and individuals. Scientometrics, 95(3): 885-894. Wooldridge J M. (2010). Econometric analysis of cross section and panel data. MIT press. Zhang C. (2016). Agglomeration of knowledge intensive business services and urban productivity. Papers in Regional Science, 95(4): 801-818. Zhang G, Lin B. (2018). Impact of structure on unified efficiency for Chinese service sector -- A two-stage analysis. Applied energy, 231: 876-886.
1604.03147	3	1604	2017-01-31T09:19:42	Graph-based Collaborative Ranking	[ "cs.SI", "cs.IR" ]	Data sparsity, that is a common problem in neighbor-based collaborative filtering domain, usually complicates the process of item recommendation. This problem is more serious in collaborative ranking domain, in which calculating the users similarities and recommending items are based on ranking data. Some graph-based approaches have been proposed to address the data sparsity problem, but they suffer from two flaws. First, they fail to correctly model the users priorities, and second, they cannot be used when the only available data is a set of ranking instead of rating values. In this paper, we propose a novel graph-based approach, called GRank, that is designed for collaborative ranking domain. GRank can correctly model users priorities in a new tripartite graph structure, and analyze it to directly infer a recommendation list. The experimental results show a significant improvement in recommendation quality compared to the state of the art graph-based recommendation algorithms and other collaborative ranking techniques.	cs.SI	cs	Graph-based Collaborative Ranking Bita Shams a and Saman Haratizadeh a a University of Tehran, Faculty of New Sciences and Technologies North Kargar Street, Tehran, Iran 1439957131 Abstract Data sparsity, that is a common problem in neighbor-based collaborative filtering domain, usually complicates the process of item recommendation. This problem is more serious in collaborative ranking domain, in which calculating the users' similarities and recommending items are based on ranking data. Some graph-based approaches have been proposed to address the data sparsity problem, but they suffer from two flaws. First, they fail to correctly model the users' priorities, and second, they can't be used when the only available data is a set of ranking instead of rating values. In this paper, we propose a novel graph-based approach, called GRank, that is designed for collaborative ranking domain. GRank can correctly model users' priorities in a new tripartite graph structure, and analyze it to directly infer a recommendation list. The experimental results show a significant improvement in recommendation quality compared to the state of the art graph-based recommendation algorithms and other collaborative ranking techniques. Keywords: Collaborative ranking, pairwise preferences, graph modelling, recommendation systems, personalized PageRank Introduction Collaborative filtering (CF) techniques are effective algorithms that help people by filtering irrelevant contents and providing personalized recommendation of useful services. These techniques seek to learn models to predict the services that a user will require in the future based on his preferences in the past. Collaborative-filtering techniques can be categorized into two classes: rating-oriented and ranking-oriented algorithms. The goal of rating-oriented algorithms is to accurately predict a user's ratings and then, recommend the items with the highest predicted rating for him. On the other hand, ranking-oriented approach, called collaborative ranking, seek to directly predict the rankings of items from the viewpoint of a target user, without explicitly predict the ratings. It has been shown that ranking-oriented collaborative filtering approach is sometimes more intuitive and applicable. To see why, notice that, recommendation is naturally a ranking task and what a recommendation algorithm really needs is to improve the quality of Top-k ranking not predicting the rates (N. Liu & Yang, 2008; Y Shi, Karatzoglou, & Baltrunas, 2012; Y Shi, Larson, & Hanjalic, 2010; Yue Shi, Larson, & Hanjalic, 2013). Moreover, in many applications, all we have is a set of implicit feedbacks while no rating data is available and hence, rating based methods can't be used in such a situation. Note that despite rating and other kinds of explicit feedback, that require the user to explicitly assess the items, implicit feedback can be automatically gathered by tracking the user's interactions with the system (e.g. click, buy, like, etc.). Ranking oriented collaborative filtering can be applied in such situations as well. Neighbor-based collaborative filtering, one of the main classes of collaborative filtering, estimates the ranking/rating of target user based on the behavior of similar users. Despite several researches in this class of algorithms, they still are not able to precisely calculate users' similarities. The reason can be explained by sparsity problem which refers to the fact that in recommender systems, users have given feedback to a small proportion of items, and consequently, they rarely have enough common items or pairwise comparisons for estimation of their true similarities/ dissimilarities (Desrosiers & Karypis, 2011). One approach to overcome this issue, is graph-based recommendation that takes advantages of heterogeneous information networks, that are information networks containing different types of nodes and edges, to refine the similarity measures (M. S. Shang, Fu, & Chen, 2008; Z.-K. Zhang, Zhou, & Zhang, 2010; Zhou, Ren, Medo, & Zhang, 2007), expand the neighborhoods, and, directly calculate the closeness of users and items(Chiluka, Andrade, & Pouwelse, 2011; Silva & Zaki, 2013; Xiang et al., 2010; Yao, He, Huang, Cao, & Zhang, 2013). Graph-based recommendation algorithms represent the relations between users and items as a bipartite graph in which there is a weighted or unweighted link between a user and each item he has rated (Li & Chen, 2013; M. S. Shang et al., 2008; M.-S. Shang, Zhang, Zhou, & Zhang, 2010; Ting, Yan, & Xiang-wei, 2013; Xiang et al., 2010; Z.-K. Zhang et al., 2010; Zhou et al., 2007). Unfortunately, this approach is basically designed for rating/binary feedbacks and has crucial insufficiencies for ranking-oriented class of neighbor-based collaborative filtering. The first problem is that current graph-based approaches are incompetent to capture the preference order of users. We refer to the example of Fig.1a to illustrate this shortcoming. Mike and Lee have the same preference order for item A, and B, while Mike, and Martin have completely opposite preference orders on all items. Current graph-based algorithms represent this data as Fig.1b (Sawant, 2013; M. S. Shang et al., 2008). Intuitively, under this graph modeling, most of well-known graph proximity measures (e.g. common neighbors, distance, Katz, and personalized PageRank) will suggest that Mike is much closer to Martin than Lee that is counterintuitive. Fig.1. An example to illustrate incompetence of current graph-based structure to capture preference data gathered in form of rating. The second shortcoming of current graph-based approaches have been proposed for binary implicit feedback and that they cannot capture the pairwise preference (i.e. choice context) of user that is generated by different implicit feedbacks. It is clear that the choice context is a valuable piece of information that can be used to improve the recommendation quality. To see how such information is lost when data is modeled by current graph representations, you can observe in the example of Fig.2 that John has preferred item A over B in one session and item B over C in another session, while, Jack has preferred item B over A. Current graph-based representation of implicit feedbacks, makes a link between the user and those items receiving the positive feedbacks (Chen, Wang, Huang, & Mei, 2012; Xiang et al., 2010; Z.-K. Zhang et al., 2010). Therefore, these algorithms cannot differentiate heterogeneous implicit feedbacks (i.e. buy, click). More importantly, they are not able to clarify the fact that Jack and John disagree when it comes to comparison of items A and B, as illustrated in Fig.2b. Fig.2. An example to illustrate incompetence of current graph-based structure to capture the choice context collected through the browsing/ purchasing history. This paper presents a novel framework, called GRank, that captures the preference of users using a new Tripartite Preference Graph (TPG) structure that demonstrates the relations between users, items, and pairwise preferences. GRank, also provides a new ranking algorithm, which extends personalized PageRank for top-k recommendation. To the best of our knowledge, this algorithm is the first graph-based approach that is able to capture the preference information provided by implicit feedbacks. Experimental results show higher accuracy of GRank compared to the state of the art collaborative ranking algorithms as well as available graph-based recommendation systems. The rest of this paper is organized as follows. In Section 2, the related work on graph- based recommendation and collaborative ranking techniques are discussed; then, we present the details of GRank's framework in Section 3. The experimental results are presented and analyzed in Section 4. In Section 5, we discuss how GRank can address some the current shortcomings of collaborative ranking and graph-based recommendation methods. Finally, in Section 6 we conclude and introduce our future works. Related Work The quality of recommendation can be analyzed from many different points of view including accuracy(Koren, Bell, & Volinsky, 2009; Weimer & Karatzoglou, 2007), coverage (Bellogin & Parapar, 2012; Cacheda, Carneiro, Fernández, & Formoso, 2011), diversity (Adomavicius & Kwon, 2012; Said, Kille, Jain, & Albayrak, 2012; Zhou et al., 2010), serendipity(Lu, Chen, Zhang, Yang, & Yu, 2012; Xiao, Che, Miao, & Lu, 2014), uncertainty (M. Zhang, Guo, & Chen, 2015), shilling attack detection(Z. Zhang & Kulkarni, 2014) and scalability (Jiang, Lu, Zhang, & Long, 2011). Although all these aspects are important factors in the success of a recommender system, the recommendation's accuracy is a key element with this regard and a core set of researches have been formed to achieve higher levels of recommendation accuracy. This paper lies in this category, presenting a novel graph-based framework that improves the accuracy of recommendation in the absence of contextual information location when the only available information is the preference data. Contextual information usually refers to the environmental state in which the interaction of the user and the system happens (e.g. time, location, emotion, etc.). On the other hand, the "choice context" reflects the options among which the user makes a choice. Here, we will review the existing researches related to the main aspects of our proposed algorithm: Collaborative ranking, and Graph-based recommendation. . Collaborative ranking Collaborative ranking is a class of collaborative filtering algorithms that seeks to predict how a user will rank items. As we mentioned before, despite some similarities, collaborative ranking algorithms are differentiated from rating-oriented collaborative filtering (i.e. collaborative rating) by the fact that collaborative rating algorithms rely on the rating data and try to minimize the rating prediction error, while collaborative ranking algorithms do not depend on rating data. They can use any kind of preference data and try to minimize the rank prediction error. There is also some similarities between two concepts of collaborative ranking and learning-to-rank problems in the information retrieval domain, as they both try to order entities of on type, e.g. documents/items, for a target entity of another type, e.g. queries/users. However these two kinds of problems are different in practice. In learning- to rank problem there exist a set of explicit common features, such as terms' frequencies among two types of entities, queries and document (Balakrishnan & Chopra, 2012; Fan & Lin, 2013; Y Shi et al., 2010; Volkovs & Zemel, 2012) while there is no such features available or used to relate user and item entities in collaborative ranking problem (Balakrishnan & Chopra, 2012; Y Shi et al., 2010; Volkovs & Zemel, 2012). Because of this important difference between the nature of the problem in these domains, different classes of algorithms have emerged for solving those problems. These approaches can be categorized into two categories: matrix factorization (MFCR) and Neighbor-based algorithms (NCR). Matrix factorization techniques in collaborative ranking, try to learn representative latent features for an accurate prediction on ranking of items for each user. CofiRank was the first algorithm that uses matrix factorization techniques to optimize a rank-oriented metric(Weimer, Karatzoglou, & Smola, 2008; Weimer & Karatzoglou, 2007). Another technique, ListRank, estimates the Top-1 probabilities to infer a ranking for items. (Y Shi et al., 2010). URM is another model that combines ListRank and probabilistic matrix factorization in order to improve system's accuracy in terms of both ranking and ratings (Yue Shi, Larson, et al., 2013). BoostMF is another matrix factorization approach that sequentially learns a set of weak matrix factorization models based on preference data(Chowdhury, Cai, & Luo, 2015). Bayesian personalized ranking and its variants, try to optimize area under the curve (AUC) for a Bayesian prediction model that is generated based on a set of prediction of pairwise comparisons between relevant and irrelevant items (Lerche & Jannach, 2014; Pan, Zhong, Xu, & Ming, 2015; Rendle, Freudenthaler, Gantner, & Schmidt-thieme, 2009). Recently, some approaches have been proposed that focus on correctly predicting the pairwise preferences for the items with the highest ranks (Christakopoulou & Banerjee, 2015; Dhanjal, Clémençon, & Gaudel, 2015). Climf (Y Shi et al., 2012) and xClimf (Yue Shi, Karatzoglou, Baltrunas, & Larson, 2013) are two other algorithms that exploit matrix factorization techniques to optimize Mean Reciprocal Rank (MRR) of the recommendation list (Y Shi et al., 2012; Yue Shi, Karatzoglou, et al., 2013). Although the collaborative ranking methods based on matrix factorization consist of a diverse set of algorithms and methods (Weimer et al., 2008; Weimer & Karatzoglou, 2007), the approach of GRank is conceptually different from them as it does not represent the data in a latent feature space. Instead, it models the rank data in the form of a graph structure that enables it to directly estimate the closeness of users and items, based on which it can do the recommendation. So, in a sense, GRank lies in another class of recommendation algorithms called neighbor-based collaborative ranking (NCR). Although this second class of algorithms has its advantages, this approach has remained less investigated, and few successful NCR algorithms have been proposed so far. EigenRank(N. Liu & Yang, 2008) is the most famous NCR technique that infers a total ranking based on pairwise preferences of users similar to the target user. EigenRank computes users' similarity using Kendall correlation that takes into accounts the agreement and disagreement of users over pairwise comparisons. After estimation of similarities, EigenRank estimates a preference matrix whose elements are a weighted linear combination of neighbors' preferences. Finally, it uses a greedy or Markov-based approach to infer a total ranking over items. To Our knowledge, all of NCR techniques follow the main approach presented by EigenRank with slight modifications. EduRank(AvSegal, Katzir, & Gal, 2014), WSRank (Meng, Li, & Sun, 2011), and Cares(Yang, Wei, Wu, Zhang, & Zhang, 2009) customized EigenRank for different applications. VSRank (Wang, Sun, & Gao, 2014) focuses to improve Kendall similarity measure via considering importance of each pairwise comparison in similarity calculation. However, this approach still suffers from the sparsity problem since it still relies on common pairwise comparisons for calculating similarities. As stated earlier, GRank aims to solve the sparsity problem of neighbor-based collaborative ranking by introducing a novel graph-based approach for modeling and analyzing data. It also differs from the current neighbor-based algorithms as it does not follow the traditional three-step framework, and directly estimates the users' preferences. 2.2. Graph-based recommendation approaches Although there is no graph-based methods designed for collaborative ranking, many recent studies have been conducted in other areas of recommender systems. Here, we will briefly review those algorithms and clarify the main differences between the current work and them. Graph-based recommendation algorithms are composed of two steps: Constructing a graph representing the data and making recommendations by analyzing the graph. These recommendation algorithms have exploited different types of graphs. However, in all of them, the main component of the graph is the relations between users and those items that have been rated by them. Therefore, the most common approach is constructing a bipartite network where the connections are from one part of the network, users, to the other part, items. Once the bi-partite graph is constructed, several approaches can be used to rank the items using the information from the neighbors of the target user. Approaches like using common neighbors, Katz similarity, diffusion scores and personalized PageRank have been used in this domain (Huang, Li, & Chen, 2005; Z.-K. Zhang et al., 2010) Recent methods have extended the bi-partite network by adding some layers to it. Some researchers (Xiang et al., 2010) have considered using a session layer to take into account the long-term and the short-term preferences of the user in order to make recommendations in a particular time. Others (Yao et al., 2013) have used different types of nodes in a multi-layer structure to make context-aware recommendation through a random walk in the graph. In (Z.-K. Zhang et al., 2010) a three-layer graph is used to improve recommendation through considering the tags assigned to items by users using a diffusion-based score introduced in (Zhou et al., 2007). In some works (Lee, Park, Kahng, & Lee, 2013; Yu, Ren, Sun, & Gu, 2014) the structure of the network has been revised. They consider a star heterogeneous network, where users and items can be connected to different types of nodes. They use this graph structure to improve the model–based recommendations (Yu et al., 2014) or to make recommendation through improvement of personalized PageRank algorithms in heterogeneous networks. (Lee et al., 2013). We emphasize that none of these algorithms are designed to capture the choice context and preferences of users. Also most of them depend on the contextual information (e.g. time, content, etc.) that does not exist or is not available to the system in all applications and may be expensive to collect. GRank: a graph-based framework for collaborative filtering It has been shown that heterogeneous information networks have strong capabilities to model the relationships among different entities of recommender systems (Cong, 2009; Sun, Han, Yan, & Yu, 2011; Yu et al., 2013, 2014). In this paper, we seek to propose an effective graph approach to ranking-oriented recommender systems, called Graph-based collaborative Ranking, or GRank. In the following, we first define the problem of graph- based collaborative ranking and its purposes. Then, we present some definitions that are needed to understand the algorithm. Next, we introduce a novel heterogeneous graph structure, called tripartite preference graph (TPG) that embeds different kinds of relations among users, preferences and items in an aggregated structure. Finally, we suggest an efficient algorithm to exploit TPG in order to rank items for each target user. . Problem definition From the collaborative ranking perspective, recommender systems can be represented by the set of users 𝑈 = {𝑢1, … , 𝑢𝑀}, set of items 𝐼 = {𝑖1, … , 𝑖𝑁} and the observation set 𝑂 = {< 𝑢, 𝑖, 𝑗 >} that is the set of preferences occasionally stated by users. Generally, we define the observation 𝑜 =< u, i, j > where u ∈ U, i ∈ I and j ∈ I denoting that the user u has preferred i over j. For simplicity, we call the first item, the desirable item and the second one the undesirable item. Note that pairwise comparison is a general form of ranking data, and, all kinds of preferences (e.g. rating, browsing history) can be converted to a set of pairwise comparisons using the following rules: Rule 1. Let L be a rating matrix in which 𝐿𝑢𝑖 represents the rating of user u for item i. The preference observation set can be obtained by 𝑂 = {< 𝑢, 𝑖, 𝑗 >𝐿𝑢𝑖 ≠ 0, 𝐿𝑢𝑗 ≠ 0 , 𝐿𝑢𝑖 > 𝐿𝑢𝑗} Rule 2. Let L be the matrix of positive feedbacks (e.g. Like) in which a non-zero element 𝐿𝑢𝑖 represents that user u likes item i. Similarly, Let D be the matrix of negative feedbacks (e.g. dislike) in which a non-zero element 𝐷𝑢𝑖 represents the user u dislikes item i. The preference observation set can be obtained by 𝑂 = {< 𝑢, 𝑖, 𝑗 >𝐿𝑢𝑖 ≠ 0, 𝐷𝑢𝑗 ≠ 0 } Rule 3. Let W be the set of sessions that are defined as 𝑊 = {𝑤1, … 𝑤𝑤}, we can create the observation set as 𝑂 = {< 𝑢, 𝑖, 𝑗 > ∃𝑤 ∈ 𝑊 𝑢 = 𝑤. 𝑢, 𝑖 ∈ 𝑤. 𝐵, 𝑗 ∈ 𝑤. 𝐶} Where w.u is the user in the session w, w.B is the set of items bought in the session w and w.C is the set of items clicked but not bought in session w Given a preference dataset O, a graph-based framework will face a key question that is how to model the information available from the preference data set in the graph structure. To answer this question, we first categorize some information that an effective graph modeling of rank-oriented recommender system ought to capture:  Users' similarities' in terms of priorities: Two users of a ranking-oriented recommender system, are assumed to be similar when they have either similar opinions about certain pairwise comparisons(N. Liu & Yang, 2008; Wang et al., 2014) or prefer a particular item A in some comparisons even if the items over which it has been preferred are different for those two users (Meng et al., 2011). A well-organized graph model of ranking data should reflect both type of similarities among users.  Correlation among comparisons: Correlated pairwise comparisons are those preferences that are similarly voted by users. These comparisons should be simply discovered by analyzing the graph representation of the data.  Items' similarities: Similar items are those that are similarly favored/disfavored by a group of similar users. An effective graph modeling should clearly reflect the closeness of these items.  Prediction of users' priorities: The ultimate goal of ranking-oriented recommender systems is to infer the total ranking of target user over unseen items and recommend the top-k items. Consequently, graph representation of these systems is responsible for the efficient and effective recommendation to the target user, and ideally, a graph representation of rank data may be used to directly predict the rankings. . Graph Construction Before we can proceed to explain how GRank constructs and exploits a graph structure based on the preference dataset, we need to define some basic concepts: Definition. 1. A pairwise preference p is a tuple < 𝑖, 𝑗 > denoting the preference of i over j. We call i as the desirable item in p, represented by p.d, and j as the undesirable item in p represented by p.u. The pairwise preference set P is formally defined as 𝑃 = {< 𝑖, 𝑗 > 𝑖 ∈ 𝐼, 𝑗 ∈ 𝐼, 𝑖 ≠ 𝑗}. Definition. 2. A user may have some certain preference over two items or not. The agreement function 𝑓: 𝑈 × 𝑃 → {0,1} indicates whether the user 𝑢𝑖 agrees with the preference 𝑝𝑗 or not and is defined as: 𝑓(𝑢𝑖, 𝑝𝑗) = { 1, < 𝑢𝑖, 𝑝𝑗. 𝑑, 𝑝𝑗. 𝑢 > ∈ 𝑂 0, 𝑂. 𝑤 Where 𝑢𝑖 ∈ 𝑈, 𝑝𝑗 ∈ 𝑃 𝑎𝑛𝑑 O is the observation set of preferences, as defined in section 3.1. Definition.3. Abstractly, each item has two sides: the desirable side and the undesirable side. We define the items' desirability set as 𝐼𝑑 = {𝑖𝑑𝑖 ∈ 𝐼} where 𝑖𝑑 represents the desirable side of item i. Also the items' undesirability set is defined as 𝐼𝑢 = {𝑖𝑢𝑖 ∈ 𝐼} where 𝑖𝑢 represents the undesirable side of item i. We also define the representative set as = 𝐼𝑑 ∪ 𝐼𝑢 , that will contain two elements for each item, one for each side of it. Definition.4. The support function 𝑠: 𝑃 × 𝑅 → {0,1} indicates whether a preference p supports the representative r or not. Formally, we define s as 𝑠(𝑝, 𝑟) = { 𝑝. 𝑢 = 𝑖 𝑎𝑛𝑑 𝑟 = 𝑖𝑢 1, 1, 𝑝. 𝑑 = 𝑖 𝑎𝑛𝑑 𝑟 = 𝑖𝑑 0, 𝑂. 𝑊 Where 𝑝 ∈ 𝑃 and 𝑟 ∈ 𝑅. Based on these definitions and concepts we can now explain how GRank models the preference data using a structure called Tripartite Preference Graph. Formally, Tripartite Preference Graph (TPG) is a tripartite graph TPG(𝑈 ∪ 𝑃 ∪ 𝑅, 𝐸𝑈𝑃 ∪ 𝐸𝑃𝑅), where U is the set of users, P is the set of pairwise preferences and R is the set of representatives. EUP = {(𝑢, 𝑝)𝑓(𝑢, 𝑝) = 1, 𝑢 ∈ 𝑈, 𝑝 ∈ 𝑃}, is the set of edges between the nodes in U and P, and 𝐸𝑃𝑅 = {(𝑝, 𝑟)𝑠(𝑝, 𝑟) = 1, 𝑝 ∈ 𝑃, 𝑟 ∈ 𝑅}, is the set of edges connecting the nodes in P to the nodes in R. More clearly, TPG contains three layers, each containing a different type of nodes:  Users: TPG's first layer contains one node for each user.  Preferences: The preference layer consist the nodes corresponding to the set of all possible pairwise preferences 𝑝 ∈ 𝑃. Foe simplicity, the corresponding node to each preference 𝑝 =< 𝑖, 𝑗 > is labeled in the form of ("𝑖 > 𝑗") that clearly demonstrates the preference of i over j.  Representatives: The representative layer contains the set of both undesirable and undesirable representative of items represented by 𝑖𝑑 and 𝑖𝑢, respectively TPG also contains two types of links:  User-Preference links: 𝐸𝑈𝑃 is the set of edges that connects each user u to his stated preference. More clearly, For each preference data < 𝑢, 𝐴, 𝐵 >∈ 𝑂, there is a link between user 𝑢 and the preference node labeled by "𝐴 > 𝐵"  Preference-representative links: 𝐸𝑃𝑅 is the set of links that connects each preference to the representatives that it supports. For instance, a preference node labeled by "𝐴 > 𝐵" is connected to the nodes corresponding to desirable case of 𝐴, i.e. 𝐴𝑑 and undesirable case of B i.e. 𝐵𝑢. This links are used to model the fact that an "𝐴 > 𝐵" preference, implicitly supports item desirable side of "𝐴" and undesirable side of "𝐵". Fig. 3. An example: A TPG constructed from a system containing 5 users, 4 items, and, 9 assigned pairwise preference. As mentioned before, TPG is a tripartite graph in which the preference layer connects to both other layers: the user layer and the item layer. Traversing TPG through different types of paths reveals different types of information in a ranking-oriented recommender systems, some examples of which are presented in Table.1. Table 1. The meta-paths and their semantics in TPG Meta-patha 𝑼 − 𝑷 − 𝑼 𝑼 − 𝑷 − 𝑰𝒅 − 𝑷 − 𝑼 𝑼 − 𝑷 − 𝑰𝒖 − 𝑷 − 𝑼 𝑷 − 𝑼 − 𝑷 𝑰𝒅 − 𝑷 − 𝑼 − 𝑷 − 𝑰𝒅 𝑰𝒖 − 𝑷 − 𝑼 − 𝑷 − 𝑰𝒖 𝑰𝒖 − 𝑷 − 𝑼 − 𝑷 − 𝑰𝒅 𝑰𝒅 − 𝑷 − 𝑼 − 𝑷 − 𝑰𝒖 Semantic Users' similarities in terms of pairwise preference Agreement of users over desirability/ undesirability of an item over different items. Correlation between pairwise comparisons Direct relations between items; items that are simultaneously favored/ disfavored by users Indirect relationships between items. Items that are contrarily ranked by user. The pseudo code for constructing TPG is presented in Algorithm.1. It first generates 𝑀 nodes corresponding to 𝑀 users and 2𝑁 nodes for desirable and undesirable case of 𝑁 items. Then, it generates 𝑁(𝑁 − 1) nodes for each possible preference data. Each preference data has desirable and an undesirable sides that are represented by p.d and p.u, respectively. The next step, is to add links between users and preferences. It scans the preference database and for each < user, item1, item2 > triple, it adds a link between the corresponding nodes in the user and the preference layers. In more details, getUserNode(u) returns the corresponding node to user u while getPreferenceNode(i,j) returns the node p representing the preference of i over j. After that, a link is created between the node u and p. Finally, the algorithm scans through the preference nodes in which, a preference node p states that "p.d > p.u". For such a node, the algorithm finds the corresponding representative nodes using getDesirableNode(p.d) and getUnDesirableNode(p.u) and inserts an edge between p and each of those representatives. Algorithm 1. Construction of Tripartite Preference Graph (TPG) Input: Set of users U, Set of items I, Observation set of preference (O) Output: Tripartite graph (G) Initialize a graph G //Initializing user layer For each item 𝑢 ∈ 𝑈 Create a node 𝑢 in user layer //Initializing representative layer For each item 𝑖 ∈ 𝐼 Create a desirable-node 𝑖𝑑 in the representative layer Create an undesirable-node 𝑖𝑢 in representative layer //Initializing preference layer For each item 𝑖 ∈ 𝐼 For each item 𝑗 ∈ 𝐼 If (𝑖 ≠ 𝑗) // Connecting user and preference layer For each o:(u,i,j) ∈ 𝑂 u = G. GetUserNode (u); p = G. GetPreferenceNode (i ,j); Connect node u to node p. //Connecting preference and item layer For each node v ∈ preference-layer 𝑟𝑑= G.GetDesirableNode (p.d) 𝑟𝑢= G.GetUnDesirableNode (p.u) Make a link between p and 𝑟𝑑 Create a node p in preference layer p.d = i; p.u=j; Make a link between p and 𝑟𝑢. Example.1. Fig.4 illustrates how TPG can reflect the preference data mentioned in Fig.1 and Fig.2. As shown in Fig.4b, TPG clearly indicates that Mike and Martin (in Fig.1a and Fig.4a) have not the same preference as they neither share any neighbors nor longer paths to each other in TPG. On the other hand, Mike and Lee share one common neighbor that denotes their agreement over comparison of A and B. The same holds for Jack and John in the example of Fig.4c where TPG representation (Fig.4d) can evidently reflect that they do not have the same opinion over the comparison of A and B. Note that current graph-based approaches are not capable to model users' preference in these samples as mentioned in Fig.1b and Fig.2b. Fig.4. TPG representation of examples in illustrated in Fig.1, and, Fig.2, respectively. . Top-k recommendation using TPG graph It is a common assumption in recommender systems that users are interested in items that are preferred by their neighbors or are similar to their favorite items. As mentioned before, TPG provides an affluent platform to determine the users' similarity and items' relations. GRank exploits the closeness of users to the desirable/undesirable representatives to estimate how much the target user likes/dislikes a specific item. In other words, given a target user u and a TPG describing the observation set, GRank defines a function 𝐺𝑅: 𝑈 × 𝐼 → ℝ for predicting the goodness of each unseen item i for each user u based on the closeness of the node u to the desirable and undesirable representative nodes of item i. Then GRank recommends to u the items with the highest GR values calculated for u. The desirability/undesirability of an item i for user u can be estimated based on two general types of paths: Desirability and Undesirability paths. Desirability paths are in form of < 𝑢, 𝑣1, 𝑣2, … , 𝑣𝑚, 𝑖𝑑 > and show the closeness of a target user u to the desirable case of an item i. Similarly, undesirability paths are in form of < 𝑢, 𝑣1, 𝑣2, … , 𝑣𝑚, 𝑖𝑢 > and depict the closeness of the target user u and the undesirable case of item i. Intuitively, there are a large number of desirability/undesirability paths between the target user u and an item i. Therefore, some form of proximity measure is required to compare the number of desirability/undesirability paths between the target user and items. Recently, some proximity measures have been proposed for analyzing heterogeneous networks. Unfortunately, these methods are heavily dependent to the definition of meta paths and require to get the importance weight of different meta-paths as an input parameter to calculate the proximity among nodes(Lee et al., 2013; Sun et al., 2011). On the other hand, a general approach for measuring nodes' proximity in a network, is personalized PageRank, or PPR (Page, Brin, Motwani, & Winograd, 1999) that is acknowledged as one of the most effective measures that ranks nodes based on their reachability from a certain set of nodes in a network. It gives high scores to items that are closer to the target user regarding a wide range of graph properties such as distance or number of paths between them (Lee et al., 2013). GRank defines a measure based on PPR to calculate proximities, but before introducing the measure, first we briefly review the concept of PPR. Formally, the personalized PageRank of a node indicates the probability that a random walker, with a given skewed restarting distribution, will jump to that node. PPR can be considered as a Markov process with restart, and, is defined by Eq.1 PPR(t) = α · T · PPR(t − 1) + (1 − α) · PV (1) where PPR(t) denotes the rank vector at the 𝑡 − 𝑡ℎ iteration, 𝑇 is the transition matrix, 𝛼 is the damping factor and "𝑃𝑉" is the user-specific personalized vector. In most of applications 𝛼 is set to 0.85. To calculate closeness of some graph nodes, such as item nodes, to a particular node, like the target user's node, personalized PageRank needs to define a personalized vector 𝑃𝑉 as in Eq.2 𝑃𝑉(𝑗) = { 1, 𝑗 = 𝑢 0, o𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 where 𝑢 is the target node. (2) Given a graph G(V, E) with the set of all nodes V and the set of all edges E, each element of the transition matrix is obtained from Eq.3 T𝑖𝑗 = { , (𝑖, 𝑗) ∈ 𝐸 1 𝑑𝑖 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (3) Where 𝑑𝑖 is the degree of the 𝑖 − 𝑡ℎ node. As aforementioned, PPR can be used to find the closeness of each node to the target user's node in TPG. More clearly, personalized PageRank of the target users in TPG, estimates the probability that a random walker, starting from the target user, will follow a path to the desirable and undesirable representative of each item. GRank defines the goodness of an item i for the target user based on PPR of the desirable and undesirable cases of i, as in Eq. 4 𝐺𝑅(𝑖) = 𝑃𝑃𝑅(𝑖𝑑) 𝑃𝑃𝑅(𝑖𝑑)+𝑃𝑃𝑅(𝑖𝑢) (4) One can expect that 𝐺𝑅(𝑖) gives top scores to those items for which personalized PageRank of their desirable case is much higher than that of their undesirable case. To make a recommendation to the target user, items are sorted according to their GR values, and the top-k items are suggested. Algorithm 2 summarizes the GRank's approach for top-k recommendation. GRank requires to calculate the personalized PageRank of nodes for each target user u. For this purpose, it first defines the transition matrix of TPG and the personalization vector for u, using Eq.3 and Eq.2, respectively. Then, GRank randomly initializes the PPR values and then normalizes them to their summation. Next, it updates PPR value using Eq.1 and iterates until convergence. After that, GR values are calculated for each item using Eq.4. Finally, the items are sorted based on their GR values and the top-k items are recommended to the target user u Algorithm 2. Top-K recommendation on TPG Input: Tripartite graph (G), Target user u, number of recommended items K, set of Items I Output: The best k items. Initialize transition matrix T through Eq.3 Initialize personalized vector PV through Eq.2 Randomly initialize 𝑃𝑃𝑅0 𝑃𝑃𝑅0 = 𝑃𝑃𝑅0/𝑠𝑢𝑚(𝑃𝑃𝑅0 ) t=1 Repeat until convergence 𝑃𝑃𝑅𝑡 = 𝛼 · 𝑇 · 𝑃𝑃𝑅𝑡−1 + (1 − 𝛼) · 𝑃𝑉 t=t+1; For each item 𝑖 ∈ 𝐼 Put items in descending order of their GR values in list L Return as the recommendation list the first k items in L Calculate GR value for items through Eq.4 . Computational Complexity GRank is composed of two phases: Graph construction and recommendation task. Assume that the 𝑀 is the number of users, 𝑁 is the number of items and 𝑆 is the number of total pairwise preference assigned by all users. Clearly to total number of possible pairwise comparison would be 𝑁(𝑁 − 1). TPG contains 𝑁(𝑁 − 1) + 𝑀 + 2𝑁 vertices and 𝑆 + 2𝑁(𝑁 − 1) edges. The time complexity of constructing the graph would depend on the implementation approach. If we use adjacency lists, then it would have a time complexity of O(𝑁2 + M + N + S + 2𝑁2) = 𝑂(𝑁2 + 𝑀 + 𝑆). . Additionally, we often know that 𝑆 = 𝑐𝑁2 where c is a small constant (e.g. 2.48 for MovieLense100K). So, the time needed for graph construction phase is in O(𝑀 + 𝑁2). The time complexity of recommendation task is equal to that of personalized PageRank (PPR) computation. Computational complexity of personalized PageRank is 𝑂(𝑡𝐸) where 𝐸, is the number of graphs' edges and 𝑡 is the number of iterations needed before personalized PageRank converges. In TPG, we have (𝑆 + 2𝑁2 − 2𝑁) edges, and as we mentioned, we expect that 𝑆 = 𝑐𝑁2, for a small constant c. So the time complexity of a recommendation in GRank, is expected to be 𝑂(𝑡𝑁2). Note that 𝑡 is a small number for sparse graphs such as TPG. In our experiments, it does not exceed 20 Note that recommendation by GRank has a better computational complexity than EigenRank, the most acknowledged memory-based CR. EigenRank's computational complexity is in the order of 𝑂(𝑀𝑁2 + 𝐾𝑁2 + 𝑁2) where 𝑂(𝑀𝑁2) is for calculating the similarity between the target user and all other users, 𝑂(𝐾𝑁2) is for estimating the preference matrix, and, 𝑂(𝑁2) for inferring the total ranking. Experimental settings and results We have conducted a series of experiments for evaluating GRank algorithm. Here, we will first give a detailed description of the experimental protocol. Then, we will analyze the ranking quality and scalability of GRank. . Experimental setting 4.1.1 Dataset We conducted all experiments on two publicly available datasets that are widely used in related work (Fan & Lin, 2013; Rendle et al., 2009; Yue Shi, Larson, et al., 2013; Volkovs & Zemel, 2012; Wang et al., 2014): Both data sets have been generated by Movielens group , but contains different number of users, items, and, ratings. The first data set, Movielens-100K, consists of 100,000 ratings (scale 1-5) assigned by 943 users to a collection of 1,682 movies. The second dataset, Movielens-1M, is composed of 3,952 movies rated by 6, 040 users. There are one million ratings in this dataset. Since GRank is designed for using pairwise preference data, we converted the rating information into a set of pairwise comparisons using Rule 1: we created a preference instance of data <u, item#1, item#2) if item#1 has been rated higher than item#2 by user u. 4.1.2 Evaluation methodology In our experiments we followed a standard protocol widely used in related work (Balakrishnan & Chopra, 2012; Fan & Lin, 2013; J. Liu, Wu, Xiong, & Liu, 2014; Rendle et al., 2009; Y Shi et al., 2012; Volkovs & Zemel, 2012). We analyzed the effectiveness of our algorithm under different conditions of user profiles, regarding to the number of user's ratings; for each user, a fixed number 𝑇 of ratings was randomly sampled and placed in the training set, and the remaining ratings went to the test set. Our experiments involve 𝑇 = 20, 30, 40, 50 items. For each 𝑇, we make sure that we can compare algorithms on at least 10 rated items per user in the test set. Therefore, the users with respectively less than 30, 40, 50, and, 60 items are dropped from both train and test sets. We generated 5 variants of both data sets via random sampling and the average performance on all variants of the test set is reported. Fig. 5. Performance comparison of algorithms in terms of NDCG where T=20. 4.1.3 Baseline algorithms Since GRank makes a connection between neighbor-based collaborative ranking and graph-based recommendation, we have compared its performance against state of the art approaches of these two classes of methods. For further analysis, we also compare our algorithm to Cofi-Rank, the state of the art matrix factorization technique that is able to do recommendation based on pairwise preference data. These algorithms are briefly described in the following:  CofiRank: CofiRank (Weimer & Karatzoglou, 2007) is one of the state of the art MFCR techniques that extracts latent representations in order to optimize a structured loss function. CofiRank has several extensions. In our experiments, we used CofiRank-Ordinal and CofiRank-NDCG as our baseline algorithms. (Weimer et al., 2008) . CofiRank-Ordinal minimizes the number of discordant preferences in the predicted ranking list while CofiRank-NDCG triers to maximize the NDCG. We used the publicly available code for CofiRank and adopted the optimal values for its parameters as suggested in (Weimer et al., 2008).  EigenRank: EigenRank(N. Liu & Yang, 2008) is another famous algorithm in the family of NCR techniques. We have implemented the random-walk version of EigenRank using neighborhood sizes of 100 and 𝜀 = 0.85, that have been reported to be the best parameter values for the algorithm (N. Liu & Yang, 2008).  Graph-based recommendation: Also, we compare GRank with a graph-based recommendation algorithm that exploits a bi-partite graph structure to model the user-item interactions (such as Fig.1). Then, a random-walk with restart is used to rank items in the bi-partite graph(Chiluka et al., 2011). We have implemented two versions of this algorithm, abbreviated by BGR and WBGR, which links the users to those items rated by him with un-weighted and weighted links, respectively. In the weighted version, the weight of the edge between user u and item i is set equal to the rating of u to i. Fig. 6. Performance comparison of algorithms in terms of NDCG where T=30 . Results 4.2.1 Accuracy Following the standard evaluation strategy applied to recommender systems, we assessed the recommendation performance of the models by comparing the quality of their top-k suggestions. Normalized Discounted Cumulative Gain (NDCG) is an evaluation metric that is widely used for assessment of CR techniques. The definition of NDCG at the top-K suggestions for a user 𝑢 can be given as: 𝑁𝐷𝐶𝐺@𝐾 = 1 𝛼𝑢 ∑ 𝐾 𝑖=1 −1 𝑢 2𝑟𝑖 log (𝑖+1) (5) Where 𝐾 is the length of the recommendation list, 𝑟𝑖 i-th item in the recommendation list, and 𝛼𝑢 is the normalization factor to ensure that NDCG of the ideal recommendation for 𝑢 is equal to 1. In this paper, we report the recommendation performance by NDCG@3, NDCG@5 and NDCG@10, averaged across all users. 𝑢 is the rating given by user u to the Fig 7. Performance comparison of algorithms in terms of NDCG where T=40. The performance of the algorithms on ML-100k and ML-1M are shown in Fig.5 Fig.6, Fig.7, and, Fig.8. In most of experiments, GRank outperforms other algorithms. The results are reported based on a set of experiments on sufficiently large samples of data, and different algorithms are tested on common sets of samples. So, to test if the differences between the performance of GRank and other algorithms are significant, we can conduct a set of paired t-tests on the results. (Fouss, Pirotte, Renders, & Saerens, 2007; Guo, Zhang, & Yorke-Smith, 2015; Kim & Ahn, 2008; Shani & Gunawardana, 2011; Ting-Peng, Hung-Jen, & Yi-Cheng, 2006; Zhen et al., 2009). Given {𝑥1, … , 𝑥𝑁} be the set of samples derived from the dataset, let 𝜇𝑑 and 𝜎𝑑 as the average and variance of differences 𝑑𝑖 = 𝑎𝑖 − 𝑏𝑖 where 𝑎𝑖 and 𝑏𝑖 denote the performance of two approaches A and B on the i-th sample. To determine whether A significantly outperforms B or not, we consider the null hypothesis is 𝜇𝑑 = 0 whereas the alternative hypothesis is 𝜇𝑑 ≠ 0. The null hypothesis is rejected in favor of the alternative hypothesis if the p-value, obtained by the t-statistic 𝑡 = , is below than the 𝜇𝑑 𝜎𝑑/√𝑁 significance threshold (e.g. 0.01). Table 2 shows the p-values indicating the significant outperformance of GRank w.r.t other algorithms. Fig 8. Performance comparison of algorithms in terms of NDCG where T=50. The experimental results can be summarized as below:  GRank significantly outperforms all algorithms in the majority of evaluation conditions. Yet, EigenRank and WBGR show an improvement of less than 1% in ML-100k and Ml-1M where T=20.  The performance of GRank is up to 6% and 8% better than WBGR and BGR, respectively. This result reemphasizes the importance of capturing users' preference and their choice context for recommendation. It is worth noting that this feature will be more renowned while increasing T. In case of small number of training data, users rarely have common items that indicate their similarity.  GRank improves EigenRank about 1%-4% in ML-100K, while, its performance is up to 8% better than EigenRank in ML-1M. This results can be explained by the natural problem of Kendall correlation measure with sparse data. At fixed T, the number of items in ML-1M is two times more than ML-100K, and consequently is two times sparser, Therefore, EigenRank faces more difficulties to effectively calculate similarities in ML-1M and so wrong users may be picked by EigenRank as neighbors of a target user. On the other hand, GRank that does not rely on direct similarity calculation among users, can handle such sparse data sets quite well. Table 2. P-values obtained for the paired t-test under each evaluation condition. P<0.01 indicates the significant out performance of GRank w.r.t other algorithm NDCG NDCG ML-1m Eigen Rank BGR WBGR Cofi- BGR WBGR Cofi- Cofi- Ordinal ML-100K Ta Kb Eigen Cofi- Ordinal Rank 20 1 0.5396 0.038 0.3653 0.0003 0.0020 0.8310 0.7465 0.5417 0.0959 0.0006 30 1 0.6952 0.0040 0.0214 0.0002 0.0012 0.0100 0.0095 0.0071 0.0035 0.0000 40 1 0.0131 0.0040 0.0143 0.0002 0.0001 0.0002 0.0000 0.0003 0.0005 0.0007 50 1 0.0112 0.0040 0.0072 0.0000 0.0000 0.0051 0.0013 0.0015 0.001 0.0009 20 3 0.2577 0.0020 0.3033 0.0044 0.0005 0.2033 0.5991 0.3539 0.0481 0.0001 30 3 0.2908 0.0003 0.0041 0.0000 0.0007 0.0008 0.0007 0.0064 0.0073 0.0000 0.0069 0.0010 0.0064 0.0000 0.0000 0.0000 0.0000 0.0001 0.0003 0.0007 40 3 50 3 0.0014 0.0002 0.0009 0.0000 0.0000 0.0015 0.0006 0.0007 0.0008 0.0010 20 5 0.9047 0.0003 0.1455 0.0099 0.0001 0.2910 0.5021 0.2865 0.0321 0.0003 30 5 0.1842 0.0000 0.0014 0.0000 0.0004 0.0021 0.0002 0.002 0.0039 0.0000 40 5 0.0015 0.0007 0.0051 0.0000 0.0000 0.0000 0.0004 0.0002 0.0003 0.0007 50 5 0.0060 0.0000 0.0005 0.0000 0.0000 0.0007 0.001 0.0008 0.0005 0.0009 20 10 0.3772 0.0009 0.1446 0.0139 0.0000 0.5587 0.6757 0.2618 0.2020 0.0002 30 10 0.0070 0.0000 0.0007 0.0000 0.0002 0.0063 0.0000 0.0039 0.0069 0.0000 40 10 0.0005 0.0000 0.0010 0.0000 0.0000 0.0000 0.0003 0.0000 0.0003 0.0005 50 10 0.0046 0.0000 0.0000 0.0000 0.0000 0.0000 0.0006 0.0004 0.0004 0.0005 4.2.2 Scalability The final experiment investigates the scalability of GRank that, as mentioned in section 3.4, depends on three factors: number of users (𝑀), number of items (𝑁), and number of assigned preferences (𝑆). To investigate the scalability of GRank under each evaluation condition, that is the number of ratings available for each user in the training set (See Section 4.1.2), we measured the running time by altering one factor while fixing the other two parameters. For example, to evaluate the effect of the number of users on the running time of the algorithm, we fix the number of items and preferences in the training dataset, while varying the number of users by randomly selecting from 20% to 100% of all available users. Then, we compute the average running time for each recommendation. Similar steps have been followed for evaluating the effect of the number of pairwise preferences and the number of items on the running time. The results are presented in Fig.9. As shown in Figure. 9, the computational complexity is almost constant while varying 𝑀 and 𝑆, but, there is a quadratic rise of the running time when the number of items increases. Note that GRank's computational complexity is quadratic to the number of items since it considers all the possible pairwise comparisons. However, the real number of pairwise comparisons is much less than 𝑁2 in many practical applications, because items usually form clusters and the comparisons often happen within those clusters. This means that in practice, we may be able to prune TPG by omitting those comparisons that have not been done by any user. Fig. 9. Scalability analysis of GRank in terms of the number of users, items, and pairwise preferences assigned by the users Discussion GRank framework was introduced to resolve the sparsity problem of NCR techniques for similarity calculation. In the following, we briefly discuss how GRank is accomplished to resolve the issues.  Using TPG, GRank implicitly aggregates different kinds of users' similarities: One type of user similarities is calculated based on their common comparisons. This type of similarity is reflected by the paths following < 𝑈 − 𝑃 − 𝑈 > that connect two users through a pairwise comparison's node. Additionally, two users are assumed to be similar if they both have preferred a particular item A in some comparisons, even if the items over which it has been preferred are different for those two users. The same holds for situations in which two users prefer different items over some particular item A. These types of similarity can be discovered through tracking the paths in form of < 𝑈 − 𝑃 − 𝐼𝑑 − 𝑃 − 𝑈 > and < 𝑈 − 𝑃 − 𝐼𝑢 − 𝑃 − 𝑈 >, respectively. Combination and replication of these meta-paths define many different relations among entities and GRank captures and aggregates them in its rank calculation process.  Following the meta-paths < 𝑃 − 𝑈 − 𝑃 > in TPG, GRank implicitly finds correlated preferences that are the preferences of similar users. Correlate preference are highly connected through paths following < 𝑃 − 𝑈 − 𝑃 >. consequently, unknown preferences are iteratively estimated by propagating the known preferences of the target user to those unknown preferences that are correlated to them. This information can be used to calculate users' similarity even in case of no common pairwise comparisons. We refer to an example to clarify the concept: As illustrated in Fig.3, 𝑢1, 𝑢2, and 𝑢5 , each one, has one link to the preference node B<A, B<D, and C<D. Therefore, the similarity between each pair of them will be zero according to the Kendall correlation. However, TPG reflects that {𝐴 > 𝐵} is highly correlated to the preference node{𝐶 < 𝐷} while it has no relations with B<D. Therefore, 𝑢1 is more similar to 𝑢5 as a consequence of following the paths passing from{𝐴 > 𝐵} and {𝐶 > 𝐷}  Taking advantages of TPG and PageRank algorithm, GRank directly estimates the users' ranking over unseen items. Personalized PageRank computation in TPG enables GRank to aggregate the ranking information obtained from different forms of meta-paths < 𝑢, 𝑣1, … , 𝑣𝑚, 𝑖𝑑 > and < 𝑢, 𝑣1, … , 𝑣𝑚, 𝑖𝑢 > for fast and accurate prediction of users' ranking. Note that this approach differs from the typical 3-step framework (calculation of similarity, generation of preference matrix, and, inference of total ranking) used in all neighbor-based approaches. Conclusion In this paper, we studied how a graph-based framework can be designed and exploited to address the shortcomings of current neighbor-based collaborative ranking algorithms. For this purpose, we suggested that modeling the preference data as a new tri-partite graph structure and then exploring it can help us to capture the different kinds of relations existing in a ranking preference dataset (e.g. users' similarities, items' similarities, etc.). We also proposed a random-walk approach to make recommendation based on the proposed structure. Experimental results showed significant improvement of the suggested framework, GRank over other state-of-the-art graph-based and neighbor-based collaborative ranking methods. It seems that the graph based approach of GRank can be beneficial both in sparse and dense data sets. In dense data sets, it can form the neighborhoods more precisely, by exploring different paths that exist among entities. In sparse data sets, that users rarely have common pairwise comparisons and direct neighborhoods are usually very small, it can still traverse the edges to find farther neighbors and use their information as well for recommendation. The proposed graph structure has been mainly used here for finding closeness between users and items, but it can also be used for other purposes like finding clusters of similar users and similar items, and also discovering correlated preferences which are some of essential concepts in the field of recommendation systems. References Adomavicius, G., & Kwon, Y. (2012). Improving Aggregate Recommendation Diversity Using Ranking-Based Techniques. IEEE Transactions On Knowledge And Data Engineering, 24(5), 896–911. AvSegal, I., Katzir, Z., & Gal, K. (2014). EduRank : A Collaborative Filtering Approach to Personalization in E-learning. In Educational Data Mining 2014 (pp. 68–75). Balakrishnan, S., & Chopra, S. (2012). Collaborative ranking. In Proceedings of the fifth ACM international conference on Web search and data mining. ACM (pp. 143–152). Bellogin, A., & Parapar, J. (2012). Using graph partitioning techniques for neighbour selection in user-based collaborative filtering. In Proceedings of the sixth ACM conference on Recommender systems (pp. 213–216). Cacheda, F., Carneiro, V., Fernández, D., & Formoso, V. (2011). Comparison of collaborative filtering algorithms. ACM Transactions on the Web, 5(1), 1–33. http://doi.org/10.1145/1921591.1921593 Chen, B., Wang, J., Huang, Q., & Mei, T. (2012). Personalized video recommendation through tripartite graph propagation. ACM Transactions on Multimedia Computing, Communications 1133. http://doi.org/10.1145/2393347.2396401 Applications, and 10(4), Chiluka, N., Andrade, N., & Pouwelse, J. (2011). A Link Prediction Approach to Recommendations in Large-Scale User-Generated Content Systems. In Advances in information retrieval (pp. 189–200). Chowdhury, N., Cai, X., & Luo, C. (2015). BoostMF : Boosted Matrix Factorisation. In Machine Learning and Knowledge Discovery in Databases (ECML/PKDD) (pp. 3– 18). http://doi.org/10.1007/978-3-319-23525-7 Christakopoulou, K., & Banerjee, A. (2015). Collaborative Ranking with a Push at the Top. In WWW 2015: Proceedings of the 24th International Conference on World Wide Web (pp. 205–215). http://doi.org/10.1145/2736277.2741678 Cong, J. (2009). RankClus : Integrating Clustering with Ranking for Heterogeneous Information Network Analysis. In Proceedings of the 12th International Conference on Extending Database Technology: Advances in Database Technology. ACM, (pp. 565–576). Desrosiers, C., & Karypis, G. (2011). A comprehensive survey of neighborhood-based recommendation methods. In Recommender systems handbook (pp. 107–144.). Dhanjal, C., Clémençon, S., & Gaudel, R. (2015). Collaborative filtering with localised ranking. In Proceedings of Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI'15) (Vol. 4, pp. 2554–2560). Fan, C., & Lin, Z. (2013). Collaborative Ranking with Ranking-Based Neighborhood. Web Technologies and Applications, 770–781. Fouss, F., Pirotte, A., Renders, J. M., & Saerens, M. (2007). Random-walk computation of similarities between nodes of a graph with application to collaborative recommendation. IEEE Transactions on Knowledge and Data Engineering, 19(3), 355–369. http://doi.org/10.1109/TKDE.2007.46 Guo, G., Zhang, J., & Yorke-Smith, N. (2015). Leveraging multiviews of trust and similarity to enhance clustering-based recommender systems. Knowledge-Based Systems, 74, 14–27. http://doi.org/10.1016/j.knosys.2014.10.016 Huang, Z., Li, X., & Chen, H. (2005). Link Prediction Approach to Collaborative Filtering. In Proceedings of the 5th ACM/IEEE-CS joint conference on Digital libraries (pp. 141–142). Jiang, J., Lu, J., Zhang, G., & Long, G. (2011). Scaling-Up Item-Based Collaborative Filtering Recommendation Algorithm Based on Hadoop. 2011 IEEE World Congress on Services, 490–497. http://doi.org/10.1109/SERVICES.2011.66 Kim, K. jae, & Ahn, H. (2008). A recommender system using GA K-means clustering in an online shopping market. Expert Systems with Applications, 34(2), 1200–1209. http://doi.org/10.1016/j.eswa.2006.12.025 Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix Factorization Techniques for Recommender Systems. Computer, 42(8), 30–37. Lee, S., Park, S., Kahng, M., & Lee, S. (2013). PathRank : Ranking nodes on a heterogeneous graph for flexible hybrid recommender systems. Expert Systems with Applications, 40, 684–697. Lerche, L., & Jannach, D. (2014). Using Graded Implicit Feedback for Bayesian Personalized Ranking. In Proceedings of the 8th ACM Conference on Recommender systems. ACM (pp. 353–356). Li, X., & Chen, H. (2013). Recommendation as link prediction in bipartite graphs: A graph kernel-based machine learning approach. Decision Support Systems, 54(2), 880–890. http://doi.org/10.1016/j.dss.2012.09.019 Liu, J., Wu, C., Xiong, Y., & Liu, W. (2014). List-wise probabilistic matrix factorization for recommendation. Information Sciences, 278, 434–447. Liu, N., & Yang, Q. (2008). EigenRank : A Ranking-Oriented Approach to Collaborative Filtering. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval. ACM (pp. 83–90). Lu, Q., Chen, T., Zhang, W., Yang, D., & Yu, Y. (2012). Serendipitous personalized IEEE/WIC/ACM 258–265. ranking International Conference on Web http://doi.org/10.1109/WI-IAT.2012.135 recommendation. Proceedings for top-N - 2012 Intelligence, WI 2012, Meng, L., Li, J., & Sun, H. (2011). WSRank: A Collaborative Ranking Approach for Web Service Selection. In 2011 IEEE 11th International Conference on Computer and Information Technology (pp. 103–108). Ieee. http://doi.org/10.1109/CIT.2011.30 Musa, A. B. (2012). Comparative study on classification performance between support vector machine and logistic regression. International Journal of Machine Learning and Cybernetics, 4(1), 13–24. http://doi.org/10.1007/s13042-012-0068-x Page, L., Brin, S., Motwani, R., & Winograd, T. (1999). The PageRank citation ranking: bringing order to the web., 1–17. http://doi.org/10.1109/ICCKE.2013.6682858 Pan, W., Zhong, H., Xu, C., & Ming, Z. (2015). Adaptive Bayesian personalized ranking for heterogeneous implicit feedbacks. Knowledge-Based Systems, 73, 173–180. Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-thieme, L. (2009). BPR : Bayesian Personalized Ranking from Implicit Feedback. In Proceedings of the Twenty-Fifth Conference 452–461). http://doi.org/10.1145/1772690.1772773 on Uncertainty in Artificial Intelligence (pp. Said, A., Kille, B., Jain, B., & Albayrak, S. (2012). Increasing Diversity Through Furthest Neighbor-Based Recommendation. In Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, 1–4. Sawant, S. (2013). Collaborative Filtering using Weighted BiPartite Graph Projection - A Recommendation System for Yelp. Proceedings of the CS224W: Social and Information Network Analysis Conference. Shang, M. S., Fu, Y., & Chen, D. Bin. (2008). Personal recommendation using weighted bipartite graph projection. In 2008 International Conference on Apperceiving Computing Intelligence Analysis, 198–202). http://doi.org/10.1109/ICACIA.2008.4770004 ICACIA 2008 (pp. and Shang, M.-S., Zhang, Z.-K., Zhou, T., & Zhang, Y.-C. (2010). Collaborative filtering with diffusion-based similarity on tripartite graphs. Physica A: Statistical Mechanics and Its Applications, 389(6), 1259–1264. http://doi.org/10.1016/j.physa.2009.11.041 Shani, G., & Gunawardana, A. (2011). Evaluating recommendation systems. Recommender Systems Handbook, 257–298. http://doi.org/10.1007/978-0-387- 85820-3_8 Shi, Y., Karatzoglou, A., & Baltrunas, L. (2012). Climf: learning to maximize reciprocal rank with collaborative less-is-more filtering. In Proceedings of the sixth ACM conference on Recommender systems. ACM (pp. 139–146). Shi, Y., Karatzoglou, A., Baltrunas, L., & Larson, M. (2013). xCLiMF : Optimizing Expected Reciprocal Rank for Data with Multiple Levels of Relevance. In Proceedings of the 7th ACM conference on Recommender systems. ACM (pp. 431– 434). Shi, Y., Larson, M., & Hanjalic, A. (2010). List-wise learning to rank with matrix factorization for collaborative filtering. In Proceedings of the fourth ACM conference on Recommender system (pp. 269–272). Shi, Y., Larson, M., & Hanjalic, A. (2013). Unifying rating-oriented and ranking-oriented collaborative filtering for improved recommendation. Information Sciences, 229, 29– 39. Silva, A., & Zaki, M. (2013). ProfileRank : Finding Relevant Content and Influential Users based on Information Diffusion. In Proceedings of the 7th Workshop on Social Network 1–9). http://doi.org/10.1145/2501025.2501033 Analysis. Mining ACM and (pp. Sun, Y., Han, J., Yan, X., & Yu, P. S. (2011). PathSim : Meta Path-Based Top-K Similarity Search in Heterogeneous Information Networks. In VLDB (pp. 992–1003). Ting, Y., Yan, C., & Xiang-wei, M. (2013). Personalized Recommendation System Based on Web Log Mining and Weighted Bipartite Graph. 2013 International Conference on 587–590. http://doi.org/10.1109/ICCIS.2013.161 Computational Information Sciences, and Ting-Peng, L., Hung-Jen, L. A. I., & Yi-Cheng, K. U. (2006). Personalized Content Recommendation and User Satisfaction: Theoretical Synthesis and Empirical Findings. Journal of Management Information Systems, 23(3), 45–70. http://doi.org/10.2753/MIS0742-1222230303 Volkovs, M. N., & Zemel, R. S. (2012). Collaborative Ranking With 17 Parameters. In Advances in Neural Information Processing Systems (pp. 2294–2302). Wang, S., Sun, J., & Gao, B. J. (2014). VSRank : A Novel Framework for Ranking-Based collaborative filtering. ACM Transactions on Intelligent Systems and Technology (TIST), 5(3), 51. Weimer, M., & Karatzoglou, A. (2007). Maximum Margin Matrix Factorization for Collaborative Ranking. Advances in Neural Information Processing Systems, 1–8. Weimer, M., Karatzoglou, A., & Smola, A. (2008). Improving maximum margin matrix factorization. Machine Learning, 72(3), 263–276. http://doi.org/10.1007/s10994-008- 5073-7 Xiang, L., Yuan, Q., Zhao, S., Chen, L., Zhang, X., Yang, Q., & Sun, J. (2010). Temporal recommendation on graphs via long- and short-term preference fusion. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 723–731). ACM Press. http://doi.org/10.1145/1835804.1835896 - KDD '10 Xiao, Z., Che, F., Miao, E., & Lu, M. (2014). Increasing serendipity of recommender system with ranking topic model. Applied Mathematics and Information Sciences, 8(4), 2041–2053. http://doi.org/10.12785/amis/080463 Yang, C., Wei, B., Wu, J., Zhang, Y., & Zhang, L. (2009). CARES: a ranking-oriented CADAL recommender system. In Proceedings of the 9th ACM/IEEE-CS joint conference on Digital libraries. ACM (pp. 203–212). Yao, W., He, J., Huang, G., Cao, J., & Zhang, Y. (2013). Personalized Recommendation on Multi-Layer context graph. In Web Information Systems Engineering–WISE 2013 (pp. 135–148). Yu, X., Ren, X., Sun, Y., & Gu, Q. (2014). Personalized Entity Recommendation : A Heterogeneous Information Network Approach. In 7th ACM international conference on Web search and data mining (pp. 283–292). Yu, X., Ren, X., Sun, Y., Sturt, B., Khandelwal, U., Gu, Q.,Han, J. (2013). Recommendation in Heterogeneous Information Networks with Implicit User Feedback. In Proceedings of the 7th ACM conference on Recommender systems. ACM (pp. 347–350). Zhang, M., Guo, X., & Chen, G. (2015). Prediction uncertainty in collaborative filtering : Enhancing personalized online product ranking. Decision Support Systems, 83, 10– 21. http://doi.org/10.1016/j.dss.2015.12.004 Zhang, Z., & Kulkarni, S. R. (2014). Detection of Shilling Attacks in Recommender Systems via Spectral Clustering. In Information Fusion (FUSION), 2014 17th International Conference on. IEEE. (pp. 1–8). Zhang, Z.-K., Zhou, T., & Zhang, Y.-C. (2010). Personalized recommendation via integrated diffusion on user–item–tag tripartite graphs. Physica A: Statistical Mechanics 179–186. http://doi.org/10.1016/j.physa.2009.08.036 Applications, 389(1), and Its Zhen, Y., Li, W. J., & Yeung, D. Y. (2009). TagiCoFi: Tag Informed Collaborative Filtering. Proceedings of the Conference on Recommender Systems (RecSys), 69–76. http://doi.org/10.1145/1639714.1639727 Zhou, T., Kuscsik, Z., Liu, J.-G., Medo, M., Wakeling, J. R., Zhang, Y.-C., & Settings, M. (2010). Solving the apparent diversity-accuracy dilemma of recommender systems. Proceedings of the National Academy of Sciences of the United States of America, 107(10), 4511–5. http://doi.org/10.1073/pnas.1000488107 Zhou, T., Ren, J., Medo, M., & Zhang, Y. C. (2007). Bipartite network projection and personal recommendation. Physical Review E - Statistical, Nonlinear, and Soft Matter Physics, 76(4), 1–7. http://doi.org/10.1103/PhysRevE.76.046115
1103.4550	1	1103	2011-03-23T15:27:57	Community Detection via Semi-Synchronous Label Propagation Algorithms	[ "cs.SI", "physics.soc-ph" ]	A recently introduced novel community detection strategy is based on a label propagation algorithm (LPA) which uses the diffusion of information in the network to identify communities. Studies of LPAs showed that the strategy is effective in finding a good community structure. Label propagation step can be performed in parallel on all nodes (synchronous model) or sequentially (asynchronous model); both models present some drawback, e.g., algorithm termination is nor granted in the first case, performances can be worst in the second case. In this paper, we present a semi-synchronous version of LPA which aims to combine the advantages of both synchronous and asynchronous models. We prove that our models always converge to a stable labeling. Moreover, we experimentally investigate the effectiveness of the proposed strategy comparing its performance with the asynchronous model both in terms of quality, efficiency and stability. Tests show that the proposed protocol does not harm the quality of the partitioning. Moreover it is quite efficient; each propagation step is extremely parallelizable and it is more stable than the asynchronous model, thanks to the fact that only a small amount of randomization is used by our proposal.	cs.SI	cs	Community Detection via Semi–Synchronous Label Propagation Algorithms Gennaro Cordasco and Luisa Gargano Dipartimento di Informatica ed Applicazioni “R.M. Capocelli” University of Salerno, Fisciano 84084, ITALY {cordasco,lg}@dia.unisa.it Abstract—A recently introduced novel community detection strategy is based on a label propagation algorithm (LPA) which uses the diffusion of information in the network to identify communities. Studies of LPAs showed that the strategy is effective in finding a good community structure. Label propagation step can be performed in parallel on all nodes (synchronous model) or sequentially (asynchronous model); both models present some drawback, e.g., algorithm termination is nor granted in the first case, performances can be worst in the second case. In this paper, we present a semi–synchronous version of LPA which aims to combine the advantages of both synchronous and asynchronous models. We prove that our models always converge to a stable labeling. Moreover, we experimentally investigate the effectiveness of the proposed strategy comparing its performance with the asynchronous model both in terms of quality, efficiency and stability. Tests show that the proposed protocol does not harm the quality of the partitioning. Moreover it is quite efficient; each propagation step is extremely parallelizable and it is more stable than the asynchronous model, thanks to the fact that only a small amount of randomization is used by our proposal. I. INTRODUCTION Collaboration networks, the Internet, the World-Wide-Web, Biological networks, Communication and Transport networks, Social networks are just some examples of wide complex networks. An interesting property to investigate, typical to many such networks, the division of networks into groups of nodes that are similar among them but dissimilar from the rest of the network. The capability of detecting the partitioning of a network into communities can give important insights into the organization and behavior of the system that the network models. is the community structure, i.e. The generally adopted notion of community structure in complex networks [GN02] refers to the fact that nodes in many real networks appear to group into subgraphs which are densely inter-connected and sparsely connected to other parts of the network. The quite challenging problem of com- munity detection has attracted ample attention in recent years and community detection methods have been applied in a wide range of scientific problems such as Social networks, Citation networks, the World Wide Web, and many others [AB02]. Several different approaches have been proposed to find community structures in networks; reviews of the various methods present in the literature can be found in [DDGDA05] and [OL09]. 978-1-4244-8998-5/10/$26.00 c(cid:13) 2010 IEEE Recently Raghavan et al. [RAK07] proposed a label prop- agation algorithm (LPA) for detecting network communities. This algorithm uses only the network structure as a guide, and can be summarized as follows: Each node in the network is first given a unique label; at each iteration, each node is updated by choosing the label which is the most frequent among its neighbors – if multiple choices are possible (as for example in the beginning), one label is picked randomly. Experiments have shown that the label propagation technique is very effective in discovering accurate community structure. In [RAK07], the authors suggest to use an asynchronous label propagation approach, since otherwise the process may result in a cyclic oscillation of the labels of some vertices which precludes the convergence (termination) of the algorithm; the use of an asynchronous algorithm means to update only one node at time which implies O(m) time for each iteration, if the network has m links. In general, while the experiments reported in [LHLC09] empirically show that the expected number of iterations grows logarithmically with respect to the size of the network, the problem of analytically proving the convergence of this method and of determining its speed is still an open problem (in both synchronous and asynchronous models). Liu and Murata [LM09] introduced a variation of the label propagation algorithm specifically developed for bipartite network; the algorithm is semi–synchronous: it allows to update simultaneously all the nodes belonging to one of the partitions of the bipartite networks. The authors experience that by running this version of the LPA, the oscillation problem disappears. However, no formal proof (or an informal hint) that shows why their proposal is able to defeat the oscillation issue has been provided. In this paper we extend the approach in [LM09] to any graph. Moreover, we show that no randomization is needed for obtaining good algorithm performance and formally prove that the proposed algorithm converges. The rest of the paper is organized as follows. Section II describes the notation used in the paper and introduce the modularity measure that will be used to evaluate the quality of the discovered community structures. In Section III, we briefly discuss on community detection strategies and introduce the LPA. Section IV introduces three variants of the standard LPA, which have been proposed to improve its features and discusses the stop criteria of LPAs. In Section V, we present and analyze our algorithm. In Section VI, we report on the experimental data. II. NOTATION: GRAPHS, PARTITIONING AND COLORING Let G = (V, E) be an undirected connected network having n = V vertices and m = E edges. Let v ∈ V , we denote by N (v) = {u : u ∈ V, {u, v} ∈ E} the neighborhood of v, by deg(v) = N (v) the degree of v and by deg(G) = max v∈V deg(v) the maximum degree over all the vertices in G. Denote by C = {C1, C2, . . . , Ck} a partition of the vertices in V . Each Ci is called community. Each community Ci is associated to an induced subgraph G(Ci) = (Ci, E(Ci)) of G, where E(Ci) = {{u, v} : u, v ∈ Ci, {u, v} ∈ E}. A good community detection strategy should strive to achieve two conflicting goals: • provide an accurate network partition and • keep low the computational complexity of the algorithm so as the algorithm can be applied to very large networks. While the latter goal is easily measured by means of a standard worldwide recognized approach (e.g., Asymptotic notation), how to evaluate the quality of a given network partition is a recently raised question. Several metrics have been proposed in the literature, e.g., edges density or NMI (Normalized mutual information) [DDGDA05]. Among these, a metric which has been widely adopted to measure the accuracy of a network par- tition is the graph modularity proposed by Newman [NG04]. This measure is obtained by summing up, over all communities of a given partition, the difference between the observed fraction of links inside the community and the expected value for a null model, that is, a randomized network having the same size and same degree sequence. Formally, the modularity of a given partition C of a graph G = (V, E) is defined as q(C) = XC∈C" E(C) m −(cid:18)Pv∈C deg(v) 2m (cid:19)2# . (1) where E(C) is the set of edges connecting two vertices of the community C ∈ C. Given a graph G, the network coloring problem is that of coloring the vertices of a network so that no two adjacent vertices share the same color. The smallest number of colors needed to color a graph G is called its chromatic number, X (G). Graph coloring is computationally hard. Especially, it is NP-hard to compute the chromatic number. However, there are several algorithms, even parallelizable [BE09], which allow to color a graph with a number of colors upper bounded by deg(G) + 1. III. RELATED WORK A. Community Detection The goal of community detection algorithms is to partition a given network into communities consisting of nodes with similar characteristics. Specifically, a community is generally defined as a subset of nodes densely interconnected relatively to the rest of the network. Several algorithm have been proposed which are typically classified in: divisive [GN02], agglomerative[NG04] (depending on whether they focus on the addition or removal of edges to or from the network) and optimization [BDG+07] which continuously update the network partition in order to maximize a given measure of the quality of the network partition (i.e., the modularity). See [DDGDA05] and [OL09] for detailed reviews. B. Label propagation algorithms The community detection strategy based on a label propaga- tion algorithm (LPA), introduced by Raghavan et al. [RAK07], in contrast with previously proposed approaches, identifies network partitions by an “epidemic” approach, i.e., it uses the diffusion of information in the network to identify communi- ties. Main idea: Initially, each vertex in the network is as- signed a unique label, which will be used to determine the community it belongs to. Subsequently, an iterative process is performed so that connected groups of vertices are able to reach a consensus on some label giving rise to a community. At each step of the process, each vertex updates its label to a new one which corresponds to the most frequent label among its neighbors. Formally, for each vertex v ∈ V , v updates it label according to lv = argmax [lu == l] (2) l Xu∈N (v) where lv denotes the label of v and [P ] =(1 0 if the statement P is true otherwise denotes the Iverson bracket. When more than one choice is possible, ties are broken randomly (different ties management schemes will be considered in the rest of the paper). This process is performed until some stop condition is met, e.g., no vertex changes its label during one step. A network community is then identified as a connected group of vertices having the same final label. The process is formally described as Algorithm 1: Label Propagation (Synchronous). In the following, we denote by lv(i) the label of vertex v at step i, for i = 0, 1, . . . and ∀ v ∈ V. Initially all labels are distinct, that is lv(0) 6= lw(0) ∀ v, w ∈ V . C. Synchronous vs Asynchronous LPA The label propagation algorithm can be either synchronous, as presented above, or asynchronous. In the synchronous model (cf. Algorithm 1) each vertex computes its label at Algorithm 1: Label Propagation (Synchronous) 1 Initialize labels: for each v ∈ V, lv(0) = v; 2 i=0; 3 while the stop criterion is not met do 4 5 6 i++; Propagation: foreach v ∈ V do lv(i) = argmax [lu(i − 1) == l], 7 l Xu∈N (v) end 8 9 end 10 return Final labeling: lv(t) for each v ∈ V , where t is the last executed step. step i based on the label of its neighbors at step i − 1. The synchronous algorithm is easy to implement and em- barrassingly parallelizable. Since there are no dependencies between labels belonging to the same step, each Propagation step can be executed in parallel over the vertices. However, it has been shown that synchronous updating may result in a cyclic oscillation of labels. This problem mainly occurs when the network considered contains a bipartite or a star-like component. We observe that label oscillations are not only due to bipartite or star subgraphs, several other kinds of graphs, see for instance Figure 1, suffer this oscillation problem. In order to avoid possible cycles and ensure termination the authors of [RAK07] suggest opting for an asynchronous approach (cf. Algorithm 2). Although, oscillation phenomenon, it exhibits some side effects: the asynchronous approach deeply reduces the • Since each vertex label is updated according to the current label of its neighbors, several dependencies need to be considered if one has to parallelize the algorithm: each vertex cannot compute his own label before each of its neighbor, which precedes it in the chosen permutation, has completed its computation. With more details, while the f oreach in line 6 of Algorithm 1 is parallelizable, this is not true for the Algorithm 2 (the computation on line 8 depends on the current labels of the considered vertices). Parallelism is very important when the network size is large. Even if the LPA requires a few propagation steps, using the asynchronous approach, each step need to be sequenced among all vertices. Accordingly, the amount of time needed by a parallelized version of the synchronous LPA scales logarithmically [LHLC09], whereas the time complexity of a parallel asynchronous LPA grows more than linearly, with respect to the network size. • Moreover, because during each iteration, the updating se- quence is randomly chosen, the whole algorithm becomes unstable: different run of the algorithm may provide different final labeling. • A major limitation of the asynchronous algorithm has been observed in [LHLC09]: It often wrongly produces a “monster” community and several small communities. Algorithm 2: Label Propagation (Asynchronous) 1 Initialize labels: for each v ∈ V, lv(0) = v; 2 i=0; 3 while the stop criterion is not met do 4 5 i++; let π = (vπ(1), vπ(2), . . . , vπ(n)) be a random permutation of the vertices. Propagation: for j = 1 to n do 6 7 8 9 lvπ(j) (i) = argmax l [lu == l], Xu∈N(vπ(j)) where lu =(lu(i) lu(i − 1) if u = vπ(k), k < j if u = vπ(k), k > j end 10 11 end 12 return Final labeling: lv(t) for each v ∈ V , where t is the last executed step. This problem is related to the fact that, due to the asyn- chronous nature of the algorithm, during the initial steps, the random permutation of the vertices tends to benefit the spread of some labels with respect to the others. For this reason certain communities do not form links which are strong enough to prevent a foreign label flooding. Several experiments confirm that the synchronous version of the algorithm slows down the formation of such monster communities, even though it does not completely prevent them [LHLC09]. D. LPA for bipartite networks A variation of the LPA for bipartite networks was proposed in [LM09]. The proposed algorithm has experimentally shown to be as effective as standard LPA, easily parallelizable, and stable (no label oscillations were observed). Let B (blue) and R (red) be the two sets of vertices which correspond to the canonical partition of a bipartite network. The algorithm divides each propagation step into two synchronized stages. The first stage computes the new labels for blue vertices according to the current labeling of red vertices. When the labels of all the blue vertices have been updated, the second stage updates the label of red vertices according to the current labeling of blue vertices. This algorithm corresponds to the asynchronous model where, for each propagation step, the permutation of vertices is such that blue vertices precede red ones. Notice that, since the network is bipartite, both the stages are parallelizable: the label propagation for different red (resp. blue) vertices is independent of each other. IV. LPA STOPPING CRITERIA AND TIE RESOLUTION STRATEGIES Before proceeding with the description of our proposal, we briefly discuss the stop criteria and tie resolution strategies for LPAs that will be considered in the following. Fact 1: The synchronous version of LPA-Max does not generate cycle of size larger than two. Fact 1 implies a simplification of the stop criterion (c1): if either lv(i) = lv(i − 1) for each v ∈ V or lv(i) = lv(i − 2) for each v ∈ V, then the algorithm is stopped. Synchronous LPA-Max is completely deterministic; it will always generate the same network partition whenever it starts with the same initial vertex labels. Anyway, by randomizing the initial labeling, or by using an asynchronous approach, it is possible to obtain different results. (c2) Notice that when the number of initial labels is larger than 2, the two variants of LPA can also be applied together, we will call this tie resolution strategy LPA-Prec-Max. V. SEMI–SYNCHRONOUS LABEL PROPAGATION In this Section we present the main contribution of this paper. Our proposal combines the advantages of both the syn- chronous and asynchronous model discussed above. Namely, our system is stable and efficient (easy parallelizable) as the synchronous model while does not undergo the oscillation problem. We present a semi-synchronous LPA which allows to over- come the oscillation problem in any network. We will also formally prove that our algorithm avoids oscillations, that is, it converges to a stable labeling1. Our work is inspired by the label propagation algorithm for bipartite networks given in [LM09]. We stress that, in general, the formal study of LPAs convergence is an open problem. In particular, no formal proof (or informal hint) that shows why the proposal in [LM09] is able to defeat the oscillation issue has been provided. We propose an algorithm which consists of two phases: 1) Coloring Phase. Color the network vertices so that no two adjacent vertices share the same color (i.e., by any distributed graph coloring algorithm). The coloring phase is easily parallelizable and efficient (only O(deg(G)) synchronous parallel steps) [BE09]; 2) Propagation Phase. Each label propagation step is di- vided into stages. Each stage is named upon a different color. At stage c, labels are simultaneously propagated to the vertices that have been assigned color c during the coloring phase 1. is easy to verify that A formal description of the algorithm is given as Algorithm 3: LPA (Semi–synchronous). It the number of stages per propagation step corresponds to the number of color needed to color the network, which is at most deg(G) + 1. Moreover, each step of the propagation phase is easily parallelizable: since there are no dependencies between vertices having the same color, each stage can be executed synchronously. Hence the amount of time needed to reach a final consensus grows like the number of propagation 1Depending on the propagation rule used, some cycles can be still obtained due to the management of ties. Fig. 1. The oscillation phenomenon on a non-bipartite network. Once one of the two configurations is entered, then the label values indefinitely oscillate between them. The simplest stop criterion consists in comparing the current labeling with the previous one, if no label change occurs, the algorithm is stopped (successive step will not change any label). Unfortunately, the LPA, even in the asynchronous model, does not prevent cycles (i.e., a sequence of steps such that the final labeling corresponds to the initial one). In such cases the stop criterion defined above will fail. For this reason the stop criterion (c1) proposed in [RAK07] is more elaborated: if in the current labeling every vertex in the network has a label to which the maximum number of its neighbors belong to, then the algorithm is stopped. (c1) Said in other words, if in the next step all the labels change, if any, will come from ties, then the procedure ends and the current labeling determines a network partition. Alternatively, in order to use a simple version of the stop criterion a change in the management of ties is required as proposed in [BC09]. When a ties occurs (i.e., there are two or more labels that maximize the sum in the equation (2)) one of the labels that maximize the sum is chosen randomly. If, in case of ties, the current label of the vertex has priority over the others, the probability of generating cycles is sensibly reduced [BC09]. We will call this version of the algorithm LPA with Precedence (LPA-Prec). Formally this variation is defined as follows: during the propagation step, if the current label satisfies equation (2), then the vertex keeps its current label, otherwise the algorithm follows the standard rules. LPA-Prec is typically more stable then the LPA algorithm (a small amount of randomization is injected into the algorithm). However in some cases the LPA-Prec stops whereas the standard LPA keeps running in search of better solutions. For this reason, in addition to be more stable, the LPA-Prec usually exhibits a faster convergence then LPA but the quality of the network partition generated can be a bit worse. Another tie resolution strategy guarantees the convergence of the algorithm. Assume that the set of labels allows defining a priority between each pair of labels. For instance, we can assume that labels are integer, and that a label l has priority over a label l′ if l > l′. In this version of the algorithm, named LPA-Max, each tie is solved deterministically by taking the label with higher priority between the set of labels that get the maximum in equation (2). Using results in [PS83] one can show that Algorithm 3: LPA (Semi–synchronous) 1 Initialize labels: for each v ∈ V, lv(0) = v; 2 Network coloring: assign a color to the vertices of the network such that no two adjacent vertices share the same color. Let D = {D1, D2, . . . , Dℓ} be the color partitioning obtained. 3 i=0; 4 while the stop criterion is not met do 5 6 7 8 i++; Propagation: for j = 1 to ℓ do foreach v ∈ Dj do lv(i) = argmax 9 10 [lu == l], l Xu∈N (v) where lu =(lu(i) lu(i − 1) if u ∈ Dk, k < j if u ∈ Dk, k > j end end 11 12 13 end 14 return Final labeling: lv(t) for each v ∈ V , where t is the last executed step. step times the number of stages per propagation step, where the former value has been experimentally shown to grow logarithmically with respect to n, while the latter is bounded by deg(G) + 1. Theorem 1: Consider a network G = (V, E). Assume that Algorithm 3 uses the stop criterion (c1), that is, it ends at the first step t such that for each v ∈ V one of the following condition holds i) lv(t) = lv(t − 1) ii) lv(t) 6= lv(t − 1) but this change is due to a tie. Then the Algorithm 3 converges, independently of the tie management rule. Proof: Let f (i) = X{u,v}∈E [lu(i) == lv(i)] be a function that computes the number of monochromatic edges of G = (V, E) at step i, for any i ≥ 1. Clearly, the function f (i) is upper bounded by E. We will show that the value of f (i) grows monotonically with the step number i. Consider any step i. If the stop criterion is not met, then at least one vertex v has changed his label without the occurrence of a tie. Since the labeling occurs by colors and neighboring vertices are assigned different colors, we get that when a vertex v is relabeled, v’s neighbors do not change their labels. This immediately implies that the number of monochromatic edges incident in v strictly increases during step i. The same reasoning applies to each other vertex in G that changes is label without the occurrence of a tie. Hence, we have that if after label propagation step i the stop criterion is not met, then f (i) > f (i − 1) and the result follows. VI. EXPERIMENTAL RESULTS We have implemented both the asynchronous and the semi– synchronous LPA with 4 tie resolution strategies: LPA, LPA- Prec, LPA-Max and LPA-Prec-Max. We compared both the quality, the timing and the stability of such algorithms on a set of real networks, for which we know the community structure. A. Analyzed Network and their properties The choice of networks to be used as a benchmark is a crucial problem. Several experiments have been executed on computer generated networks with a community structure known by construction [GN02], [PSSL09]. However generated networks cannot model real networks. Several studies have been developed in order to devise a class of benchmark graphs, having a known community structure and able to resemble real networks [LFR08]. Results are rather preliminary and deserve much attention. Hence our tests have been performed over real networks having a known community structure [New06b]. Table I reports some known properties of such networks: 1) Zachary’s Karate: Social network of friendships between 34 members of a karate club at a US university in the 1970s [Zac77]; 2) Dolphins: Social network of frequent associations be- tween 62 dolphins in a community living off Doubtful Sound, New Zealand [LSB+03]; 3) Football: Network of American football games between Division IA colleges during regular season Fall 2000 [GN02]; 4) NetScience: Coauthorship network of scientists working on network theory and experiment, as compiled by M. Newman in May 2006 [New06a]; 5) Power: A network representing the topology of the Western States Power Grid of the United State [WS98]; 6) Internet: A symmetrized snapshot of the structure of the Internet at the level of autonomous systems, recon- structed from BGP tables posted by the University of Oregon Route Views [New06b]; 7) Cond-Mat: Coauthorships between scientists posting preprints between Jan 1, 1995 and June 30, 2003 on the Condensed Matter E-Print Archive [New01]. B. Test Settings We have performed our tests by considering 56 different test settings. Each test setting is characterized by the choice of the label propagation timing (asynchronous or semi–synchronous), the network and the tie resolution strategy. Since LPAs involve a certain amount of randomization, we execute each test 100 times and use the means and the variances of their “performances” for our comparisons. Each test setting is composed as follows: i) We randomly assign the initial network labeling and ex- ecute a greedy graph coloring algorithm which considers the vertices in increasing order of their labels. Notice that different initial assignment may correspond to different network coloring. # of vertices # of Edges Modularity (Known real partitioning) Maximum modularity Karate Dolphins Football 34 62 115 78 159 613 NetScience 1, 589 2, 742 Power Internet 4, 941 6, 594 22, 963 48, 436 Cond-Mat 31, 163 120, 029 0.383 0.492 − 0.955 − − 0.688 TABLE I NETWORKS’ PROPERTIES 0.431 0.527 0.588 0.917 0.807 0.516 0.657 ii) We execute LPA using the stop criterion (c1) proposed in [RAK07]. During each test we collect the number of phases required to reach the final labeling (timing). iii) Then we use a simple propagation algorithm which identifies the communities as connected group of nodes having the same final label. Notice that disconnected group of nodes having the same label will be considered as different communities. iv) We measure the average modularity of the partitioning obtained (quality). v) We also collect information about the standard deviation of the modularity of the partitioning obtained, the average number of communities and the size of the greatest com- munity. Such data will be used to compare the stability of the algorithms (stability). In the following each test setting is identified by a triple (P, N, T ) where • P ∈ {Asynchronous, Semi–synchronous}, indicates the label propagation timing, • N ∈ {Karate, Dolphins, Football, NetScience, Power, Internet, Cond-Mat}, indicates the network and • T ∈ {LPA, LPA-Prec, LPA-Max, LPA-Prec-Max}, indi- cates the tie resolution strategy. C. Results Quality: Figure 2 depicts the average modularity observed for each test setting. The results are quite similar, and reproduce the known values presented in Table I. In particular the semi– synchronous approach slightly outperforms the asynchronous one on Karate, Dolphins and Football networks while it is slightly worse on NetScience, Power and Cond-Mat networks. On the Internet network performances are almost identical. We stress that our goal here is not to improve the quality of the results obtained by the asynchronous approach, rather we want to show that the semi-synchronous approach does not degrade any quality. The accuracy of the results seems to be independent of the strategy used for the management of ties: except for the Power network, where strategies LPA and LPA-Max provide a much better modularity (≈ 0.8) compared with LPA-Prec and LPA- Prec-Max (≈ 0.6), the other results are comparable. Timing: The efficiency of our proposal is shown in Figure 3. The improvement grows with the network size, in our tests it ranges from ≈ 5.7: (Asynchronous, Karate, LPA-Max) required 80 stages, on average, while (Semi–synchronous, Karate, LPA-Max) only 14 stages, to ≈ 1802: (Asynchronous, Cond-Mat, LPA-Prec) required 528772 stages, on average, while (Semi–synchronous, Cond-Mat, LPA-Prec) required 290 stages. This means that our algorithm would greatly benefit from parallel implementations. Timing results also confirm that the number of steps needed to reach a final labeling grows logarithmically with respect to the size of the network and is independent from the model used (asynchronous or semi– synchronous). A careful comparison of this results also shows that all the variants of the standard LPA (that is, LPA-Max, LPA-Prec and LPA-Prec-Max) converge faster to the final labeling. As an example, the number of propagation steps required by (, Power, LPA) is 33.15, on average, while (, Power, LPA-Max) is only 10.54. Notice that, in Figure 3 results are shown with logarithmic scale in order to represent both the results of the asynchronous and semi-synchronous approaches. Stability: In order to compare the stability of LPAs, we report, in Figures 2 and 4, the standard deviation (σ) of the modularity obtained for each test setting. The results indicate that the semi–synchronous approach is more stable than the asynchronous one. the stability of Furthermore, by an accurate analysis of the results, it is possible to infer that whatever the algorithm used (semi– synchronous or asynchronous) results is strongly influenced by the strategy used for the management of ties. For this reason we show these results in four different graphs, one for each LPA variant (cf. Figure 4). In particular, the LPA-Prec is more stable than the others, while the LPA- Max is very unstable, especially in Karate, Dolphins, Internet and Cond-mat networks. The instability of LPA-Max is mainly synchronous LPA-Max converges to a period 2 [PS83]. What about the asynchronous and semi-synchronous approaches? What about the LPA-Prec and the LPA-Prec-Max variants? REFERENCES [AB02] [BC09] R. Albert and A.-L. Barab´asi. Statistical mechanics of complex networks. Rev. Mod. Phys., 74(1):47–97, Jan 2002. M. J. Barber and J. W. Clark. Detecting network communities by propagating labels under constraints. Phys. Rev. E 80, 2:1– 16, 2009. [BDG+07] U. Brandes, D. Delling, M. Gaertler, R. Gorke, M. Hoefer, Z. Nikoloski, and D. Wagner. On finding graph clusterings with maximum modularity. In Graph-Theoretic Concepts in Computer Science, volume 4769 of Lecture Notes in Computer Science, pages 121–132. Springer Berlin / Heidelberg, 2007. L. Barenboim and M. Elkin. Distributed (δ+1)-coloring in linear (in δ) time. In STOC ’09: Proceedings of the 41st annual ACM symposium on Theory of computing, pages 111–120, New York, NY, USA, 2009. ACM. [BE09] due to the fact that before each test the initial labeling is randomized and consequently the way in which ties are solved changes too. On the other hand, LPA-Prec solves each tie in favor of its own label irrespective of the value of the labels involved. We have also collected information about the size and the number of the communities generated during each test (cf. Figure 5 and 6). The rationale behind this analysis is to understand whether the semi–synchronous algorithm avoids the generation of a “monster” community (cf. Section III-C). The results confirm that the semi–synchronous LPA slightly reduces this phenomenon: the average size of the biggest community is always smaller while the average number of community is bigger. Results show also that several networks (e.g., Internet) suffer this problem much more than others (e.g., NetScience). To conclude we also note that even in this experiment, the stability of the algorithm is influenced by the strategy used to manage ties. All tests have shown that LPA-Prec is more stable than LPA and LPA-Prec-Max is more stable than LPA-Max. D. Discussion We have showed that our strategy provides good perfor- mances, with respect to several quality metrics. In particular we showed that our approach is easily parallelizable and would greatly benefit, in terms of convergence timing, of a parallel implementation. At the same time, neither the accuracy of the partitioning nor the stability of results are harmed by our strategy. The set of developed tests also gives us the chance to perform a detailed comparison of 4 different LPA tie reso- lution strategies. Results show that each strategy has some peculiarity: LPA-Max is faster than the other approaches, LPA-Prec is more stable and LPA-Prec-Max seems to be a good tradeoff. Results shows also that the peculiarity of LPA variants are not influenced by the approach (asynchronous or semi–synchronous) used during the label propagation step. VII. CONCLUSION We have presented a novel semi–synchronized LPA for detecting network communities. Our proposal has shown to be • effective: the accuracy of our approach is comparable with standard LPAs; • efficient: it is easy to parallelize and consequently much faster than asynchronous approaches; • stable: our proposal has the advantage of “limiting” randomization to such an extent that results are quite uniform. We showed that using a quite simple stop criterion, our algorithm is able to defeat the oscillation phenomena that preclude the use of synchronous strategies. Several problems still remain open. In particular, charac- terizing the convergence of LPAs and determining its speed is an open question. For instance, it has been shown that [LM09] [New01] [GN02] [LFR08] [New06a] [LSB+03] [LHLC09] [DDGDA05] L. Danon, A. Duaz-Guilera, J. Duch, and A. Arenas. Compar- ing community structure identification. Journal of Statistical Mechanics: Theory and Experiment, P09008:1–10, 2005. M. Girvan and M. E. J. Newman. Community structure in social and biological networks. Proc. Natl. Acad. Sci. USA, 99(12):7821–7826, 2002. A. Lancichinetti, S. Fortunato, and F. Radicchi. Benchmark graphs for testing community detection algorithms. Physical Review E78, 046110:1–5, 2008. I. X. Y. Leung, P. Hui, P. Li`o, and J. Crowcroft. Towards real- time community detection in large networks. Phys. Rev. E, 79(6):1–10, Jun 2009. X. Liu and T. Murata. How does label propagation algorithm work in bipartite networks? In WI-IAT ’09: Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, pages 5–8, Washington, DC, USA, 2009. IEEE Computer Society. D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and S. M. Dawson. The bottlenose dolphin community of doubtful sound features a large proportion of long-lasting associations. can geographic isolation explain this unique trait? Behavioral Ecology and Sociobiology, 54:396–405, 2003. M. E. J. Newman. The structure of scientific collaboration networks. Proc. Natl. Acad. Sci. USA, 98:404–409, 2001. M. E. J. Newman. Finding community structure in networks us- ing the eigenvectors of matrices. Phys. Rev. E 74, 036104(3):1– 22, Sep 2006. M. Network http://www-personal.umich.edu/∼mejn/netdata/, 2006. M. E. J. Newman and M. Girvan. Finding and evaluating com- munity structure in networks. Phys. Rev. E 69, 026113(2):1–16, Feb 2004. G. Keziban Orman and V. Labatut. A comparison of community detection algorithms on artificial networks. In Discovery Science, volume 5808 of Springer Berlin / Heidelberg, pages 242–256. Springer Berlin / Heidelberg, 2009. S. Poljak and M. Sura. On periodical behaviour in societies with symmetric influences. Combinatorica, Volume 3, Number 1:119–121, 1983. C. Pang, F. Shao, R. Sun, and S. Li. Detecting community structure in networks by propagating labels of nodes. In Advances in Neural Networks - ISNN 2009, volume 5553 of Lecture Notes in Computer Science, pages 839–846. Springer Berlin – Heidelberg, 2009. U. N. Raghavan, R. Albert, and S. Kumara. Near linear time algorithm to detect community structures in large-scale networks. Phys. Rev. E 76, 036106(3):1–12, Sep 2007. D. J. Watts and S. H. Strogatz. Collective dynamics of ’small- world’ networks. Nature, 393:440–442, 1998. W. W. Zachary. An information flow model for conflict and fission in small groups. Journal of Anthropological Research, 33:452–473, 1977. [New06b] E. J. Newman. [WS98] [Zac77] [NG04] [OL09] [PS83] [PSSL09] [RAK07] data. Fig. 2. Average modularity and standard deviation measured on each test setting. Each test has been executed 100 times. Fig. 3. Average number of stages required by each test setting. The y-axis, which denotes the number of stages, is on a logarithmic scale. Fig. 4. Standard deviation of the modularity measured on each test setting. Each test has been executed 100 times. Fig. 5. Average size of the biggest community. The y-axis, which denotes the size of the biggest community, is on a logarithmic scale. Fig. 6. Average number of communities. The y-axis, which denotes the number of communities, is on a logarithmic scale.
1108.4675	1	1108	2011-08-23T19:22:25	Category-Based Routing in Social Networks: Membership Dimension and the Small-World Phenomenon (Short)	[ "cs.SI", "cs.DS", "physics.soc-ph" ]	A classic experiment by Milgram shows that individuals can route messages along short paths in social networks, given only simple categorical information about recipients (such as "he is a prominent lawyer in Boston" or "she is a Freshman sociology major at Harvard"). That is, these networks have very short paths between pairs of nodes (the so-called small-world phenomenon); moreover, participants are able to route messages along these paths even though each person is only aware of a small part of the network topology. Some sociologists conjecture that participants in such scenarios use a greedy routing strategy in which they forward messages to acquaintances that have more categories in common with the recipient than they do, and similar strategies have recently been proposed for routing messages in dynamic ad-hoc networks of mobile devices. In this paper, we introduce a network property called membership dimension, which characterizes the cognitive load required to maintain relationships between participants and categories in a social network. We show that any connected network has a system of categories that will support greedy routing, but that these categories can be made to have small membership dimension if and only if the underlying network exhibits the small-world phenomenon.	cs.SI	cs	Category-Based Routing in Social Networks: Membership Dimension and the Small-World Phenomenon David Eppstein∗ [email protected] Maarten Loffler∗ [email protected] ∗Deptartment of Computer Science, University of California, Irvine, USA Michael T. Goodrich∗ [email protected] Darren Strash∗ [email protected] Lowell Trott∗ [email protected] 1 1 0 2 g u A 3 2 ] I S . s c [ 1 v 5 7 6 4 . 8 0 1 1 : v i X r a Abstract -- A classic experiment by Milgram shows that individuals can route messages along short paths in social networks, given only simple categorical information about recipients (such as "he is a prominent lawyer in Boston" or "she is a Freshman sociology major at Harvard"). That is, these networks have very short paths between pairs of nodes (the so-called small-world phenomenon); moreover, participants are able to route messages along these paths even though each person is only aware of a small part of the network topology. Some sociologists conjecture that participants in such scenarios use a greedy routing strategy in which they forward messages to acquaintances that have more categories in common with the recipient than they do, and similar strategies have recently been proposed for routing messages in dynamic ad-hoc networks of mobile devices. In this paper, we introduce a network property called membership dimension, which characterizes the cognitive load required to maintain relationships between participants and categories in a social network. We show that any connected network has a system of categories that will support greedy routing, but that these categories can be made to have small membership dimension if and only if the underlying network exhibits the small-world phenomenon. I. INTRODUCTION In a pioneering experiment in the 1960's, Stanley Milgram and colleagues [14], [20], [24] studied message routing in real-world social networks. 296 randomly chosen people in Nebraska and Kansas were asked to route a letter to a lawyer in Boston by forwarding it to an acquaintance, who would receive the same instructions. Messages that reached their destinations typically passed between at most six acquaintances;1 The observation that acquaintance graphs have such short paths has come to be called the small-world phenomenon [11], [25]. Even more surprising than the existence of these short paths is that participants are able to efficiently route mes- sages using only local information and simple facts about targets, such as ethnicity, occupation, name, and location. As a way to model the methods used by humans to route such messages, sociologists have studied the importance of categories, that is, various groups to which people belong, 1This observation has also led to the concept of "six degrees of separation" between all people on earth and the trivia game, "Six Degrees of Kevin Bacon," where players take turns trying to link performers to the actor Kevin Bacon via at most six movie collaborations. in the small-world phenomenon. In the early 1970's, Hunter and Shotland [8] found that messages routed between people in the same university category (such as student, faculty, etc.) had shorter paths than messages routed across categories. Killworth and Bernard [10] performed experiments in the late 1970's that they called reverse small-world experiments in which each participant was presented with a list of mes- sages for hundreds of targets, identified by the categories of town, occupation, ethnic background, and gender, and asked to whom they would send each of these messages. The study concluded that the choices people make in selecting routes are overwhelmingly categorical in nature. In the late 1980's, Bernard et al. [3] extended this work to identify which of twenty categories are most important for message routing to people from various cultures. More recently, Watts et al. [26] present a hierarchical model for categorical organization in social networks for the sake of message routing. They propose groups as the leaves of rooted trees, with internal nodes defining groups-of-groups, and so on. They define an ultrametric on sets of such overlapping hierarchies and conjecture that people use the minimum distance in one of their trees to make message routing decisions. That is, they argue that individuals can understand their "social distance" to a target as the minimum distance between them and the target in one of their categories. Such a determination requires some global knowledge about the structures of the various group hierarchies. Although this previous work shows the importance of categories and of hierarchies of categories in explaining the small world phenomenon, it does not explain where the categories come from or what properties they need to have in order to allow greedy routing to work. Hence, this prior work leaves open the following questions: • Which social networks support systems of categories that allow participants to route messages using the sim- ple greedy rule of sending a message to an acquaintance who has more categories in common with the target? • How complicated a system of categories is needed for this purpose, and what properties of the underlying network can be used to characterize the complexity of the category system? of a social network, its path length and its membership dimension. In particular: • We show that the membership dimension of (G,S) must be at least the diameter of G, diam(G), for a local, greedy, category-based routing strategy to work. • We show that every connected graph G = (U, E), has a collection S of categories such that local, greedy, category-based routing always works, with membership dimension O((diam(G) + log U)2). Since Milgram's work [14], [20], [24], social scientists have believed that real-world social networks have diameters bounded by constants or slowly growing functions of the network size. Under a weak form of this assumption, that the diameter is O(log U), our results provide a natural model for how participants in a social network could efficiently route messages using a local, greedy, category-based routing strategy while remembering an amount of information that is only polylogarithmic in the size of the network. B. Previous Related Work Geometric greedy routing [6], [15] uses geographic lo- cation rather than categorical data to route messages. In this method, vertices have coordinates in a geometric metric space and messages are routed to any neighbor that is closer to the target's coordinates. Greedy routing may not succeed in certain geometric networks, so a number of techniques have been developed to assist such greedy routing schemes when they fail [4], [9], [16]. Introduced by Rao et al. [23], virtual coordinates can overcome the shortcomings of real- world coordinates and allow simple greedy forwarding to function without the assistance of fallback algorithms. This approach has been explored by other researchers [1], [12], [17], [22], who study various network properties that allow for greedy routing to succeed. Several researchers also study the existence of succinct greedy-routing strategies [5], [7], [18], [21], where the number of bits needed to represent the coordinates of each vertex is polylogarithmic in the size of the network; this notion of succinctness for geometric greedy routing is closely analogous to our definition of the membership dimension for categorical greedy routing. Recent work by Mei et al. [19], studies category-based greedy routing as a heuristic for performing routing in dynamic delay-tolerant networks. Mei et al. assume that the network nodes have been organized into pre-defined categories based on the users' interests. Experiments suggest that using these categories for greedy routing is superior to routing heuristics based on location or simple random choices. One can interpret the categorical greedy routing techniques of Mei et al. and of this paper as being geometric routing schemes using virtual coordinates, where the coordi- nates represent category memberships. In this interpretation, the membership dimension of an embedding corresponds to the number of nonzero coordinates of each node, and our (a) (b) Figure 1. A set of elements U (drawn arbitrarily as points in the plane). (a) The graph G on U. (b) The categories S on U. In this example, the membership dimension is 4, because no element is contained in more than 4 groups. Our goal therefore, in this paper, is to address these questions by studying the existence of mathematical and algorithmic frameworks that demonstrate the feasibility of local, greedy, category-based routing in social networks. A. Our Results Inspired by the work of Watts et al. [26], we view a social network as an undirected graph G = (U, E), whose vertices represent people and whose edges represent relationships, taken together with a collection, S ⊂ 2U , of categories defined on the vertices in G. Figure 1 shows an example. In addition, given a network G = (U, E) and category system S, we define the membership dimension of S to be max u∈U {C ∈ S : u ∈ C}, that is, the maximum number of groups to which any one person in the network belongs. The membership dimension characterizes the cognitive load of performing routing tasks in the given system of categories -- if the membership di- mension is small, each actor in the network only needs to know a proportionately small amount of information about his or her own categories, his or her neighbors' categories, and the categories of each message's eventual destination. Thus, we would expect real-world social networks to have small membership dimension. In this paper, we provide a constructive proof that a category system can support greedy routing. Our results are not intended to model the actual formation of social categories, and we take no position on whether categories are formed from the network, the network is formed from categories, or both form together. Rather, our intention is to show the close relation between two natural parameters 2 GSFigure 2. Illustration of the routing rule. results show that such greedy routing schemes can be done succinctly in graphs with small diameter. Similarly to the work of this paper, Kleinberg [11] studies the small-world phenomenon from an algorithmic perspec- tive. However, his approach is orthogonal to ours: He focuses on location rather than categorical information as the critical factor for the ability to find short routes efficiently, and constructs a random network based on that information, whereas our approach takes the network as a given and studies the kinds of categorical structures needed to support category-based greedy routing. In addition, it is worth noting that small world networks exhibit scale-free properties. II. ROUTING BASED ON CATEGORICAL INFORMATION In this section, we introduce a mathematical model of categorical greedy routing, and provide basic definitions and properties that guarantee the success of this strategy. A. Basic definitions Abstracting away the social context, let U be the universe of n people defining the potential sources, targets, and intermediates for message routes, and let G = (U, E) be an undirected graph whose m edges represent pairs of people who can communicate. For any two elements s, t ∈ U, let sp(s, t) be the length of the shortest path in G from s to t. The diameter diam(G) = maxs,t∈U sp(s, t) is the maximum length of any shortest path. For s ∈ U, define the neighborhood of s to be the set of neighbors N (s) = {u ∈ U {s, u} ∈ E} of s in G. Now let S ⊂ 2U be a set of subsets of U, which represent the abstract categories that elements of U belong to. For a given u ∈ U, we define cat(u) ⊂ S to be the set of groups to which u belongs: cat(u) = {C ∈ S u ∈ C}. Definition 1 (membership dimension): The membership dimension of S is the maximum number of elements of S that any element of U is contained in, that is, memdim(S) = max u∈U cat(u). As discussed, there is evidence that in real world social networks and group structures (G,S), both diam(G) and memdim(S) are significantly smaller than U. B. The routing strategy We now describe a simple greedy category-based strategy to route a message from one node to another. We clarify the distance function immediately following the rule definition. (a) (b) Figure 3. Two networks with the same elements and categories. (a) An example that is internally connected, but not shattered: no category contains y and a neighbor of v but not v itself. (b) An example that is shattered, but not internally connected: the induced graph of category {u, w, x, z} is not connected. Definition 2 (greedy routing rule): If a node u receives a message M intended for a destination w (cid:54)= u, then u should forward M to a neighbor v ∈ N (u) that is closer to w than u is, that is, for which d(v, w) < d(u, w). The category-based distance function used by this rule is d(s, t) = cat(t) \ cat(s), which measures the number of categories of the target that the current node does not share.2 This number decreases as the number of shared groups of S between the current node and the target increases. We refer to the greedy routing strategy that uses this distance function as ROUTING (see Figure 2). For category systems with low membership dimension, this strategy is easy to evaluate using only local knowledge about the categories of each neighbor of the current node and the categories of the target node. C. Successful routing We now investigate conditions under which ROUTING can successfully route messages between all pairs of nodes in a network. We identify several properties of a graph G and associated group structure S that directly influence the feasibility of routing. For routing to succeed, G must be connected. It seems natural to consider a stronger property: Definition 3 (internally connected): (G,S) is internally connected if for each C ∈ S, G restricted to C is connected. Figure 3(a) shows an example of an internally connected pair (G,S). This is a very natural property for sociological groups to exhibit. People belonging to the same group will have greater cohesiveness, and if a group is not internally connected then it may be redefined to be the set of groups defined by its connected components. Definition 4 (shattered): A pair (G,S) is shattered if, for all s, t ∈ U, s (cid:54)= t, there is a neighbor u ∈ N (s) and a set C ∈ S such that C contains u and t, but not s. Figure 3(b) shows an example of a shattered pair. Note that in this definition, u and t could be the same node. This property falls out naturally from the instructions given in the 2Note that d is not a metric, since it is not necessarily symmetric. 3 wuvyvzxuwyvzxuwFigure 5. The sets Bv for each vertex v in the path. The sets Av are constructed symmetrically. A. Constructing S given G Given a connected graph G = (U, E) as input, we would like to construct a family S ⊂ 2U so that ROUTING works, and the membership dimension of S is small. We concentrate foremost on constructions of category collections that are internally connected and shattered, because of the social significance of these properties. Nevertheless, even without these properties, we have the following lower bound. Lemma 3: Let G and S be a graph and a category system, respectively, such that ROUTING works for G and S. Then memdim(S) ≥ diam(G). Proof: By definition of the diameter, there are two vertices s, t ∈ U such that sp(s, t) = diam(G). Let P be the path that ROUTING follows from s to t, and note that the length of P must be at least diam(G). An edge (u, v) can only be on P if d(v, t) < d(u, t). Since d(·,·) can only take integer values, d(u, t) ≥ d(v, t) + 1. Therefore, d(s, t) ≥ P. By definition, d(s, t) = cat(t) \ cat(s), and memdim(S) is the maximum of cat(·) over all elements; hence memdim(S) ≥ cat(t) ≥ cat(t) \ cat(s) = d(s, t) ≥ P ≥ diam(G), as claimed. 1}, and let S = (cid:83) For paths, this bound is tight: Lemma 4: If G is a path, then there exists an S s.t. (G,S) is shattered and internally connected with memdim(S) = diam(G). Proof: Arbitrarily pick one of the two end vertices of G and let us refer to the vertices in G by their distance, 0 to n− 1, from this vertex. For each vertex i, form two sets Ai and Bi, where Ai = {0, . . . , i−1} and Bi = {i+1, . . . , n− v∈U{Av, Bv}. Figure 5 illustrates this construction. Each set in S consists of a path of vertices and therefore S is internally connected. S is also shattered, since for all s and t, s has a neighbor that shares either As or Bs with t, but s is not in these sets. Considering memdim(S), note that each vertex i is contained in sets Aj for 0 ≤ j < i and Bk for k < i ≤ n − 1. Therefore, each vertex is in exactly n − 1 sets, which is diam(G). It follows from Lemmas 2 and 4 that, if G is a path, one can construct S with memdim(S) = diam(G), so that ROUTING works in G. There are other graphs for which it is relatively easy to set up a category set that is shattered and internally connected in a way that supports the ROUTING algorithm. For example, in a tree of height 1 (i.e., a star graph), with root r, we could create for each leaf of the tree two categories, one containing the leaf itself and one containing both the leaf and the root. Every path in this tree supports ROUTING. However, the Figure 4. ROUTING does not work in this graph, even though it is internally connected and shattered. Routing from v to x fails: v, u, and w are all at distance 2 from x, so v has no neighbor that is closer than it to x. real-world routing experiments of Milgram and others. In order for someone to advance a letter toward a target, there must be an acquaintance that shares additional interests with the target. Indeed, we now show that the shattered property is necessary for ROUTING to work. Lemma 1: If (G,S) is not shattered, ROUTING fails. Proof: Since (G,S) is not shattered, there exists a pair of vertices s and t, where s's neighbors are not in sets with t that do not contain s. Therefore, s's neighbors cannot share strictly more sets with t as s does, and ROUTING will fail to route from s to t. If G is a tree, then these two properties together are sufficient for the routing strategy to always work: Lemma 2: If G is a tree, and (G,S) is internally con- nected and shattered, then ROUTING is guaranteed to work. Proof: Let s and t be vertices in G. Since G is a tree, there is one simple path from s to t. Let (u, v) be an edge on the path from s to t. First, we claim that every set in S that contains both u and t also contains v. This follows from (G,S) being internally connected: any set C ∈ S with u, t ∈ C must also contain v, since v is on the only path between u and t. Therefore, v is contained in at least as many sets in S with t as u is. However, by the shattered property, v is in a set in S with t that does not contain u. Therefore v is in strictly more sets with t than u is. This property holds for every simple path; hence, ROUTING always works. Although sufficient for routing in trees, the internally connected and shattered properties are not sufficient for ROUTING to work on arbitrary connected graphs. Figure 4 shows a counter-example -- ROUTING is unable to route a message from the leftmost to the rightmost node, since there is no neighbor whose distance to the target is smaller. III. EXISTENCE OF CATEGORIES In this section, we consider the following question: Is it possible to construct the family S so that ROUTING always works and S has low membership dimension? We show that such a construction is always possible if we are given a connected graph as input. We also show that it is impossible to construct an S such that ROUTING will work if the graph is not known in advance. 4 vuwx(cid:16)(cid:80) O for any v. (cid:17) u∈ancestors(v) height(u) sets, which is O(diam2(G)) We now extend this result to arbitrary trees by applying weight-balanced binary trees [2], [13]. Definition 5 (weight balanced binary tree): A weight balanced binary tree is a binary tree that stores weighted items in its leaves. If item i has weight wi, and all items have a combined weight of W then item i is stored at depth O(log (W/wi)). Lemma 6: Let T be an n-node rooted tree with height h. We can embed T into a binary tree such that the ancestor -- descendant relationship is preserved, and the resulting tree has height O(h + log n). Proof: Let nu be the number of descendants of vertex u in T . For each vertex u in T that has more than two children, we expand the subtree consisting of u and u's children into a binary tree as follows. Construct a weight balanced binary tree B on the children of u, where the weight of a child v is nv. We let u be the root of B. Each child v of u in the original tree is then a leaf at depth log(nu/nv) in B. Performing this construction for each vertex u in the tree expands T into a binary tree with the ancestor -- descendant relationship preserved from T . Furthermore, each path from root to leaf in T is only expanded by log(n) nodes, which we can see as follows. Each parent-to-child edge (u, v) in T is replaced by a path of length O(log(nu/nv)). Therefore for each path P from root r to leaf l in T , our construction expands P by length (u,v)∈P log(nu/nv)), which is a sum telescoping to O(log(nr/nl)) = O(log n). Therefore, the height of the new binary tree is O(h + log n). O((cid:80) Combining this lemma with Lemma 2, we get the follow- ing theorem. Theorem 1: Given a tree T , it is possible to construct a family S of subsets such that ROUTING works for T and memdim(S) = O((diam(T ) + log n)2). Proof: Arbitrarily root T and embed T in a binary tree B using the method in Lemma 6. Then B has height O(diam(T ) + log n), and diameter diam(B) = O(diam(T ) + log n). Applying the construction from Lemma 5 to B gives us a family SB with memdim(SB) = O((diam(T ) + log n)2). We then construct a family ST , by removing vertices that are in B but not T from the sets in SB. By construction, (T,ST ) is shattered and in- ternally connected, and memdim(ST ) ≤ memdim(SB) = O((diam(T )+log n)2). By Lemma 2, ROUTING works on T with category sets from ST . We can further extend this theorem to arbitrary connected graphs, which is the main upper bound result of this paper. Theorem 2: If G is connected, there exists S s.t. ROUT- ING works and memdim(S) = O((diam(G) + log(n))2). Proof: Compute a low-diameter spanning tree T of G. This step can easily be done using breadth-first search, producing a tree with diameter at most 2 diam(G). We then Figure 6. The collection of sets Lv for an example subtree at v. membership dimension of this category system is high, since the root belongs to a linear number of categories. So even in this simple example, supporting ROUTING and achieving low membership dimension is a challenge. Moreover, this challenge becomes even more difficult already for a tree of height 2, since navigating from any leaf, x, to another leaf, y, requires that the parent of x belong to more categories with y than x -- and this must be true for every other leaf, y. Thus, it is perhaps somewhat surprising that we can construct a set of categories, S, for an arbitrary binary tree that causes this network to be shattered and internally connected (so the ROUTING strategy works, by Lemma 2) and such that S has small membership dimension. Lemma 5: If G is a binary tree, then there exists an S s.t. (G,S) is shattered and internally connected with memdim(S) = O(diam2(G)). Proof: We show how to construct S from G. Arbitrarily pick a vertex r ∈ U of degree at most 2 and root the binary tree at r, so each vertex v has left and right children, left(v) and right(v), and let height(v) be the length of the longest simple path from v to any descendant of v. For each vertex v, we create a set Sv, containing v's descendants (which includes v). We further construct two families, Lv and Rv, using helper sets Lv,i and Rv,i. Let Lv,i (resp., Rvi) consist of v, the vertices in v's left (right) subtree down to depth i, and all vertices in v's right (left) subtree. Then define Lv = {Lv,i depth(v) ≤ i ≤ depth(v) + height(left(v))}. Figure 6 illustrates this. The family Rv is defined symmet- rically. Our S is then defined as S = {Sv} ∪ Lv ∪ Rv. (cid:91) v∈U Each set in S is a connected subgraph of G, so S is internally connected. As the following argument shows, S is shattered: If s is an ancestor of t, then s's child u on the path to t is in set Su which contains u and t but not s. Otherwise, let v be the lowest common ancestor of s and t, and assume without loss of generality that s in v's left subtree; then Lv,depth(s)−1 contains s's parent and t but not s. We now analyze the membership dimension of this con- struction. Let v be a vertex, and let ancestors(v) be the set of v's ancestors. For u ∈ ancestors(v), v ∈ Su, and v belongs to O(height(u)) sets of Lu and Ru. Then v belongs to 5 vLvuse the construction from Theorem 1 on T . For greedy routing to work in a graph G, note that it is sufficient to show that it works in a spanning tree of G. Therefore, since ROUTING works in T , ROUTING also works in G. IV. CONCLUSION AND OPEN PROBLEMS We have presented a construction of groups S on a connected graph G that allows a simple greedy routing algo- rithm, utilizing a notion of distance on group membership, to guarantee delivery between nodes in G. Such a construction will have membership dimension O((diam(G) + log n)2), demonstrating a small cognitive load for the members of G. There are several directions for future work. For example, while we have shown that the membership dimension must be minimally the diameter of G, it remains to be shown if the membership dimension must be the square of the diameter plus a logarithmic factor for arbitrary graphs. We conjecture that the square term is not strictly needed in the membership dimension in order for ROUTING to work. Our group construction is performed for a general graph by selecting a low diameter spanning tree and using the presented tree construction, so it may be possible that there is a group construction that has lower membership dimension and more efficient routing if it is constructed directly in G. In this paper all categories are given equal weight with respect to routing tasks and that participants use a simple greedy routing algorithm based solely on increasing the number of categories in common with the target. Future work could include study of a category-based routing strat- egy that allows participants to weight various categories higher than others, as in the work of Bernard et al. [3]. REFERENCES [1] P. Angelini, F. Frati, and L. Grilli. An algorithm to construct greedy drawings of triangulations. In Graph Drawing, LNCS 5417, pages 26 -- 37. Springer, 2009. [2] S. W. Bent, D. D. Sleator, and R. E. Tarjan. Biased search trees. SIAM J. Computing, 14(3):545 -- 568, 1985. [3] H. R. Bernard, P. D. Killworth, M. J. Evans, C. McCarty, and G. A. Shelley. Studying social relations cross-culturally. Ethnology, 27(2):155 -- 179, 1988. [4] P. Bose, P. Morin, I. Stojmenovic, and J. Urrutia. Routing with guaranteed delivery in ad hoc wireless networks. Wireless Networks, 7:609 -- 616, 2001. [5] D. Eppstein and M. T. Goodrich. Succinct greedy graph drawing in the hyperbolic plane. In Graph Drawing, LNCS 5417, pages 14 -- 25. Springer, 2009. [6] G. G. Finn. Routing and addressing problems in large metropolitan-scale internetworks. Research report, ISI, 1987. [7] M. Goodrich and D. Strash. Succinct greedy geometric rout- ing in the Euclidean plane. In Algorithms and Computation, LNCS 5878, pages 781 -- 791. Springer, 2009. [8] J. E. Hunter and R. L. Shotland. Treating data collected by the "small world" method as a Markov process. Social Forces, 52(3):321 -- 332, 1974. [9] B. Karp and H. T. Kung. GPSR: greedy perimeter stateless routing for wireless networks. In 6th ACM MobiCom, pages 243 -- 254, 2000. [10] P. Killworth and H. Bernard. Reverse small world experiment. Social Networks, 159(1), 1978. [11] J. Kleinberg. The small-world phenomenon: an algorithm perspective. In 32nd ACM STOC, pages 163 -- 170, 2000. [12] R. Kleinberg. Geographic routing using hyperbolic space. In 26th IEEE INFOCOM, pages 1902 -- 1909, 2007. [13] D. E. Knuth. Optimum binary search trees. Acta Informatica, 1:14 -- 25, 1971. [14] C. Korte and S. Milgram. Acquaintance networks between J. racial groups: Application of the small world method. Personality and Social Psychology, 15(2):101 -- 108, 1970. [15] E. Kranakis, H. Singh, and J. Urrutia. Compass routing on geometric networks. In 11th CCCG, pages 51 -- 54, 1999. [16] F. Kuhn, R. Wattenhofer, Y. Zhang, and A. Zollinger. Geo- metric ad-hoc routing: of theory and practice. In 22nd ACM PODC, pages 63 -- 72, 2003. [17] T. Leighton and A. Moitra. Some results on greedy embed- dings in metric spaces. Discrete Comput. Geom., 44:686 -- 705, 2010. [18] P. Maymounkov. Greedy embeddings, trees, and Euclidean vs. Lobachevsky geometry, 2006. http://pdos.csail.mit.edu/ ∼petar/papers/maymounkov-greedy-prelim.pdf. [19] A. Mei, G. Morabito, P. Santi, and J. Stefa. Social-aware In 30th stateless forwarding in pocket switched networks. IEEE INFOCOM, 2011. [20] S. Milgram. The small world problem. Psychology Today, 1(May):61 -- 67, 1967. [21] R. B. Muhammad. A distributed geometric routing algorithm for ad hoc wireless networks. In 3rd ITNG, pages 961 -- 963, 2007. [22] C. H. Papadimitriou and D. Ratajczak. On a conjecture related to geometric routing. Theor. Comput. Sci., 344:3 -- 14, November 2005. [23] A. Rao, S. Ratnasamy, C. Papadimitriou, S. Shenker, and I. Stoica. Geographic routing without location information. In 9th ACM MobiCom, pages 96 -- 108, 2003. [24] J. Travers and S. Milgram. An experimental study of the small world problem. Sociometry, 32(4):425 -- 443, 1969. [25] D. J. Watts. Networks, dynamics, and the small-world phe- nomenon. The American Journal of Sociology, 105(2):493 -- 527, 1999. [26] D. J. Watts, P. S. Dodds, and M. E. J. Newman. Identity and search in social networks. Science, 296:1302 -- 1305, 2002. 6
1809.02711	1	1809	2018-09-07T23:34:11	BLAG: Bandit On Large Action Set Graph	[ "cs.SI" ]	Information diffusion in social networks facilitates rapid and large-scale propagation of content. However, spontaneous diffusion behavior could also lead to the cascading of sensitive information, which is neglected in prior arts. In this paper, we present the first look into adaptive diffusion of sensitive information, which we aim to prevent from widely spreading without incurring much information loss. We undertake the investigation in networks with partially known topology, meaning that some users' ability of forwarding information is unknown. Formulating the problem into a bandit model, we propose BLAG (Bandit on Large Action set Graph), which adaptively diffuses sensitive information towards users with weak forwarding ability that is learnt from tentative transmissions and corresponding feedbacks. BLAG enjoys a low complexity of O(n), and is provably more efficient in the sense of half regret bound compared with prior learning method. Experiments on synthetic and three real datasets further demonstrate the superiority of BLAG in terms of adaptive diffusion of sensitive information over several baselines, with at least 40 percent less information loss, at least 10 times of learning efficiency given limited learning rounds and significantly postponed cascading of sensitive information.	cs.SI	cs	BLAG: Bandit On Large Action Set Graph Yucheng Lu, Xudong Wu, Jingfan Meng, Luoyi Fu, Xinbing Wang 8 1 0 2 p e S 7 ] I S . s c [ 1 v 1 1 7 2 0 . 9 0 8 1 : v i X r a Abstract Information diffusion in social networks facilitates rapid and large-scale propagation of content. However, spontaneous diffusion behavior could also lead to the cascading of sen- sitive information, which is neglected in prior arts. In this paper, we present the first look into adaptive diffusion of sensitive information, which we aim to prevent from widely spreading without incurring much information loss. We un- dertake the investigation in networks with partially known topology, meaning that some users' ability of forwarding in- formation is unknown. Formulating the problem into a ban- dit model, we propose BLAG (Bandit on Large Action set Graph), which adaptively diffuses sensitive information to- wards users with weak forwarding ability that is learnt from tentative transmissions and corresponding feedbacks. BLAG enjoys a low complexity of O(n), and is provably more efficient in the sense of half regret bound compared with prior learning method. Experiments on synthetic and three real datasets further demonstrate the superiority of BLAG in terms of adaptive diffusion of sensitive information over sev- eral baselines, with at least 40% less information loss, at least 10 times of learning efficiency given limited learning rounds and significantly postponed cascading of sensitive informa- tion. Introduction The prevalence of massive social medias facilitates informa- tion diffusion, which plays an important role in content shar- ing and spreading (Bakshy et al. 2012). While the low-cost diffusion can easily lead to large-scale propagation called in- formation cascading, the unconstrained cascading behavior could meanwhile cause the sensitive information to be in- cautiously propagated over the network. Here the sensitive information can refer to any kind of information that needs to be prohibited from cascading such as rumors, personal in- formation, trade secrets, and etc. Generally, sensitive infor- mation has two features: a) Hard to capture, meaning that it is hard for classification because of being user-specific and hard for detection due to huge volume of daily information consumption. b) Time-sensitive, meaning that it can quickly fade out with time elapse (e.g., rumors about US presidential election in 2016 is no longer sensitive in 2017). The two above features, consequently, render it rather difficult to prevent the sensitive information from wide Copyright c(cid:13) 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. spread by targeting the sensitive information itself. How- ever, it will be a lot easier if we instead target those indi- viduals who may carry the sensitive information. In other words, we can adaptively adjust the diffusion of information based on users' attributes in hope of subsiding the cascad- ing of sensitive information while preserving the propaga- tion of non-sensitive one. Such adaptive transmission in the sense of sensitive information has unfortunately received no prior attention amongst the intensively studied topic of in- formation diffusion (Guille and Hacid 2012)(Wang, Wang, and Xu 2012)(Yang and Leskovec 2010)(Yang et al. 2015). The only two recent attempts that share the closet correla- tion with the issue belong to (Xu, Chen, and others 2015) and (Giakkoupis et al. 2015). Particularly, considering the time sensitiveness of information, (Xu, Chen, and others 2015) points out that decreasing users' information forward- ing probability with time results in limited size of cascad- ing. However, the method may blindly block a lot of non- sensitive information, thus leading to severe information loss. (Giakkoupis et al. 2015) aims not to reveal much about the users' own opinions on items by reposting them based on users' opinions and followers. The idea nevertheless re- quires that users have a full knowledge of network topology, which violates real practices. Regarding both concerns, we are motivated to under- take the first investigation of adaptive diffusion of sensi- tive information with partially known network topology, which, in our context, refers to that some users have no prior knowledge of their neighbors. Assuming that users have dif- ferent levels of forwarding ability that are positively related to their degrees, we intend to adaptively balance their out- ward transmission behaviors by constraining the sensitive information transmission from those with strong forward- ing ability while promoting more transmissions from users with weak forwarding ability. The idea of guiding the infor- mation diffusion towards users with weak forwarding ability can potentially incur less information loss than simply lim- iting the diffusion. Since some users' forwarding ability is unknown due to lack of prior knowledge of their neighbors, adaption should be based on tentative tranmissions and cor- responding feedbacks. Technically, we convert the above adaptive diffusion into a Constrained Combinatorial Multi-Arm Bandit (CCMAB) model, where we propose a learning algorithm BLAG (Bandit on Large Action set Graph). While we de- fer more details of BLAG later, here we briefly delineate its main ingredients of: BLAG makes transmission policy based on current estimation of users' forwarding ability and update the policy based on the new feedback, which is calcu- lated by the expected number of nodes receiving information under adopted policy with observation noise. The update of policy balances the trade-off between trying to get the op- timal policy and minimizing the feedback in the next trial. BLAG turns out to enjoy a low complexity of O(n), and provably returns half of the time-limited regret bound com- pared to previously proposed CUCB(Combinatorial Upper Confidence Bound) strategies (Gai, Krishnamachari, and Jain 2010)(Chen, Wang, and Yuan 2013). Our key contributions can be summarized as follows: • We take the first look into adaptive diffusion of sensitive information in networks with partially known topology, and formulate it into a novel constrained combinatorial bandit problem. • We propose a learning algorithm BLAG that involves bandit on large action set for adaptive diffusion. BLAG theoretically reduces the time-limited regret bound by half compared to classic CUCB bandit algorithms. • We perform extensive experiments on both synthetic and real data sets, which confirm the superiority of BLAG over several baselines in terms of at least 40% less infor- mation loss, at least 10 times higher learning efficiency as well as postponed cascading of sensitive information. Preliminary Network Model We model the social network as an undirected graph G = (V, E) with neither self-loops nor multiple edges between any two nodes. Each node is classified as either a sensitive node or a non-sensitive one. In correspondence to real so- cial networks, sensitive nodes can refer to the individuals who hold plenty of sensitive information. A strict criterion for classifying nodes into sensitive and non-sensitive ones is not our focus and will not affect our later analysis. In each time slot, sensitive information can only be transmitted from a sensitive node to a non-sensitive one, and there can be multiple transmissions within one slot. In addition, the trans- mission of non-sensitive information also occurs between nodes. Here our focus falls primarily on extracting and ana- lyzing the transmission of sensitive information, which, dif- ferent from non-sensitive information, would be limited by the desired adaptive diffusion that we will introduce shortly. Each edge e ∈ E has a weight representing the transmission probability of sensitive information between the two nodes it connects. We assume that the weight follows a uniform distribution U (0, ξ), where ξ is a sufficiently small number, implicitly implying that sensitive information is hard to cap- ture. A non-sensitive node may turn into a sensitive one in a time slot as long as it receives enough amount1 of sensitive 1The exact amount is is beyond the focus of this paper. In the actual social network, only a user receiving tons of sensitive infor- mation should be monitored. information. Let G(cid:48) denote the subgraph induced by sensi- tive nodes and G\G(cid:48) the subgraph induced by non-sensitive nodes. As mentioned earlier, we consider the network with par- tially known topology, which, in our setting, is associated with whether non-sensitive nodes are informed or unin- formed, defined as follows: Definition 1. Informed nodes refer to the non-sensitive nodes with known one-hop neighbors, while uninformed nodes are non-sensitive nodes with unknown one-hop neigh- bors. Based on Definition 1, we further give the definition of partially known topology, which we call semi-informed network throughout the paper: Definition 2. A semi-informed network is a network where both informed and uninformed non-sensitive nodes coexist along with sensitive nodes. For a better understanding, Fig. 1 further illustrates an ex- ample of information diffusion in a semi-informed network. Figure 1: An illustration of a semi-informed network. The red nodes represent sensitive nodes connected by two types of non- sensitive nodes, i.e., informed nodes (white) and uninformed ones (grey). Information transmission occurs between nodes as long as there is an undirected path in between, and their sensitivity states subject to the foregoing conditions. Here node j turns from a non- sensitive node into a sensitive one after receiving sufficient infor- mation from node i. αij is the probability i sending sensitive infor- mation to j. Problem Statement Let E denote the set of edges connecting nodes from G(cid:48) and uninformed nodes from G\G(cid:48). The problem of interest in semi-informed network is to adaptively diffuse information among nodes by adjusting parameter on edges in E based on destination's forwarding ability. In the present work, we simply associate a node's forwarding ability with its degree, with a high degree implying a strong forwarding ability. Our target is to adaptively diffuse sensitive information towards low degree nodes, while avoiding high degree nodes. As the degrees of the uninformed nodes are unknown, our target, equivalently, is to make trials to these uninformed nodes and adaptively diffuse information based on the feed- back. We denote edges in E with uninformed nodes being destination as target edges and these connected uninformed nodes as target nodes. In a time slot, let the m-dimensional −→ vector D(i) representing the degree of target node i. As fomulated ear- −→ D(i) are unknown. The m- lier, exact values of elements in −→ β0 denotes original transmission proba- dimensional vector −→ bility on target edges, with β0(i) representing original trans- −→ D denote the degree set of m target nodes, with ...............Sensitive nodesUninformed nodesInformed nodesG G\G ijαij−→ ∆β denote the variation mission probability on edge i. Let −→ of probability on target edges, with ∆β(i) represents vari- ation of probability on edge i. Thus, our target above is to learn an optimal −→ ∆β∗ within the time slot that minimizes −→ D · ( −→ β0 + −→ ∆β∗) This process, as we will demonstrate in sequel, is similar to a bandit problem, which technically interprets the adap- tive diffusion problem in semi-informed networks. Problem Formulation −→ β0 + −→ ∆βt) where Mapping Diffusion Problem into Bandit Model We assume there are multiple but limited rounds of tri- als in one time slot. During the trials, we try to minimize −→ −→ D · ( ∆βt represents variation vector in round t. We adapt information transmission probability to those uninformed nodes based on trials and feedbacks. To this end, we map the problem into a Constrained Com- binatorial Multi-Arm Bandit (CCMAB) model. Particu- larly, we summarize below the mapping between CCMAB components and the key elements in the formed problem: • Trial: A trial in round t refers to a transmission policy −→ determined by ∆βt. The total number of trials between two time-slots is determined. Besides, we assume that the statuses of both sensitive and non-sensitive nodes remain unchanged in a time slot. • Base-action: In a combinatorial bandit model, a base- action (Chen, Wang, and Yuan 2013) is an atomic action in the action set. In our problem, base-actions are vec- tors with pair-wise non-zero elements. In adaptive diffu- sion, to maintain the overall probability on target edges, for any probability decreasing on one edge, there must be a corresponding increasing probability of same amount on another edge. Based on this fact, there are pair-wise edges containing probability variation with zero summa- tion. We regard these vectors representing pair-wise vari- −→ ation of edges as base-actions βi, each of which has the following characteristics: -- −→ βi only has two non-zero elements. -- (cid:80)m −→ βi (j) = 0. j=1 −→ βi (j) ∈ [−1, 1], 1 ≤ j ≤ m. -- • Super-action: A super-action (Chen, Wang, and Yuan 2013) is combination of several base-actions. However, in our problem, arbitrary combinations may not be valid. To determine whether two arms can combine in valid, we give Definition 3. It not only works for base-actions but for any base-action or super-action. Definition 3. Combination of any two arms valid if and only if ∀i, 1 ≤ i ≤ m, 0 ≤ −→ −→ β2(i) ≤ 1 The restriction holds due to the natural bound of probabil- ity. An illustrative example of valid combination of two base actions is demonstrated in Fig. 2. −→ −→ β2 is β1 and −→ β1(i) + β0(i) + Figure 2: A toy example of combining base-actions a super-action −→ ∆β. −→ β1 and −→ β2 into D ·−→ −→ • Reward: In our problem, reward refers to feedback of an adopted super-action, in the form of ∆β + σ, where σ is the noise in observation subjected to Gaussian distribu- tion. We assume reward of base-actions combined in the observed super-action can be observed individually. Other than that, we assume that each base-action contains an un- reliable estimation before the first trial. (In the real world, it could be context or estimation of a user's forwarding ability) −→ ∆βt) from super-action Learning Process Upon the mapping from the diffusion problem into the ban- dit model, the main focus in learning process is to get the −→ minimum ∆βt selected in round t. This target is equivalent to minimizing the re- gret, which means the variation of the reward between the selected super-action and the optimal super-action. Since a lower regret demonstrates better performance of a bandit policy, the learning process is to obtain as low time-limited regret bound as possible within the total rounds in a slot. −→ D · ( −→ β0 + Algorithm 2 The mapping and learning that we have just mentioned may provoke two major issues. One is the huge size of base- action set. Based on our assumption, the size of the base- action set could reach O(m2) (The number of vectors with (cid:1) and each pair could have multiple pair-wise elements is(cid:0)m variations). Another issue is the limited number of trials in a slot. Here we assume that the size of the base-action set is too large to be trialed all over in one slot. To tackle the large base-action set problem, we map base- actions into a weighted graph and denote the graph as Ac- tion Set Graph (ASG). As for the limited learning rounds problem, we let our algorithm globally follow ε − greedy process. Based on techniques, we propose a learning method called BLAG (Bandit on Large Action set Graph). Action Set Graph We start with the action set graph, which represents whether combination of two base-actions is valid. The topology of ASG follows two rules: (1) Each node in ASG represents −→ a base-action βi, and (2) the weight of each nodes rep- resents the current estimated reward of the corresponding base-action. If the combination of the two base-actions is valid, their corresponding nodes in ASG are connected by an unweighted edge. Obviously, any valid combination is a clique in ASG, and nodes in a clique are pair-wise con- nected. Limited Learning Rounds The key point in the learning process is to balance the exploration-exploitation trade-off while considering limited 213-0.060.08-0.02213-0.060.062130.02-0.02=+learning rounds. Two widely adopted methods of tackling the issue belong to ε − greedy and U CB (Upper Confi- dence Bound)(Auer, Cesa-Bianchi, and Fischer 2002)-based approaches. In our algorithm, we adopt ε − greedy with ε decreasing with time. As we will theoretically demonstrate later, such strategy is provably twice efficient than the al- ternatives in the learning process. Here we summarize the two key procedures, i.e., exploration and exploitation, as fol- lows: • Exploration: The target in this procedure is performing trial to get as much context of the model as possible. Since each base-action can only be observed once in each round, our strategy here is to select as many base-actions as pos- sible. • Exploitation: The target in this procedure is to get mini- mum reward under current context. That is, it selects the combination with the minimum summation of reward. Exact Bandit Algorithm Now we introduce our designed algorithm BLAG, the global process of which follows an ε − greedy process.That is, at the beginning of each round, we decide exploration or ex- ploitation by a given time-decreasing ε. Later we will show that our algorithm performs better than Combinatorial UCB algorithm when the number of learning rounds is limited. Generally, BLAG can be decomposed into three parts: Exploration, Exploitation and Update parameters. As noted earlier, the target in exploration procedure is to get a large size combination. Regarding ASG, this can be treated as a Maximum Clique like problem in the graph the- ory. However, the issue is further complicated in our sce- nario since the algorithms proposed in the Maximum Clique problems cannot guarantee the valid combination expected in our problem. We thus tackle the issue by designing a BF S-like procedure in our algorithm. That is, we randomly choose a node in ASG and iterate its neighbors, if the neigh- bor is not conflicting with the current combination, we add it to the combination. A difference between BF S and our procedure here is we only iterate the one-hop neighbor of the chosen node. Because any multi-hop node is conflicting with the chosen node, thus it must not be chosen. The pseudo code of exploration procedure is given in Algorithm 1. Comparatively, the target in exploitation procedure is to select combination with minimum cumulative reward. Re- garding ASG, this is a Maximum Weighted Clique like problem in the graph theory. Similar to the exploration procedure, algorithms proposed in the Maximum Weighted Clique problems cannot guarantee valid combination here. In our algorithm, we use a greedy strategy, which is to say, we iterate nodes in ASG based on their estimated reward. If a node is not conflicting with current combination and can help us get a smaller reward summation, we add it to the combination. Algorithm 2 provides the pseudo code of the exploitation procedure. Combined with the updating procedure, the whole pseudo code of BLAG is shown in Algorithm 3. After jumping out of procedure of forming combination, BLAG has a final pro- cedure of updating estimation of reward and selected time of each nodes selected. Algorithm 1: Exploration procedure Input: ASG Output: A super-action 1 u ←RANDOM(ASG, 1); /* RANDOM(S, n) returns n random nodes in graph S. /; 2 combination ← {u}; 3 iteration ← 1; 4 for v in Γ(u) do 5 6 7 8 Break; iteration ← iteration + 1; if iteration > m then end / VALID(a1, a2) returns a boolean value of whether the combination of vectors a1 and a2 is valid or not. /; if VALID(combination, v) then combination ← combination ∪ {v}; 9 10 11 12 end 13 return combination; end / MIN(S) returns item with smallest Algorithm 2: Exploitation procedure Input: ASG 1 ActionP ool ←RANDOM(ASG,(cid:98)√ Output: A super-action 2 combination ← ∅; 3 while ActionP ool (cid:54)= ∅ do v =MIN(ActionP ool); 4 m(cid:99)); value in set S. */; if µv,t > 0 then end ActionP ool ← ActionP ool\{v}; if VALID(combination, v) then Continue; 5 6 7 8 9 10 11 12 end 13 return combination; end combination ← combination ∪ {v}; Algorithm 3: BLAG Input: ASG, initialized ε0, learning round T Output: a sequence of super-actions 1 for t = 1 to T do 2 combination ←Exploration(ASG); combination ←Exploitation(ASG); εt ← ε0√ ; t if εt then else end for i in combination do µi,t ← [(t − 1) ∗ µi,t−1 + reward(i)]/t; Ti,t ← Ti,t−1 + 1; 3 4 5 6 7 8 9 10 11 12 end end Complexity In the exploration procedure, we jump out of iteration when the iteration times reaches m. Thus in this procedure the complexity is O(m). In the exploitation procedure, we it- m(cid:99) nodes in ASG and MIN contains a sorting sub- procedure, which brings O(m) complexity. In the last step, we iterate the selected nodes, which is upper bounded by m, erate (cid:98)√ also brings O(m) complexity. As a result, the overall com- plexity of our algorithm is O(m). Regret Analysis2 t j=1 E[ µi,t µi,t i∈St i∈S∗ − Ti,tδ2 2σ2 D · −→ −→ −→ D · −→ −→ ∆βt We proceed to analyze the regret bound brought by BLAG, and meanwhile provably show BLAG exhibits half expected regret than previous approach. Time-Limited Regret Bound of BLAG In ε − greedy, the total reward in round t can be written as: β i − µi,t > δ (cid:17) ≤ 2e D · −→ −→ ep] + (1 − ε)E[ ∆βt et] D · −→ −→ ∆βt] = εE[ where ∆βt ep is the super-action adopted from exploration and ∆βt et is the super-action adopted from exploitation. Let µi,t be estimated reward of base-action i at round t, then we have (cid:3) ≤ αE(cid:104)(cid:80) (cid:105) β 0(j), B∗ (cid:44)(cid:80)B0 D(j) −(cid:80)B0 P(cid:16)−→ isfies: E(cid:2)(cid:80) D and let B0 (cid:44)(cid:80)m and B× (cid:44)(cid:80)m j=m−B0+1 j=1 Theorem 1. For any valid combination D · −→ D · −→ β − −→ −→ B0: Lemma 1. For any two valid combinations (cid:26) −→ D · −→ β 1 − −→ β 2 ≤ B×. D · −→ −→ Let event Ft be where Ti,t is the time base-action i has been selected by round t. The base-actions choosen in exploitation satisfy sat- , where α is the approximation factor of Algorithm 2, St is the output com- bination of Algorithm 2, and S∗ t is the set of base-actions in the optimal super-action. Without loss of generality, we sort D ·−→ −→ D(j)−−→ −→ β 0 β ,(cid:80)m −→ β (j) = −→ −→ D · β 2: , where V is the base-action set and c is a contant. Based on Theorem 1 and Lemma 1, we can get −→ j=1 D(j). −→ ×P(Ft) (1) β i − µi,t ≤ cσ√ 1(cid:112)Ti,t (cid:104)(cid:80) (cid:48) µi,t ≥(cid:80) (cid:104)(cid:80) (cid:105) ≥ EFt (cid:35) is the (cid:48) be the set of base-actions in opti- µi,t, we have i∈S∗ . Now, (cid:48) µi,t we can calculate regret bound from exploitation in T rounds: D · −→ D · −→ −→ −→ The only undetermined element in EFt mal super-action. From(cid:80) observation noise. Let S∗ D · −→ −→ (cid:34)(cid:88) D · −→ ·−→ Let c ≥ √ B×(cid:88) (cid:105) (cid:105) ×(cid:88) 2 ln 2B×M T , thus we can get ·P(Ft) ≤ B×M T · 2e− c2 ∆β∗ = EFt √ T + B β 0 ≥ B∗ D · −→ −→ (cid:104)(cid:80) ≤ 2cσM ∀i ∈ V (cid:88) 2 ≤ 1 −→ β 1, et − α et − α ] ≤ cσ (cid:27) i∈S∗ i∈S∗ i∈S∗ i∈S∗ t ∆βt ∆βt i∈St µi,t µi,t + B ∗ ∆β ∗ ∆β , Ti,t t t E[ j=1 t t t=1 t=1 E t ·P(Ft) (2) t=1 2For page restrictions, rem 1, Lemma 1, Eqs. https://github.com/EugeneLYC/AAAI18. the detailed deduction of Theo- (2) and (3) are available at (1), , and we can t (cid:34)(cid:88) Thus this term can be ignored. Let εt = ε0√ finally get the regret bound of BLAG: E[RBLAG] = D · −→ ·−→ √ ≤ 2cσM T + B D · −→ −→ ·εt + 1 ≤ 2cσM ×(cid:88) et − αT (cid:88) ·εt √ ∆βt ∆β t=1 t=1 E + ∗ (cid:16)−→ D · −→ ep − −→ ∆βt ×√ T + 2B T + 1 D · −→ ∆βt et (cid:17)(cid:35) t=1 The last step holds because m is a sufficiently large number. Superiority of the Derived Regret We further prove BLAG has superior performance on regret bound to the alternative algorithm CUCB (Combinatorial Upper Confidence Bound), an extension of previously pro- posed algorithm (Auer, Cesa-Bianchi, and Fischer 2002). CUCB selects super-actions based on their estimation and variation, i.e., µi,t + cσ√ . Next, we are going to prove that CUCB has twice regret bound than BLAG with given round T . Similar to BLAG, we have Ti,t D · −→ −→ ∆βt − α D · −→ −→ ∆β ∗ E[ ×P(Ft) (3) (cid:88) ] ≤ 2cσ D · −→ −→ i∈St 1(cid:112)Ti,t β i − 2 cσ√ + B ≤ µi,t − Under the condition of Ft, cσ√ β i. Thus the minimum reward in second term must be smaller than the actual minimun. Here c and Ft fol- low the same definitions as those in BLAG. Now we can obtain the regret bound of CUCB as follows: Ti,t Ti,t D · −→ ≤ −→ E[RCUCB] = E(cid:104)−→ D · −→ ×(cid:88) + B ·P(Ft) ≤ 4cσM ∗(cid:105) ≤ 2cσ T(cid:88) (cid:88) t=1 i∈St 1(cid:112)Ti,t D · −→ −→ ∆βt − α √ T + 1 ∆β t=1 √ Considering that m is a sufficiently large number, and M = O(m2) while B× = O(m), we can see that E[RCUCB] = √ √ T + 1 ≈ 4cσM T and E[RBLAG] = 2cσM 2B×√ √ T + 4cσM T + 1 ≈ 2cσM T . Comparing to E[RBLAG], we can see that BLAG has half of the regret bound of CUCB when T is limited. This is because CUCB spends most of its first few rounds on exploration, thus receiving almost ran- dom results. Meanwhile, BLAG can limit the exploration procedures by given ε0, thus having a comparatively stable performance on getting smaller reward. Experimental Results3 Datasets We empirically evaluate the performance of BLAG on one synthetic and three real world datasets4, of which the basic descriptions and statistics are summarized as follows: • B.A. graph (short for Barabasi Albert graph): This is a synthetic graph that forms with newly added nodes pref- erentially attached to existing nodes of higher degrees. It 3Implementation codes and datasets can be found at https://github.com/EugeneLYC/AAAI18. 4The three real datasets are adopted from http://snap.stanford.edu/data/index.html well captures the attribute of power-law degree distribu- tion in real social networks. Each graph has two inputs: n (total number of nodes), p (number of edges each new node attaches to existing nodes) with seed nodes fixed. All the B.A. graphs in our experiment are generated using python package 'networkx'. • Facebook: This dataset consists of 'circles' (or 'friends lists') from Facebook. Facebook data was collected from survey participants using this Facebook app. The dataset includes 4,039 node and 88,234 edges. • Livejournall: LiveJournal is a free on-line community with almost 10 million members. This datasets contains 4,847,571 nodes and 68,993,773 edges. • Pokec: As the most popular on-line social network in Slo- vakia, this dataset contains anonymized data of the whole network, with 1,632,803 nodes and 30,622,564 edges. Information Loss We first evaluate the information loss brought by BLAG. In this experiment, we randomly choose 10 connected seeds as sensitive nodes and randomly block 50% of their neighbors. Here we simulate the learning rounds in one time-slot, thus sensitive nodes are static during the experiment. Each sen- sitive node transmits signal to its uninformed neighbors in each round, with some random rounds dedicated to sensitive information transmission that is represented by transmitting labeled signal. Let info be the cumulative summation of de- gree of the nodes receiving labels in each round. And we calculate variation of info between strategy used and origi- nal transmitting per adaptive diffused labeled signal, record- ing as info loss with normalized results. Info loss demon- strates to what extent a strategy lose information with same amount of sensitive information being adaptively diffused. To demonstrate the superiority of BLAG in terms of less info loss, we compare it with the two baselines: • Monotone Decreasing: This strategy is originated from (Xu, Chen, and others 2015), which shows that forward- ing information with decreasing probability vs. time lim- its the cascading size of information. We fit this strategy into our scenario by decreasing transmission probability in a monotone manner with time decreasing amount re- duce with time as well. • RIPOSTE-like: This strategy is originated from the previously proposed RIPOSTE(Giakkoupis et al. 2015). Typically, RIPOSTE forwards an item with probability slightly larger than a given amount if a user likes the item, or with probability slightly smaller than that amount oth- erwise. To fit this strategy into our scenario, we randomly choose some rounds and decrease the transmission prob- ability by a static amount. We graphically plot the results of info loss in Fig. 4, where BLAG exhibits the best performance on saving information loss under the same amount of adaptively diffused sensi- tive information. Due to the noise brought by the learning process, we can see that in some cases depicted by Fig. 4(b)(c)(e), BLAG does not show its effectiveness in the first few rounds, but the trend of info loss in BLAG is decreas- ing. In contrast, RIPOSTE-like strategy has unstable perfor- mance. This is caused by the random choice on the rounds. Figure 3: Information loss incurred by different strategies. In other words, RIPOSTE-like strategy generally does not consume too much information loss, but meanwhile fails to accurately capture the sensitive information. This result ac- cords with previous assertion that sensitive information is hard to capture. Among all the three strategies, monotone decreasing turns out to perform the worst in the sense that it incurs most information loss in adaptive diffusion of each piece of sensitive information. Furthermore, Fig. 4(a)(b)(c) reveal with the increase of network scale and complexity, that is, more nodes (increas- ing of m) and larger density (increasing of p), BLAG goes through a longer sequence of rounds with unpredictable info loss. This is caused by increasing of m, thus leading to larger size of ASG. That results in randomness in combination adopted from each round and consumption of more rounds to get close to the optimal action. From Fig. 4(a)-(f) we can roughly calculate the least info loss saved by BLAG and other two algorithms. The least amount is depicted in Fig. 4(e), where we get 0.25 normal- ized info loss from BLAG and 0.4 from REPOSTE-like. Their difference is about 40%. Regret in Learning Process Now we proceed to evaluate the regret of BLAG and CUCB with limited rounds and large base-action set. Particularly, as it is time consuming to calculate the regret bound, we calculate the cumulative reward of each round instead. Note that the two evaluations are equivalent since the regret can be deduced from the reward. First we extract m nodes from a B.A. graph (n = 10K, p = 5) as target nodes and record their degrees. Then we generate 2(cid:0)m in previous sections. Thus 2(cid:0)m (cid:1) base-actions by ran- (cid:1) is the size of base-action set. domly allocating element values under conditions discussed And we set ε0 = 1 for the initialization of global ε− greedy process. The comparison results are summarized in Table 3 2 2 From Table 3 we observe a superior learning perfor- Rounds Base-actions CUCB -0.6e-2 -0.6e-3 -0.4e-3 -0.2e-5 -1.4e-2 -0.8e-3 -0.5e-3 -1.6e-4 BLAG -6.7e-2 -32.0e-3 -4.8e-3 -12.3e-4 -8.4e-2 -22.3e-3 -109.9e-3 -14.8e-4 1000 3000 200 5K 20K 2M 200 5K 20K 2M Table 1: Comparison of cumulative reward between CUCB and BLAG within limited rounds. The results are normalized. mance of the global ε−greedy process adopted in BLAG to CUCB. This is because in the first few rounds, ε − greedy can allocate rounds for exploration and exploitation by the given ε0 while CUCB allocates rounds by the observed re- ward and variation of each base-action. Due to the limited rounds, most of the base-actions in ASG are chosen with few times. That means, CUCB spends most of its first few rounds on exploration, thus leading to unstable (bad) perfor- mance. By comparison, ε−greedy benefits from the current estimation and artificially limited exploration probability. In real life, the initiated ε0 can be determined by algorithm ex- cuters to deceide whether stick to the current estimation or slightly tolerate more explorations. We can see from Table 3 that given same rounds, BLAG gets reward at least 10 times than CUCB (remember in our scenario, the smaller reward, the better). Furthermore, with number of base-actions increase from 200 to 2M, CUCB has an obvious deteriorating perfor- mance, increasing from -0.6e-2 to -0.2e-5 and from -1.4e- 2 to -1.6e-4 when learning rounds are 1,000 and 3,000, re- spectively. In contrast, BLAG has a stable performance on getting reward. We can see that when given 3,000 rounds, cumulative reward from BLAG decrease from -22.3e-3 to - 109.9e-3 when base-actions increase from 5K to 20K. This is also attributed to the increased randomness in CUCB and BLAG sticking to the current estimation controlled by given ε. And we can slightly make the conclusion that with greater extent on large base-action set and limited rounds, BLAG will show even more superior performance than CUCB. Cascading Scale Last but not least, we also take a look into the effective- ness of adaptive diffusion on limiting cascading of sensitive information. In doing so, in each experiment, we first ran- domly select a source node to be the sensitive node. We ini- tialize the transmission probability on each edge connecting a sensitive node and non-sensitive node be 5e − 5, which means the sensitive nodes have 5e − 5 probability to trans- fer its non-sensitive nodes to be sensitive ones. At each time slot, we sort the degree of neighbors and let transmission probability to those first half nodes be 0 and to those last half be 1e − 4. In that case, we adaptively diffuse the sensi- tive information but maintain the overall probabilities, which accords to the main idea of adaptive diffusion. Since the learning process is not the focus in this experiment, we let total number of round in one time-slot be 1. And we cal- culate the rate of sensitive nodes divided by total number of nodes in the network. Fig. 4 plots the cascading scale of both Figure 4: Cascading size under adaptive diffusion of BLAG and spontaneous spreading. BLAG and spontaneous spreading under B.A. graphs (Fig. 4(a)-(c)) and three real datasets (Fig. 4(d)-(f)). A common observation from Fig. 4(a)-(f) is that the cascading behavior undergoes a transition, i.e., the cascading scale increases ex- plosively after a certain time threshold. Adaptive diffusion has strong effect on postponing such threshold. Considering of the time-sensitive attribute of sensitive information, if the threshold is postponed after the timeliness of sensitive in- formation, no sensitive information will be widely spread. Related Works Back in 2003, (Kempe, Kleinberg, and Tardos 2003) ana- lyzed two classic diffusion models: LTM and ICM. Differ- ent works have been done on information diffusion. (Cheng et al. 2014)(Prakash et al. 2012)(Venkatramanan and Ku- mar 2012) discuss how different behavior in network may affect the information diffusion. (Iribarren and Moro 2009) and (Karsai et al. 2011) discuss human activities patterns that would slow down diffusion. (Myers, Zhu, and Leskovec 2012) and (Wu et al. 2011) convey investigation on how roles of users influence information diffusion. Our work is also connected to combinatorial multi-arm bandit problem. In our combinatorial bandit setting, feed- back is observed on each played arm individually during each round of play. Other types of feedback (Audibert, Bubeck, and Lugosi 2011)(Mannor and Shamir 2011) in- clude observing the outcomes of all arms, and only ob- serveing the final reward but no outcome of any individual arm. (Anantharam, Varaiya, and Walrand 1987)(Caro and Gallien 2007) consider simultaneous plays of arms. (Auer 2002)(Hanawal et al. 2015) discuss the case where only the cumulative reward of the combination can be observed in a single round. Conclusion In this paper, we propose an adaptive diffusion strategy in social network to protect sensitive information while de- creasing information loss. We design BLAG algorithm to tackle the formed constrained combinatorial Multi-Armed Bandit problem. We show that our algorithm reduces time- limited regret bound by half compared to previous proposed CUCB algorithms. Mannor, S., and Shamir, O. 2011. From bandits to experts: In Advances in Neural On the value of side-observations. Information Processing Systems, 684 -- 692. Myers, S. A.; Zhu, C.; and Leskovec, J. 2012. Information In Proceed- diffusion and external influence in networks. ings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, 33 -- 41. ACM. Prakash, B. A.; Chakrabarti, D.; Valler, N. C.; Faloutsos, M.; and Faloutsos, C. 2012. Threshold conditions for arbitrary cascade models on arbitrary networks. Knowledge and in- formation systems 33(3):549 -- 575. Venkatramanan, S., and Kumar, A. 2012. Co-evolution of content popularity and delivery in mobile p2p networks. In INFOCOM, 2012 Proceedings IEEE, 3006 -- 3010. IEEE. Wang, F.; Wang, H.; and Xu, K. 2012. Diffusive logis- tic model towards predicting information diffusion in online social networks. In Distributed Computing Systems Work- shops (ICDCSW), 2012 32nd International Conference on, 133 -- 139. IEEE. Wu, S.; Hofman, J. M.; Mason, W. A.; and Watts, D. J. 2011. Who says what to whom on twitter. In Proceedings of the 20th international conference on World wide web, 705 -- 714. ACM. Xu, X.; Chen, X.; et al. 2015. Modeling time-sensitive in- formation diffusion in online social networks. In Computer Communications Workshops (INFOCOM WKSHPS), 2015 IEEE Conference on, 408 -- 413. IEEE. Yang, J., and Leskovec, J. 2010. Modeling information dif- fusion in implicit networks. In Data Mining (ICDM), 2010 IEEE 10th International Conference on, 599 -- 608. IEEE. Yang, Y.; Tang, J.; Leung, C. W.-k.; Sun, Y.; Chen, Q.; Li, J.; and Yang, Q. 2015. Rain: Social role-aware information diffusion. In AAAI, 367 -- 373. References Anantharam, V.; Varaiya, P.; and Walrand, J. 1987. Asymp- totically efficient allocation rules for the multiarmed bandit problem with multiple plays-part i: Iid rewards. IEEE Trans- actions on Automatic Control 32(11):968 -- 976. Audibert, J.-Y.; Bubeck, S.; and Lugosi, G. 2011. Minimax policies for combinatorial prediction games. In Proceedings of the 24th Annual Conference on Learning Theory, 107 -- 132. Auer, P.; Cesa-Bianchi, N.; and Fischer, P. 2002. Finite- time analysis of the multiarmed bandit problem. Machine learning 47(2-3):235 -- 256. Auer, P. 2002. Using confidence bounds for exploitation- exploration trade-offs. Journal of Machine Learning Re- search 3(Nov):397 -- 422. Bakshy, E.; Rosenn, I.; Marlow, C.; and Adamic, L. 2012. The role of social networks in information diffusion. In Pro- ceedings of the 21st international conference on World Wide Web, 519 -- 528. ACM. Caro, F., and Gallien, J. 2007. Dynamic assortment with demand learning for seasonal consumer goods. Management Science 53(2):276 -- 292. Chen, W.; Wang, Y.; and Yuan, Y. 2013. Combinatorial multi-armed bandit: General framework and applications. In International Conference on Machine Learning, 151 -- 159. Cheng, J.; Adamic, L.; Dow, P. A.; Kleinberg, J. M.; and Leskovec, J. 2014. Can cascades be predicted? In Pro- ceedings of the 23rd international conference on World wide web, 925 -- 936. ACM. Gai, Y.; Krishnamachari, B.; and Jain, R. 2010. Com- binatorial network optimization with unknown variables: Multi-armed bandits with linear rewards. arXiv preprint arXiv:1011.4748. Giakkoupis, G.; Guerraoui, R.; J´egou, A.; Kermarrec, A.- M.; and Mittal, N. 2015. Privacy-conscious information diffusion in social networks. In International Symposium on Distributed Computing, 480 -- 496. Springer. Guille, A., and Hacid, H. 2012. A predictive model for the temporal dynamics of information diffusion in online social networks. In Proceedings of the 21st international confer- ence on World Wide Web, 1145 -- 1152. ACM. Hanawal, M.; Saligrama, V.; Valko, M.; and Munos, R. 2015. Cheap bandits. In Proceedings of the 32nd Interna- tional Conference on Machine Learning (ICML-15), 2133 -- 2142. Iribarren, J. L., and Moro, E. 2009. Impact of human activity patterns on the dynamics of information diffusion. Physical review letters 103(3):038702. Karsai, M.; Kivela, M.; Pan, R. K.; Kaski, K.; Kert´esz, J.; Barab´asi, A.-L.; and Saramaki, J. 2011. Small but slow world: How network topology and burstiness slow down spreading. Physical Review E 83(2):025102. Kempe, D.; Kleinberg, J.; and Tardos, ´E. 2003. Maximizing the spread of influence through a social network. In Proceed- ings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, 137 -- 146. ACM.
1611.05561	3	1611	2018-08-28T07:37:57	A Fast and Provable Method for Estimating Clique Counts Using Tur\'an's Theorem	[ "cs.SI", "cs.DS" ]	Clique counts reveal important properties about the structure of massive graphs, especially social networks. The simple setting of just 3-cliques (triangles) has received much attention from the research community. For larger cliques (even, say 6-cliques) the problem quickly becomes intractable because of combinatorial explosion. Most methods used for triangle counting do not scale for large cliques, and existing algorithms require massive parallelism to be feasible. We present a new randomized algorithm that provably approximates the number of k-cliques, for any constant k. The key insight is the use of (strengthenings of) the classic Tur\'an's theorem: this claims that if the edge density of a graph is sufficiently high, the k-clique density must be non-trivial. We define a combinatorial structure called a Tur\'an shadow, the construction of which leads to fast algorithms for clique counting. We design a practical heuristic, called TUR\'AN-SHADOW, based on this theoretical algorithm, and test it on a large class of test graphs. In all cases,TUR\'AN-SHADOW has less than 2% error, in a fraction of the time used by well-tuned exact algorithms. We do detailed comparisons with a range of other sampling algorithms, and find that TUR\'AN-SHADOW is generally much faster and more accurate. For example, TUR\'AN-SHADOW estimates all cliques numbers up to size 10 in social network with over a hundred million edges. This is done in less than three hours on a single commodity machine.	cs.SI	cs	A Fast and Provable Method for Estimating Clique Counts Using Turán's Theorem University of California, Santa Cruz University of California, Santa Cruz C. Seshadhri Santa Cruz, CA [email protected] Shweta Jain Santa Cruz, CA [email protected] 8 1 0 2 g u A 8 2 ] I S . s c [ 3 v 1 6 5 5 0 . 1 1 6 1 : v i X r a reveal important properties about ABSTRACT the Clique counts structure of massive graphs, especially social networks. The simple setting of just 3-cliques (triangles) has received much attention from the research community. For larger cliques (even, say 6-cliques) the problem quickly becomes intractable because of combinatorial explosion. Most methods used for triangle counting do not scale for large cliques, and existing algorithms require massive parallelism to be feasible. We present a new randomized algorithm that provably for any constant approximates the number of k-cliques, k. The key insight is the use of (strengthenings of) the classic Tur´an's theorem: this claims that if the edge density of a graph is sufficiently high, the k-clique density must be non-trivial. We define a combinatorial structure called a Tur´an shadow, the construction of which leads to fast algorithms for clique counting. We design a practical heuristic, called Tur´an-shadow, based on this theoretical algorithm, and test it on a large class of test graphs. In all cases, Tur´an-shadow has less than 2% error, and runs in a fraction of the time used by well-tuned exact algorithms. We do detailed comparisons with a range of other sampling algorithms, and find that Tur´an-shadow is generally much faster and more accurate. For example, Tur´an-shadow estimates all clique numbers up to size 10 in social network with over a hundred million edges. This is done in less than three hours on a single commodity machine. CCS Concepts •Theory of •Mathematics theory; computation → Social networks; graph of computing → Extremal Keywords Cliques, sampling, graphs, Tur´an's theorem c(cid:13)2017 International World Wide Web Conference Committee (IW3C2), published under Creative Commons CC BY 4.0 License. ACM 978-1-4503-4913-0/17/04. http://dx.doi.org/10.1145/3038912.3052636 1. INTRODUCTION Pattern counting is an important graph analysis tool in many domains: anomaly detection, social network analysis, bioinformatics among others [21, 27, 10, 29, 22, 17]. Many real world graphs show significantly higher counts of certain patterns than one would expect in a random graph [21, 46, 27]. This technique has been referred to with a variety of names: subgraph analysis, motif counting, graphlet analysis, etc. But the fundamental task is to count the occurrence of a small pattern graph in a large input graph. In all such applications, it is essential to have fast algorithms for pattern counting. It is well-known that certain patterns capture specific semantic relationships, and thus the social dynamics are reflected in these graph structures. The most famous such pattern is the triangle, which consists of three vertices connected to each other. Triangle counting has a rich history in the social sciences and network science [21, 46, 10, 47]. We focus on the more general problem of clique counting. A k-clique is a set of k vertices that are all connected to each other; thus, a triangle is a 3-clique. Cliques are extremely significant in social network analysis (Chap. 11 of [20] and Chap. 2 of [23]). They are the archetypal example of a dense subgraph, and a number of recent results use cliques to find large, dense subregions of a network [32, 41, 28, 43]. 1.1 Problem Statement Given an undirected graph G = (V, E), a k-clique is a set S of k vertices in V with all pairs in S connected by an edge. The problem is to count the number of k-cliques, for varying values of k. Our aim is to get all clique counts for k ≤ 10. The primary challenge is combinatorial explosion. An autonomous system network with ten million edges has more than a trillion 10-cliques. Any enumeration procedure is doomed to failure. Under complexity theoretical assumptions, clique counting is believed to be exponential in the size k [11], and we cannot hope to get a good worst-case algorithm. Our aim is to employ randomized sampling methods for clique counting, which have seen some success in counting triangles and small patterns [42, 35, 25]. We stress that we make no distributional assumption on the graph. All probabilities are over the internal randomness of the algorithm itself (which is independent of the instance). 1.2 Main contributions Our main theoretical result is a randomized algorithm Tur´an-shadow that approximates the k-clique count, for any constant k. We implement this algorithm on a (i) Percent relative error for k=7 (ii) Timings for k=7 and k=10 (iii) Speedup for k=7 Figure 1: Summary of behavior of Tur´an-shadow over several datasets. Fig. 1i shows the percent relative error in the estimates for k=7 given by Tur´an-shadow. We only show results for graphs for which we were able to obtain exact counts using either brute force enumeration, or from the results of [18]. The errors are always < 2% and mostly < 1%. Fig. 1ii shows the time taken by Tur´an-shadow for k=7 and k=10. Fig. 1iii shows the speedup (time of algorithm/time of Tur´an-shadow) over other state of the art algorithms for k=7. The red line indicates a speedup of 1. We could not give a figure for speedup for k=10 because for most instances no competing algorithm terminated in min(7 hours, 100 times Tur´an-shadow time). commodity machine and get k-clique counts (for all k ≤ 10) on a variety of data sets, the largest of which has 100M edges. The main features of our work follow. Extremal combinatorics meets sampling. Our novelty is in the algorithmic use of classic extremal combinatorics results on clique densities. Seminal results of Tur´an [44] and Erdos [16] provide bounds on the number of cliques in a sufficiently dense graph. Tur´an-shadow tries to cover G by a carefully chosen collection of dense subgraphs that contains all cliques, called a Tur´an-shadow. It then uses standard techniques to design an unbiased estimator for the clique count. Crucially, the result of Erdos [16] (a quantitative version of Tur´an's theorem) is used to bound the variance of the estimator. We of Tur´an-shadow, proving correctness and analyzing its time complexity. The running time of our algorithm is bounded by the time to construct the Tur´an-shadow, which as we shall see, is quite feasible in all the experiments we run. theoretical detailed provide a analysis Extremely fast. In the worst case, we cannot expect the Tur´an-shadow to be small, as that would imply new theoretical bounds for clique counting. But in practice on a wide variety of real graphs, we observe it to be much smaller than the worst-case bound. Thus, Tur´an-shadow can be made into a practical algorithm, which also has provable bounds. We implement Tur´an-shadow and run it on a commodity machine. Fig. 1ii shows the time required for Tur´an-shadow to obtain estimates for k = 7 and k = 10 in seconds. The as-skitter graph is processed in less than 3 minutes, despite there being billions of 7-cliques and trillions of 10-cliques. All graphs are processed in minutes, except for an Orkut social network with more than 100M edges (Tur´an-shadow handles this graph within 2.5 hours). To the best of our knowledge, there is no existing work that gets comparable results. An algorithm of Finocchi et al. also computes clique counts, but employs MapReduce on the same datasets [18]. We only require a single machine to get a good approximation. We tested Tur´an-shadow against a number of state of the art algorithmic techniques (color coding [4], edge sampling [42], GRAFT [30]). For 10-clique counting, none of these algorithms terminate for all instances even in 7 hours; Tur´an-shadow runs in minutes on all but one instance (where it takes less than 2.5 hours). For 7-clique counting, Tur´an-shadow is typically 10-100 times faster than competing algorithms. (A notable exception is com-orkut, where an edge sampling algorithm runs much faster than Tur´an-shadow.) Excellent accuracy. Tur´an-shadow has extremely small variance, and computes accurate results (in all instances we could verify). We compute exact results for 7-clique numbers, and compare with the output of Tur´an-shadow. In Fig. 1i, we see that the accuracy is well within 2% (relative error) of the true answer for all datasets. We do detailed experiments to measure variance, and in all cases, Tur´an-shadow is accurate. The efficiency and accuracy of Tur´an-shadow allows us to get clique counts for a variety of graphs, and track how the counts change as k increases. We seem to get two categories of graphs: those where the count increases (exponentially) with k, and those where it decreases with k, see Fig. 5. This provides a new lens to view social networks, and we hope Tur´an-shadow can become a new tool for pattern analysis. 1.3 Related Work The importance of pattern counts gained attention in bioinformatics with a seminal paper of Milo et al. [27], though it has been studied for many decades in the social sciences [21]. Triangle counting and its use has an incredibly rich history, and is used in applications as diverse as spam detection [6], graph modeling [34], and role detection [10]. Counting four cliques is mostly feasible using some recent developments in sampling and exact algorithms [25, 2]. Clique counts are an important part of recent dense subgraph discovery algorithms [32, 41]. Cliques also play an important role in understanding dynamics of social [24], and their importance in the social sciences capital is well documented [20, 23]. In topological approaches to network analysis, cliques are the fundamental building blocks used to construct simplicial structures [36]. From an algorithmic perspective, clique counting has received much attention from the theoretical computer science community [12, 4, 11, 45]. Maximal clique loc-gowweb-StanamazonyoutubeGoogleBerkStanas-skitterPatentssoc-pokecGraphs0.00.51.01.52.0Percentrelativeerrork=7loc-gowweb-StanamazonyoutubeGoogleBerkStanas-skitterPatentssoc-pokeccom-ljcom-orkutGraphs10−1100101102103104105106TimeinsecondsTimingsk=7k=10loc-gowweb-StanamazonyoutubeGoogleBerkStanas-skitterPatentssoc-pokeccom-ljcom-orkutGraphs10−1100101102103Speedupk=7ESGRAFTenumeration has been an important topic [3, 37, 15] since the seminal algorithm of Bron-Kerbosch [9]. Practical algorithms for finding the maximum clique were given by Rossi et al. using branch and bound methods [31]. Most relevant to our work is a classic algorithm of Chiba and Nishizeki [12]. This work introduces graph orientations to reduce the search time and provides a theoretical connection to graph arboricity. We also apply this technique in Tur´an-shadow. The closest result to our work is a recent MapReduce algorithm of Finocchi et al. for clique counting [18]. This result applies the orientation technique of [12], and creates a large set of small (directed) egonets. Clique counting overall reduces to clique counting in each of these egonets, and this can be parallelized using MapReduce. We experiment on the same graphs used in [18] (particularly, some of the largest ones) and get accurate results on a single, commodity machine (as opposed to using a cluster). Alternate MapReduce methods using multi-way joins have been proposed, though this is theoretical and not tested on real data [1]. A number of randomized techniques have been proposed for pattern counting, and can be used to design algorithms for clique counting. Most prominent are color coding [4, 22, 7, 48] and edge sampling methods [42, 40, 30]. (MCMC methods [8] typically do not scale for graphs with millions of vertices [25].) We perform detailed comparisons with these methods, and conclude that they do not scale for larger clique counting. 2. MAIN IDEAS The starting point for our result is a seminal theorem if the edge density of a graph is more than of Tur´an [44]: 1− 1 k−1 , then it must contain a k-clique. (The density bound is often called the Tur´an density for k.) Erdos proved a stronger version [16]. Suppose the graph has n vertices. Then in this case, it contains Ω(nk−2) k-cliques! Consider the trivial randomized algorithm to estimate k-cliques. Simply sample a uniform random set of k vertices and check if they form a clique. Denote the number of k-cliques by C, then the success probability is C/(cid:0)n we can estimate this probability using (cid:0)n (cid:1). Thus, (cid:1)/C samples. By k Erdos' bound, C = Ω(nk−2). Thus, if the density of a graph (with n vertices) is above the Tur´an density, one can estimate the number of k-cliques using O(n2) samples. k Of course, the input graph G is unlikely to have such a high density, and O(n2) is a large bound. We try to cover all k-cliques in G using a collection of dense subgraphs. This collection is called a Tur´an shadow. We employ orientation techniques from Chiba-Nishizeki to recursively construct a shadow [12]. We take the degeneracy (k-core) ordering in G [33]. It is well-known that outdegrees are typically small in this ordering. To count k-cliques in G, it suffices to count (k − 1)-cliques in every outneighborhood. (This is the main idea in the MapReduce algorithms of Finocchi et al [18].) If an outneighborhood has density higher than the Tur´an density for (k − 1), we add this set/induced subgraph to the Tur´an shadow. If not, we recursively employ this scheme to find denser sets. When the process terminates, we have a collection of sets (or induced subgraphs) such that each has density above the Tur´an threshold (for some appropriate k(cid:48) for each set). Furthermore, the sum of cliques (k(cid:48)-cliques, for the same k(cid:48)) is the number of k-cliques in G. Now, we can hope to use the randomized procedure to estimate the number of k(cid:48)-cliques in each set of the Tur´an shadow. By a theorem of √ Chiba-Nishizeki [12], we can argue that number of vertices 2m (where m is in any set of the Tur´an shadow is at most the number of edges in G). Thus, O(m) samples suffices to estimate clique counts for any set in the Tur´an shadow. But the Tur´an shadow has many sets, and it is infeasible to spend O(m) samples for each set. We employ a randomized trick. We only need to approximate the sum of clique counts over the shadow, and can use random sampling for that purpose. Working through the math, we effectively set up a distribution over the sets in the Tur´an shadow. We pick a set from this distribution, pick some subset of random vertices, and check if they form a clique. The probability of this event can be related to the number of k-cliques in G. Furthermore, we can prove that O(m) samples suffice to estimate this probability. All in all, after constructing the Tur´an shadow, k-clique counting can be done in O(m) time. 2.1 Main theorem and significance The formal version of the main theorem is Theorem 5.6. It requires a fair bit of terminology to state. So we state an informal version that maintains the spirit of our main result. This should provide the reader with a sense of what we can hope to prove. We will define the Tur´an shadow formally in later sections. But it basically refers to the construct described above. Theorem 2.1. [Informal] Consider graph G = (V, E) with n vertices, m edges, and maximum core number α. Let S be the Tur´an k-clique shadow of G, and let S be the number of sets in S. Given any δ > 0, ε > 0, k, with probability at least 1−δ, the procedure Tur´an-shadow outputs a (1 + ε)-multiplicative approximation to the number of k-cliques in G. The running time is linear in S and mα log(1/δ)/ε2. The storage is linear in S. Observe that the size of the shadow is critical to the procedure's efficiency. As long as the number of sets in the Tur´an shadow is small, the extra running time overhead is only linear in m. And in practice, we observe that the Tur´an shadow scales linearly with graph size, leading to a practically viable algorithm. Outline: In §3, we formally describe Tur´an's theorem and set some terminology. §4 defines (saturated) shadows, and shows how to construct efficient sampling algorithms for clique counting from shadow. §5 describes the recursive In §5.1, we describe construction of the Tur´an shadow. the final procedure Tur´an-shadow, and prove (the formal in §6, we detail our version of) Theorem 2.1. Finally, empirical study of Tur´an-shadow and comparison with the state of the art. 3. TURÁN'S THEOREM Ci(H)/(cid:0)V (H) For any arbitrary graph H = (V (H), E(H)), Ci(H) denote the set of cliques in H, and ρi(H) (cid:1) is the i-clique density. Note that ρ2(H) let := is the standard notion of edge density. i The following theorem of Tur´an is one of the most important results in extremal graph theory. Theorem 3.1. (Tur´an [44]) For any graph H, if ρ2(H) > 1 − 1 k−1 , then H contains a k-clique. This is tight, as evidenced by the complete (k − 1)-partite graph Tn,k−1 (also called the Tur´an graph). In a remarkable generalization, Erdos proved that if an n-vertex graph has even one more edge than Tn,k−1, it must contain many k-cliques. One can think of this theorem as a quantified version of Tur´an's theorem. Algorithm 1: sample(S, γ, k, ε, δ) S is γ-saturated k-clique shadow ε, δ are error parameters 1 For each (S, (cid:96)) ∈ S, set w(S) =(cid:0)S p(S) = w(S)/(cid:80) 2 Set probability distribution D over S where 3 For r ∈ 1, 2, . . . , t = 20 (S,(cid:96))∈S w(S) ; (cid:1) ; γε2 log(1/δ); (cid:96) Independently sample (S, (cid:96)) from D; Choose a u.a.r. (cid:96)-tuple A from S; If A forms (cid:96)-clique, set indicator Xr = 1. Else, 4 5 6 Xr = 0 ; (cid:80) (cid:80) r Xr t (S,(cid:96))∈S (cid:96) (cid:0)S (cid:1) as estimate for Ck(G); Theorem 3.2. (Erdos [16]) For any graph H over n k−1 , then H contains at least vertices, if ρ2(H) > 1 − 1 (n/(k − 1))k−2 k-cliques. 7 Output It will be convenient to express this result in terms on k-clique densities. We introduce some notation: let f (k) = kk−2/k!. By Stirling's approximation, f (k) is well √ approximated by ek/ 2πk5. Note that f (k) is some fixed constant, for constant k. This corollary will be critical to our analysis. Corollary 3.3. For any graph H over n vertices, ρ2(H) > 1 − 1 k−1 , then ρk(H) ≥ 1/f (k)n2. if Proof. By Theorem 3.2, H has at least ( n (k−1) )k−2 k-cliques. Thus, (cid:0)n (cid:1) (k−1) )k−2 ρk(H) ≥ ( n k ≥ nk−2/nk×k!/(k−1)k−2 ≥ 1/(f (k)n2) 4. CLIQUE SHADOWS A key concept in our algorithm is that of clique shadows. Consider graph G = (V, E). For any set S ⊆ V , we let C(cid:96)(S) denote the set of (cid:96)-cliques contained in S. that: there is a bijection between Ck(G) and(cid:83) Definition 4.1. A k-clique shadow S for graph G is a multiset of tuples {(Si, (cid:96)i)} where Si ⊆ V and (cid:96)i ∈ N such (S,(cid:96))∈S C(cid:96)(S). Furthermore, a k-clique shadow S is γ-saturated if ∀(S, (cid:96)) ∈ S, ρ(cid:96)(S) ≥ γ. Intuitively, it is a collection of subgraphs, such that the sum of clique counts within them is the total clique count of G. Note that for each set S in the shadow, the associated clique size (cid:96) is different (for different S). Observe that {(V, k)} is trivally a clique shadow. But it is highly unlikely to be saturated. It is important to define the size of S, which is really the storage required to represent it. size(S), and is(cid:80) (S,(cid:96))∈S S. Definition 4.2. The representation size of S is denoted When a k-clique shadow S is γ-saturated, each (S, (cid:96)) ∈ S has many (cid:96)-cliques. Thus, one can employ random sampling within each S to estimate C(cid:96)(S), and thereby estimate Ck(G). We use a sampling trick to show that we do not need to estimate all C(cid:96)(S); instead we only need O(1/γ) samples in total. The Theorem 4.3. Suppose S is a γ-saturated k-clique shadow for G. The procedure sample(S) outputs an estimate C such C − Ck(G) ≤ εCk(G) with probability > 1 − δ. is O(size(S) + time of sample(S) 1 γε2 log(1/δ)). Proof. We remind the reader that w(S) = (cid:0)S α = Ck(G)/(cid:80) S∈S w(S). Observe that (cid:1). Set running (cid:96) Pr[Xr = 1] = Pr[(S, (cid:96)) is chosen] (cid:88) (S,(cid:96))∈S (cid:88) (cid:96) (cid:1) = C(cid:96)(S)/w(S). So, × Pr[(cid:96)-clique chosen in S(S, (cid:96)) is chosen] S∈S w(S), and The former probability is exactly w(S)/(cid:80) the latter is exactly C(cid:96)(S)/(cid:0)S (cid:88) (cid:88) Since S is a k-clique shadow, (cid:80) γ, equivalent to C(cid:96)(S) ≥ γw(S). So (cid:80) γ(cid:80) r≤t Xr] =(cid:80) expectation, E[(cid:80) (S,(cid:96))∈S C(cid:96)(S) = Ck(G). Thus, Pr[Xr = 1] = α. By the saturation property, ρ(cid:96)(S) ≥ S∈S C(cid:96)(S) ≥ S∈S w(S). That implies that α ≥ γ. By linearity of r≤t E[Xr] ≥ γt. C(cid:96)(S)/ Pr[Xr = 1] = (S,(cid:96))∈S w(S) S∈S Note that all the Xrs come from independent trials. (The graph structure plays no role, since the distribution of each Xr does not change upon conditioning on the other Xrs.) By a multiplicative Chernoff bound (Thm 1.1 of [13]), Pr[ Xr/t ≤ α(1 − ε)] ≤ exp(−ε2E[ Xr]/3) r r ≤ exp(−ε2γt/3) = exp(−5 log(1/δ)) ≤ δ/5. By an analogous upper tail bound, Pr[(cid:80) α(1 − ε) ≤ (cid:80) C = ((cid:80) r Xr/t)(cid:80) on(cid:80) r Xr/t by(cid:80) r Xr/t ≥ α(1+ε)] ≤ δ/5. By the union bound, with probability at least 1− 2δ/5, r Xr/t ≤ α(1 + ε). Note that the output S∈S w(S). We multiply the bound above S∈S w(S) = S∈S w(S), and note that α(cid:80) Ck(G) to complete the proof. We stress the significance of Theorem 4.3. Once we get a γ-saturated clique shadow S, Ck(G) can be approximated in time linear in size(S). The number of samples chosen only depends on γ and the approximation parameters, not on the graph size. But how to actually generate a saturated clique shadow? Saturation appears to be extremely difficult to enforce. This is where the theorem of Erdos (Theorem 3.2) saves the day. (cid:88) It merely suffices to make the edge density of each set in the clique shadow high enough. The k-clique density automatically becomes large enough. Theorem 4.4. Consider a k-clique shadow S such Let γ = that ∀(S, (cid:96)) ∈ S, ρ2(S) > 1 − 1 (cid:96)−1 . 1/ max(S,(cid:96))∈S f ((cid:96))S2. Then, S is γ-saturated. Proof. By Corollary 3.3, for every (S, (cid:96)) ∈ S, ρ(cid:96)(S) ≥ 1/(f ((cid:96))S2). We simply set γ to be the minimum such density over all (S, (cid:96)) ∈ S. 5. CONSTRUCTING CLIQUE SHADOWS SATURATED We use a refinement process to construct saturated clique shadows. We start with the trivial shadow S = {(V, k)} and iteratively "refine" it until the saturation property is satisfied. By Theorem 4.4, we just have to ensure edge densities in each set are sufficiently large. For any set S ⊂ V , let GS be the subgraph of G induced by S. Given an unsaturated k-clique shadow S, we find some (S, (cid:96)) ∈ S such that ρ2(S) ≤ 1 − 1 (cid:96)−1 . By iterating over the vertices, we replace (S, (cid:96)) by various neighborhoods in GS to get a new shadow. We would like the edge densities of these neighborhoods to increase, in the hope of crossing the threshold given in Theorem 4.4. The key insight is to use the degeneracy ordering to construct specific neighborhoods of high density that also yield a valid shadow. This is basically the classic graph theoretic technique of computing core decompositions, which is widely used in large-graph analysis [33, 19]. As mentioned earlier, this idea is used for fast clique counting as well [12, 18]. Definition 5.1. For a (labeled) graph G = (V, E), a degeneracy ordering is a permutation of V given as v1, v2, . . . , vn such that: for each i ≤ n, vi is the minimum degree vertex in the subgraph induced by vi, vi+1, . . . , vn. (As defined, this ordering is not unique, but we can enforce uniqueness by breaking ties by vertex id.) The degree of vi in G{vi,...,vn} is the core number of vi. The largest core number is called the degeneracy of G, denoted α(G). The degeneracy DAG of G, denoted D(G) is obtained In other words, by orienting edges in degeneracy order. every edge (u, v) ∈ G is directed from lower to higher in the degeneracy ordering. The degeneracy ordering is the deletion time of the standard linear time procedure that computes the degeneracy [26]. It is convenient for us to think of the degeneracy in terms of graph orientations. As defined earlier, any permutation on V can be used to make a DAG out of G. We use this idea for generating saturated clique shadows. Essentially, while G may be sparse, out-neighborhoods in G are typically dense. (This has been observed in numerous results on dense subgraph discovery [5, 38, 32].) We now define the procedure Shadow-Finder(G, k), which works by a simple, iterative refinement procedure. Think of T as the current working set, and S as the final output. We take a set (S, (cid:96)) in T , and construct all outneighborhoods in the degeneracy DAG. Any such set whose density is above 4 3 s denote the outneighborhood (within Algorithm 2: Shadow-Finder(G, k) 1 Initialize T = {(V, k)} and S = ∅; 2 While ∃(S, (cid:96)) ∈ T such that ρ2(S) ≤ 1 − 1 (cid:96)−1 ; Construct the degeneracy DAG D(GS); Let N + D(GS)) of s ∈ S; Delete (S, (cid:96)) from T ; For each s ∈ S; If (cid:96) ≤ 2 or ρ2(N + Add (N + s , (cid:96) − 1) to S; s ) > 1 − 1 (cid:96)−2 ; s , (cid:96) − 1) to T ; Else, add (N + 5 6 7 8 9 10 Output S; the Tur´an threshold goes to S (the output), otherwise, it goes to T (back to the working set). It is useful to define the recursion tree T of this process as follows. Every pair (S, (cid:96)) that is ever part of T is a node in s , (cid:96) − 1) T . The children of (S, (cid:96)) are precisely the pairs (N + added in Step 8. (At the point, (S, (cid:96)) is deleted from T , and s , (cid:96) − 1) are added.) Observe that the root of T all the (N + is (V, k), and the leaves are precisely the final output S. between C(cid:96)(S) and (cid:83) a γ where k-clique shadow, γ-saturated Theorem 5.2. The output S of Shadow-Finder(G, k) = is 1/ max(S,(cid:96))∈S(f ((cid:96))S2). Proof. We first prove by induction the following loop invariant for Shadow-Finder: T ∪ S is always a k-clique shadow. For the base case, note that at the beginning, T = {(V, k)} and S = ∅. For the induction step, assume that T ∪S is a k-clique shadow at the beginning of some iteration. s , (cid:96) − 1) is The element (S, (cid:96)) is deleted from T . Each (N + added to S or to T . s ). s∈S C(cid:96)−1(N + Thus, it suffices to prove that there is a bijection mapping (By the induction hypothesis, we can then construct a bijection between Ck(G) and the appropriate cliques in T ∪ S.) Consider an (cid:96)-clique K in S. Set s to be the minimum vertex according to the degeneracy ordering in D(GS). Observe that the remaining vertices form an ((cid:96)−1)-clique in N + s , which we map the K to. This is a bijection, because every clique K can be mapped to a (unique) ((cid:96)−1)-clique, and furthermore, every ((cid:96)−1)-clique Thus, when Shadow-Finder terminates, T ∪S is a k-clique shadow. Since T must be empty, S is a k-clique shadow. Furthermore, a pair (S, (cid:96)) is in S iff ρ2(S) > 1 − 1 (cid:96)−1 . By Theorem 4.4, S is 1/ max(S,(cid:96))∈S(f ((cid:96))S2)-saturated. s ) is in the image of this mapping. in(cid:83) s∈S C(cid:96)−1(N + We have a simple, but important claim that bounds the size of any set in the shadow by the degeneracy. Claim 5.3. Consider non-root (S, (cid:96)) ∈ T . Then S ≤ α(G). Proof. Suppose the parent of (S, (cid:96)) is (P, (cid:96)+1). Observe that S is the outneighborhood of some node p in the DAG D(GP ). Thus, S ≤ α(GP ). The degeneracy can never be larger in a subgraph. (This is apparent by an alternate definition of degeneracy, the maximum smallest degree of an induced subgraph [26].) Hence, α(GP ) ≤ α(G). time of Shadow-Finder(G, k) is O(α(G)size(S) + m + n). The total storage is O(size(S) + m + n). Theorem 5.4. The running (cid:88) (cid:88) s Proof. Every time we add (N + S , (cid:96) − 1) (Step 8) to T , we explicitly construct the graph GN + . Thus, we can guarantee that for every (S, (cid:96)) present in T , we can make queries in the graph GS. This construction takes O(S2) time, to query every pair in S. (This is not required when S = V , since GV = G.) Furthermore, this construction is done for every (S, (cid:96)) ∈ T , except for the root node in T . Once we have GS, the degeneracy order can be computed in time linear in the number of edges in GS [26]. (S,(cid:96))∈T :S(cid:54)=V S2 = O(α(G)(cid:80) be by (S,(cid:96))∈T :S(cid:54)=V S2 + m + n). By Claim 5.3, we can (S,(cid:96))∈T S). We split the sum over leaves and non-leaves. The sum over leaves is precisely a sum over the sets in S, so It suffices to prove that that yields O(α(G)size(S)). (S,(cid:96))∈T :S non-leaf S = O(size(S)), which we show next. Observe that a non-leaf node (S, (cid:96)) in T has exactly S O((cid:80) bound (cid:80) (cid:80) bounded running Thus, time can children, one for each vertex s ∈ S. Thus, the S = # children of (S, (cid:96)) (S,(cid:96))∈T :(S,(cid:96))non-leaf (S,(cid:96))∈T = # edges in T All internal nodes in T have at least 2 children, so the number of edges in T is at most twice the number of leaves in T . But this is exactly the number of sets in the output S, which is at most size(S). (S,(cid:96))∈T S + m + n), which is The total storage is O((cid:80) O(size(S) + m + n) by the above arguments. We now formally define the Tur´an shadow to be output of this procedure. Definition 5.5. The k-clique Tur´an shadow of G is the output of Shadow-Finder(G, k). 5.1 Putting it all together Algorithm 3: Tur´an-shadow(G, k, ε, δ) 1 Compute S = Shadow-Finder(G, k); 2 Set γ = 1/ max(S,(cid:96))∈S(f ((cid:96))S2); 3 Output Ck = sample(G, k, γ, ε, δ); Theorem 5.6. Consider graph G = (V, E) with m edges, n vertices, and degeneracy α(G). Assume m ≤ n2/4. Let S be the Tur´an k-clique shadow of G. With probability at least 1 − δ (this probability is over the randomness of Tur´an-shadow; there is no stochastic assumption on G), Ck − Ck(G) ≤ εCk(G). The running time of Tur´an-shadow is O(α(G)size(S) + f (k)m log(1/δ)/ε2 + n) and the total storage is O(size(S) + m + n). Proof. By Theorem 5.2, S is γ-saturated, for γ = 1/ max(S,(cid:96))∈S f ((cid:96))S2. Since m ≤ n2/4, the procedure Shadow-Finder(G, k) cannot just output {(V, k)}. All leaves in the recursion tree must have depth at least 2, and by Claim 5.3, for all (S, (cid:96)) ∈ S, S ≤ α(G). A classic bound on the degeneracy asserts that α(G) ≤ √ 2m (Lemma 1 of [12]). Since f ((cid:96)) is increasing in (cid:96), max(S,(cid:96))∈S f ((cid:96))S2 ≤ 2f (k)m. Thus, γ = Ω(1/(f (k)m)). By Theorem 4.3, the running time of sample is O(size(S)+ log(1/δ)/(γε2)), which is O(size(S) + f (k)m log(1/δ)/ε2). Theorem 4.3 also asserts the accuracy of the output. Adding the bounds of Theorem 5.4, we prove the running time and storage bounds. 5.2 The shadow size The practicality of Tur´an-shadow hinges on size(S) It is not hard to prove a worst-case bound, being small. using the degeneracy. Claim 5.7. size(S) = O(nα(G)k−2). Proof. By arguments in the proof of Theorem 5.4, we can show that size(S) is at most the number of edges in T . In T , the degree of the root is n, and by Claim 5.3, the degree of all other nodes is at most α(G). The depth of the tree is at most k − 1, since the value of (cid:96) decreases every step down the tree. That proves that nαk−2 bound. is not that This bound interesting, the Chiba-Nishizeki algorithm for exact clique enumeration matches this bound [12]. Indeed, we can design instances where Claim 5.7 is tight (a set of n/α Erdos-R´enyi graphs Gα,1/3). In any case, beating an exponential dependence on k for any algorithm is unlikely [11]. and The key empirical insight of this paper is that Tur´an clique shadows are small for real-world graphs. We explain in more detail in the next section; Fig. 3 shows that the shadow sizes are typically less than m, and never more than 10m. 6. EXPERIMENTAL RESULTS Preliminaries: We implemented our algorithms in C++ and ran our experiments on a commodity machine equipped with a 3.00GHz Intel Core i7 processor with 8 cores and 256KB L2 cache (per core), 20MB L3 cache, and 128GB memory. We performed our experiments on a collection of graphs from SNAP [49], the largest with more than 100M edges. The collection includes social networks, web networks, and infrastructure networks. Each graph is made simple by ignoring direction. Basic properties of these graphs are presented in Tab. 1. In the implementation of Tur´an-shadow, there is just one parameter to choose: the number of samples chosen in Step 2 in sample. Theoretically, it is set to (20/γε2) log(1/δ); in practice, we just set it to 50K for all our runs. Note that γ is not a free parameter and is automatically set in Step 1 of Tur´an-shadow. We focus on counting k-cliques for k ranging from 5 to 10. We ignore k = 3, 4, since there is much existing (scalable) work for this setting [35, 25, 2]. For the sake of presentation, we showcase results for k = 7, 10. We focus on k = 10 since no existing algorithm produces results for 10-cliques in reasonable time. We also show specifics for k = 7, to contrast with k = 10. Convergence of Tur´an-shadow: We picked two smaller graphs amazon0601 and web-Google for which the exact k-clique count is known (for all k ∈ [5, 10]). We choose both k = 7, 10. for sample size in [10K,50K,100K,500K,1M], we perform 100 runs of the algorithm. We plot the spread of the output of Tur´an-shadow, over all these runs. The results are shown in Fig. 2. The red line denotes the true answer, and there is For each graph, Table 1: Graph properties k=5 k=7 k=10 vertices 1.97E+05 2.82E+05 4.03E+05 1.13E+06 8.76E+05 6.85E+05 1.70E+06 3.77E+06 1.63E+06 4.00E+06 3.07E+06 edges 9.50E+05 1.99E+06 4.89E+06 2.99E+06 4.32E+06 6.65E+06 1.11E+07 1.65E+07 2.23E+07 3.47E+07 1.17E+08 51 71 10 51 44 201 111 64 47 360 253 graph time loc-gowalla 3 web-Stanford 52 amazon0601 1 com-youtube 8 web-Google 2 web-BerkStan 262 as-skitter 180 cit-Patents 9 soc-pokec 45 com-lj 252 com-orkut 9298 Table 2: Table shows the sizes, degeneracy, maximum degree of the graphs, the counts of 5, 7 and 10 cliques obtained using Tur´an-shadow, the percent relative error in the estimates, and time in seconds required to get the estimates. Some of the exact counts were obtained from [18] (where available). This is the first such algorithm that obtains these counts with < 2% error without using any specialized hardware. degen max degree 14730 38625 2752 28754 6332 84230 35455 793 14854 14815 33313 estimate % error 0.20 1.46E+07 0.00 6.21E+08 0.93 3.64E+06 1.08 7.29E+06 0.10 1.05E+08 2.19E+10 0.00 0.01 1.17E+09 0.34 3.05E+06 0.13 5.29E+07 - 2.46E+11 1.57E+10 0.00 estimate % error 0.36 4.78E+07 0.13 3.47E+10 0.95 9.98E+05 1.38 7.85E+06 0.09 6.06E+08 9.30E+12 1.05 0.23 7.30E+10 0.83 1.89E+06 0.48 8.43E+07 - 4.48E+14 3.61E+11 1.97 estimate % error 1.63 1.08E+08 - 5.82E+12 0.01 9.77E+03 0.20 1.83E+06 0.82 1.29E+10 5.79E+16 - - 1.42E+13 4.46 2.55E+03 0.01 1.98E+08 - 1.47E+19 3.03E+13 - time 2 20 1 7 2 101 153 10 42 106 3119 time 2 43 1 8 2 214 164 9 45 153 5587 a point for the output of every single run. Even for 10-clique counting, the spread of 100 runs is absolutely minimal. For 50K samples, the range of values is within 2% of the true answer. This was consistent with all our runs. Accuracy of Tur´an-shadow: For many graphs (and values of k), it was not feasible to get an exact algorithm to run in reasonable time. The run time of exact procedures can vary wildly, so we have exact numbers for some larger graphs but could not generate numbers for smaller graphs. We collected as many exact results as possible to validate Tur´an-shadow. For the sake of presentation, we only show a snapshot of these results here. For k = 7, we collected exact results for a collection of graphs, and for each graph, compared the output of a single run of Tur´an-shadow (with 50K samples) with the true answer. We compute relative error : true - estimate/true. These results are presented in Fig. 1i. Note that the errors are within 2% in all cases, again consistent with all our runs. In Tab. 1, we present the output of our algorithm for a single run on all instances and k = 5, 7, 10. For every graph where we know the true value, we present the relative error. Barring one example (cit-Patents for k = 10), all errors are less than 2%. Even in the worst case, the error is at most 5%. Running time: All runtimes are presented in Tab. 1. (We show the time for a single run, since there was little variance for different runs on the same graph.) In all instances except com-orkut, the runtime was a few minutes, even for graphs with tens of millions of edges. We stress that these are all on a single machine. For com-orkut, the runtime is at most 2.5 hours. Previously, such graphs were processed with MapReduce on clusters [18]. 6.1 Comparison with other algorithms This for cliques. Our exact brute-force procedure is a well-tuned algorithm that uses the degeneracy ordering and exhaustively searches outneighborhoods is basically the procedure of Finetti et al. [18], inspired by the algorithm of Chiba-Nishizeki [12]. We compare with the following algorithms. • Color coding: This is a classic algorithmic technique [4]. For counting k-cliques, the algorithm randomly colors vertices with one of k colors. Then, the algorithm uses a brute-force procedure to count polychromatic k-cliques (where each vertex has a different color). This number is scaled to give an unbiased estimate, and the coloring helps cut down the search time of the brute-force procedure. This method has been applied in practice for numerous pattern counting problems [22, 7, 48]. • Edge sampling: Edge sampling was discussed by Tsourakakis et al. in the context of triangle counting [42, 39, 40], though the idea is flexible and can be used for large patterns [14]. The idea here is to sample each edge independently with some probability p, and then count k-cliques in the down-sampled graph. This number is scaled to give an unbiased estimate for the number of k-cliques. For clique counting, we observe that minor differences in p (by 0.1) have huge effects on runtime and accuracy. To do a fair comparison, we run multiple experiments with varying p (increments of 0.1), until we reach the smallest p that consistently yields less than 5% error. (Note that the error of Tur´an-shadow is significantly smaller that this.) Timing comparisons are done with runs for that value of p. • GRAFT [30]: Rahman et al. give a variant of edge sampling with better performance for large pattern counts [30]. The idea is to sample some set of edges, and exactly count the number of k-cliques on each of these edges. This can be scaled into an unbiased estimate for the total number of k-cliques. As with edge sampling, we increase the number of edge samples until we get consistently within 5% error. Timing comparisons are done with this setting. Typical settings seems to be in the range of 100K to 1M samples. Beyond that, GRAFT is infeasible, even for graphs with 10M edges. In all cases, We focus on k = 7, 10 for clarity. takes more we simply terminate the algorithm if than the minimum of 7 hours and 100 times the time required by Tur´an-shadow. We present the speedup of Tur´an-shadow with respect to all these algorithms in Fig. 1iii for k=7. For k=10, for most instances, no competing algorithm terminated. • k = 7 (Fig. 1iii): Tur´an-shadow outperformed Color Coding and GRAFT across all instances. Color Coding never gave good accuracy, so we ignore it in our speedup plots. We do note that Edge Sampling gives extremely good performance in some instances, but can be very slow in others. For amazon0601, com-youtube, cit-Patents, and soc-pokec, Edge Sampling is faster than Tur´an-shadow. it Figure 2: Figure shows convergence over 100 runs of Tur´an-shadow using 10K, 50K, 100K, 500K and 1M samples each. Tur´an-shadow has an extremely low spread and consistently gives very accurate results. Figure 3: Figures show the sizes of the Tur´an shadows generated for k=7 and k=10 in all the graphs. The runtime of the algorithm is proportional to the size of the shadow and crucially, the sizes scale only linearly with the number of edges. Figure 5: Figures show the trends in clique counts of some graphs. While cit-Patents, com-youtube and amazon0601 show a decreasing trend, all other graphs show an exponential increase in the number of cliques with clique size. Figure 4: Figures show the success ratio (probability of finding a clique) obtained in the sampling experiments in all the graphs. But Tur´an-shadow handles all these graphs with a minute. The only exception is com-orkut, where GRAFT is much faster than Tur´an-shadow. We note that all other algorithms can perform extremely poorly on fairly small graphs: Edge Sampling is 10-100 times slower on a number of graphs, which have only millions of edges. On the other hand, Tur´an-shadow always runs in minutes for these graphs. • k = 10 : No competing algorithm is able to handle 10 cliques for all datasets, even in 7 hours (giving a speedup of anywhere between 3x to 100x). They all generally fail for at least half of the instances. Tur´an-shadow gets an answer for com-orkut within 2.5 hours, and handles all other graphs in minutes. 6.2 Details about Tur´an-shadow Shadow size: In Fig. 3, we plot the size of the k-clique Tur´an shadow with respect to the number of edges in each instance. This is done for k = 7, 10. (The line y = x is drawn Success The final probability: as well.) As seen from Theorem 5.6, the size of the shadow controls the storage and runtime of Tur´an-shadow. We see how in almost all instances, the shadow size is around the number of edges. This empirically explains the efficiency of Tur´an-shadow. The worst case is com-orkut, where the shadow size is at most ten times the number of edges. estimate of Tur´an-shadow is generated through sample. We asserted (theoretically) that O(m) samples suffice, and in practice, we use 50K samples. In Fig. 4, we plot (for k = 7, 10) the empirical probability of finding a clique in Step 6 of sample. The higher this is, the fewer samples we require and the more confidence in the statistical validity of our estimate. Almost all (empirical) probabilities are more than 0.1, and 50K samples are more than enough for convergence. Trends in clique numbers: Fig. 5 plots the number of k-cliques (as computed by Tur´an-shadow) versus k. (We do not consider all graphs for the sake of clarity.) Interestingly, there are some graphs where the number of cliques grows exponentially. This is probably because of a large clique/dense-subgraph, and it would be interesting to verify this. For another class of graphs, the clique counts are consistently decreasing. This seems to classify graphs into one of two types. We feel further analysis of these trends would be interesting, and Tur´an-shadow can be a useful tool for network analysis. 104105106Numberofsamples0.00.51.0Cliques×106amazon0601,k=7104105106Numberofsamples0.00.51.0Cliques×104amazon0601,k=10104105106Numberofsamples0.00.5Cliques×109web-Google,k=7104105106Numberofsamples0.00.51.01.5Cliques×1010web-Google,k=101051061071081091010Numberofedges1051061071081091010ShadowsizeShadowsize,k=71041051061071081091010Numberofedges1041051061071081091010ShadowsizeShadowsize,k=10loc-gowweb-StanamazonyoutubeGoogleBerkStanas-skitterPatentssoc-pokeccom-ljcom-orkutNumberofedges10−310−210−1100SuccessratioSuccessratio,k=7loc-gowweb-StanamazonyoutubeGoogleBerkStanas-skitterPatentssoc-pokeccom-ljcom-orkutNumberofedges10−310−210−1100SuccessratioSuccessratio,k=105678910Cliquesize100102104106108101010121014101610181020CliquesTrendsincliquecountsamazon0601com-youtubeweb-BerkStanas-skittercit-Patentscom-ljcom-orkut7. REFERENCES [25] M. Jha, C. Seshadhri, and A. Pinar. Path sampling: A fast and provable method for estimating 4-vertex subgraph counts. In World Wide Web (WWW), pages 495 -- 505, 2015. [1] F. N. Afrati, D. Fotakis, and J. D. Ullman. Enumerating [26] D. W. Matula and L. L. Beck. Smallest-last ordering and subgraph instances using map-reduce. In International Conference on Data Engineering (ICDE), pages 62 -- 73, 2013. clustering and graph coloring algorithms. Journal of the ACM (JACM), 30(3):417 -- 427, 1983. [2] N. K. Ahmed, J. Neville, R. A. Rossi, and N. Duffield. Efficient [27] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, graphlet counting for large networks. In Proceedings of International Conference on Data Mining (ICDM), 2015. and U. Alon. Network motifs: Simple building blocks of complex networks. Science, 298(5594):824 -- 827, 2002. [3] E. A. Akkoyunlu. The enumeration of maximal cliques of large [28] M. Mitzenmacher, J. Pachocki, R. Peng, C. E. Tsourakakis, graphs. SIAM J. Comput., 2:1 -- 6, 1973. [4] N. Alon, R. Yuster, and U. Zwick. Color-coding: A new method for finding simple paths, cycles and other small subgraphs within large graphs. In Symposium on the Theory of Computing (STOC), pages 326 -- 335, 1994. [5] R. Andersen and K. Chellapilla. Finding dense subgraphs with size bounds. In Workshop on Algorithms and Models for the Web-Graph (WAW), pages 25 -- 37, 2009. [6] L. Becchetti, P. Boldi, C. Castillo, and A. Gionis. Efficient semi-streaming algorithms for local triangle counting in massive graphs. In KDD'08, pages 16 -- 24, 2008. [7] N. Betzler, R. van Bevern, M. R. Fellows, C. Komusiewicz, and R. Niedermeier. Parameterized algorithmics for finding connected motifs in biological networks. IEEE/ACM Trans. Comput. Biology Bioinform., 8(5):1296 -- 1308, 2011. [8] M. Bhuiyan, M. Rahman, M. Rahman, and M. A. Hasan. Guise: Uniform sampling of graphlets for large graph analysis. In Proceedings of International Conference on Data Mining, pages 91 -- 100, 2012. [9] C. Bron and J. Kerbosch. Algorithm 457: Finding all cliques of an undirected graph. Commun. ACM, 16(9):575 -- 577, Sept. 1973. [10] R. Burt. Structural holes and good ideas. American Journal of Sociology, 110(2):349 -- 399, 2004. [11] J. Chen, X. Huang, I. A. Kanj, and G. Xia. Linear FPT reductions and computational lower bounds. In L. Babai, editor, Symposium on the Theory of Computing (STOC), pages 212 -- 221. ACM, 2004. [12] N. Chiba and T. Nishizeki. Arboricity and subgraph listing algorithms. SIAM J. Comput., 14:210 -- 223, 1985. [13] D. Dubhashi and A. Panconesi. Concentration of Measure for the Analysis of Randomized Algorithms. Cambridge University Press, 2009. [14] E. R. Elenberg, K. Shanmugam, M. Borokhovich, and A. G. Dimakis. Distributed estimation of graph 4-profiles. In J. Bourdeau, J. Hendler, R. Nkambou, I. Horrocks, and B. Y. Zhao, editors, World Wide Web (WWW), pages 483 -- 493. ACM, 2016. [15] D. Eppstein and D. Strash. Listing all maximal cliques in large sparse real-world graphs. In P. M. Pardalos and S. Rebennack, editors, Symposium of Experimental Algorithms, volume 6630 of Lecture Notes in Computer Science, pages 364 -- 375. Springer, 2011. [16] P. Erdos. On the number of complete subgraphs and circuits contained in graphs. Casopis Pest. Mat., 94:290 -- 296, 1969. [17] K. Faust. A puzzle concerning triads in social networks: Graph constraints and the triad census. Social Networks, 32(3):221 -- 233, 2010. [18] I. Finocchi, M. Finocchi, and E. G. Fusco. Clique counting in mapreduce: Algorithms and experiments. ACM Journal of Experimental Algorithmics, 20, 2015. [19] C. Giatsidis, F. Malliaros, D. M. Thilikos, and M. Vazirgiannis. Corecluster: A degeneracy based graph clustering framework. In IAAA: Innovative Applications of Artificial Intelligence, 2014. [20] R. A. Hanneman and M. Riddle. Introduction to social network methods. University of California, Riverside, 2005. http://faculty.ucr.edu/hanneman/nettext/. [21] P. Holland and S. Leinhardt. A method for detecting structure in sociometric data. American Journal of Sociology, 76:492 -- 513, 1970. [22] F. Hormozdiari, P. Berenbrink, N. Przulj, and S. C. Sahinalp. Not all scale-free networks are born equal: The role of the seed graph in ppi network evolution. PLoS Computational Biology, 118, 2007. [23] M. O. Jackson. Social and Economic Networks. Princeton University Press, 2010. and S. C. Xu. Scalable large near-clique detection in large-scale networks via sampling. In SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 815 -- 824, 2015. [29] N. Przulj, D. G. Corneil, and I. Jurisica. Modeling interactome: scale-free or geometric?. Bioinformatics, 20(18):3508 -- 3515, 2004. [30] M. Rahman, M. A. Bhuiyan, and M. A. Hasan. Graft: An efficient graphlet counting method for large graph analysis. IEEE Transactions on Knowledge and Data Engineering, PP(99), 2014. [31] R. A. Rossi, D. F. Gleich, and A. H. Gebremedhin. Parallel maximum clique algorithms with applications to network analysis. SIAM Journal on Scientific Computing, 37(5):C589 -- C616, 2015. [32] A. E. Sariyuce, C. Seshadhri, A. Pinar, and U. V. ¸Catalyurek. Finding the hierarchy of dense subgraphs using nucleus decompositions. pages 927 -- 937, 2015. [33] S. B. Seidman. Network structure and minimum degree. Social Networks, 5(3):269 -- 287, 1983. [34] C. Seshadhri, T. G. Kolda, and A. Pinar. Community structure and scale-free collections of Erdos-R´enyi graphs. Physical Review E, 85(5):056109, May 2012. [35] C. Seshadhri, A. Pinar, and T. G. Kolda. Fast triangle counting through wedge sampling. In Proceedings of the SIAM Conference on Data Mining, 2013. [36] A. Sizemore, C. Giusti, and D. S. Bassett. Classification of weighted networks through mesoscale homological features. Journal of Complex Networks, 10.1093, 2016. [37] E. Tomita, A. Tanaka, and H. Takahashi. The Worst-Case Time Complexity for Generating All Maximal Cliques, pages 161 -- 170. Springer Berlin Heidelberg, Berlin, Heidelberg, 2004. [38] C. Tsourakakis, F. Bonchi, A. Gionis, F. Gullo, and M. Tsiarli. Denser than the densest subgraph: Extracting optimal quasi-cliques with quality guarantees. In Knowledge Data and Discovery (KDD), 2013. [39] C. Tsourakakis, P. Drineas, E. Michelakis, I. Koutis, and C. Faloutsos. Spectral counting of triangles in power-law networks via element-wise sparsification. In ASONAM'09, pages 66 -- 71, 2009. [40] C. Tsourakakis, M. N. Kolountzakis, and G. Miller. Triangle sparsifiers. J. Graph Algorithms and Applications, 15:703 -- 726, 2011. [41] C. E. Tsourakakis. The k-clique densest subgraph problem. In Proceedings of the Conference on World Wide Web WWW, pages 1122 -- 1132, 2015. [42] C. E. Tsourakakis, U. Kang, G. L. Miller, and C. Faloutsos. Doulion: counting triangles in massive graphs with a coin. In Knowledge Data and Discovery (KDD), pages 837 -- 846, 2009. [43] C. E. Tsourakakis, J. W. Pachocki, and M. Mitzenmacher. Scalable motif-aware graph clustering. CoRR, abs/1606.06235, 2016. [44] P. Tur´an. On an extremal problem in graph theory. Mat. Fiz. Lapok, 48(436-452):137, 1941. [45] V. Vassilevska. Efficient algorithms for clique problems. Information Processing Letters, 109(4):254 -- 257, 2009. [46] D. Watts and S. Strogatz. Collective dynamics of 'small-world' networks. Nature, 393:440 -- 442, 1998. [47] B. Welles, A. Van Devender, and N. Contractor. Is a friend a friend?: Investigating the structure of friendship networks in virtual worlds. In CHI-EA'10, pages 4027 -- 4032, 2010. [48] Z. Zhao, G. Wang, A. Butt, M. Khan, V. S. A. Kumar, and M. Marathe. Sahad: Subgraph analysis in massive networks using hadoop. In Proceedings of International Parallel and Distributed Processing Symposium (IPDPS), pages 390 -- 401, 2012. [49] Stanford Network Analysis Project (SNAP). Available at [24] M. O. Jackson, T. Rodriguez-Barraquer, and X. Tan. Social http://snap.stanford.edu/. capital and social quilts: Network patterns of favor exchange. American Economic Review, 102(5):1857?1897, 2012.
1609.07580	2	1609	2016-10-03T17:24:05	Classification of infrastructure networks by neighborhood degree distribution	[ "cs.SI", "physics.soc-ph" ]	A common way of classifying network connectivity is the association of the nodal degree distribution to specific probability distribution models. During the last decades, researchers classified many networks using the Poisson or Pareto distributions. Urban infrastructures, like transportation (railways, roads, etc.) and distribution (gas, water, energy, etc.) systems, are peculiar networks strongly constrained by spatial characteristics of the environment where they are constructed. Consequently, the nodal degree of such networks spans very small ranges not allowing a reliable classification using the nodal degree distribution. In order to overcome this problem, we here (i) define the neighborhood degree, equal to the sum of the nodal degrees of the nearest topological neighbors, the adjacent nodes and (ii) propose to use neighborhood degree to classify infrastructure networks. Such neighborhood degree spans a wider range of degrees than the standard one allowing inferring the probabilistic model in a more reliable way, from a statistical standpoint. In order to test our proposal, we here analyze twenty-two real water distribution networks, built in different environments, demonstrating that the Poisson distribution generally models very well their neighborhood degree distributions. This result seems consistent with the less reliable classification achievable with the scarce information using the standard nodal degree distribution.	cs.SI	cs	Classification of infrastructure networks by neighborhood degree distribution 1Orazio Giustolisi, 1Antonietta Simone, 2 Luca Ridolfi 1Politecnico di Bari, Bari, Italy. [email protected] - [email protected] 2Politecnico di Torino, Torino, Italy. [email protected] Abstract A common way of classifying network connectivity is the association of the nodal degree distribution to specific probability distribution models. During the last decades, researchers classified many networks using the Poisson or Pareto distributions. Urban infrastructures – like transportation (railways, roads, etc.) and distribution (gas, water, energy, etc.) systems – are peculiar networks strongly constrained by spatial characteristics of the environment where they are constructed. Consequently, the nodal degree of such networks spans very small ranges not allowing a reliable classification using the nodal degree distribution. In order to overcome this problem, we here (i) define the "neighborhood" degree, equal to the sum of the nodal degrees of the nearest topological neighbors, the adjacent nodes and (ii) propose to use "neighborhood" degree to classify infrastructure networks. Such "neighborhood" degree spans a wider range of degrees than the "standard" one allowing inferring the probabilistic model in a more reliable way, from a statistical standpoint. In order to test our proposal, we here analyze twenty-two real water distribution networks, built in different environments, demonstrating that the Poisson distribution generally models very well their "neighborhood" degree distributions. This result seems consistent with the less reliable classification achievable with the scarce information using the "standard" nodal degree distribution. 1. INTRODUCTION The complex network theory (CNT) is becoming one of the most powerful and versatile tool to investigate, describe, and understand the world (Barabasi [1]). Networks allow studying and interpreting a huge number of physical, biological and social processes. The examples range from social relationships to neural connections, from multi-species ecological interactions to financial and economic exchanges. Although each network exhibits its own topological and 1 structural peculiarities, apparently very different networks share amazing similar features (Albert and Barabasi [2], Buchanan [3]). In the last decades, CNT had an unrestrained development and researchers proposed novel approaches, metrics and theories to explore and disentangle network features (e.g., Boccaletti et al. [4]; Newman [5]). Nevertheless, the nodal degree distribution remains a relevant concept to classify networks. The nodal degree distribution describes the probability distribution of the number of edges incident each node of the network. Several network features are associated to the shape of such distribution. Erdos and Rény [6] [7] were the first to study the nodal degree distribution of real networks introducing the random networks against regular networks. In the former case, the nodal degree distribution is random around an average value and the network is characterized by a high homogeneity; differently, regular networks have a constant degree of internal nodes and are characterized by an absolute homogeneity. The Poisson model is often used to describe nodal degree distribution of random networks, which are able to capture some features of real networks better than regular networks: e.g., being highly ordered, the shortest paths linking two nodes of regular networks are too large, while random networks show lower and more realistic values (e.g., Milgram [8]). Later on, starting from regular networks, Watts and Strogatz [9] introduced the small world networks. They demonstrated the existence of the so-called small world effect, i.e. a behavior between the regular and random networks. The model by Watts and Strogatz [9] covers from regular networks to random networks using the Poisson distribution of nodal degrees and a probability of connections between two nodes: for small world networks, that probability is greater than zero of regular networks but lower than values typical of random networks. In the same period, Barabasi and Albert [10] proposed the scale-free networks in order to describe real networks characterized by non-homogeneous nodal degree distributions, where many nodes have a low degree and few nodes (called hubs) have a high degree. These distribution typically exhibits a Pareto (or power law) behavior. The interest in classifying networks according to connectivity structures is related to capturing the emerging behaviors of real systems (e.g., Albert and Barabasi [2], Newman [5] [11]) For instance, the nodal degree distribution affects the network vulnerability to random failures and intentional threats. In fact, regular small world networks and random networks present a significant structural resistance to both random failures and intentional threats, while scale free networks show a very high structural resistance to random failures but a weak resistance to 2 intentional threats (Albert et al. [12]). In this sense, the classification of real networks by associating their degree distribution to the Poisson or Pareto models is useful to assess network vulnerability. In spite of the nodal degree distribution has a crucial role to assess the network behavior, for a relevant class of real systems this information is poorly informative; they are the so-called infrastructure networks. Remarkable examples are transportation networks (railways, roads, etc.), distribution networks (energy, water, gas, etc.), and urban street patterns. These networks are one of the most widespread and important examples of spatial networks, and they are characterized to be (i) located on a two-dimensional space and (ii) constrained by a number of spatial impediments (Barthelemy, [13]). This entails that the maximum nodal degree is generally very low, consequently the nodal degree distribution spans over a very limited range and, therefore, statistical inference results unreliable. It follows a quite paradoxical situation: on one hand, many theoretical and applicative works underline the key information embedded in the degree distribution; on the other hand, infrastructure networks have characteristics that make standard nodal degree distribution evaluation very elusive and difficult to interpret. In order to overcome this problem, we introduce the neighborhood nodal degree spanning over a range of values wider than the standard nodal degree, resulting in a statistically more reliable classification of infrastructure networks (hereinafter we will specify as standard the usual nodal degree when a possible misunderstanding with neighborhood nodal degree can occur). The neighborhood nodal degree corresponds, for each node, to the sum of the degrees of nearest topological neighbors, namely of the adjacent nodes. We will demonstrate that the approach based on the neighborhood nodal degree makes more reliable the association of the connectivity structure of an infrastructure network to a Poisson or a Pareto distribution. We will show that the neighborhood nodal degree can be interpreted as a weighting of the network edges incident to each node by the degree of the nearest neighbors or adjacent nodes. In order to show the effectiveness of the neighborhood nodal degree to classify spatial networks, we will focus on water distribution networks (WDNs). This is due to several reasons. Firstly, WDNs are planar urban infrastructure networks, strongly constrained by external geometrical/environmental factors – like the landscape topography, the structure of the cities (e.g., street network, building size, etc.) – decreasing the connectivity probability with the distance between two nodes. It follows that WDNs classification by the standard nodal degree distribution is very difficult and uncertain (Yazdani and Jeffrey [14]) because the very low 3 values of the maximum standard nodal degree characterizing those networks (generally around five and quite always lower than ten). Secondly, and somewhat surprisingly, CNT literature reports very few investigations on WDNs in spite of their ubiquity, topological variety, and importance in everyday life. Thirdly, the availability of real WDNs is quite rare due to the sensible data associated with water supply service. Thus, the unusual availability of detailed topological data of twenty-two real WDNs built in different urban environments, allows a robust and well representative databased investigation about the promising use of neighborhood nodal degree to classify real infrastructure networks. In the following, we will demonstrate that the network connectivity structure of WDNs generally follows the Poisson distribution. The result is consistent with the fact that technicians generally design the network connectivity structure of WDNs using a criterion of redundancy beyond the hydraulic capacity requirements. This fact provides a low vulnerability of the connectivity structure with respect to any kind of failure event. 2. DEGREE DISTRIBUTION MODELS The Empirical nodal degree distribution P(k) is defined as the fraction of nodes in the network having degree k. Hence (1) where nk is the number of nodes having degree k and n is the total number of nodes in the network. The formulation of the Poisson distribution, approximating the binomial one, for nodal degrees (Watts and Strogatz [9]) is (2) where p is the probability of connection between two nodes and <k> is the average nodal degree of the network. The Poisson distribution models random networks and it exhibits a peak value at <k> and a high probability of nodal degree around <k>, i.e. a high homogeneity of the degree distribution. The nodal degree distribution of small world networks has essentially the same features as the random networks (Barthelemy [13]). Therefore, the Poisson distribution can 4 knPkn111!kknkknekPkpppkapply to small world networks arguing a narrow distribution of degrees, i.e. a very high homogeneity of the nodal degree distribution, which degenerates to a single degree (considering only internal nodes) for the case of regular networks, characterized by an absolute homogeneity. In fact, Watts-Strogatz [9] introduced the small world networks starting from a ring lattice regular network, increasing the probability of connection, p. In the present study, we will show that a Poisson distribution models very well the neighborhood nodal degree distributions of WDNs allowing arguing that it also models the standard nodal degree. It is to note that Watts-Strogatz [9] assumed the probability p, determining the average degree of the network equal to (3) In the case of the infrastructure networks, the average nodal degree strongly depends on spatial constraints and, generally, it is very low. E.g., in the case of WDNs, <k> ranges from 2 to 3. Hence, we can write (4) Namely, the probability of connection results inversely proportional to the size of the network because of the spatial constraints strongly limiting the range of variability for <k>. Consequently, the Watts-Strogatz [9] scheme has to be carefully used with respect to the probability p, because Eq. (4) shows that p is generally very low for the connectivity structure typical of infrastructure networks and it decreases with the network size. Finally, the formulation of the Pareto (or power law) model for the nodal degree distribution is (5) where γ is a constant generally ranging from 1.5 to 3 (Newman [5]; Barthelemy [13]). The model of Eq. (5) applies to the scale free networks, which are non-homogeneous with respect to the nodal degree distribution. In other words, few nodes with a large degree (hubs) characterize scale free networks. 5 1kpn12.511kppnnnPkk3. NEIGHBORHOOD NODAL DEGREE DISTRIBUTION The starting point of our reasoning is that a reliable classification of spatial networks by means the structure of the standard degree distribution (e.g., if it exhibits a Pareto or Poisson law behaviour) is generally hampered by the very low maximum nodal degree characterizing this class of networks. This fact is a consequence of the spatial limits constraining the network topology at local scale (i.e., at the node scale) (Barthelemy [13]; Lämmer et al., [15]). E.g., the topographic structure of a city typically involves the crossing of three-four roads and rarely the nodal degree of a node-square exceeds six-eight. It follows that infrastructure networks inheriting these urban constrains (e.g., water and gas distribution networks) exhibit a very narrow range of nodal degrees, which makes very difficult to model the degree distribution in a statistically meaningful way. In order to classify infrastructure networks still using the degree distribution concept, we here introduce a more suitable idea: the neighborhood nodal degree. I.e., we define a degree distribution involving the nearest neighbors or adjacent nodes. The neighborhood degree distribution of each node is the sum of the standard degrees of the topologically nearest (i.e., adjacent) nodes. We propose the neighborhood nodal degree distribution to classify infrastructure networks. The approach allows significantly increasing the maximum nodal degree and, consequently, the range of values spanned by neighborhood nodal degrees. This fact makes much more reliable the classification, because a greater number of points supports the robust identification of a specific statistical distribution. The formulation of the neighborhood degree is (6) where kn(i) is the "neighborhood" degree (involving adjacent nodes) of the i-th node, Aij are the elements of the adjacency matrix, k(j) is the standard degree of the j-th node, and N(i) is the topological neighborhood of i-th node, i.e. the set of adjacent nodes. Therefore, kn(j) is the product between the standard nodal degree and the i-th row of the adjacency matrix providing a non-null value for the nearest/adjacent nodes only. Figure 1 shows an illustrative example of the neighbourhood degree evaluation. Consider the portion of a network reported in the Figure 1 and focus on the node 1, Figure 1(b). Its neighbourhood is constituted by nodes labelled {2,3,4}, whose nodal degrees are equal to 6 nijjNikiAkjk(2)=4, k(3)=3, and k(4)=3, respectively. In this case, the neighbourhood degree of node 1 results equal to kn(1)=4+3+3=10, i.e. the sum of the degree of the three adjacent nodes. Similarly, the strategy assigns to the node 2 (see Figure 1(c)) a degree kn(2)=3+3+3+4=13, i.e. the sum of the degree of the four adjacent nodes {1,3,5,6}. Figure 1. (a) Example of a portion of a network. Panels (b) and (c) show the neighbourhood of node 1 and node 2, respectively. The neighborhood nodal degree can be interpreted as the node strength (Boccaletti et al. [4]) of a suitable oriented weighted network, which has the structure of the infrastructure network adjacency matrix. Each non-null cell i,j has a value corresponding to the link weight between i and j which is assumed equal to degree of the starting node. E.g., the Figure 2(a) and (b) reports the weights to be assigned to the links incident node 1 and 2 in order to interpret the proposed 7 12365(a)412365(b)412365(c)4strategy as a weighting of the adjacent matrix. In the Figure 2(a) the link 1-2 has a weight, w12, equal to 3 considering the direction from node 1 to node 2 because k(1)=3, while the link 2-1 has a weight w21=4 (considering the direction from node 2 to node 1) being k(2)=4; similarly for the links 1-3, 3-1, 1-4, and 4-1. The strategy is similarly applied to node 2. Figure 2. Weights assigned to links incident nodes 1 (a) and 2 (b). Introducing such oriented weighted network, the neighbourhood degree can be seen as the node strength (Boccaletti et al. [4]), namely (7) where wij are the previously defined topological weights. In this way, the weights can be interpreted as a sort of flow of topological information, which travels on each link, and the neighbourhood degree is the overall information that flows into each node from the nearest neighbours. 4. CLASSIFYING THE NETWORK STRUCTURE: NEIGHBORHOOD VERSUS STANDARD NODAL DEGREE DISTRIBUTION In order to test the effectiveness of neighbourhood nodal degree to classify infrastructure networks, we start showing some results about a small but realistic infrastructure network (see Figure 3) while in the next section results about a significant number of real networks will be described. Figure 3 shows the network layout of the so-called BBLAWN network (Giustolisi et al. [16]). The network is composed of 390 nodes and 439 links and it is frequently used as 8 ()njijNikiw1234W14=3W41=3(a)W32=312365W23=4W25=4W52=4(b)benchmark network in studies about WDNs (Ostfeld et al., [17]). Figure 4 reports the Empirical density distributions, P(k) and P(kn), and the corresponding cumulative distributions, Pcum(k) and Pcum(kn), of the BBLAWN related both to the standard nodal degree, k (Figures 4(a, c)), and the neighbourhood nodal degree, kn (Figures 4(b, d)). The same figure reports the corresponding theoretical Poisson and Pareto distributions, in order to compare them to the Empirical ones. In order to allow an easier comparison in the case of Pareto distribution, Figures 4(e, f) report the cumulative distributions also in logarithmic scale. The diagrams (a), (c) and (e) of Figure 4 reveal that the identification of a specific probabilistic model is very difficult when using the standard nodal degree: the maximum degree is five and, consequently, only five points are available to fit the distribution. It follows that to discern between the Poisson and Pareto distributions is difficult and statistically unreliable. The only reasonable conclusion arguable from the diagrams (a), (c) and (e) of Figure 4 is that the standard nodal degree distribution seems to be qualitatively more similar to the Poisson model than the Pareto one. Figure 3. BBLAWN layout. 9 00,050,10,150,20,250,30,350,40,450123456P(k) k EmpiricalPoissonPareto(a)00,050,10,150,20,250,30,3502468101214P(kn)knEmpiricalPoissonPareto(b) Figure 4. Empirical standard and neighbourhood nodal degree versus Poisson and Pareto distributions. On the contrary, diagrams (b), (d) and (f) on the right side of Figure 4 (related to the neighbourhood nodal degree) allows discerning between the Poisson and Pareto distributions being the maximum degree equal to 12 and, therefore, a wider interval of degrees is now available. In fact, the eleven points (degrees) make reliable the statistical analysis as will be reported in Table I. About the patterns of the diagrams in the right side of Figure 4, it is evident that the Empirical distributions follow very well the Poisson distributions especially referring to the cumulative ones. The different pattern between Empirical data and the Pareto distribution is more evident in the logarithmic scale of the cumulative distribution. We systematically compared standard and neighbourhood nodal degree distributions for twenty-two networks corresponding to existing WDNs. Table I reports their relevant characteristics. In particular, it is worth noting that network size spans two order of magnitude; in fact, the number of nodes ranges from 390 to 26.761, while the number of links varies from 439 to 32.096. In spite of the size variability, the average nodal degree remains always very low – ranging from 2.08 to 2.59 – as well as the maximum nodal degree, ranging from 5 to 11. These values confirm the situation reported in the previous paragraphs about the BBLAWN. The small number of points of the Empirical distribution frustrates any tentative to infer a statistical model. In contrast, the neighbourhood nodal degree depicts a very different picture: the average and maximum values range from 4.68 to 7.98 and from 12 to 33, respectively. The 10 00,10,20,30,40,50,60,70,80,910123456Pcum(k)kEmpiricalPoissonPareto(c)(c)00,10,20,30,40,50,60,70,80,9102468101214Pcum(kn)knEmpiricalPoissonPareto(d)1,E-021,E-011,E+001248Pcum(k) k EmpiricalPoissonPareto(e)1,E-031,E-021,E-011,E+00124816Pcum(kn)knEmpiricalPoissonPareto(f)minimum neighbourhood nodal degree results equal to two as it corresponds to the extreme nodes of network branches composed by serial pipes/links. Furthermore, in logarithmic diagrams, Empirical curves does not report any final value. This is because the log of zero does not admit finite value. Table I. Relevant data of the 22 networks (for few benchmark real WDNs the inhabitants are not available) N D W e m a N e d o N # e p i P # BBLAWN Big Town Apulia 1 BWSN Apulia 2 Apulia 3 Norway 1 Apulia 4 Apulia 5 Apulia 6 Apulia 7 Apulia 8 Apulia 9 Apulia 10 Piedmont 1 Norway 2 Apulia 11 Apulia 12 Exnet Apulia 13 Apulia 14 Apulia 15 390 26,761 18,718 12,518 5,288 5,036 5,035 4,242 4,188 3,547 3,000 2,968 2,895 2,810 2,784 2,520 2,403 1,918 1,776 1,762 1,263 1,263 439 32,096 19,990 14,314 6,116 5,848 5,292 4,940 4,727 3,881 3,189 3,400 3,333 3,307 2,894 2,651 2,820 2,153 2,300 2,098 1,428 1,406 n a e M d r a d n a t S 2.25 2.40 2.14 2.29 2.31 2.32 2.10 2.33 2.26 2.19 2.13 2.29 2.30 2.35 2.08 2.10 2.35 2.25 2.59 2.38 2.26 2.23 h t g n e L ] m K [ 57 2,054 678 1,844 277 199 239 161 261 113 91 130 85 175 171 129 145 108 594 70 46 30 s t n a t i b a h n I ] 0 0 0 1 x [ - 1,347 326 - 151 107 - 70 94 33 49 60 30 398 49 24 94 56 - 25 17 16 n i M d r a d n a t S x a M d r a d n a t S n a e M t s e r a e n r o b h g i e N t s e r a e n n i M r o b h g i e N x a M t s e r a e n r o b h g i e N t s e t S K ] % [ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 5 8 7 6 6 6 6 6 5 5 5 5 5 5 5 7 5 6 11 5 5 5 5.69 6.18 5.01 5.80 6.01 5.98 4.68 6.06 5.82 5.37 5.12 5.89 5.86 6.26 4.89 4.87 6.11 5.74 7.98 6.38 5.78 5.56 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 13 17 16 17 17 16 13 17 14 17 13 16 16 16 12 18 15 16 33 17 15 15 78.6 91.2 99.8 99.9 99.9 99.8 78.6 100.0 99.5 100.0 99.1 100.0 99.8 99.8 98.5 99.8 99.7 100.0 25.8 100.0 99.7 99.5 Figure 5 refers to the exemplifying case of the Big Town WDN. The panel 5(a) shows the network structure, which retraces clearly the urban structure, while the other figures compare the Empirical density and cumulative distributions of neighbourhood nodal degree versus the theoretical Poisson and Pareto distributions. These latter are fitted using the average value of the neighbourhood nodal degree and a calibrated value of  of Eq. (5), respectively. As in Figure 4 reports the cumulative distributions using both arithmetical and logarithmic scales. The comparisons clearly show that the Empirical distribution of neighbourhood nodal degree is very similar to the Poisson distribution, while it substantially differs from the Pareto one. 11 Figure 5. Layout of the Big Town network and Empirical versus Poisson and Pareto distributions. In the Appendix, the layouts and the cumulative distributions in logarithmic scale for all the WDNs listed in Table I, excluding BBLAWN, Big Town and Exnet, are reported. Apart the case of Exnet network, all the comparisons demonstrate that the Poisson distribution models very well the Empirical distribution of the neighbourhood nodal degree. The visual inspection is confirmed by the two-sample Kolmogorov-Smirnov (KS) goodness-of-fit hypothesis test (Massey [18]), determining if the distributions of the values in the Empirical and Poisson samples are drawn from the same underlying population. The results of the test are reported in the third last column of Table I and clearly show a high level of significance. It is worth noting that the Exnet network is the only one having a low significance level (about 26%). Figure 6. Layout of the Exnet network and Empirical versus Poisson and Pareto cumulative distributions. 12 00,050,10,150,20,25024681012141618P(kn) knEmpiricalPoissonPareto(b)00,10,20,30,40,50,60,70,80,91024681012141618Pcum(kn)knEmpiricalPoissonPareto(c)1,E-051,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn)knEmpiricalPoissonPareto(d)(a)1,E-111,E-101,E-091,E-081,E-071,E-061,E-051,E-041,E-031,E-021,E-011,E+001248163264Pcum(kn)knEmpiricalPoissonPareto(b)Figure 6 shows the Exnet case. It is to observe that Empirical cumulative distribution (in log scale) partially follows the Poisson model up to value of kn equal to about 13 and then it becomes linear, similarly to the Pareto model. This is the only case, among the analysed network connectivity structures, which shows the features of both Poisson and Pareto models consistently with Boccaletti et al. [4]. Actually, it is to note that (see Table I) Exnet network has the greatest maximum standard and neighbourhood nodal degree, respectively equal to 11 and 33, revealing in Figure 6 the presence of few hubs. 5. DISCUSSION AND CONCLUDING REMARKS Infrastructure networks are one of most representative cases of spatial networks. The term spatial characterizes networks in which nodes are located in a space equipped with a metric (Barthelemy [13]). For most of the infrastructure networks, the space is two-dimensional and the metric is the usual Euclidean distance. Spatial information comes in the network topology typically by the construction process, which is strongly affected by spatial constraints. In fact, infrastructure networks are generally manmade and they progressively grow filling the space and balancing connection costs and nodal distances, but also constrained by the impracticality of some connections (Buhl et al. [19]). The existence of such constrains explain why spatial networks generally are not scale free networks (Barthélemy [13] [20]) and, in this line, our findings indicate that the Poisson distribution is generally the best model to describe the network connectivity (as seen by the neighborhood nodal degree) of WDNs. The work by Barthélemy and Flammini [21] about the temporal evolution of urban road networks (see Figure 7) is helpful to understand our finding about WDNs, being these latter networks strongly influenced by the topology of urban patterns. The initial network (see the top-left panel in Figure 7) has a low size and appears quite regular (a ring with some branches is the typical initial configuration). Afterwards, the network evolves and new connections to customer proprieties are built. During this growth phase, the network is affected by the spatial constrains but a certain level of randomness emerges also related to the decision of technicians to design considering system redundancy for WDN management; this is due to the different local shapes of such constrains: e.g., different buildings, house blocks, districts, etc. During its evolution and increasing its size, the network can be increasingly classified into the category of random networks (and possibly as small world network). It follows that the Poisson distribution is the best suited to model the network structure of WDNs because the evolution 13 of such systems, although constrained, introduces random characteristics increasing the network size. The aim of the present work was twofold. Firstly, we proposed the neighborhood nodal degree. It is a novel quantity geared to describe the topology in the neighborhood of each node of a network. The second aim has been to show that neighborhood nodal degree distribution is suitable to classify infrastructure networks. Differently from the standard nodal degree distribution – which generally is uninformative, due to the very limited range of nodal degree values in the infrastructure networks – the neighborhood nodal degree exploits the topological information of the nearest neighbors and allows inferring reliable probabilistic models. In particular, we have investigated twenty-two real water distribution networks having different sizes and characteristics. In almost all cases, a Poisson distribution fits the Empirical neighborhood nodal degree distribution very well. This result appears in agreement with the characteristics of networks that evolve constrained by urban patterns. Figure 7. Snapshots of a urban network at different times of its evolution (from Barthélemy and Flammini [21]). 14 Appendix Layout and Empirical cumulative distribution of data versus the theoretical Poisson and Pareto distributions for the nineteen WDNs, which are not reported in the main text. 15 1,E-051,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-051,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn)knEmpiricalPoissonPareto 16 1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn)knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto 17 1,E-051,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn)knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto 18 1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-061,E-051,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto 19 1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn)knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+0012481632Pcum(kn) knEmpiricalPoissonPareto1,E-041,E-031,E-021,E-011,E+00124816Pcum(kn) knEmpiricalPoissonParetoReferences [1] A.-L Barabási, The network takeover. Nature Physics 8, 14 (2012). [2] R. Albert and A.-L. Barabasi, Statistica mechanics of complex networks. Rev. Mod. Phys. 74, 47 (2002). [3] M. Buchanan, Nexus: Small Worlds and the Groundbreaking Science of Networks, edited by W. W. Norton & Company. (2003). [4] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez and D.-U. Hwa, Complex networks: structure and dynamics. Phys. Rep. 424, 175 (2006). [5] M.E.J. Newman, Networks: An Introduction, edited by Oxford University Press, Oxford, (2010). [6] P. Erdös and A. Rényi, On random graphs, Publ. Math. 6, 290 (1959). [7] P. Erdös and P. Rényi, On the evolution of random graphs, Publ. Math. Inst. Hung. Acad. Sci. 5, 17 (1960). [8] S. Milgram, The Small World Problem, Psychology today, 1, 61 (1967). [9] D.J. Watts and D.H. Strogatz, Collective dynamics of small-world networks, Nature 393, 440 (1998). [10] A.-L. Barabási and R. Albert, Emergence of scaling in random networks, Science 286, 509 (1999). [11] M.E.J. Newman, The structure and function of complex networks, SIAM Rev. 45, 167 (2003). [12] R. Albert, H. Jeong and A.-L. Barabasi. Error and attack tolerance of complex networks. Nature. 406, 378 (2000). [13] M. Barthélemy, Spatial networks, Physics Reports 499, 1 (2011). [14] A. Yazdani and P. Jeffrey, Complex network analysis of water distribution systems, Chaos, 21 (2011). [15] S. Lammer, B. Gehlsen, and D.Helbing, Scaling laws in the spatial structure of urban road networks, Physica A 363, 89 (2006). [16] O. Giustolisi, L. Berardi, D. Laucelli, D. Savic and Z. Kapelan, Operational and Tactical Management of Water and Energy Resources in Pressurized Systems: Competition at WDSA 2014. J. Water Resour. Plann. Manage 142 (2016). [17] A. Ostfeld et al., The Battle of the Water Sensor Networks, A design challenge for engineers and algorithms, J. Water Resour. Plann. Manage., 134, 556 (2008). [18] F. J. Massey, The Kolmogorov-Smirnov Test for Goodness of Fit, J. American Statistical Association, 253, 68 (1951). 20 [19] J. Buhl, J. Gautrais, R.V. Solé, P. Kuntz, S. Valverde, J.L. Deneubourg and G. Theraulaz, Efficiency and robustness in ant networks of galleries, Eur. Phys. J. B 42, 123 (2004). [20] M. Barthelemy, Crossover from Scale-Free to Spatial Networks, Europhys. Lett. 63, 915 (2003). [21] M. Barthélemy and A. Flammini, Modeling Urban Street Patterns. Phys. Rev. Lett. 100, (2008). 21
1406.0516	4	1406	2015-06-11T12:04:59	Modeling Adoption and Usage of Competing Products	[ "cs.SI", "physics.soc-ph" ]	The emergence and wide-spread use of online social networks has led to a dramatic increase on the availability of social activity data. Importantly, this data can be exploited to investigate, at a microscopic level, some of the problems that have captured the attention of economists, marketers and sociologists for decades, such as, e.g., product adoption, usage and competition. In this paper, we propose a continuous-time probabilistic model, based on temporal point processes, for the adoption and frequency of use of competing products, where the frequency of use of one product can be modulated by those of others. This model allows us to efficiently simulate the adoption and recurrent usages of competing products, and generate traces in which we can easily recognize the effect of social influence, recency and competition. We then develop an inference method to efficiently fit the model parameters by solving a convex program. The problem decouples into a collection of smaller subproblems, thus scaling easily to networks with hundred of thousands of nodes. We validate our model over synthetic and real diffusion data gathered from Twitter, and show that the proposed model does not only provides a good fit to the data and more accurate predictions than alternatives but also provides interpretable model parameters, which allow us to gain insights into some of the factors driving product adoption and frequency of use.	cs.SI	cs	Modeling Adoption and Usage of Competing Products Isabel Valera MPI for Software Systems [email protected] Manuel Gomez-Rodriguez MPI for Software Systems [email protected] 5 1 0 2 n u J 1 1 ] I S . s c [ 4 v 6 1 5 0 . 6 0 4 1 : v i X r a Abstract—The emergence and wide-spread use of online social networks has led to a dramatic increase on the availability of social activity data. Importantly, this data can be exploited to investigate, at a microscopic level, some of the problems that have captured the attention of economists, marketers and sociologists for decades, such as, e.g., product adoption, usage and competition. In this paper, we propose a continuous-time probabilistic model, based on temporal point processes, for the adoption and frequency of use of competing products, where the frequency of use of one product can be modulated by those of others. This model allows us to efficiently simulate the adoption and recurrent usages of competing products, and generate traces in which we can easily recognize the effect of social influence, recency and competition. We then develop an inference method to efficiently fit the model parameters by solving a convex program. The problem decouples into a collection of smaller subproblems, thus scaling easily to networks with hundred of thousands of nodes. We validate our model over synthetic and real diffusion data gathered from Twitter, and show that the proposed model does not only provides a good fit to the data and more accurate predictions than alternatives but also provides interpretable model parameters, which allow us to gain insights into some of the factors driving product adoption and frequency of use. I. INTRODUCTION the Bass diffusion model [2], There is a long history of work in economics and marketing on studying product adoption, product competition and product life cycle [1]. This work has typically developed mathematical models, such as, the probit model [3], or models based on information cascades [4], that attempt to capture the macroscopic evolution of a product or set of competing products but not its (their) microscopic evo- lution, i.e., product adoptions and recurrent usage by specific individuals. This has been in part due to the lack of large-scale fine-grained product adoption and product usage data, which would allow to propose and validate microscopic models at scale. The emergence and wide-spread use of online social networks has led to a dramatic increase on the availability of large-scale fine-grained social activity data, in which all individual adoptions and subsequent uses of a product are often observable. This availability has opened up a great opportunity for a paradigm shift, where we investigate both the macro- scopic and microscopic dynamics of product adoption and usage. As a consequence, there recent empirical studies have investigated the adoption and usage of competing products [5] in online social networks. Therefore, it is necessary to develop microscopic models of product adoption and usage, which are largely non-existent to date. A. Our approach We introduce a continuous-time probabilistic model, based on temporal point processes, for the adoption and frequency of use of competing products in online social networks. The model captures several intuitive key factors, which has been previously discussed in the literature: I. Social Influence. Whenever a user observes that one or more of her neighbors are using a product, she may decide to start using the product [6]. II. Recency. If a user has used a product in the past, she is more likely to use the product in the future than if she has never used it [7]. III. Competition. If a user who is currently using a product p1 observes that one or more of her neighbors use a competing product p2, she may decide to switch from product p1 to p2 [8]. To simultaneously capture the above mentioned factors, which have been typically studied separately, our model lever- ages multivariate Hawkes processes [9]. Although Hawkes processes have been used for modeling positive social influ- ence before [10], [11], [12], [13], [14], the key innovation here is that we explicitly model recency and competition, and allow for both positive and negative social influence. As a consequence, our design is especially well fitted to model adoption and recurrent usage of competing products. Perhaps surprisingly, the flexibility of our model of com- peting products does not prevent us from efficiently simulate from the model, and learn the model parameters from real world product fine-grained usage data: • Efficient simulation. We develop an efficient sampling method that leverages Ogata’s method [15], and scales linearly in the number of products. Our method exploits the sparsity of social networks to reduce the overall complexity (refer to Algorithm 1). • Efficient parameter estimation. We fit the parameters of the model using historical data by solving a maximum likelihood estimation problem, which reduces to solving a convex program. Importantly, the problem decouples in several smaller problems, allowing for natural paralleliza- tion so that we can fit the parameters of the model for networks with hundred of thousands of nodes. Moreover, we validate our model using both synthetic and real social activity data, and show that: products a user may repeatedly keep using for long periods of time. 1) Our model can generate product use events that obey temporal patterns described in related literature [16], [17], such as temporal bursts of events or product switching. 2) Our regularized maximum likelihood estimation method can recover the parameters of the model with an accuracy that increases as we record more product uses over time. 3) Our model provides a significantly better fit and more accurate product use event predictions in real data gath- ered from Twitter [18] than several alternatives [7], [16]. In particular, we experiment with one type of competing product, url shortening services [5], and one type of conventions, the way Twitter users indicated a tweet was a retweet back in 2009 [19]. 4) Our model allows us to gain insights into some of the factors driving product adoption and frequency of use. For example, we find that the usage of popular products is typically driven by recency, however, an exposure to a less popular product can have a strong inhibiting effect on future uses of popular products. B. Related work Several models of adoption of competing products in online social networks have been proposed in the computer science literature in recent years [8], [20], [21]. In particular, Bharathi et al. [8] extend the independent cascade model, Dubey et al. [20] propose a local quasilinear model, and Goyal et al. [21] consider a broad class of local influence processes. Moreover, competition has been also studied in other con- texts: meme diffusion [22], rumor propagation [23], spread of misinformation [24], opinion formation [25], [26] and disease spread [27], [28]. However, most previous work share the following limitations, which we tackle in our work. First, they consider discrete-time sequential models of competition, which are difficult to estimate accurately from real world data [29], [30], [12]. Instead, we model competition using a continuous-time asynchronous model, designed to naturally fit the event data we record (i.e., the times in which users use products). Second, they consider each user to adopt only one contagion (be it a product, a meme, rumor, opinion or disease), the one she adopted first, and use it only once. In contrast, we allow users to use it once or several times and adopt different competing contagions simultaneously. That is, we model users’ frequency of use, and how this frequency depends on what users observe over time. Hence, we are able to model non-progressive phenomena [6]. Third, they only allow for pairwise interactions between contagions, being able to study competition between only two different products. In contrast, we consider n-ary interactions, which consider the interactions/competition among all the competing products in the network. Repeat online behavior have been studied in the specific contexts of repeat queries in web search [31], [32] or repeat website visits [33], [34]. Only very recently, Anderson et al. [7] propose a fairly general discrete-time model of online recon- sumption of goods, such as songs or movies, and experiences. However, this line of work does not model social influence nor competition and typically focus on goods, experiences or behaviors a user may get satiated or bored of, rather than Last, Hawkes processes have been also applied to a large variety of problems in which self-excitement plays a fundamental role: earth quake prediction [35], crime predic- tion [36], computational neuroscience [37], or high frequency trading [38]. However, to the best of our knowledge, Hawkes processes have not been applied in the context of adoption and usage of competing products. II. PROPOSED MODEL ?fiIn this section, we formulate our probabilistic model of adoption and frequency of use of competing products. We start by revisiting the foundations of temporal point processes, and then describe the specifics of our model. A. Temporal Point Processes A temporal point process is a random process whose realization consists of a list of discrete events localized in time, {ti} with ti ∈ R+ and i ∈ Z+. Many different types of data produced in online social networks and the Web can be represented as temporal point processes, e.g., the times when a user uses a product. A temporal point process can be equivalently represented as a counting process, N (t), which records the number of events before time t, e.g., the number of product use events before time t. Let the history H(t) be the list of times of events {t1, t2, . . . , tn} up to but not including time t. Then, in a small time window dt between [0, t), the number of observed event is dN (t) = δ(t − ti) dt, (1) (cid:88) ti∈H(t) (cid:82) t where δ(t) is a Dirac delta function, and hence N (t) = 0 dN (s). It is often assumed that only one event can happen in a small window of size dt, and hence dN (t) ∈ {0, 1}. An important way to characterize temporal point processes is via the conditional intensity function — the stochastic model for the time of the next event given all the times of previous events. The conditional intensity function λ∗(t) (intensity, for short) is the conditional probability of observing an event in a small window [t, t + dt) given the history H(t), i.e., λ∗(t)dt = P{event in [t, t + dt)H(t)} , (2) where ∗ means that the function λ∗(t) may depend on the history H(t). Based on the intensity, one can obtain the conditional expectation of the number of events in the windows [t, t + dt) and [0, t), respectively, as E[dN (t)H(t)] = λ∗(t) dt, and E[N (t)H(t)] = λ∗(τ ) dτ (cid:90) t (3) (4) 0 The functional form of the intensity λ∗(t) is often designed to capture the phenomena of interests [39]. For example, the following forms have been argued for in the information propagation and social influence literature [40], [11], [30], [41]: • Poisson process. The intensity is assumed to be indepen- dent of the history H(t) and constant, i.e., (5) λ∗(t) = λ, where λ (cid:62) 0 is a scale parameter. • Weibull renewal process. The intensity is assumed to be dependent only on the last event tn ∈ H(t) before t, i.e., (6) where k > 0 is a shape parameter and λ > 0 is a scale parameter. Remarkably, if the shape parameter is less than 1, then the inter-event time distribution is heavy-tailed, as observed in real social activity data [16]. • Hawkes Process. The intensity captures a mutual excita- λ∗(t) = k(t − tn)k−1λk (cid:62) 0, tion phenomena between events, i.e., λ∗(t) = µ + α gω(t − ti), (7) (cid:88) ti∈H(t) where gω(t) is a nonnegative triggering kernel such that gω(t) = 0 for t < 0, µ (cid:62) 0 is a baseline intensity inde- pendent of the history, and the summation of kernel terms is history dependent and a stochastic process by itself. Here, the occurrence of each historical event increases the intensity by a certain amount α (cid:62) 0. Now, given a time t(cid:48) (cid:62) t, we can also characterize the conditional probability that no event happens during [t, t(cid:48)) and the conditional probability density that an event occurs at time t(cid:48) using the intensity λ∗(t) by means of the following well- known relationships [39]: (cid:32) (cid:90) t(cid:48) (cid:33) λ∗(τ ) dτ , and (8) − S∗(t(cid:48)) = exp f∗(t(cid:48)) = λ∗(t(cid:48)) S∗(t(cid:48)). t n(cid:88) (cid:90) T (9) With these two quantities, we can express the log-likelihood of a list of event times {t1, t2, . . . , tn} in an observation window [0, T ) with T (cid:62) tn as L = log λ∗(ti) − λ∗(τ ) dτ, (10) i=1 0 which will be useful for learning the parameters of our model from observed data. With the above background in temporal point processes, we can now proceed to detail the formulation of the proposed model for product adoption and frequency of use, including our choice of specific functional form for the intensity λ∗(t). B. Model for Product Adoption Given a social network G = (V,E) and a set of (related) competing products P, we model the generation of product use events by users in the network using temporal point processes. Here, we record each event as a triple product e := ( u , ↑ user ↓ p, ), t ↑ time (11) where the triplet means that the user u ∈ V used product p ∈ P at time t. We assume that we can observe the triple e for every product use and there are not attribution problems. Given a list Hu(t) = {ei = (ui, pi, ti)ui = u and ti < t}, of use events {e1 = (u1, p1, t1), . . . , en = (un, pn, tn)} up to time t, the history Hu(t) of product use events by user u is (12) and the history Hup(t) of use events of product p by user u is Hup(t) = {ei = (ui, pi, ti)ui = u, pi = p and ti < t}. (13) Finally, the entire history of product use events is denoted as H(t) := ∪u∈VHu(t). Note that each node u does not necessarily use all products P and can use a specific product p once or several times. Now, we use a set of counting processes, one for each user and product, to record the generated product use events. More specifically, we denote the set of counting processes as a matrix N (t) of size G × P for each fixed time point t. The (u, p)-th entry in the matrix, Nup(t) ∈ {0} ∪ Z+, counts the number of times user u used product p up to time t. That is, Nup(t) = Hup(t) is the size of the history Hup(t). Then, we can characterize the users’ product usage over time using their corresponding intensities as E[dN (t)H(t)] = Λ∗(t) dt, (14) where the matrix dN (t) := ( dNup(t) )u∈V,p∈P contains the number of product use events per user and product in the window [t, t + dt), and the matrix Λ∗(t) := ( λ∗ up(t) )u∈V,p∈P contains the intensities associated to all the user-product pairs at time t. The sign ∗ means that the intensity matrix Λ∗(t) depends on the history H(t) and each particular in- tensity λ∗ up(t) depends on the history node u is exposed to, ∪v∈N (u)∪{u}Hv(t), where N (u) = {v ∈ V(v, u) ∈ E} denotes the nodes u follows (i.e., her neighbors). In the next section, we specify the functional form of the intensities that accounts for both recency and social influence. C. Intensity for Product Adoption To model the intensities, Λ∗(t), for product use events, we leverage the multivariate version [9] of the Hawkes process defined by Eq. 7. However, in our work, we modify the original definition of Hawkes processes to allow for both excitation and inhibition phenomena between product use events [42]. In particular, each intensity λ∗ up(t) takes the following form: λ∗ up(t) := spontaneous use µp(cid:124)(cid:123)(cid:122)(cid:125) P(cid:88) (cid:124) l=1 blp + + (cid:88) (cid:88) (cid:123)(cid:122) ei∈Hul(t) recency alp l=1 P(cid:88) (cid:124) (cid:88) (cid:123)(cid:122) v∈N (u) ei∈Hvl(t) social influence gω(t − ti) , (cid:125) gω(t − ti) (cid:125) (15) The first term µp (cid:62) 0 denotes the base event intensity, which models the spontaneous adoption of a product by a user on her own initiative. In other words, it models adoptions due to other factors different than recency or social influence. The second term models recency, where alp ∈ R denotes the influence that a previous use of a product l by the user has on her intensity for product p. The third term models the (a) Burstiness (b) Product switch Fig. 1. Example of event burstiness and product competition. Our model is able to generate bursts and product switches due to considering mutual excitation and inhibition between events in the intensities, given by Eq. 15, while Poisson processes fail to do so. Since product uses are realizations of a (stochastic) temporal point process, one cannot expect to precisely match the timing of each individual event. propagation of peer influence over the network. Here, blp ∈ R denotes the influence that a previous use of a product l by one of her neighbors has on the user’s intensity for product p. Intuitively, one may expect that app, bpp (cid:62) 0, since previous uses of a product may increase the probability that a user in the future, while alp, blp (cid:54) 0, uses the same product l (cid:54)= p, since previous uses of a product may decrease the probability that a user use a competing product in the future. For conciseness, we stack all the model parameters in a vector and two matrices, i.e., µ := (µ)p∈P, A = (alp)l,p∈P and B = (blp)l,p∈P. Finally, the triggering kernel gω(t) models the decay of influence over time. Here, for simplicity, we opt for an exponential kernel gω(t) := exp(−ωt)I[t (cid:62) 0]. However, since our inference method does not depend on this particular choice, one may decide to use more complicated kernels with some desirable characteristics [14] or perform a non-parametric estimation of the kernels [43]. Remarkably, using exponential kernels allow for slightly more efficient likelihood evaluations and, as a result, more efficient model simulation and estimation, as discussed in Sections III and IV. III. MODEL SIMULATION In this section, we first describe an efficient sampling me- thod for our model of product adoption and then elaborate on the temporal patterns described in related literature [17] that our model is able to reproduce. A. Simulation method We simulate samples (product usages) from our model by leveraging Ogata’s thinning algorithm [15]. However, an off-the-shelf implementation of Ogata’s algorithm would need O(n2V2) operations to draw n samples. Here, we first exploit the sparsity of social networks: whenever we sample a new product use event from the model, only a small number of intensity functions, in the local neighborhood of the node that used the product, changes. As a consequence, we do not have to re-evaluate all intensity functions. Moreover, for exponential trigger kernels, each time we re-evaluate an intensity function, we exploit the properties of the exponential function to do so in O(1): let t1 and t2 be two consecutive events, then, we can up(t1)−µp) exp(−ω(t2− compute each intensity λ∗ t1)) + µp. up(t2) as (λ∗ The complete simulation algorithm is summarized in Algo- rithm 1, which needs O(ndV) operations to draw n samples, where d is the maximum number of neighbors per node. Note that, since the model parameters (alp)l,p∈P and (blp)l,p∈P may take negative values, it may happen that λ∗ up(t) < 0 for some time t. In those cases, as proposed by Ogata [42], we set λ∗ up(t) to zero. B. Properties of the simulated product adoption By modeling recency and social influence, our model is able to generate realistic collections of product use events, which obey several temporal patterns observed in real social activity data [16], [17]. Here, we pay attention to two of these temporal patterns, which alternative models based on Poisson processes or Weibull renewal processes fail to capture: • Bursts. Social events are characterized by bursts of rapidly occurring events separated by long periods of inactivity [16]. In our Twitter dataset (Refer to Sec- tion VI), we also find such bursts of events, for example, Figure 1(a) shows the times (in black) in which a user used url shortening services (e.g., ‘bit.ly’). Remarkably, our model (in red) is able to mimic such bursts of events while alternative models, such as Poisson processes (in blue), fail to do so. Here, we trained the parameters of both models using the user’s real events (Refer to Section IV). Our model is able to generate bursts due to considering mutual excitation and inhibition between events in the intensities, given by Eq. 15. • Product Switches. A person who uses a product may decide to switch to a competing product due to multiple factors, including social influence [17]. In our Twitter dataset (Refer to Section VI), we often find users that decide to switch products, for example, Figure 1(b) shows the times (in the center) in which a user switch from one retweet convention, ‘RT’ (in red), to another, ‘via’ (in blue). Our model (in the top) is able to generate such a product switch while a Poisson process (in the bottom) does not succeed. Again, we trained the parameters of both models using the user’s real events (Refer to Sec- tion IV). Our model is able to generate product switches because it models competition by considering inhibition between events in the intensities, given by Eq. 15. We remark that here our goal here is to show that our model can generate product use events obeying several temporal patterns observed in real social activity . However, since product uses are realizations of a (stochastic) continuous-time 020406080PoissonTrueOur modelday020406080PoissonTrueOur modelday RTviaAlgorithm 1 Efficient Model Simulation Initialization: nup = 0 for u = 1, . . . ,V, p = 1, . . . ,P 1: I∗ ← IV×P(t0) ←(cid:80)V (cid:80)P I∗ ln(q) u p (t0) p λu Generate first event: 2: Generate q ∼ U[0,1] and s ← t0 − 1 3: if s > T , then go to last step. 4: else Attribution Test: i) Sample d ∼ U[0,1] ii) Choose u and p such that I up−1(t0) iii) Set t1 ← tup General subroutine: 5: while s < T do 6: 7: 8: 9: 10: Generate q ∼ U[0,1] Update s ← ti − 1 if s > T , then go to last step else Attribution-Rejection Test: I∗ ← IV×P(ti) +(cid:80)P pl +(cid:80) I∗ ln(q) pl au I∗ 1 ← s, i ← 1 and nup ← 1 u(cid:48)∈N (u) bu(cid:48) pl < d ≤ I up(t0) I∗ i) Sample d ∼ U[0,1] ii) if d ≤ IV×P(s) , then I∗ I up−1(t0) - Choose u and p such that < d ≤ I up(t0) - Set ti+1 ← tup nup+1 ← s, i ← i + 1 I∗ I∗ and nup ← nup + 1 iii) else - Update I∗ ← IV×P(s) and go to step 8. end while ({tup 11: Output: Retrieve the simulated process i })u=1,...,V,p=1,...,P on [t0, T ] temporal point process, one cannot expect to precisely match the timing of each individual event. IV. MODEL PARAMETER ESTIMATION In this section, we show how to efficiently learn the param- eters of our model using a set of historical product use events. Given a collection of product use events H(T ) = {(ui, pi, ti)} recorded during a time period [0, T ) in a social network G = (V,E), our goal is to find the optimal parameters µ, A and B by solving a regularized maximum likelihood estimation (MLE) problem. To this end, we first compute the log-likelihood of the recorded events, using Eq. 10, as follows: L(µ, A, B) = log λ∗ uipi λ∗ up(τ ) dτ. (cid:88) ei∈H(T ) (ti) −(cid:88) (cid:88) (cid:90) T u∈G p∈P 0 Fig. 2. Performance of our model estimation method in three Kronecker networks. We set the number of nodes in the three networks to 512 users and the number of edges to 2,040 (Core-Periphery), 4,608 (Hierarchical) and 7,669 (Random). Then, we can formulate the regularized MLE problem as: − L(µ, A, B) + β R(µ, A, B), µ,A,B minimize subject to µp ≥ 0, (18) where the first term is the negative log-likelihood of the events and the second term is the regularization term, being β the parameter that controls the trade-off between these two terms. As long as we choose a convex regularizer, it is easy to show that this regularized MLE problem is jointly convex in µ, A and B by using linearity, composition rules for convexity, and the concavity property of the logarithm. Hence, the global opti- mum can be found by many well-known algorithms [13], [14]. In practice, we solved Eq. 18 with CVX, a software package for specifying and solving convex programs [44], and used a quadratic regularizer, R(µ, A, B) = (cid:107)µ(cid:107)2 2 + (cid:107)B(cid:107)2 2, which we found to work well in practice. 2 + (cid:107)A(cid:107)2 For simplicity, we have described the model using the same model parameters, µ, A and B, for all users. However, in our experiments, we use different model parameters, µu, Au and Bu, for each user. By doing so, we can decompose the model estimation procedure into V × P independent maximum likelihood estimation problems, one per user u and product p. We can solve each subproblem in parallel, obtaining local solutions that are globally optimal. Additionally, under exponential kernels, we can precompute all sums of triggering kernels and integrals of triggering kernels for each subproblem i.e., O(Hu(T ) + ∪v∈N (u) Hv(T )), by in linear time, exploiting the properties of the exponential function, similarly as described in Section III. As a result, our model estimation procedure scales to networks on the order of hundreds of thousands of nodes. V. EXPERIMENTS ON SYNTHETIC DATA In this section, we validate our model using synthetic networks and product use event data. In particular, we will demonstrate that our model estimation method can accurately recover the true model parameters from historical event data using synthetic data. where we can further express the integral terms as: λ∗ up(τ ) dτ = T µp + Gω(T − ti) (cid:90) T 0 P(cid:88) P(cid:88) l=1 l=1 alp blp (cid:88) (cid:88) ei∈Hul(t) ei∈Hvl(t) (cid:88) (16) (17) + where Gω(t) =(cid:82) t v∈N (u) 0 gω(t(cid:48))dt(cid:48). Gω(T − ti), A. Experimental setup We generate three types of synthetic networks using a well-known model of directed social networks, the Kro- necker graph model [45]: (i) core-periphery networks [46] 0501001502000.160.180.20.220.240.26Average # of events per userMSE Core−peripheryHierarchicalRandom(a) Average prediction probability across all users. Data Our model Poisson Weibull Recency [7] URL RT 0.83 0.82 0.64 0.65 0.45 0.78 0.78 0.81 (b) Percentage of users for which each model achieves the maximum prediction probability. Data Our model Poisson Weibull Recency [7] URL RT 32.8% 58.2% 45.2% 76.3% 81.6% 89.1% 41.3% 57.3% TABLE I. PREDICTION PROBABILITY. VI. EXPERIMENTS ON REAL DATA In this section, we validate our model using real-world networks and product use event data. We first show that our model can accurately predict product use events in real-world data gathered from Twitter [18], significantly outperforming a state of the art method [7] and two baselines. We then use our model to derive insights into some of the factors driving product adoption and frequency of use, i.e., social influence, recency and product competition. A. Experimental Setup We use data gathered from Twitter as reported in previous work [18], which comprises of 1.7 billion public tweets posted by 52 million users during a three year period, from March 2006 to September 2009. Importantly, this dataset is complete, meaning that it contains all public tweets posted before the end of September 2009. Based on this raw data, we build two dataset of product use events. In the first dataset, we track every use event of the seven most popular (used) url shorteners [5]: Bitly, TinyURL, Isgd, TwURL, SnURL, Doiop and Eweri. Url shorteners existed before Twitter, however, by constraining the number of characters per message, Twitter increased the proliferation of new url shorteners as well as their usage. In the second dataset, we track every use event of the seven most popular retweet conventions1 [19]: RT, via, HT, retweet, retweeting, the symbol and R/T. Retweet conventions emerged organically during the first few years of Twitter and until November 2009, when Twitter rolled out an official built-in retweet button. In both cases, the functionality (or quality) of each competing product is similar, therefore, one may expect social influence, recency and competition factors to play a more significant role than product quality [7]. This allows us to directly observe and quantify the impact of each of these factors in the behavior pattern of each user in the network. It would be also interesting to consider competing products with clearly different quality and augment our model to incorporate product quality as a covariate in the product intensity defined by Eq. 15. For each of the two datasets, we experiment with four consecutive (arbitrary) months of product event data2, where 1The way Twitter users indicated back in 2009 that a tweet was being retweeted (or forwarded). Here, we think of different retweet conventions as competing products. 2January 15 to May 15 for url shorteners; February 1 to May 31 for retweet conventions. (a) Events (b) Products Fig. 3. Product use event statistics. Panel (a) shows the number of users that generated (solid lines) or got exposed to (dashed lines) at least x use events during the considered 4-month period. Panel (b) shows the number of users that used (solid lines) or got exposed to (dashed lines) at least x different products during the considered 4-month period. pp} and {bu pp} from U (0, 1) and {au [0.9, 0.5; 0.5, 0.3]), (ii) hierarchical net- (parameter matrix: works [47] ([0.9, 0.1; 0.1, 0.9]), and (iii) random networks ([0.5, 0.5; 0.5, 0.5]). Then, for each network, we assume there are two competing products, and set the model parameters for each product and node in the networks as follows. First, lp}l(cid:54)=p and we draw {au {bu lp}l(cid:54)=p from U (−1, 1). Then, we allow for a small set of users to have a baseline parameter greater than zero, and draw their baseline parameters from the uniform distribution U (0, 1). Finally, for each network, we generate and record a set of 100,000 events using our efficient sampling method, summarized in Algorithm 1. Now, given the times when each user used any of the two products, our goal is to find the true model parameters µu, Au, and Bu for each user u, by solving the regularized maximum likelihood estimation (MLE) problem defined in Eq. 18. (cid:104) (x − x)2(cid:105) B. Model Estimation We evaluate our model estimation method by comparing the inferred and true parameters in terms of the mean squared error (MSE), E , where x is the true parameter and x is the estimated parameter. Figure 2 shows the MSE against average number of events per user for three Kronecker networks, where we set the regularization parameter β = 10 and the parameter of the triggering kernels ω = 1. We observe that the MSE decreases as the average number of events per user increases, reaching values below 0.18 for the three networks once we observe ∼200 product use events per user in average. It is important to note that even though the networks have very different global network structures, the performance of our estimation method is remarkably stable and does not seem to depend on the structure of the network. 1001021041060100020003000# of events# of users URL usesURL exposuresRT usesRT exposures012345670100020003000# of products# of users URL usesURL exposuresRT usesRT exposures(a) Average test log-likelihood per event across all users. Data URL RT Our model −1.09 2.11 Weibull −1.18 −1.76 Poisson −3.28 −3.96 (b) Percentage of users for which each model achieves log- likelihood per event. the maximum average test Data URL RT Our model 80.2% 94% Poisson 4.1% 0% Weibull 15.7% 6% TABLE II. AVERAGE TEST LOG-LIKELIHOOD. we filter out users that join Twitter less than one month before the start of the period, do not use any product after the period, or generate less than 100 product use events during the whole three years of data, and focus on the 3,000 most active users (within the period) and their neighbors. We build each user’s neighborhood using the interactions via @-messages; we create a directed edge (i, j) as soon as user j mentions user i in a tweet, since this provides evidence that user j is paying attention to user i [48]. Then, we assume node i got exposed only to product use events from node j that occurred later than this first mention, as argued in previous work [49]. Here, one could think of building each user’s neighborhood by simply considering all the users she follows, however, this would require obtaining the times in which each user starts following others, and these times are currently not provided by the Twitter public API [50]. In the remainder of the section, for each dataset, we employ the first three months of data as training set, and the last month as test set. Figure 3 provides general statistics on the number of events users generate and are exposed to through her neighbors in both datasets. We find that ∼ 35% of the url shortener users and ∼ 75% of the retweet convention users use three or more products. Moreover, retweet conventions users are exposed to a larger number of product use events and products than url shorteners users. B. Implementation and scalability We developed an efficient distributed implementation of our model estimation method, described in Section IV, and de- ployed it in a cluster with 1,000 CPU cores. This allowed us to efficiently model thousands of users. Although we considered five url shorteners and retweet conventions, in practice, due to the cross validation step of the regularization parameter β and the triggering kernel parameter ω, we experimented with 38 different configurations per product and user. Therefore, since we modeled 3,000 users for both url shortener and retweet conventions, the analysis involved estimating over 50 million model parameters by solving 1,140,000 (3,000 × 5 × 38 × 2) convex optimization problems, in which the number of events in the objective function spans from 0 to hundreds of events. Remarkably, for each optimization problem, we can precom- pute the sums of triggering kernels and integrals of triggering kernels in a time linear in the number of events a user is exposed to, as discussed in Section IV. As a consequence, the overall fitting procedure for a very active user never took more than a few seconds and, for a less active user, it often took only a few milliseconds. C. Product use event prediction We evaluate the performance of our model in comparison with one state of the art method [7] (‘Recency’), which is the only approach in the literature that deal with recurrent usages of products and not only for single adoptions, and therefore allows for fair comparison with ours. We also compare our model to two baselines, a memoryless Poisson process (‘Pois- son’) [51] and a Weibull renewal process (‘Weibull’) [16]. The Recency model is a discrete time model that considers a product from t events ago to be reused with a probability proportional to a function of t. Poisson and Weibull are continuous time models that consider, respectively, a constant (memory-less) intensity, as defined in Eq. 5, and an intensity that depends on the last event, as defined in Eq. 6. For each user, we fit all models using maximum likelihood estimation. In our model, we set the regularization parameter β and the decay function parameter ω via cross-validation. In the Recency model, we consider a memory equal to five events. Then, we use three different evaluation measures to com- pare our model to the other three models: I. Prediction Probability: For each user and (trained) model, we predict the identity of the product associated to each true event in her test set by evaluating and rank- ing the likelihoods (probability for the Recency model) associated to each product at the test event and taking the top-1 product. The prediction probability is then defined as the probability that the true and the predicted identity for each product use test event coincide. This evaluation measure has been used previously by the authors of the Recency method [7]. II. Average Test Log-likelihood: For each user and continuous-time (trained) model, we compute the average log-likelihood per event in the test set. This allows us to assess how well each continuous-time model fits the true times of each test event – in other words, its goodness of fit. Note that we compute the log-likelihood over the to measure how well test data, not the training data, each model generalizes to the data. Here, we cannot compute the average test log-likelihood for the discrete- time Recency model, since it does not model the actual time of each product use event. III. AIC: For each user and continuous-time (trained) model, we compute the AIC [52] (Akaike information criterion). The AIC is given by AIC = 2Np − 2L, where Np is the number of model parameters and L is the average log- likelihood evaluated in the training set. AIC quantifies the trade-off between the goodness of fit for the true times of each training event and the complexity of the model and thus has been typically used for model selection. The preferred model is the one with the minimum AIC value. As with the average test log-likelihood, we cannot compute the AIC for the discrete-time Recency model, since it does not model the actual time of each product use event. Table I(a) shows the average production prediction proba- bility across all users. Our model outperforms the Recency (a) Average AIC across all users. Lower is better. Data URL RT Our model −430.4 −44.6 Poisson −6.4 25.5 Weibull −66.6 24.5 (b) Percentage of users for which each model achieves the maximum AIC. Data URL RT Our model 65% 66.8% Poisson 2.5% 6.1% Weibull 32.5% 27.1% TABLE III. AKAIKE INFORMATION CRITERION (AIC). (a) URL shortening services Product Bitly Tiny URL Isgd TwURL SnURL µu p > 0 100% 100% 0% 0% 0% au pp > 0 39.2% 89.1% 16.3% 11.2% 2.9% bu pp > 0 39.4% 82.3% 23.6% 19.9% 3.6% (b) Retweet conventions Convention RT via HT retweet retweeting µu p > 0 20.5% 15% 25.4% 11.1% 3.9% au pp > 0 94.9% 59.6% 60.2% 4% 3.2% bu pp > 0 89.2% 66.5% 67.2% 7.5% 4.8% TABLE IV. PERCENTAGE OF USERS WITH PARAMETERS µu pp > 0 AND au au pp > 0. p > 0, (a) Twitter user’s url shortening service usage pattern (b) Average test log-likelihood per event and user Fig. 4. An example of external intervention. On May 6th, 2009, Twitter changed the default url shortener from ‘tinyurl.com’ to ‘bit.ly’. model and both baselines for both url shorteners and retweet conventions. At first, one may think that the improvement with respect to the Recency model is small3, however, Table I(b) demonstrates that our model consistently achieves the highest prediction probability for each individual user by showing the percentage of users for which each model provides the maximum product prediction probability across all models. In terms of individual users, our model beats the second best, the Weibull model, by more than 20% for url shorteners and more than 10% for retweet conventions while the Recency and the Poisson models achieve an underwhelming performance. Here, note that the percentages sum up more than 100% because whenever we encounter a user that generate a few use events during the test period, several models often provide the same maximum prediction probability. Table II(a) shows the average test log-likelihood per event across users. Our model outperforms both continuous-time baseline models for url shorteners and retweet conventions. Moreover, Table II(b) demonstrate that our model consis- tently achieves the highest average log-likelihood across the continuous-time models for each individual user by showing the percentage of users for which each model provides the maximum test log-likelihood. In particular, our model is the best fit for as much as 80.2% and 92% of url shortening 3The difference is statistically significant (p =). services and the retweet conventions users, respectively. Note that probability densities of continuous random variables can be larger than 1, and thus test log-likelihoods be larger than 0 (e.g., our model achieves an average test log-likelihood value of 2.11 for retweet conventions). Table III(a) shows the average AIC across users and Table III(b) the percentage of users for which each model provides the minimum AIC. Our model beats both continuous- time baseline models in more than 65% of the users. Hence, we can conclude that our model is not only the model with the best performance in terms of product prediction probability and test log-likelihood, but it also provides the best trade-off between goodness of fit and model complexity. D. External interventions We have shown the effectiveness of our model at predicting the adoption and frequency of use of competing products. However, what if there is a sudden exogenous change, which cannot be explained through social influence, recency and product competition? For example, on May 6th, 2009, Twitter changed the default url shortener from ‘tinyurl.com’ to ‘bit.ly’, triggering a significant increase (decrease) on the amount of new adopters of ‘bit.ly’ (‘tinyurl.com’) shortly afterwards, as illustrated by Fig. 4(a). Can we at least detect such a change once it happened? As Figure 4(b) demonstrates for the above mention example, such change is easily identifiable by means of the average test log-likelihood per event across users, which drops dramatically. Therefore, once a change is detected, by means of the average log-likelihood per event, one may decide to re-train the model. E. Social influence, recency and competition We conclude our experiments by exploring the properties of the inferred users’ parameters on real data both qualitatively and quantitatively. By doing so, we aim to gain insights into some of the factors driving adoption and frequency of use of competing products. Here, we focus on the five most popular url shorteners and retweet conventions, since the parameters of the two less popular shorteners and retweet conventions are hardly ever different than zero. 15th April6th May15th May BitlyTiny URL15th − 21st April22nd − 28th April29th April− 5th May6th May − 13th May−3−2−1012Test log−likelihoodHHHHl p Bitly Tiny URL Isgd TwURL SnURL HHHHl p RT via HT Retweet Retweeting (a) URL shortening services Bitly Tiny URL Isgd TwURL SnURL 2.64% 28.2% 33% 36.6% 58.8% 80.8% 81.8% 87.9% 9.7% 1.1% 7.2% 0.5% 9.2% 21.6% 24.6% 13.8% 1.3% 0.2% 1.4% 2.8% (b) Retweet conventions RT via 1.6% 25.8% 22.2% 12.9% 45.2% 52% 54.3% 62% HT 1% 10.7% 40% 62.5% Retweet Retweeting 0.1% 3.3% 2.5% 4.4% 0.1% 2.6% 2.3% 1.6% HHHHl p Bitly Tiny URL Isgd TwURL SnURL (a) URL shortening services Bitly Tiny URL Isgd TwURL SnURL 6.5% 30.1% 29.7% 31.5% 30.6% 86.4% 82.6% 81.9% 2.6% 1.8% 4.6% 4% 7.2% 15.6% 17.9% 10.5% 0.6% 0.3% 0.9% 1.6% HHHHl p (b) Retweet convention RT via HT Retweet Retweeting RT via HT Retweet Retweeting 2.1% 1% 1.8% 27.6% 26.5% 1.3% 56.6% 47.2% 19.6% 64.3% 61.1% 55.5% 0.2% 0.6% 0.5% 2.6% 0.1% 0.6% 0.6% 1.6% TABLE V. PERCENTAGE OF USERS WITH PARAMETERS au lp < 0. TABLE VI. PERCENTAGE OF USERS WITH PARAMETERS bu lp < 0. p, au pp and bu pp) and (positive) social influence (bu We start by comparing spontaneous adoption (by means pp) on p), recency (au of µu the adoption and frequency of use of a product. Table IV summarizes the results, by showing the percentage of users for pp are greater than zero for url shorteners which µu and retweet conventions. We find clear differences between url shorteners and retweet conventions, which we summarize next. On the one hand, Table IV(a) indicates that, for the 3,000 users under study, spontaneous adoption dominates over recency or (positive) social influence to explain the use of popular shorteners (‘bitly’ and ‘tinyurl’). In contrast, less popular shorteners (‘isgd’, ‘twurl’ and ‘snurl’) are never used by spontaneous adoption but first used by (positive) social influence and then re-used by recency. On the other hand, Table IV(b) indicates that recency and (positive) influence dominate over spontaneous adoption to explain the use of popular retweet conventions (‘RT’, ‘via’ and ‘HT’). In contrast, spontaneous adoption, social influence and recency have a comparable strength on the adoption and use of less popular retweet conventions (‘retweet’ and ‘retweeting’). lp and bu Next, we investigate competition (or inhibition) between products. Tables V and VI summarize the results, by showing lp (for l (cid:54)= p) are the percentage of users for which au smaller than zero. Here, each value can be viewed as a measure of the degree of inhibition that an event of the product in the row entails on the usage of the products in each column. We find several interesting patterns, which we summarize next. First, if we pay attention to the last rows of each table, which correspond to less popular products and conventions, we conclude that using one of these products or conventions has a strong inhibitory effect on using more popular product or conventions. Second, we find that ‘bit.ly’ has a stronger inhibitory effect on ‘tinyurl.com’ but ‘tinyurl.com’ does not have almost any effect on ‘bit.ly’. This may indicate that even before the change of default service from ‘tinyurl.com’ to ‘bit.ly’ on May 6th, ‘bit.ly’ was already a strong competitor to ‘tinyurl.com’. Finally, if we compare the inhibitory effect of previous product events by a user herself (au lp) and her neighbors (bu lp), we find qualitatively similar values. VII. CONCLUSIONS We have developed a probabilistic model, based on tem- poral point processes, for the adoption and frequency of use of competing products in online social networks. By modeling recency, social influence and product competition, our model is able to generate realistic collections of product use events, obeying several temporal patterns observed in real online social activity, as well as predict product use events more accurately than alternatives. Our work also opens many interesting venues for future work. For example, we have modeled the decay of the influ- ence of an event over time using exponential triggering kernels gω(·) and considered linear intensities with constant baseline rates. However, this is implicitly considering one characteristic time scale, which may be unrealistic in some scenarios. A natu- ral follow-up to potentially improve the accuracy of our model would be considering more complex triggering kernels [14], performing a non-parametric estimation of the kernels [43], assuming time-varying baselines rates [39] or extending the model to incorporate nonlinearities [53]. We have considered considered products and contagions with similar functionalities (or qualities) in order to study the impact of social influence, recency and competition play in the users’ decisions. How- ever, as future work, it would be also interesting to consider competing products with clearly different quality and augment our model to explicitly incorporate product quality. Moreover, we have built each user’s neighborhood using the interactions via @-messages and considered each neighbor to have the same influence on a user. It would be interesting to build each user’s neighborhood using the actual follower network and consider each neighbor to have different influence4. Finally, our experimental validation relies on data gathered exclusively from Twitter, it would be interesting to evaluate our model in other microblogging (Weibo, Tumblr) and social networking (Facebook, G+) sites. 4This would require obtaining the times in which each user starts following others and, unfortunately, these times are currently not provided by the Twitter public API [50]. REFERENCES [1] V. Mahajan, E. Muller, and F. M. Bass, “New product diffusion models in marketing: A review and directions for research,” The Journal of Marketing, 1990. [2] F. M. Bass, “A new product growth for model consumer durables,” Management Science, vol. 15, no. 1, p. 215, 1969. [3] D. McFadden, “Econometric models for probabilistic choice among products,” Journal of Business, vol. 53, no. 3, pp. S13–S29, 1980. [4] S. Bikhchandani, D. Hirshleifer, and I. Welch, “A theory of fads, fashion, custom, and cultural change as informational cascades,” The Journal of Economic Perspectives, vol. 12, no. 3, pp. 151–170, 1998. [5] D. Antoniades, I. Polakis, G. Kontaxis, E. Athanasopoulos, S. Ioannidis, E. P. Markatos, and T. Karagiannis, “we. b: The web of short urls,” in 20th International Conference on World Wide Web, 2011. [6] D. Kempe, J. M. Kleinberg, and E. Tardos, “Maximizing the spread of influence through a social network,” in 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2003. [7] A. Anderson, R. Kumar, A. Tomkins, and S. Vassilvitskii, “The dynam- ics of repeat consumption,” in 23rd International Conference on World Wide Web, 2014. [8] S. Bharathi, D. Kempe, and M. Salek, “Competitive influence max- imization in social networks,” Internet and Network Economics, pp. 306–311, 2007. [9] T. Liniger, “Multivariate hawkes processes,” Ph.D. dissertation, Swiss Federal Institute of Technology Zurich, 2009. [10] C. Blundell, J. Beck, and K. A. Heller, “Modelling reciprocating relationships with hawkes processes,” in Neural Information Processing Systems, 2012. [11] M. Farajtabar, N. Du, M. Gomez-Rodriguez, I. Valera, L. Song, and H. Zha, “Shaping social activity by incentivizing users,” in Neural Information Processing Systems, 2014. [12] T. Iwata, A. Shah, and Z. Ghahramani, “Discovering latent influence in online social activities via shared cascade poisson processes,” in 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2013. [13] K. Zhou, L. Song, and H. Zha, “Learning social infectivity in sparse low-rank networks using multi-dimensional hawkes processes,” in 16th International Conference on Artificial Intelligence and Statistics, 2013. [14] K. Zhou, H. Zha, and L. Song, “Learning triggering kernels for multi- dimensional hawkes processes,” in 30th International Conference on Machine Learning, 2013. [15] Y. Ogata, “On lewis’ simulation method for point processes,” IEEE Transactions on Information Theory, vol. 27, no. 1, pp. 23–31, 1981. [16] A.-L. Barab´asi, “The origin of bursts and heavy tails in human dynam- ics,” Nature, vol. 435, p. 207, 2005. [17] P. Shukla, “Effect of product usage, satisfaction and involvement on brand switching behaviour,” Asia Pacific Journal of Marketing and Logistics, vol. 16, no. 4, pp. 82–104, 2004. [18] M. Cha, H. Haddadi, F. Benevenuto, and P. K. Gummadi, “Measuring User Influence in Twitter: The Million Follower Fallacy,” in 4th AAAI Conference on Weblogs and Social Media, 2010. [19] F. Kooti, H. Yang, M. Cha, P. K. Gummadi, and W. A. Mason, “The emergence of conventions in online social networks.” in 6th AAAI Conference on Weblogs and Social Media, 2012. [20] P. Dubey, B. De Meyer, and R. Garg, Competing for customers in a social network. Cowles Foundation Discussion Paper, 2006. [21] S. Goyal and M. Kearns, “Competitive contagion in networks,” in 44th ACM Symposium on Theory of Computing, 2012. [22] S. Myers and J. Leskovec, “Clash of the Contagions: Cooperation and Competition in Information Diffusion,” in 12th IEEE International Conference on Data Mining, 2012. J. Kostka, Y. A. Oswald, and R. Wattenhofer, “Word of mouth: Rumor dissemination in social networks,” in SIROCCO, 2008. [23] [24] C. Budak, D. Agrawal, and A. El Abbadi, “Limiting the Spread of Misinformation in Social Networks,” in 4th ACM International Conference on Web Search and Data Mining, 2011. [25] M. H. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, vol. 69, no. 345, pp. 118–121, 1974. [26] N. E. Friedkin and E. C. Johnsen, “Social influence and opinions,” Journal of Mathematical Sociology, vol. 15, no. 3-4, pp. 193–206, 1990. [27] A. Beutel, B. A. Prakash, R. Rosenfeld, and C. Faloutsos, “Interacting viruses in networks: can both survive?” in 18th ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining, 2012. [28] B. A. Prakash, A. Beutel, R. Rosenfeld, and C. Faloutsos, “Winner takes all: competing viruses or ideas on fair-play networks,” in 21st International Conference on World Wide Web, 2012. [29] N. Du, L. Song, A. Smola, and M. Yuan, “Learning networks of heterogeneous influence,” in Neural Information Processing Systems, 2012. [31] [30] M. Gomez-Rodriguez, D. Balduzzi, and B. Scholkopf, “Uncovering the Temporal Dynamics of Diffusion Networks,” in 28th International Conference on Machine Learning, 2011. J. Teevan, E. Adar, R. Jones, and M. Potts, “History repeats itself: repeat queries in yahoo’s logs,” in 29th annual international ACM SIGIR conference on Research and development in information retrieval, 2006. J. Teevan, E. Adar, R. Jones, and M. A. Potts, “Information re-retrieval: repeat queries in yahoo’s logs,” in 30th annual international ACM SIGIR conference on Research and development in information retrieval, 2007. [33] E. Adar, J. Teevan, and S. T. Dumais, “Large scale analysis of web revisitation patterns,” in SIGCHI conference on Human Factors in Computing Systems, 2008. [32] [34] ——, “Resonance on the web: web dynamics and revisitation patterns,” in SIGCHI Conference on Human Factors in Computing Systems, 2009. [35] D. Marsan and O. Lengline, “Extending earthquakes’ reach through cascading,” Science, vol. 319, no. 5866, pp. 1076–1079, 2008. [36] M. Egesdal, C. Fathauer, K. Louie, J. Neuman, G. Mohler, and E. Lewis, “Statistical modeling of gang violence in los angeles,” SIAM Undergraduate Research Online, vol. 3, 2010. [37] M. Krumin, I. Reutsky, and S. Shoham, “Correlation-based analysis and generation of multiple spike trains using hawkes models with an exogenous input,” Frontiers in computational neuroscience, 2010. [38] E. Bacry, S. Delattre, M. Hoffmann, and J.-F. Muzy, “Modelling mi- crostructure noise with mutually exciting point processes,” Quantitative Finance, vol. 13, no. 1, 2013. [39] O. Aalen, Ø. Borgan, and H. Gjessing, Survival and event history analysis: a process point of view. Springer Verlag, 2008. [40] N. Du, L. Song, M. Gomez-Rodriguez, and H. Zha, “Scalable influence estimation in continuous-time diffusion networks,” in Neural Informa- tion Processing Systems, 2013. [41] M. Gomez-Rodriguez, J. Leskovec, and B. Scholkopf, “Modeling Information Propagation with Survival Theory,” in 30th International Conference on Machine Learning, 2013. [42] Y. Ogata and H. Akaike, “On linear intensity models for mixed doubly stochastic poisson and self-exciting point processes,” in Selected Papers of Hirotugu Akaike. Springer, 1998, pp. 269–274. [43] E. Bacry, K. Dayri, and J.-F. Muzy, “Non-parametric kernel estimation for symmetric hawkes processes. application to high frequency financial data,” The European Physical Journal B, vol. 85, no. 5, pp. 1–12, 2012. [46] [44] M. Grant and S. Boyd, “CVX,” http://cvxr.com/cvx, 2013. [45] J. Leskovec, D. Chakrabarti, J. Kleinberg, C. Faloutsos, and Z. Ghahra- mani, “Kronecker graphs: An approach to modeling networks,” Journal of Machine Learning Research, 2010. J. Leskovec, K. J. Lang, A. Dasgupta, and M. W. Mahoney, “Statistical properties of community structure in large social and information networks,” in 17th International Conference on World Wide Web, 2008. [47] A. Clauset, C. Moore, and M. E. J. Newman, “Hierarchical structure and the prediction of missing links in networks,” Nature, vol. 453, no. 7191, pp. 98–101, 2008. [48] B. A. Huberman, D. M. Romero, and F. Wu, “Social networks that matter: Twitter under the microscope,” First Monday, vol. 14, no. 1, 2009. [49] D. M. Romero, B. Meeder, and J. Kleinberg, “Differences in the mechanics of information diffusion across topics: idioms, political hashtags, and complex contagion on twitter,” in 20th International Conference on World Wide Web, 2011. [50] S. A. Myers and J. Leskovec, “The bursty dynamics of the twitter information network,” in 23rd International Conference on World Wide Web, 2014, pp. 913–924. J. Kingman, Poisson Processes, ser. Oxford studies in probability. Clarendon Press, 1992. [51] [52] H. Akaike, “A new look at the statistical model identification,” IEEE Transactions on Automatic Control, vol. 19, no. 6, pp. 716–723, Dec. 1974. [53] P. Br´emaud and L. Massouli´e, “Stability of nonlinear hawkes pro- cesses,” The Annals of Probability, pp. 1563–1588, 1996.
1302.6109	1	1302	2013-02-25T14:47:50	Social Resilience in Online Communities: The Autopsy of Friendster	[ "cs.SI", "physics.soc-ph" ]	We empirically analyze five online communities: Friendster, Livejournal, Facebook, Orkut, Myspace, to identify causes for the decline of social networks. We define social resilience as the ability of a community to withstand changes. We do not argue about the cause of such changes, but concentrate on their impact. Changes may cause users to leave, which may trigger further leaves of others who lost connection to their friends. This may lead to cascades of users leaving. A social network is said to be resilient if the size of such cascades can be limited. To quantify resilience, we use the k-core analysis, to identify subsets of the network in which all users have at least k friends. These connections generate benefits (b) for each user, which have to outweigh the costs (c) of being a member of the network. If this difference is not positive, users leave. After all cascades, the remaining network is the k-core of the original network determined by the cost-to-benefit c/b ratio. By analysing the cumulative distribution of k-cores we are able to calculate the number of users remaining in each community. This allows us to infer the impact of the c/b ratio on the resilience of these online communities. We find that the different online communities have different k-core distributions. Consequently, similar changes in the c/b ratio have a different impact on the amount of active users. As a case study, we focus on the evolution of Friendster. We identify time periods when new users entering the network observed an insufficient c/b ratio. This measure can be seen as a precursor of the later collapse of the community. Our analysis can be applied to estimate the impact of changes in the user interface, which may temporarily increase the c/b ratio, thus posing a threat for the community to shrink, or even to collapse.	cs.SI	cs	David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ 3 1 0 2 b e F 5 2 ] I S . s c [ 1 v 9 0 1 6 . 2 0 3 1 : v i X r a Social Resilience in Online Communities: The Autopsy of Friendster David Garcia, Pavlin Mavrodiev, Frank Schweitzer Chair of Systems Design, ETH Zurich, Weinbergstrasse 56/58, 8092 Zurich, Switzerland [email protected],[email protected],[email protected] February 22, 2013 Abstract We empirically analyze five online communities: Friendster, Livejournal, Facebook, Orkut, Myspace, to identify causes for the decline of social networks. We define social resilience as the ability of a community to withstand changes. We do not argue about the cause of such changes, but concentrate on their impact. Changes may cause users to leave, which may trigger further leaves of others who lost connection to their friends. This may lead to cascades of users leaving. A social network is said to be resilient if the size of such cascades can be limited. To quantify resilience, we use the k-core analysis, to identify subsets of the network in which all users have at least k friends. These connections generate benefits (b) for each user, which have to outweigh the costs (c) of being a member of the network. If this difference is not positive, users leave. After all cascades, the remaining network is the k-core of the original network determined by the cost-to-benefit (c/b) ratio. By analysing the cumulative distribution of k-cores we are able to calculate the number of users remaining in each community. This allows us to infer the impact of the c/b ratio on the resilience of these online communities. We find that the different online communities have different k-core distributions. Consequently, similar changes in the c/b ratio have a different impact on the amount of active users. As a case study, we focus on the evolution of Friendster. We identify time periods when new users entering the network observed an insufficient c/b ratio. This measure can be seen as a precursor of the later collapse of the community. Our analysis can be applied to estimate the impact of changes in the user interface, which may temporarily increase the c/b ratio, thus posing a threat for the community to shrink, or even to collapse. 1 Introduction Online Social Networks (OSN), such as Facebook or Friendster, can quickly become popular, but can also suddenly lose large amounts of users. The appearance of competing OSN, with dif- ferent functionalities and designs, create unexpected shifts of users that abandon one community for another [13]. While the dynamics of growth in these online communities are an established research subject [4, 18], there are still many open questions regarding the decline of online com- munities, in particular related to large OSN [30]. What are the reasons behind the decision of users to stop using an OSN? What is the role of the social network in keeping user engagement, 1/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ or in the spreading of user dissatisfaction? Are there network structures that lead to higher risks of massive user departures? In this article, we assess the question of the relation between the topology of the user network, and the cascades of user departures that threaten the integrity of an online community. We build on previous theoretical work on network effects [5], providing the first empirical study of this phenomenon across successful, failed, and declining OSN. The most successful OSN attract millions of users, whose interactions create emergent phenom- ena that cannot be reduced back to the behavior of individual users. The OSN is a communication medium that connects a large amount of people, which would stay together only if their interac- tion dynamics leads to the emergent entity that we call the community. The OSN and its users form a socio-technical system in which the persistence of the community depends on both the social interaction between users, and the implementation and design of the OSN. In this context, the social resilience [2] of an online community is defined as “The ability of the community to withstand external stresses and disturbances as a result of environmental changes”. In particular, the technological component of the OSN can change the environment of the users, and create stress that threatens the cohesion of the community. As an example, changes in the user interface pose a general risk for user engagement in OSN. The fast pace of the Internet society has already led to the total disappearance of some very large online communities. The most paradigmatic example is Friendster, one of the first and largest OSN, which suffered a massive exodus of users towards competing sites. This led to its closure in 2011, to reopen as an online gaming without its profile data. As a reaction, the Internet Archive 1 crawled as much information as possible, creating a timeless snapshot of Friendster right before its closure. If, on the other hand, Friendster was still an alive and active community, this data would have been kept private and never made accessible at such scale. Before closure, users were warned and offered to delete their data from the site, leaving all the remaining data from this community as one of the largest publicly available datasets on social behavior. The decay of Friendster is commented in a comedy video of the Onion News, in which a fictitious “Internet archaelogist” explains Friendster as an ancient civilization 2. While proposed as a satire of the speed of Internet culture, this video illustrates the opportunities that a failed OSN offers for research. The users of such a community leave traces that allow us to investigate its failure. In this sense, we can name our work as Internet Archeology, because we analyze non-written traces of a disappeared society, aiming at understanding the way it worked and the reasons for its demise. In this paper, we provide a quantitative approach to the collective departure of users from OSN. We start from a theoretical perspective that, under the assumption of rational user behavior, al- lows us to define a new metric for the relation between network topology and massive user leaves. 1http://archive.org/details/friendster-dataset-201107 2https://www.youtube.com/watch?v=7mFJdOsjJ0k 2/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ We apply this metric to high quality datasets from Friendster and Livejournal, comparing their social resilience with partial datasets from Facebook, Orkut, and Myspace. The research presented here is based on publicly available datasets, allowing the independent validation of our results, as well as the extension to further analyses [17]. In addition, we focus on the time evolution of Friendster, tracking the changes in its social resilience and investigating how it decayed to a complete collapse. We finish by commenting on the limitations and extensions of our approach, and outline possible future applications. 2 Related work Recent research has focused on the question of growth and decay of activity or interest-based social groups [22]. This line of research analyzes social groups as subcommunities of a larger community, tied together due to underlying common features of their members. Such approach can be equally applied to scientific communities and online social networks [4, 32], revealing patterns of diffusion and homophily that respectively spread group adoption, and increase internal connectivity. In particular, the big datasets provided by online communities allow the study of group creation and maintenance [18], as well as the patterns of their internal network structure across communities [24, 31]. These results lead to applied techniques to predict the fate of interest- based groups, and to improve clustering analysis of social networks. Our work differs from these previous results in the scope of our analysis: Instead of looking at small to medium sized groups within larger communities, we look at the OSN as a whole. In our approach, users are not connected to each other due to certain common interest or affiliation, but through an online platform that maintains their social links and serves as communication medium. Another research topic close to our work is the analysis of individual churn, defined as the decision of a user to stop using a service in favor of a competitor. This topic has received significant attention due to its business applications in the telephony sector [9], studying how the social environment of an individual can influence the switching to a competing company. In addition, further studies explored how individual users disconnect from P2P networks[16], and stop using massive multiplayer online games [20]. Regarding social networks, the question of user departure and churn has special relevance [19]. As an emerging topic, a recent study shows the relation between social interaction and user departure in the online community Yahoo answers [10]. Furthermore, the same question has been addressed in a recent article [30], analyzing a mysterious online social network of which nor the name, size, nor purpose is explained. While these results are relevant for the question of user engagement, it is difficult to consider them in further research if we do not have information about the nature of the studied network. Social networks can have very different roles in online communities, requiring a differentiation between traditional social networking sites [24], and 3/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ online communities with a social network component, but where social interaction is mediated through other channels. The results of [30] reveal that 65% of the users that have no friends still remain active after three months, indicating that such social network is not precisely necessary for a user use the site. As an example, a Youtube user does not need to create and maintain social contacts to interact with other users, which can be done through videos and comments independently of the social network. Our work complements the previous results on individual user departures mentioned above, as we analyze the social resilience of the online community at the collective level. We build on these empirically validated microscopic rules of churn, to focus on cascades of departures through large OSN. We analyze the macroscopic topology of the social network and its role in the survival of the community. This kind of macroscopic effects are relevant to study the emergence of social conventions [23], an dynamics of politically aligned communities [8, 12], in addition to the case of OSN we address here. The particular problem of enhancing resilience by fixing nodes of a social network has been proposed and theoretically analyzed [5], aiming to prevent the unraveling of a social network. This implies that social resilience can be analyzed through the k-core decomposition of the social network, as explained in Section 3.1 . In addition, k-core centrality is the current state-of-the art metric to find influential nodes in general networks [21], and information spreading in politically aligned communities [8]. Regarding social media in general, the k-core decomposition was applied for a global network of instant messaging [25], as well as for the Korean OSN Cyworld [3, 6], motivated by user centrality analysis rather than social resilience. To our knowledge, this article introduces the first empirical analysis of social resilience, relating changes in user environment with cascades of departing users, through analysis based on the the k-core decomposition of different OSN. 3 Social Resilience in OSN 3.1 Quantifying Social Resilience A characteristic property of any online social network is the presence of influence among friends. In particular, individual decisions regarding participating or leaving the network are, to a large extent, determined by the number of one’s friends and their own engagement [4]. Therefore, users leaving a community have negative indirect effects on their friends [30]. This may trigger the latter to also leave, resulting in further cascades of departing users which may ultimately endanger the whole community. Social resilience acts to limit the spread of such cascades. One approach to quantify social resilience is by natural removal of nodes based on some local property, for example degree [25]. By studying the network connectivity after such removals, one 4/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ can identify nodes with critical importance for keeping the community connected. Importantly, by focusing on local properties we can only quantify the direct effects that a node removal has on the connectivity of the network. In this paper, we propose an extension based on the k-core decomposition [28]. A k-core of a network is a sub-network in which all nodes have a degree ≥ k. The k-core decomposition is a procedure of finding all k-cores, ∀k > 0, by repeatedly pruning nodes with degrees k. Therefore, it captures not only the direct, but also the indirect impact of users leaving the network. As an illustration consider Figure 1, which shows targeted removal of nodes with degrees < 3. On one Figure 1: Effects of node removals on network connectivity as captured by degree only (A → B) and k-core decomposition (A → C → D → E) hand, starting from the network in A and removing all nodes with degrees < 3, produces the network in B. The light-grey nodes in B have been removed (and thus are disconnected), and the final network consists of the 9 white nodes. The transition A → B shows only the direct effects of users with < 3 friends leaving. On the other hand, starting again from A, and applying the k-core procedure, will repeatedly remove nodes until only those with degrees ≥ 3 remain. The first step, A → C, removes the same light-grey nodes as before. Continuing, C → D, removes those nodes that have been left with < 3 neighbours in C, and disconnects them as well. The final step, D → E, finishes the process by disconnecting the last white node in D that was left with < 3 friends. As a result, the final network is the fully connected network of the 4 white nodes. Hence, supposing that users leave a community when they are left with less than 3 friends, the k-core decomposition captures the full cascading effect that departing users have on the network as a whole. 5/22 BCDEADavid Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ We proceed by formalizing social resilience based on a generalized k-core decomposition. To this end, we present a theoretical model in which rational users decide simultaneously either to stay in the network or to leave it. These decisions are based on maximizing a utility function that weighs the benefits of membership against the associated costs. We show that the equilibrium network which maximizes the total payoff in the community, corresponds to a generalized k-core decomposition of the network. 3.2 Generalized k-core decomposition Following [15], we extend the traditional k-core decomposition by recognizing that the pruning criterion need not be limited to degree only. Let us define a property function Bi(H) that given a sub-network H ⊆ G associates a value, ni ∈ R, to node i. A generalized k-core of a network G is, then, defined as a sub-network H ⊆ G, such that Bi(H) ≥ k, ∀i ∈ H and k ∈ Z. The general form of Bi allows us to model different pruning mechanisms. For example, the traditional definition of the k-core can be recovered in the following way – for every node i take its immediate neighbourhood, Ni, and fix Bi(H) := Ni, ∀H ⊆ G. Other authors have also shown that considering weighted links in Bi can more accurately reveal nodes with higher spreading potential in weighted networks [11]. Note that by definition higher order cores are nested within lower order cores. We use this to define that a node i has coreness ks if it is contained in a core of order k, but not in a core of order k(cid:48) > k. 3.3 A rational model for OSN users Here, we model the cost-benefit trade-off of OSN users in the following way. Assume that users in a given network, G, incur a constant integer cost, c > 0, for the effort they must spend to remain engaged. Accordingly, they receive a benefit or payoff from their friends in the network. Let the benefit of player i be the property function Bi(H) with i ∈ H. Assume non-increasing marginal i (H) ≤ 0, otherwise costs are irrelevant as any cost benefits with respect to the size of H, i.e. B(cid:48)(cid:48) level could be trivially overcome by increasing the size of H. This assumption is also supported by other empirical investigations of large social networks which show that the probability of a user to leave is concave with the number of friends who left [4, 30]. Players can choose one of two possible moves – stay or leave. The utility of player i, is Ui = 0, if he played leave or Ui = Bi(H) − c, for stay. Finally, users are utility-maximizing, therefore they will choose stay as long as Ui > 0. It is easily seen that the equilibrium network, G∗, which maximizes the total utility, U (G) = s, and leave otherwise. In other words, i Ui, is composed of users who choose stay when c < ki (cid:80) 6/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ s, and by definition, Bj(H)−kj node i should remain engaged in the network as long as the cost, c, does not exceed its generalized coreness, ks. In this sense, G∗ corresponds to the generalized k-core of G. To illustrate that G∗ is indeed an equilibrium network, we need to show that no user has an incentive to unilaterally join it or leave it. Consider a node, j ∈ G∗ who chooses stay. This node s ≥ 0. Since, j stayed in the would belong to a generalized k-core, kj s, therefore Bj(H) − c > 0. So, j will be forfeiting positive utility, network, it must be that c < kj had he decides to leave. In the same manner, consider another node l /∈ G∗ who chooses leave, s ≤ c. All his friends with the same coreness would have left the network, thus his coreness kl therefore the only benefit that l could obtain from staying would come from his connections with nodes in higher cores. The benefit, Bl, from such connections must not exceed kl s, otherwise s ≤ c we have Bl < c. This l would have belonged to a higher core in the first place. Since kl implies that l necessarily obtains negative utility from staying, so he has no incentives to do so. Moreover, G∗ is optimal, as we showed that any change from the equilibrium actions of any user inevitably lowers his utility and decreases the total utility in the network. We also argue that it is reasonable to expect this equilibrium network to be reached in an actual setting, since it maximizes the utility of all users simultaneously, as well as the welfare of the network provider. In the rest of the paper, we approximate Bi as proportional to the number of i’s direct friends, Ni, i.e. Bi = bNi, for some b ∈ Z. Taking ki s to be the coreness of i, by definition it holds that bNi ≥ ki s. The maximum cost, c, that i would tolerate as a member of the community must be strictly smaller than its coreness, hence bNi > c and Ni > c/b. The last result implies that the minimum number of friends that a node i needs to remain engaged must be strictly larger than s ≤ K, where c/b. Therefore, the coreness of a participating user i must be at least c/b + 1, i.e. ki capital K = (c/b) + 1. Based on the above discussion, we see that an user will remain in a network with a high c/b ratio if its coreness ks is high. This is because, by definition i is part of a connected network of nodes with large minimum degrees and hence large benefits. In contrast, simply having a large degree does not imply that an user will obtain large utility from staying. Note that a high-degree node may nevertheless have low coreness. This means that i would be part of a sub-network in which all nodes have low minimum degrees. As a result a lower c/b ratio would suffice to start a cascade of users departing, that can quickly leave i with no friends and thus drive it to leave too. Finally, we define social resilience of a community as the size of the K core. In other words, this is the size of the network that remains after all users with ks ≤ c/b have been forced out. This definition allows us to quantify social resilience and reliably compare it across communities even for unknown c/b ratios, as shown in Section 6. 7/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ 4 Data on Online Social Networks For our empirical study of social network resilience, we use datasets from five different OSN. The choice of these datasets aims at spanning a variety of success stories across OSN, including successful and failed communities, as well as communities currently in decline. The size, data gathering methods, and references are summarized in Table 1, and outlined in the following. name Livejournal Friendster Myspace Orkut Facebook status successful failed Table 1: Outline of OSN and datasets date 1999 2002 2003 2004 2004 users links 5.2M 28M 117M 2580M Internet Archive 100K 3M 3M in decline in decline successful 6.8M 223M 23M source [26] [3] [26] [14] Friendster The most recent dataset we take into account is the one retrieved by the Internet Archive, with the purpose of preserving Friendster’s information before its discontinuation. This dataset provides a high-quality snapshot of the large amount of user information that was publicly available on the site, including friend lists and interest-based groups [31]. In this article, we provide the first analysis of the social network topology of Friendster as a whole. Since some user profiles in Friendster were private, this dataset does not include their connec- tions. However, these private users would be listed as contacts in the list of their friends who were not private. We symmetrized the Friendster dataset by adding these additional links. Due to the large size of the Friendster dataset, we symmetrized the data by using Hadoop, which we distribute under a creative commons license 3. Livejournal In Livejournal, users keep personal blogs and define different types of friendship links. The information retrieval method for the creation of this dataset combined user id sampling with neighborhood exploration [26], covering more than 95% of the whole community. We choose this Livejournal dataset for its overall quality, as it provides a view of practically the whole OSN. Note that the desing of Livejournal as an OSN deviates from the other four communities analyzed here. First, Livejournal is a blog community, in which the social network functionality plays a secondary role. Second, Livejournal social links are directed, in the sense that one user can be friend of another without being friended back. In our analysis, we only take include reciprocal links, referring to previous research on its k-core decomposition [21]. By including 3web.sg.ethz.ch/users/dgarcia/Friendster-sim.tar.bz2 8/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ this dataset, we aim at comparing how different interaction mechanisms and platform designs influence social resilience. Orkut Among declining social networking sites, we include a partial dataset on Orkut [26], which was estimated to cover 11.3% of the whole community. Far from the quality of the two previous datasets, we include Orkut in our analysis due to its platform design, as this dataset includes users that did not have a limit on their amount of friends. Furthermore, Orkut has a story of local success in Brazil, losing popularity against other sites at the time of writing of this article. MySpace One of the most famous OSN in decline is Myspace, which was the leading OSN before Facebook’s success [13]. We include a relatively small dataset of 100000 users of MySpace [3], which was aimed to sample its degree distribution. This dataset was crawled through a Breadth-First Search method, providing a partial and possibly biased dataset of Myspace. We include this dataset as an exercise to study the influence of sampling biases in the analysis of social resilience. Facebook We want to complete the spectrum of success of OSN, from the collapse of Friendster to the big success of Facebook. The last dataset we include is a special crawl which aims at an unbiased, yet partial dataset as close as possible to the whole community [14]. This dataset was retrieved with a special technique based on random walks, keeping unvaried some network statistics, including Facebook’s degree ditribution. 5 Not power-law degree distributions The first step in our analysis explores the degree distributions of each OSN. The reason to do so is the epidemic properties of complex networks. Under the assumptions of epidemic models, networks with power-law degree distributions do not have an epidemic threshold [27], i.e. a “sickness” would survive within the network for an unbound amount of time and eventually infect most of the nodes. Such sickness could be a meme or a social norm, but could also be the decision of leaving the community. Therefore, we need to assess the posibility of a lower-law degree distribution, as it would pose an alternative explanation for the masive cascades of user departures. Numerous previous works have reported power-law degree distributions in social networks [3, 6, 25, 26]. Nevertheless, most of these works rely on goodness of fit statistics, and do not provide a clear test of the power-law hypothesis. It states that the degree distribution follows the following for d ≥ degmin. This is usually described as p(d) ∝ d−α, and equation p(d) = α−1 degmin often argued as valid if metrics such as R2, or F are high enough. While a high goodness of fit (cid:16) d (cid:17)−α degmin 9/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ could be sufficient for some practical applications, the empirical test of the power-law hypothesis can only be tested, and eventually rejected, through the result of a statistical test, assuming a reasonable confidence level. We followed the state-of-the-art methodology to test power laws [7], which roughly involves the Figure 2: Complementary cumulative density function and probability density functions of node degree in the five considered communities. Red lines show the ML power-law fits from (cid:100)degmin following steps. First, we created Maximum Likelihood (ML) estimators α and (cid:100)degmin for p(d). Second, we tested the empirical data above (cid:100)degmin against the power law hypothesis and we datasets that follow the fitted power law above (cid:100)degmin. The p-value is then the fraction of the have the ML estimates (cid:100)degmin and α, which define the best case in terms of the KS test, with an associated D value, and the p-value. Ultimately, a power law hypothesis cannot be rejected if (i) the p-value of the KS-test is above a chosen significance level [7], and (ii) there is a sufficiently large amount of datapoints from recorded the corresponding KS-statistics (D). Third, we repeated the KS test for 100 synthetic synthetic D values that are larger than the empirical one. Thus, for each degree distribution, we 10/22 Facebook100101102103104degree10-610-410-2100Prob[degree = d]Orkut100101102103104105degree10-810-610-410-2100Prob[degree = d]Friendster100101102103104degree10-810-610-410-2100Prob[degree = d]LiveJournal100101102103104degree10-810-610-410-2100Prob[degree = d]MySpace100101102103104105degree10-610-410-2100Prob[degree = d]100101102103104105degree10-810-610-410-2100Prob[degree > d]OrkutLiveJournalMySpaceFacebookFriendsterDavid Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ degmin to degmax [29]. We found that the degree distributions of Facebook, Friendster, Orkut and Livejournal have p-values well below any reasonable significance threshold, showing an extremely reliable empirical support to reject the power-law hypothesis (Table 2). For the case of Myspace, a KS test gives a p-value of 0.22, which can be considered high enough to not reject the power-law hypothesis [7]. Therefore Myspace satisfies the first criterion, but when looking at the range of values from degmin to degmax (roughly one order of magnitude), and the low amount of datapoints included, this KS-test composes a merely anecdotal evidence of the extreme tail of Myspace. If accepted, the power-law distribution would explain just 0.623% of the Myspace dataset. In addition, the unsupervised BFS crawling method used for this dataset has been shown to have a bias that creates artificial power-law tails [1]. This leads to the conclusion that, while we cannot fully reject the power-law hypothesis, we can safely state that the dataset does not support the hypothesis otherwise. Figure 2 shows the degree distributions and their CCDF. For each OSN, we show how the typical log-log plot of the PDF is misleading, as a simple eye inspection would suggest power-law distributions, but a robust statistical analysis disproves this possibility. dataset (cid:100)degmin Table 2: Power law fits of the analyzed datasets. p ntail α 3.6 3.3 4.6 3 3.6 2.9 × 105 81141 4918 2.8 × 105 623 Friendster LiveJournal Facebook Orkut MySpace 1311 88 423 171 2350 D 4.59 < 10−15 0.02 < 10−15 0.14 < 10−15 0.02 < 10−15 0.22 0.03 6 Empirics of OSN Resilience 6.1 K-core decomposition We computed the k-core decomposition for each of the OSN datasets we introduced in Section 4. Among those datasets, Friendster and Livejournal cover the vast majority of their respec- tive communities. Figure 3 shows a schematic representation of the k-cores of Friendster and Livejournal. Each layer of the circles corresponds to the nodes with coreness ks, with an area proportional to the amount of nodes with that coreness value. The color of each layer ranges from light blue for ks = 1, to red for ks = 304. The distribution of colors reveals a qualitative difference between both communities: Friendster has many more nodes of high coreness than Livejournal, which has a similar color range but a much larger fringe, i.e. the set of nodes with low ks. This difference indicates that, to keep together as a community, Livejournal needs to 11/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ have a much lower c/b than Friendster. This scenario is rather realistic, as Livejournal is a blog community in which users create large amounts of original content. This leads to high benefits per social link as long as users have similar interests, which seems to be the key of Livejournal’s relative success. Figure 3: Overview of the k-core decomposition for Friendster and Livejournal. Layers are colored according to ks, with areas proportional to the amount of nodes with such ks. Our theoretical argumentation, presented in Section 3.3, indicates that node coreness is a more reasonable estimator for resilience than node degree. A degree of at least ks is a necessary condition for a coreness of ks, but a high degree does not necessarily mean a high coreness. Taking Friendster an example, Figure 4 shows the boxplot for the distribution of ks versus node degree, indicating the spread of ks for nodes of similar degree. The empirical data shows that a high degree does not necessarily mean a high ks, even finding nodes with very low ks and very high degree. Nevertheless, it is clear that ks is likely to increase with degree, but mapping degree to coreness would misestimate the resilience of the community as a whole. By measuring coreness, whe can detect that some nodes belong to the fringe despite their high degree, as the coreness integrates global information about the centrality of the node. 6.2 Resilience comparison Extending the above observations, we computed the k-core decomposition of the three additional OSN, aiming at comparing their relation between their environment, measured through c/b, and the amount of users expected to be active under such conditions. We focus our analysis on the Complementary Cumulative Density Function (CCDF) of each network, defined as P (ks > K). As shown in Section 3.3, the cost-benefit-ratio c/b corresponds to a value K that determines the nodes that leave the network, which are those ks coreness below K. Under this conditions, the CCDF of ks measures the amount of nodes that will remain in 12/22 ksLivejournalFriendster1100200304David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ Figure 4: Left: boxplot of k-shell indices by degree for Friendster. Dark lines represent the mean, and dashed bars show extreme values. Boxes are arranged in the x-axis according to the middle value of their bin. Right: CCDF of ks. The horizontal dashed line shows the cut at 0.2. the network under a given c/b, allowing us to compare how each OSN would withstand the same values of cost and benefit. The right panel of Figure 4 shows the log-log CCDF of the five OSN. The first two communities to compare are Livejournal and Friendster, as the datasets on these two are the most reliable. First, the CCDF of Friendster is always above the CCDF of Livejournal. This is consistent with the structure shown in Figure 3, where it can be appreciated that Livejournal has many more nodes in the fringe than Friendster. Second, both CCDF reach comparable maximum values, regardless of the fact that Friendster was 20 times larger than Livejournal. Such skewness in the coreness of Livejournal can be interpreted as a result of a higher competition for attention, as expected from a blog community in comarison with a pure social networking site, like Friendster was. Focusing on the tails of the distributions, we can compare the patterns of resilience for envi- ronments with high K. The comparison between the resilience of these communities is heavily dependent of the value of K, as for example, Livejournal is less resilient than Facebook for values of K between 10 and 50, but more resilient below and above such interval. A similar case can be seen between Friendster and Orkut, as their CCDFS cross at 60 and 200. Thus, Friendster would be more resilient than Orkut if K lies in that interval, while Orkut would have a larger fraction of active nodes if K < 60 or K > 200. It is important to note that these comparisons are made between the reliable datasets of Friendster and Livejournal, compared with partial datasets from the other communities. 13/22 3821662086582081degreeks050100150200250300100101102ks10-410-310-210-1100P(ks>K)FriendsterFacebookOrkutMySpaceLiveJournalDavid Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ While our conclusions on the first two OSN can be seen as global findings on the community as a whole, the rest are limited to the size of the datasates available. A particularily clear example of the effect of the crawling bias is the distribution of coreness for Myspace, which shows an extreme resilience in comparison to all the other datasets, with the exception of Orkut for K < 50. As commented in Section 4 , the method used for Myspace was very biased towards nodes of high degree, leaving an unrealistic picture of the resilience of the whole community. Additionally, the method used for Facebook seems to have delivered a degree distribution close to a random sample of Facebook users, but its restarting of random walkers leaves tendrils of nodes that accumulate on the 1-core. Hence the low starting value of the CCDF of Facebook could be an artifact of this crawling method. Regardless of any crawling bias, we found that these networks have maximum coreness numbers much higher than previous results. The maximum ks found for the network of instant messaging was limited to 68 [25], and close to 100 for the OSN Cyworld [6]. Livejournal has a maximum ks of 213, Friendster of 304, Orkut of 253, and Myspace as a very deep core of ks = 414. The exception lies in the Facebook dataset, where we find a maximum ks of 74. This evidence shows that OSN can have much tigther cores than the ones found in previous research, revealing that they contain small communities with very high resilience. As a final comparison, we focus on the values of K for the catastrophic case of the networks losing 80% of their nodes, i.e. where the CCDF has a value of 0.2. The data shows that both Facebook and Livejournal would lose 80% of their users under a value of K close to 10. For the case of the unsuccessful communities of Orkut and Friendster, it requires a much worse environment, with values of K above 60. This way, the emprical data supports the idea that, under the same environmental conditions, both successful communities is less resilient than the three unsuccessful ones. This means that the topology of their social network is not enough to explain their collapse, but indicates that bad decisions in design and interface changes can spread through the network and drive many users away. 7 The Time Evolution of Friendster In this section, we describe a post hoc case study of the way how Friendster rised and collapsed, using the avaliable timing information in the dataset. 7.1 Social growth mechanism The Friendster dataset does not provide the date of creation of user accounts or social links, but it includes a user id that increased sequentially since the creation of the site. We analyzed the time series of Friendster in an event time scale, where each timestamp corresponds to the 14/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ id of each user. We measured the time distance of an edge e, which connects users u1 and u2, as the difference between the ids of these users d(e(u1, u2)) = id(u1)− id(u2). In the following, we how early users connected to later users, making the network grow. We divided the network in time slices of a width of 10 million user ids, with a last smaller slice of 7 million ids. Each of these 12 slices contains a set of nodes that have connections i) to nodes that joined the community before, ii) to nodes that joined the network afterwards, and iii) internally within the slice. This way, for the slice of time period t we can calculate its internal average degree 2Ein(t)/N (t), where Ein(t) is the set of edges between nodes in the slice t, noted as N (t). As an extension, we define Ep(t) and Ef (t) as the sets of edges towards nodes that joined the community before t (past nodes), and nodes that joined after t (future nodes). We measured the time range of connections to the past P (t) as the mean distance of the edges in Ep(t), and the rage of connections to the future F (t) as the mean distance of their future counterpart Ef (t). By definition, the amount of past nodes for the first slice is 0, equally to the amount of future nodes for the last slice. If the process of edge creation was purely random, the network would resemble an Erdös-Renyi graph with an arbitrary sequence of node ids. In such network, P (t) would steadily increase with each slice, having an expected value of N/2 for the last one, where N is the size of the network. Similarly, F (t) would decrease from N/2 at the first slice, converging at 0 in the last one. Figure 5: Schema of connectivity of Friendster users across time. Each circle represents a slice of the network of width of 10 million user ids. Blue squares represent past users and red squares represent future users, with a distance from their slice according to P (t) and F (t) respectively. The dashed lines show the expectation of these two metrics for a random graph. 15/22 time [id] -4 10 -2 10 02 104 107777xxxx2 107x4 107x 1076 107x8 107x 108id distanceDavid Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ The time evolution of the range of connections to past and future is shown in Figure 5. Each circle represents a slice of the network, with growing t from top to bottom. Their vertical alignment represents the present with respect to the slice, and each circle is connected to a blue square on the left that represents past nodes, and a red square on the right that represents future nodes. Balls have a size proportional to N (t), which keep approximately constant throughout time. The darkness of each circle is proportional to its internal connectivity Ein(t), and the width of the connections from circles to past and future squares are proportional to Ep(t) and Ef (t) respectively. Internal connectivities decrease through time, as early slices had significantly higher Ein(t). This indicates that the initial root of users of Friendster was much more tightly connected among themselves than towards other nodes, creating a denser subcommunity of old users. A possible explanation for this pattern is that Friendster started as an OSN for dating, and its design was later shifted towards generalized networking as it became popular. The squares of Figure 5 are positioned according to the mean past P (t) and future F (t) dis- tances of each slice. As a comparison with random network construction, dashed lines show their expected values as explained above. For early slices, the mean future distance is significantly lower than its random counterpart, revealing a pattern of time cohesion that limits the range of future connections. This shows a decay in the diffussion process through the offline social network, where the potential of a user to bring new users decreases through time. This suggests a possible “user expiration date” after which a user of a OSN cannot be expected to bring new users. 7.2 Resilience and decline of Friendster We combined the sequence of user ids with the k-core decomposition of Friendster to study how its resilience changed over time. In particular, we explored the relation between the coreness of users and the time when they joined the community. We divided users along the median of the distribution of coreness values, ¯ks = 6. This way, for each period of time, there is a set of users in the lower half of the distribution (ks < ¯ks), which are nodes at risk of leaving the OSN. We measure the resilience of these time-dependent parts of network as the ratio between users at risk of leaving, and the total amount of users in the slice. We created slices of 100000 user ids, calculating a point sample estimate of P (ks < ¯ks). Inset of Figure 6 shows the time evolution of this ratio, with a dark area showing 99% confidence intervals. First, we notice that the skewness of ks does not affect our statistic, as the confidence intervals are sufficiently concentrated around the point estimates. Second, whe can identify certain time periods when the new users of Friendster only connected to its fringe, having larger ratios of nodes at risk. The first moment with a peak is at the very beginning, to drop to ratios around 0.3 soon after. This shows that the set of very early users did not fully exploit the social network, 16/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ and it took a bit of time for the OSN to become more resilient. The second peak is shortly after having 22 million users, which coincides with the decay of popularity of Friendster in the US. Finally, the ratio of users at risk went above 0.5 before the community had 80 million accounts, showing a lack of cohesion as its shutdown approaches, as new users do not manage to connect to the rest. To conclude our analysis, we explored how the spread of departures captured in the k-core decomposition (see Section 3.3) can describe the collapse of Friendster as an OSN. As we do not have access to the precise amount of active users of Friendster, we proxy its value through the Google search volume of www.friendster.com. The inset of Figure 6 shows the relative weekly search volume from 2004, where the increase of popularity of Friendster is evident. At some point in 2009, Friendster introduced changes in its user interface, coinciding with some technical problems, and the rise of popularity of Facebook 4. This led to the fast decrease of active users in the community, ending on its discontinuation in 2011. Figure 6: Weekly Google search trend volume for Friendster. The red line shows the estima- tion of the remaning users in a process of unraveling. Inset: time series of fraction of nodes with ks < 6. We scale the search volumes fixing 100% as the total amount of users with coreness above 0, 68 million. At the point when the collapse of Friendster started, the search volume indicates a popularity of 78% of its maximum. We take this point to start the simulation of a user departure cascade, with an initial amount of 58 million active users, i.e. users with coreness above 3. The second reference point we take is June 2010, when Friendster was reported to have 10 million 4www.time.com/time/business/article/0,8599,1707760,00.html 17/22 dateGooglesearchvolume(%)0204060802Jul20091Jan20102Jul20101Jan201178% (52M)15% (10M)P(ks<6)0.30.50.70.9t [id]040M80M120MDavid Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ active users 5, corresponding to 15% of the 68 million user reference explained above. The search volume on that date is 14%, showing the validity of the assumption that the maximum amount of active users corresponds to those with coreness above 0. Thus, these 10 million remaining users correspond to nodes with ks > 67. Given these two reference points, we can approximate the collapse of the network through its “unraveling” per k-core. Our assumption is that a critical coreness Kt starts at 3 and increases by 1 at a constant rate. Such Kt is the result of an increasing cost-to-benefit ratio, and thus all the nodes with ks < Kt would leave the community. Then, for each timestep, the amount of remaining users would correspond to the CCDF shown in Figure 4. In our analysis, K increases at a rate of 6 per month, i.e. from 3 to 67 between our two reference point. The red line of Figure 6 shows the remaining users under this process, with dashed values after the second reference point of June 2010. We can observe that this process approximates well the decay of Friendster from the start of its decline, to its total shutdown in 2011. The R2 value for this fit is 0.972, leaving some slight underfit through 2009. This fit show the match between two approximations: on one side the search volume as an estimation of the amount of active users, and on the other side the amount of remaining users when the c/b ratio increases constantly through time. 8 Discussion In this article, we have presented the first empirical analysis of social resilience in OSN. We approached this question using a theoretical model that relates the environment of the OSN with the cascades of user departures. We showed how a generalized version of the k-core decomposition allows the empirical measurement of resilience in OSN. We provided an empirical study of social resilience across five influential OSN, including successful ones like Facebook and unsuccessful ones like Friendster. We have shown that the hypothesis of a power-law degree distribution cannot be accepted for any of these communities, discarding the epidemic properties of complex networks as a possible explanation for large-scale cascades of user departures. Our k-core analysis overcomes this limitation, quantifying social resilience as a collective phenomenon using the CCDF of node coreness. We found that the topologies of two successful sites, Livejournal and Facebook, are less resilient than the unsuccessful Friendster and Orkut. This indicates that the environmental condition of an OSN play a major role for its success. Thus, we conclude that the topology of the social network alone cannot explain the stories of success and failure of the studied OSN, and it is necessary to focus future empirical analysis in measuring these costs and benefits. Additionally, we found very high maximum coreness numbers 5en.wikipedia.org/wiki/Friendster 18/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ for most of the OSN we studied. The existence of these superconnected cores indicates that information can be spread efficiently through these OSN [21]. As a case study, we provided a detailed post hoc analysis of the changes in Friendster through time. We detect that the range of connections towards future nodes is much lower than the expectation from a random process. Using the coreness of the nodes, we could track the time dependence of the risk of leaving for new users. We found shocks that indicate periods of lower resilience of the whole community. Finally, we apply all our findings to Friendster’s collapse, fitting an approximated time series of active users through the spread of user departures predicted by the k-core decomposition. We estimated the amount of active users through search volumes, but other sources can provide more reliable data, like Alexa ranks, or last login times if provided by the site. Such datasets would allow further validations of the k-core decomposition as a measure of social resilience. Our analysis focused on the macroscopic resilience of OSN, but additional research is necessary to complete our findings. Microscopic data on user activity and churn can provide estimators for the benefits and costs of each network, to further validate the work presented here. Furthermore, the generalized k-core can be applied when user decisions are more complex than just staying or leaving the network, for example introducing heterogeneity of benefits or weights in the social links. Another open question is the role of directionality in the social network, and how to measure resilience when asymmetric relations are allowed. The benefits of users of these networks would be multidimensional, representing both the reputation of a user and the amount of information it receives from its neighborhood. The work presented here is theoretically limited to the study of monotonously increasing, convex objective functions of benefit versus active neighborhood. While empirical studies support this assumption [4, 30], it is possible to imagine a scenario where information overload decreases the net benefit of users with very large neighborhoods, creating nonlinearities where the generalized k-core is not a stable solution. We leave this questions open for further research, and the study of social resilience in other types of online communities. 9 Acknowledgements The authors would like to thank the Internet Archive for their work in crawling and curating Friendster. 19/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ References [1] D. Achlioptas, A. Clauset, D. Kempe, and C. Moore. On the bias of traceroute sampling. In STOC ’05. [2] W. Adger. Social and ecological resilience: are they related? Progress in Human Geography, 24, 2000. [3] Y.-Y. Ahn, S. Han, H. Kwak, S. Moon, and H. Jeong. Analysis of topological characteristics of huge online social networking services. In WWW ’07, 2007. [4] L. Backstrom, D. Huttenlocher, J. Kleinberg, and X. Lan. Group formation in large social networks. In KDD ’06, 2006. [5] K. Bhawalkar, J. Kleinberg, K. Lewi, T. Roughgarden, and A. Sharma. Preventing unrav- eling in social networks: the anchored k-core problem. In ICALP’12. [6] H. Chun, H. Kwak, Y.-H. Eom, Y.-Y. Ahn, S. Moon, and H. Jeong. Comparison of online social relations in volume vs interaction. In IMC ’08, 2008. [7] A. Clauset, C. R. Shalizi, and M. E. J. Newman. Power-Law Distributions in Empirical Data. SIAM Review, 51(4):661, 2009. [8] M. D. Conover, B. Gonc, A. Flammini, and F. Menczer. Partisan Asymmetries in Online Political Activity. EPJ DataScience, 1(6), 2012. [9] K. Dasgupta, R. Singh, B. Viswanathan, D. Chakraborty, S. Mukherjea, A. A. Nanavati, and A. Joshi. Social ties and their relevance to churn in mobile telecom networks. In EDBT ’08, 2008. [10] G. Dror, D. Pelleg, O. Rokhlenko, and I. Szpektor. Churn prediction in new users of Yahoo! answers. In WWW ’12 Companion, 2012. [11] A. Garas, F. Schweitzer, and S. Havlin. A k-shell decomposition method for weighted networks. New Journal of Physics, 14:083030, 2012. [12] D. Garcia, F. Mendez, U. Serdült, and F. Schweitzer. Political polarization and popularity in online participatory media. In PLEAD ’12, 2012. [13] M. Giles. A world of connections - A special report on social networking. The Economist, Jan 28th, 2010. [14] M. Gjoka, M. Kurant, C. T. Butts, and A. Markopoulou. Walking in Facebook: A Case Study of Unbiased Sampling of OSNs. In INFOCOM ’10. 20/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ [15] A. Harkins. Network Games with Perfect Complements. Warwick University Draft, unpub- lished. [16] O. Herrera and T. Znati. Modeling Churn in P2P Networks. In ANSS ’07, 2007. [17] B. A. Huberman. Big data deserve a bigger audience. Nature, 482(7385):308, 2012. [18] S. R. Kairam, D. J. Wang, and J. Leskovec. The life and death of online groups. In WSDM ’12, 2012. [19] M. Karnstedt, T. Hennessy, J. Chan, and C. Hayes. Churn in Social Networks: A Discussion Boards Case Study. In SocialCom ’10, 2010. [20] J. Kawale, A. Pal, and J. Srivastava. Churn Prediction in MMORPGs: A Social Influence Based Approach. In ICCSE ’09, 2009. [21] M. Kitsak, L. K. Gallos, S. Havlin, F. Liljeros, L. Muchnik, H. E. Stanley, and H. A. Makse. Identification of influential spreaders in complex networks. Nature Physics, 6(11):888–893, 2010. [22] J. Kleinberg. Analysis of large-scale social and information networks. Philosophical trans- actions of the Royal Society A, 371, 2013. [23] F. Kooti, K. P. Gummadi, and W. A. Mason. The Emergence of Conventions in Online Social Networks. In ICWSM ’12, Dublin, Ireland, 2012. [24] M. S. S. Laine and G. Ercal. User Groups in Social Networks: An Experimental Study on YouTube. In HICSS ’11, 2011. [25] J. Leskovec and E. Horvitz. Planetary-scale views on a large instant-messaging network. In WWW ’08, 2008. [26] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee. Measurement and analysis of online social networks. In IMC ’07, 2007. [27] R. Pastor-Satorras and A. Vespignani. Epidemic dynamics in finite size scale-free networks. Physical Review E, 65(3):1–4, 2002. [28] S. B. Seidman. Network structure and minimum degree. Social Networks, 5(3):269–287, 1983. [29] M. P. H. Stumpf and M. A. Porter. Critical truths about power laws. Science, 335(6069):665– 6, 2012. 21/22 David Garcia, Pavlin Mavrodiev, Frank Schweitzer: Social Resilience in Online Communities: The Autopsy of Friendster Submitted on February 22, 2013. http://www.sg.ethz.ch/research/publications/ [30] S. Wu, A. Das Sarma, A. Fabrikant, S. Lattanzi, and A. Tomkins. Arrival and departure dynamics in social networks. In WSDM ’13, 2013. [31] J. Yang and J. Leskovec. Defining and evaluating network communities based on ground- truth. In MDS ’12, 2012. [32] E. Zheleva, H. Sharara, and L. Getoor. Co-evolution of social and affiliation networks. In KDD ’09, 2009. 22/22
1111.0920	1	1111	2011-11-03T17:20:26	Extracting spatial information from networks with low-order eigenvectors	[ "cs.SI", "physics.soc-ph" ]	We consider the problem of inferring meaningful spatial information in networks from incomplete information on the connection intensity between the nodes of the network. We consider two spatially distributed networks: a population migration flow network within the US, and a network of mobile phone calls between cities in Belgium. For both networks we use the eigenvectors of the Laplacian matrix constructed from the link intensities to obtain informative visualizations and capture natural geographical subdivisions. We observe that some low order eigenvectors localize very well and seem to reveal small geographically cohesive regions that match remarkably well with political and administrative boundaries. We discuss possible explanations for this observation by describing diffusion maps and localized eigenfunctions. In addition, we discuss a possible connection with the weighted graph cut problem, and provide numerical evidence supporting the idea that lower order eigenvectors point out local cuts in the network. However, we do not provide a formal and rigorous justification for our observations.	cs.SI	cs	EXTRACTING SPATIAL INFORMATION FROM NETWORKS WITH LOW-ORDER EIGENVECTORS MIHAI CUCURINGU†, VINCENT D. BLONDEL ‡, AND PAUL VAN DOOREN‡ Abstract. We consider the problem of inferring meaningful spatial information in networks from incomplete information on the connection intensity between the nodes of the network. We consider two spatially distributed networks: a population migration flow network within the US, and a network of mobile phone calls between cities in Belgium. For both networks we use the eigenvectors of the Laplacian matrix constructed from the link intensities to obtain informative visualizations and capture natural geographical subdivisions. We observe that some low order eigenvectors localize very well and seem to reveal small geographically cohesive regions that match remarkably well with political and administrative boundaries. We discuss possible explanations for this observation by describing diffusion maps and localized eigenfunctions. In addition, we discuss a possible connection with the weighted graph cut problem, and provide numerical evidence supporting the idea that lower order eigenvectors point out local cuts in the network. However, we do not provide a formal and rigorous justification for our observations. Key words. Eigenvector localization, diffusion maps, visualization of large data sets, social-economic networks, community detection, gravity model. AMS subject classifications. 15A18, 92-08, 91C20, 90B18 1. Introduction. Extensive research over the last decades has greatly increased our un- derstanding of the topology and the spatial distribution of many social, biological and tech- nological networks. This paper considers the problem of inferring meaningful spatial and structural information from incomplete data sets of pairwise interactions between nodes in a network. The way people interact in many aspects of everyday life often reflect surprisingly well geopolitical boundaries. This inhomogeneity of connections in networks leads to natural divisions, and identifying such divisions can provide valuable insight into how interactions in a network are influenced by its topology. The problem of finding the so-called network communities, i.e., groups of tightly connected nodes, has been extensively studied in recent years and many community detection algorithms exist with different levels of success [10]. In this paper, we consider two particular networks: a county-to-county migration network constructed from 1995-2000 US Census data, and a city-to-city communication network built from mobile phone data over a six months period in Belgium. Communities in these net- works emerge naturally and are revealed, often at different scales, by the eigenvectors of a normalized matrix constructed from the weighted adjacency matrix of the network. We dis- cuss possible explanations for this observation by describing diffusion maps and localized eigenfunctions. In the remaining part of this introduction we report on some related contributions that deal with communities in networks and spectrum of matrices. However, in none of these contributions we were able to find an explanation of why low order eigenvectors localize so well and seem to identify meaningful geographical boundaries. One example of a recent study that is related to our work both in terms of the technique and end goal, is a paper by Ratti et al. [23]. Starting from measures of the communication intensities between counties in the UK, the authors propose a spectral modularity∗ optimiza- 1 1 0 2 v o N 3 ] I S . s c [ 1 v 0 2 9 0 . 1 1 1 1 : v i X r a Road, Princeton NJ 08544-1000 USA, email: [email protected] emails: {vincent.blondel, paul.vandooren}@uclouvain.be †Program in Applied and Computational Mathematics (PACM), Princeton University, Fine Hall, Washington ‡Department of Applied Mathematics, Universit´e catholique de Louvain, B-1348 Louvain-la-Neuve, Belgium, ∗Many popular methods for community detection in networks are based on the optimization of the modularity function, a measure of the quality of a network partition into communities. 1 2 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel tion algorithm that partitions the country into small non-overlapping geographically cohesive regions that correspond remarkably well with administrative regions. In [27], the authors Shi and Malik develop a spectral-based algorithm that solves the perceptual grouping problem in computer vision by treating the task of image segmentation as a graph partitioning problem. Their approach is to segment the graph by introducing a new global criterion called normalized cut, that measures not just the dissimilarity between different groups but also the total similarity within the groups themselves. They successfully extract global impressions of a scene and provide a hierarchical description of it. In another recent paper [24], the authors connect mobile data from Telecom Italia Mobile to a series of human activities derived from data on commercial premises advertised through the Italian version of "Yellow Pages". The eigendecomposition of a specific correlation ma- trix provides a top eigenvector which clearly indicates a common underlying pattern to mobile phone usage in Rome, while the second and third eigenvectors indicate spatial variation that is very suggestive of temporally-related and activity-related patterns. Another line of work where lower order eigenvectors provide useful information comes from the community detection literature. Newman [20] shows that the modularity of a net- work can be expressed in terms of the top eigenvalues and eigenvectors of a matrix called the modularity matrix, which plays a role in the maximization of the modularity equivalent to that played by the Laplacian in standard spectral partitioning. In related work, Richardson et al. [25] extend previously available methods for spectral optimization of modularity by in- troducing a computationally efficient algorithm for spectral tripartitioning of a network using the top two eigenvectors of the modularity matrix. Recent work [9], co-authored by one of the authors of this paper, investigates the con- straints imposed by space on the network topology, and focuses on community detection by proposing a modularity function adapted to spatial networks. The proposed methods were tested on a large mobile phone network and computer-generated benchmarks, and showed that it is possible to factor out the effect of space in order to reveal more clearly any hidden structural similarities between the nodes. Finally, we point out a recent paper of Onnela et al. [21] who investigate social networks of individuals whose most frequent geographic locations are known. The authors classify the members into groups using community detection algorithms, and explore the relationship between their topological and geographic positions. This paper is organized as follows: Section 2 is an introduction to the diffusion map technique and some of its underlying theory. Section 3 contains the results of numerical sim- ulations in which we applied diffusion maps and eigenvector colorings to the US migration data set. In Section 4 we present the outcome of similar experiments on the Belgium mobile phone data set. In Section 5, we explore the connection with localized eigenfunctions, a phe- nomenon observed before in the mathematics and physics community. Finally, the last section is a summary and a discussion of possible extensions of our approach and its usefulness in other applications. 2. Diffusion Maps and Eigenvector Colorings. This section is a brief introduction to the diffusion maps literature and references therein. We also clarify the notion of eigen- vector localizations and eigenvector coloring, that we use in subsequent sections. Diffusion maps were introduced in S. Lafon's Ph.D. Thesis [14] in 2004 as a dimensionality reduction tool, and connected data analysis and clustering techniques based on eigenvectors of similar- ity matrices with the geometric structure of non-linear manifolds. In recent years, diffusion maps have gained a lot of popularity. A nonexhaustive list of references to its underlying theory and applications includes [1, 6, 7, 8, 14]. Often called Laplacian eigenmaps, these manifold learning techniques identify significant variables that live in a lower dimensional Extracting spatial information from networks with low-order eigenvectors 3 space, while preserving the local proximity between data points. Consider a set of N points V = {x1, x2, . . . , xN} in an n-dimensional space Rn, where each point (typically) character- izes an image (or an audio stream, text string, etc.). If two images xi and xj are similar, then xi − xj is small. A popular measure of similarity between points in Rn is defined using the Gaussian kernel wij = e−xi−xj2/, for some constant , so that the closer xi is from xj, the larger wij. The matrix W = (wij)1≤i,j≤N is symmetric and has positive coefficients. To normalize W , we define the diagonal matrix D, with Dii =(cid:80)N j=1 wij and define A by A = D−1W such that every row of A sums to 1. Next, one may also define the symmetric matrix S = D−1/2W D−1/2, which can also be written as S = D1/2AD−1/2 and hence is similar to A. As a symmetric matrix, S has an orthogonal basis of eigenvectors v0, v1, . . . , vN−1 associated to the N real ordered eigenvalues 1 = λ0 ≥ λ1 ≥ . . . ≥ λN−1. If we decompose S as S = V ΛV T with V V T = V T V = I and Λ = Diag(λ0, λ1, . . . , λN−1), then A becomes A = ΨΛΦT where Ψ = D−1/2V and Φ = D1/2V . Therefore AΨ = ΨΛ and the columns of Ψ form a D-orthogonal basis of eigenvectors with columns ψ0, ψ1, . . . , ψN−1 (i.e. (cid:104)ψi, Dψj(cid:105) = 0,∀i (cid:54)= j) associated to the N real eigenvalues λ0, λ1, . . . , λN−1 such that Aψi = λiψi, for i = 0, 1, . . . N − 1. Also, ΦT A = ΛΦT implies that the columns of Φ are left eigenvectors of A, which we denote by φ0, φ1, . . . , φN−1. Since ΦT Ψ = I, it follows that the vectors φi and ψj are bi-orthonormal (cid:104)φi, ψj(cid:105) = δi,j. Note that since A is a row-stochastic matrix, λ0 = 1 and ψ0 = (1, 1, . . . , 1)T , and we disregard this trivial eigenvalue/eigenvector pair as irrelevant. Using the stochasticity of A, we can interpret it as a random walk matrix on a weighted graph G = (V, E, W ), where the set of nodes consists of the points xi, and there is an edge between nodes i and j if and ony if wij > 0. Taking this perspective, Aij denotes the transition probability from point xi to xj in one step time ∆t = P r{x(t + ) = xjx(t) = xi} = Aij. √ √ The parameter can now be interpreted in two ways. On the one hand, it is the squared radius of the neighborhood used to infer local geometric and density information, in particular wij , but it is exponentially small for points that is O(1) when xi and xj are in a ball of radius are more than apart. On the other hand, represents the the discrete time step at which the random walk jumps from one point to another. Interpreting the eigenvectors as functions over our data set, the diffusion map (also called Laplacian eigenmap) maps points from the original space to the first k eigenvectors, L : V (cid:55)→ Rk, is defined as (2.1) Lt(xj) = (λt 2ψ2(j), . . . , λt 1ψ1(j), λt kψk(j)) Aij =(cid:80)N−1 where the meaning of the exponent t will be made clear in what follows. r=0 λrφr(i)ψr(j), and note that At Using the left and right eigenvectors denoted earlier, we now write the entries of A as rφr(i)ψr(j). However, recall that the probability distribution of a random walk landing at location xj after exactly t steps, given ij = P r{x(t) = xjx(0) = xi}. that is starts at point xi is precisely given by the expression At Given the random walk interpretation, it is natural to quantify the similarity between two points according to the evolution of their probability distributions r=0 λt ij =(cid:80)N−1 N(cid:88) k=1 D2 t (i, j) = (At ik − At jk)2 1 dk , 4 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel where the weight 1 takes into account the empirical local density of the points by giving dk larger weight to the vertices of lower degree. Since Dt(i, j) naturally depends on the random walk on the graph, it is denoted as the diffusion distance at time t. In the diffusion map introduced above, it is a matter of choice to tune the parameter t corresponding to the number of time steps of the random walk. Note that we used t = 1 in the diffusion maps embeddings throughout our simulations, and that using different values of t corresponds to rescaling the axis. The Euclidean distance between two points in the diffusion map space introduced in (2.1) is given by N−1(cid:88) (cid:0)λt rψr(j)(cid:1)2 . (2.2) L(xi) − L(xj)2 = rψr(i) − λt r=1 Notice that the first eigenvalue λ0 does not enter this expression, since it cancels out. More- over, as shown in [19], the expression (2.2) equals the diffusion distance D2 t (i, j), when k = N − 1, i.e., when all N − 1 eigenvectors are considered. For ease of visualization, we used the top k = 2 eigenvectors for the projections shown in Figures 3.1, 3.2 and 4.1. Finally, we denote by Ci the coloring of the N data points given by the eigenvector ψi, where the color of point xk ∈ V is given by the j-th entry in ψi, i.e. Ci(xk) = ψi(k), for all i = 1, . . . , N and k = 0, . . . , N − 1. 1 ≥ λt k > δ but λt 2 . . . ≥ λt We refer to Ci as an eigenvector coloring† of order i. The top left plot in Figure 3.4 shows the eigenvector coloring of order k = 1, together with the associated colorbar. In practice, only the first k eigenvectors are used in the diffusion map introduced in (2.1), with k << N − 1 chosen such that λt k+1 < δ, where δ is a chosen tolerance. Typi- cally, only the top few eigenvectors of A are expected to contain meaningful information, but as illustrated by the eigenvector colorings shown in this paper, one can extract relevant infor- mation from eigenvectors of much lower order. The phenomenon of eigenvector localization occurs when most of the components of an eigenvector are zero or close to zero, and almost all the mass is localized on a relatively small subset of nodes. On the contrary, delocalized eigenvectors have most of their components small and of roughly the same magnitude. Fur- thermore, note there is no issue with the fact that the eigenvectors are defined up to a scalar. Since each of them is normalized and real, we can just consider eigenvectors of different sign, however this can only reverse the color map used, and does not change the localization phenomenon. 3. US Census Migration Data. We apply the diffusion map technique to the 2000 US Census that reports the number of people that migrated from every county to every other county in the US during the 1995-2000 time frame [5, 22]. We denote by M = (Mij)1≤i,j≤N the total number of people that migrated between county i and county j (so Mij = Mji), where N = 3107 denotes the number of counties in mainland US. We let Pi denote the population of county i. Figures 3.1 and 3.2 show the results of the diffusion map technique M 2 for longitude and latitude colorings when the following kernels are used: W (1) , ij PiPj W (2) . The diffusion map resulting from these kernels place the Midwest closer to the west coast (Figure 3.1), but further from the east coast. Similarly, the colorings based on latitude reveal the north-south separation. Note that kernel W (1) does a better job at separating the east and west coasts, Figure 3.1 (b), while kernel W (2) highlights best the separation between north and south as shown in Figure 3.2 (c). Figure 3.3 shows the ij = 5500 Mij PiPj ij = Mij Pi+Pj , and W (3) ij = †Not to be confused with the "coloring" terminology from graph theory, where the colors are integers. Extracting spatial information from networks with low-order eigenvectors 5 histogram of the top 500 eigenvalues of the diffusion matrix A, when different kernels are used. Our kernel of choice for the eigenvector colorings in Figures 3.4 and 3.5 was W (1), as it produced more visually appealing results in terms of state boundary detection. For the same reason, we omit the numerical simulations where we used exponential weights to compute the similarity between the nodes. Note also that the spectrum of A = D−1W (1) in the left of Figure 3.3 is rather different from the other two spectra, with many more large eigenvalues and without a visible spectral gap. For the rest of this section, we drop the superscript from matrix W (1) and refer to it as W . (a) Map of USA, colored by longitude (b) Kernel W (1) ij = M 2 ij PiPj (c) Kernel W (2) ij = Mij Pi+Pj (d) Kernel W (3) ij = Mij PiPj FIG. 3.1. Diffusion map reconstructions from the top two eigenvectors, for various kernels, with nodes colored by longitude. In Figure 3.6 we plot the histograms of the entries of several eigenvectors of A. Note that the top eigenvector provides a meaningful partitioning that separates the East from the Midwest, and has its entries spread in the interval [−0.03, 0.03] with few entries of zero magnitude. On the other hand, the eigenvectors φ7, φ28 and φ83 are localized in the sense that they have their larger entries localized on a specific subregion of the US map (highlighted in blue or red in the eigenvector colorings), while taking small values in magnitude on the rest of the domain. We explore in Section 5 the connection with the phenomenon of "localized eigenfunctions" of the Laplace operator. We use the rest of this section to provide a possible interpretation of the color coded regions that stand out in the eigenvector colorings in Figures 3.4 and 3.5. By interpreting the matrix W as a weighted graph, we explore a possible connection of such geographically cohesive colored subgraphs with the graph partitioning problem. In general, the graph par- titioning problem seeks to decompose a graph into K disjoint subgraphs (clusters), while minimizing the sum of the weights of the "cut" edges, i.e., edges with endpoints in different 6 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel (a) Map of USA, colored by latitude (b) Kernel W (1) ij = M 2 ij PiPj (c) Kernel W (2) ij = Mij Pi+Pj (d) Kernel W (3) ij = 5500 Mij PiPj FIG. 3.2. Diffusion map reconstructions from the top two eigenvectors, for various kernels, with nodes colored by latitude. (a) A = D−1W (1) (b) A = D−1W (2) (c) A = D−1W (3) FIG. 3.3. Histogram of the top 500 eigenvalues of matrix A for different kernels. clusters. Given the number of clusters K, the Weighted-Min-Cut problem is an optimization problem that computes a partition P1, . . . ,PK of the vertex set, by minimizing the weights of the cut edges Weighted Cut(P1, . . . ,Pk) = Ew(Pi,Pi), k(cid:88) i=1 (3.1) where Ew(X, Y ) =(cid:80) i∈X,j∈Y Wij, and X denotes the complement of X. For an extensive literature survey on spectral clustering algorithms we refer the reader to [30], and point out the popular spectral relaxation of (3.1) introduced by Shi and Malik [27]. When dividing a graph into two smaller subgraphs, one wishes to minimize the sum of the weights on the edges across two different subgraphs, and simultaneously, maximize the sum of the weights on the edges within the subgraphs. Alternatively, one tries to maximize the Extracting spatial information from networks with low-order eigenvectors 7 FIG. 3.4. Top eigenvector colorings for the similarity matrix Wij = M 2 ij PiPj . 8 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel FIG. 3.5. Selected colorings by lower order eigenvectors for the similarity matrix Wij = M 2 ij PiPj . Extracting spatial information from networks with low-order eigenvectors 9 (a) φ1 (b) φ7 (c) φ28 (d) φ83 FIG. 3.6. Histogram of the entries in the eigenvectors φ1, φ7, φ28 and φ83 of matrix A = D−1W (1) ratio between the latter quantity and the former, i.e., between the weights of the inside edges and the weights of the outside edges. To that end, we perform the following experiment, where we regard the US states as the clusters, and investigate the possibility that the isolated colored regions that emerge correspond to local cuts in the weighted graph. We denote by S the matrix of size N × N (N = 49 the number of mainland US states) that aggregates the similarities between counties at the level of states. In particular, if state i has k counties with indices x1, . . . , xk, and state j has l counties with indices y1, . . . , yl, then we consider the k × l submatrix (3.2) Wi,j = W{x1,...,xk},{y1,...,yl} and denote by Sij the sum of the kl entries in Wi,j. In other words, matrix S is a "state- collapsed" version of the matrix W , and gives a measure of similarity between pairs of states. The heatmap in figure 3.7 shows the components of the matrix S on a logarithmic scale, where the intensity of entry (i, j) denotes the aggregated similarity between states i and j. (cid:80)N We refer to the diagonal entry Sii as the "inside degree" of state i, din measures the internal similarity between the counties of state i. We denote by dout i = Sii, which i = u=1,u(cid:54)=i Si,u (i.e., the sum of the non-diagonal elements in row i) the "outside degree" of node i, which measures the similarity/migration between the counties of state i and all other counties outside of state i. Finally, we denote by dratio , the "ratio degree" of node i which straddles the boundary between intra-state and inter-state migration. A large ratio degree is a good indicative that a state is very well connected internally, and has little connectivity with the outside world, and thus is a good candidate for a cluster. In other words, a large "ratio degree" of a cluster (i.e., state) denotes a high measure of separation between that cluster and its environment, which is something discovered by the localization properties of the low-order eigenvectors. Table 3 ranks the top 15 states within the US in terms of their ratio degree. i = din dout i i Next, we examine the top several eigenvector colorings in Figure 3.4, and point out the individual states on which the eigenvectors localize, together with its rank in terms of "ratio degree". Note that the entries of large magnitude are colored in red and blue, while the rest of the spectrum denotes values of smaller magnitude or very close to zero. The top three eigenvectors correspond to global cuts between various coasts within the US. The only state that stands out individually is Michigan (MI) for k = 3, which has rank 2. For k = 4, the largest entries correspond to counties in Virginia (VA) which is also ranked 1st, and similarly for Wisconsin (WI) for k = 5, ranked 14. For k = 6, the states colored in dark red and dark blue are Georgia (GA) with rank 3, and Missouri (MO) of rank 8. When k = 7, Michigan (MI), of rank 2, stands out as the only dark blue colored state. For k = 8, we point out Georgia, rank 3, together with Mississippi (MS) of rank 11, and Louisiana (LA) of rank 10. Eigenvector k = 9 localizes mostly on Maine (ME) of rank 6, and the New York (NY) area 10 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel FIG. 3.7. Heatmap of the inter-state migration flows, where the rows and columns of the matrix are sorted by the ratio degrees of the states. The intensity of entry (i, j) denotes, on a logarithmic scale, the similarity between states i and j, i.e., the sum of all entries in the submatrix Wi,j defined in 3.2. Table 3 lists the top 15 states in terms of ratio degree. with rank 7. Finally, eigenvector k = 10 localizes on a combination of states we already pointed out. We have thus enumerated nine states that stand out in the top ten eigenvector colorings, and all nine of them appear in Table 3 that ranks the top fifteen states in terms of "ratio degree". Although this experiment does not provide a formal justification for the eigenvector localization phenomenon, we believe it is a first step in providing evidence that the low order eigenvectors point out local cuts in the network. 4. Belgium Mobile Network. In a recent work [13], we studied the anonymized mobile phone communication from a Belgian operator and derived a statistical model of interaction between cities, showing that inter-city communication intensity is characterized be a gravity model: the communication intensity between two cities is proportional to the product of their sizes divided by the square of their distance. In this section, we briefly describe the Belgium mobile data set, summarize the results in [13], and apply the diffusion map technique. We refer the reader to [2] for more information on the mobile phone data set. The data set contains anonymous communication patterns of 2.5 million mobile phone customers, grouped in 571 cities in Belgium over a period of six months in 2006 (see also [15] for a description of the data set). Every customer is associated with the ZIP code of her/his billing address. Note that calls involving other operators were filtered out, meaning that both the calling and receiving individuals in the data set are customers of the mobile phone company. Also, there is a link between two customers if at least three calls were made Extracting spatial information from networks with low-order eigenvectors 11 state ratio degree rank VA 26.7 1. MI 20.4 2. GA 19.9 3. IN 19.7 4. TX 19.0 5. ME 18.9 6. NY 18.7 7. MO 18.5 8. 17.1 CO 9. 16.6 10. LA 11. MS 16.1 15.7 CA 12. 15.6 13. OH 14.5 14. WI NC 14.4 15. Left: top 15 states within the US, ordered by ratio degree. Right: map of mainland US showing the names of TABLE 3.1 the states. in both directions during the six month interval. After this pre-processing, the network has 2.5 million nodes and 38 million links. For every pair of customers we associate a commu- nication intensity by computing the total communication time in seconds. After grouping the customers into their corresponding cities, we compute Tij, the aggregated communica- tion time in seconds between the customers of city i and j, and denote the resulting matrix by T = (Tij)1≤i<j≤n. We denote by Nij the number of phone calls between cities i and j, by Rij = Tij the average duration of a call, and by Pi the number of customers that Nij have the zip code billing address of city i (from now on, we refer to Pi as the population of city i). Furthermore, the normalized number of phone calls with respect to the population of the cities is denoted by ¯Nij = Nij , and similarly the normalized communication time by PiPj ¯Tij = Tij . Finally, D = (dij)1≤i<j≤n represents the distances between the centroids of the PiPj areas of cities i and j. Using these quantities, we now consider the following three kernels: ij = e−(Rij ¯Tij)2 W (1) Figure 4.1 shows the diffusion map reconstructions for various matrices W that relate cities based on their communication intensities and population sizes. For W (2) and W (3), there is an obvious separation between the north and south parts of Belgium, which stems from the fact that the two regions belong to different linguistic groups. The same separation is emphasized by the colorings associated to the top eigenvector of matrix A, shown in Figure 4.2. The remaining eigenvector colorings in Figure 4.2 clearly isolate various subregions in Belgium. For example, eigenvectors ψ1 and ψ11 highlight language communities (French, Dutch and German), while ψ3 and ψ5 isolate the regions of Li`ege and Limburg. (cid:18) R0.16 /0.22, W (2) , and W (3) ij = ij = e = Nij PiPj . (cid:19)2 − ij ¯N 0.26 ij ¯Tij Rij 5. Localized eigenfunctions. Let us first make more precise what is meant by a lo- calized eigenfunction. This phenomenon of localization occurs when there exist eigenfunc- tions supported by small regions of the domain, i.e. they are localized in these regions. An eigenfunction localized on a domain Ω1 has support on Ω1 significantly larger than on the complement Ω\Ω1, and yet it cannot vanish on Ω\Ω1 since eigenfunctions of isolated eigen- values are real analytic functions and cannot vanish on any open set. This is also observed 12 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel (a) Map of Belgium, colored by latitude (b) Kernel W (1) ij = e−(Rij ¯Tij)2/0.22 (cid:32) R0.16 ij ¯N 0.26 ij − (c) Kernel W (2) ij = e (cid:33)2 (d) Kernel W (3) ij = ¯Tij Rij = Nij PiPj FIG. 4.1. Diffusion map reconstructions from the top two eigenvectors, for various kernels. Tij denotes the Tij aggregated communication time in seconds, Nij the number of phone calls between cities i and j, Rij = Nij the average duration of a call, and Pi the population of city i. We normalize by the population size by defining ¯Nij = and ¯Tij = . Nij PiPj Tij PiPj in the histogram of the entries of the eigenvectors φ7, φ28 and φ83 shown in Figure 3.6 and the corresponding colorings in the figures 3.4 and 3.5. In contrast, eigenfunctions that do not localize have their support "uniformly" distributed across the domain, similar to the case of eigenvector φ1 from Figure 3.6. For example, in the case of the unit interval, the eigenfunc- tions of the Laplacian are the sine or cosine functions (depending on the boundary conditions) with the larger eigenvalues corresponding to higher oscillations, and they are not localized in the sense that there is no specific subinterval that carries the most (potential) energy of the eigenfunction, and any subinterval supports an amount of energy that is proportional to its length. In other words, the energy of the top eigenfunctions is distributed uniformly across the domain, and similar results are known to hold for the disk and the sphere, where the Laplacian eigenvalues and eigenfunctions are explicitly known. The behavior observed in the eigenvector colorings from figures 3.4, 3.5 and 4.2 is related to the notion of localized eigenfunctions, a phenomenon observed before in the mathematics and physics community. The spectrum of the (continuous) Laplace operator has been exten- sively studied, and there exists a rich literature on the relationship between the spectrum and the geometry of the domain. As more complicated objects, eigenfunctions are more difficult Extracting spatial information from networks with low-order eigenvectors 13 FIG. 4.2. Colorings by the top 18 eigenvectors of A = D−1W (3), where W (3) ij = ¯Tij Rij = Nij PiPj to analyze than the spectrum, and less is known about them. Most of the literature is fo- cused on high frequency eigenfunctions (associated to larger eigenvalues), such as [3, 4, 12], although recent studies such as [11] advocated localized eigenfunctions associated to small eigenvalues. In our experiments, we found the bottom eigenvectors uninteresting as they did not contain any meaningful geometric information. In his work, Sapoval [26] studied localized eigenfunctions in different domains and pointed out their importance for physical applications, such as designing efficient noise-protective walls. Finally, considering that A is a stochastic matrix, one may further explore ideas from 14 Mihai Cucuringu, Paul Van Dooren, Vincent D. Blondel the theory of nearly completely decomposable matrices developed in 1961 by Nobel laureate Herbert Simon and his collaborator Albert Ando to describe and identify the short, medium and long-term behaviors of a dynamical system [28]. Very recent work [18] explores this idea in the context of stochastic data clustering, and proposes a technique that uses the evolution of the system to infer information on the initial structure. 6. Summary and Discussion. We have shown how the diffusion map technique can be used to obtain informative visualizations and capture natural subdivisions within two different real networks. We find surprising that some low order eigenvectors localize very well and seem to reveal small geographically cohesive regions; it is natural to ask for an explanation for our observation. In looking at figures 3.4 and 3.5 many more questions come to mind. Are the state boundaries a consequence of people migrating within the same state or not? In other words, do states emerge as communities because of people migrating from one county to the other within the state, or because of similar migration patterns directed outside the state? Prelim- inary analysis on the migration data set in the context of local clustering on graphs supports the idea that the localized low-order eigenvectors highlight local cuts in the network. This is perhaps counter-intuitive since such low-order eigenvectors must satisfy the global require- ment of exact orthogonality with respect to all of the earlier delocalized eigenvectors, and they must do so while keeping most of their components zero or close to zero. Another ques- tion to consider is whether, besides the state boundary detection, the eigenvector colorings reveal any extra information on the intensity of the migration from one region to the other. Furthermore, inter-county migration is most common among young adults and declines as people age, and one may ask how the age composition (or income level) of individual US counties impacts the migration pattern. In answering these questions, one needs to complement the mathematical description of diffusion maps and clustering by eigenvectors with a socio-demographic behavioral inter- pretation of migration trends, as considered for example in [16, 17]. A more recent paper by Slater [29] is of particular interest since it analyzes migration patterns in the US Census data from 1965-1970 and 1995-2000. Amongst others, it highlights cosmopolitan or hub like regions, as well as isolated regions that emerge when there is a high measure of separation between a cluster and its environment. Another interesting direction worth exploring is seeing how the diffusion map recon- structions and colorings change when the matrices used are no longer symmetric. In the case of the US migration data, it may be the case that there are many states for which the most common migration destination is the major city/capital of that state (although there might be other destinations spread across the US that attract people migrating out from that state). It is therefore natural to expect that major cities will stand out in the colorings, however this is not the case in our simulations since we symmetrize the migration matrix and take into account both the in- and out-migration from a given state. 7. Acknowledgments. We would like to thank Amit Singer for introducing M.C. to diffusion maps and suggesting this as a possible approach for analyzing the Belgium mobile data set; Etienne Huens and Gautier Krings for useful discussions and help with processing the same data set; Michael Mahoney, Mauro Maggioni, Peter Mucha, Mark Newman and Mason Porter for useful discussions on community detection and references to the literature on localized eigenfunctions; Thomas Espenshade, Kevin O'Neil, John Palmer, and Ashton Verdery for references and useful discussions on demographic trends in US migration. REFERENCES Extracting spatial information from networks with low-order eigenvectors 15 [1] M. BELKIN AND P. NIYOGI, Laplacian eigenmaps for dimensionality reduction and data representation, Neural Computation, 6 (2003), pp. 1373 -- 1396. [2] V. D. BLONDEL, J.-L. GUILLAUME, R. LAMBIOTTE, AND E. LEFEBVRE, Fast unfolding of communites in large networks, Journal of Statistical Mechanics: Theory and Experiment, 1742-5468 (2008). P10008. [3] N. BURQ AND M. ZWORSKI, Bouncing ball modes and quantum chaos, SIAM Rev., 47 (2005), pp. 43 -- 49. [4] V. M. B. C AND V. F. LAZUTKIN, The eigenfunctions which are concentrated near a closed geodesic, Prob- lems of Mathematical Physics, Spectral Theory, Diffraction Problems (Russian), 2 (1967), pp. 15 -- 25. [5] U. S. CENSUS BUREAU, 2002. www.census.gov/population/www.cen2000/ctytoctyflow/ index.html. [6] R. R. COIFMAN, I. G. KEVREKIDIS, S. LAFON, M. MAGGIONI, AND B. NADLER, Diffusion maps, reduc- tion coordinates and low dimensional representation of stochastic systems, SIAM Multiscale modeling and simulation, 7 (2008), pp. 842 -- 864. [7] R. R. COIFMAN AND S. LAFON, Diffusion maps, Appl Comput Harmonic Anal, 21 (2006), pp. 5 -- 30. [8] R. R. COIFMAN, S. LAFON, A. B. LEE, M. MAGGIONI, B. NADLER, F. WARNER, AND S. W. ZUCKER, Geometric diffusions as a tool for harmonic analysis and structure definition of data: Diffusion maps, PNAS, 102 (2005), pp. 7426 -- 7431. [9] P. EXPERT, T. EVANS, V. BLONDEL, AND R. LAMBIOTTE, Uncovering space-independent communities in spatial networks, PNAS (Proceedings of the National Academy of Sciences), 108 (2011), pp. 7663 -- 7668. [10] S. FORTUNATO, Community detection in graphs, Physics Reports, 486 (2010), pp. 75 -- 174. [11] S. M. HEILMAN AND R. S. STRICHARTZ, Localized eigenfunctions: Here you see them, there you don't, Notices Amer. Math. Soc., 57 (2010), pp. 624 -- 629. [12] P. W. JONES, M. MAGGIONI, AND R. SCHUL, Manifold parameterizations by eigenfunctions of the laplacian and heat kernels, Proc. Natl. Acad. Sci. USA, 6 (2008), pp. 1803 -- 1808. [13] G. KRINGS, F. CALABRESE, C. RATTI, AND V. D. BLONDEL, Urban gravity: a model for inter-city telecommunication flows, Journal of Statistical Mechanics: Theory and Experiment, L07003 (2009). [14] S. LAFON, Diffusion maps and geometric harmonics, Ph.D. Thesis, (2004). [15] R. LAMBIOTTE, V. D. BLONDEL, C. DE KERCHOVE, E. HUENS, C. PRIEUR, Z. SMOREDA, AND P. V. DOOREN, Geographical dispersal of mobile communication networks, Physica A: Statistical Mechanics and its Applications, 387 (2008), pp. 5317 -- 5325. [16] E. S. LEE, A theory of migration, Population Association of America, Demography, 3 (1966), pp. 47 -- 57. [17] D. S. MASSEY, J. A. G. HUGO, A. KOUAOUCI, A. PELLEGRINO, AND J. E. TAYLOR, Theories of interna- tional migration: A review and appraisal, Population and Development Review, 19 (1993), pp. 431 -- 466. [18] C. D. MEYER AND C. D. WESSELL, Stochastic data clustering, (2010). submitted, arXiv:1008.1758. [19] B. NADLER, S. LAFON, R. COIFMAN, AND I. KEVREKIDIS, Diffusion maps, spectral clustering and eigen- functions of Fokker-Planck operators, vol. 18, 2005. [20] M. E. J. NEWMAN, Finding community structure in networks using the eigenvectors of matrices, Physical Review E, 74 (2006). 036104. [21] J. ONNELA, S. ARBESMAN, M. C. GONZALEZ, A. L. BARABASI, AND N. A. CHRISTAKIS, Geographic constraints on social network groups, PLoS one, 6 (2011). e16939. [22] M. J. PERRY, State-to-State Migration Flows: 1995 to 2000, Census 2000 Special Reports, (2003). [23] C. RATTI, S. SOBOLEVSKY, F. CALABRESE, C. ANDRIS, J. READES, M. MARTINO, R. CLAXTON, AND S. H. STROGATZ, Redrawing the map of Great Britain from a network of human interactions, PLoS ONE, 5 (2010). [24] J. READES, F. CALABRESE, AND C. RATTI, Eigenplaces: analyzing cities using the space-time structure of the mobile phone network, Environment and Planning B, 36 (2009), pp. 824 -- 836. [25] T. RICHARDSON, P. J. MUCHA, AND M. A. PORTER, Spectral tripartitioning of networks, Physical Review E, 80 (2009). 036111. [26] S. RUSS, B. SAPOVAL, AND O. HAEBERL ´E, Irregular and fractal resonators with Neumann boundary con- ditions: Density of states and localization, Phys. Rev. E, 55 (1997), pp. 1413 -- 1421. [27] J. SHI AND J. MALIK, Normalized cuts and image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 22 (2000), pp. 888 -- 905. [28] H. A. SIMON AND A. ANDO, Aggregation of variables in dynamic systems, Econometrica, 29 (1961), pp. 111 -- 138. [29] P. B. SLATER, Hubs and clusters in the evolving united states internal migration network, ISBER, University of California, Santa Barbara, (2008). arXiv:0809.2768v3. [30] U. VON LUXBURG, A tutorial on spectral clustering, Statistics and Computing, Springer Netherlands, 17 (2007), pp. 395 -- 416. ISSN 0960-3174.
1701.04253	1	1701	2017-01-16T11:44:53	The Quantified City: Sensing Dynamics in Urban Setting	[ "cs.SI", "cs.CY" ]	The world is witnessing a period of extreme growth and urbanization; cities in the 21st century became nerve centers creating economic opportunities and cultural values which make cities grow exponentially. With this rapid urban population growth, city infrastructure is facing major problems, from the need to scale urban systems to sustaining the quality of services for citizen at scale. Understanding the dynamics of cities is critical towards informed strategic urban planning. This paper showcases QuantifiedCity, a system aimed at understanding the complex dynamics taking place in cities. Often, these dynamics involve humans, services, and infrastructures and are observed in different spaces: physical (IoT-based) sensing and human (social-based) sensing. The main challenges the system strives to address are related to data integration and fusion to enable an effective and semantically relevant data grouping. This is achieved by considering the spatio-temporal space as a blocking function for any data generated in the city. Our system consists of three layer for data acquisition, data analysis, and data visualization; each of which embeds a variety of modules to better achieve its purpose (e.g., data crawling, data cleaning, topic modeling, sentiment analysis, named entity recognition, event detection, time series analysis, etc.) End users can browse the dynamics through three main dimensions: location, time, and event. For each dimension, the system renders a set of map-centric widgets that summarize the underlying related dynamics. This paper highlights the need for such a holistic platform, identifies the strengths of the "Quantified City" concept, and showcases a working demo through a real-life scenario.	cs.SI	cs	The Quantified City: Sensing Dynamics in Urban Setting 7 1 0 2 n a J 6 1 ] I S . s c [ 1 v 3 5 2 4 0 . 1 0 7 1 : v i X r a Tahar Zanouda, Noora Al Emadi, Sofiane Abbar, and Jaideep Srivastava Qatar Computing Research Institute Hamed Bin Khalifa University {tzanouda, nalemadi, sabbar, jsrivastava}@hbku.edu.qa Doha, Qatar ABSTRACT The world is witnessing a period of extreme growth and urbanization; cities in the 21st century became nerve centers creating economic opportunities and cultural values which make cities grow exponentially. With this rapid urban population growth, city infrastructure is facing major problems, from the need to scale urban systems to sustaining the quality of services for citizen at scale. Understanding the dynamics of cities is critical towards informed strategic urban planning. This paper showcases QuantifiedCity , a system aimed at under- standing the complex dynamics taking place in cities. Often, these dynamics involve humans, services, and in- frastructures and are observed in different spaces: phys- ical (IoT-based) sensing and human (social-based) sens- ing. The main challenges the system strives to address are related to data integration and fusion to enable an effective and semantically relevant data grouping. This is achieved by considering the spatio-temporal space as a blocking function for any data generated in the city. Our system consists of three layer for data acquisition, data analysis, and data visualization; each of which em- beds a variety of modules to better achieve its purpose (e.g., data crawling, data cleaning, topic modeling, sen- timent analysis, named entity recognition, event detec- tion, time series analysis, etc.) End users can browse the dynamics through three main dimensions: location, time, and event. For each dimension, the system ren- Copyright is held by the author/owner(s). ders a set of map-centric widgets that summarize the underlying related dynamics. This paper highlights the need for such a holistic platform, identifies the strengths of the "Quantified City" concept, and showcases a work- ing demo through a real-life scenario. 1. INTRODUCTION 'We shape our buildings, and afterwards, our build- ings shape us.' wrote Winston Churchill []. After the revolutionary growth of the modern indus- try in the late 18th century, cities have became a niche of new opportunities that attracted huge numbers of lo- cal and international migrants. In 2015, we were 53.89% to live in urban areas and this number is projected to reach 70% by the end of 2050. This fast urbaniza- tion has exposed different services and infrastructures in cities to an increasing stress, leading to a plethora of urban challenges related to human mobility (e.g., traffic congestion), public health (e.g, air pollution), neighbor- hood deprivation (e.g., lack of parks), and economical difficulties (e.g., unemployment). The situation is such that the United Nation has declared managing urban areas as one of the most important development chal- lenges of the 21st century. Luckily, the new cities that we are building are big data generators, which unleashed many initiatives to use data-driven methods in order to tackle the afore- mentioned challenges and understand more generally how cities work and operate [2]. A new body of lit- erature related to Urban Computing has been enabled by the availability of data at large scales [8]. On one hand, the advent of the Internet of Things [1] made it easier to collect real-time data about the city dynamics including: infrastructure usage, power consumption, en- vironment, health related issues, human behavior, and commutes. At the other hand, the high penetration of smart connected devices (e.g., smartphones) and the rapid adoption of social media platforms, has turned hu- mans into mobile and real-time data generators, lead- ing to what is known as social sensing [7]. Thus, re- searchers have studied different types of dynamics by mining spatio-temporal data generated either by users or sensors. Examples include using call data records (CDR) to understand mobility and commute patterns in cities [3, 4], using crowd-sourced taxi GPS data to estimate gas consumption and air pollution [6], and fi- nally using check-ins data from location-based social networks (LSBNs) to infer traffic congestion [7]. Using IoT-based sensing data has long time tradi- tion in building smart city applications such as parking management systems, whereas using social-based sens- ing data has been extensively used in building life style applications such as recommender engines for restau- rants. We aim in this paper to bridge the gap between these two spaces, by designing QuantifiedCity -- a comprehensive platform that makes it possible to lever- age data from different spaces to build smarter appli- cations. QuantifiedCity aims at quantifying the dy- namics involving humans, services and infrastructures, taking place in cities. A good example to showcase the relevance of this combination are road traffic monitoring systems. Indeed traditional systems rely on the deploy- ment of physical sensors on road segments and inter- sections to count the number of cars passing by. This allows a nice "real-time" visualization of the traffic sta- tus in different parts of the city. However, relying solely on physical sensors, makes it impossible to proceed with the diagnosis of the traffic. A red congested segment, can be so for many reasons (e.g., accident, malfunction- ing traffic light, weather condition, etc.) that are com- pletely hidden to the system. The social space provides a nice complementary as many users may have posted online in a completely unsolicited way about the event causing the traffic. Thus, combining the physical space with the social space, modulo the deployment of full- fledged intelligence to link data in the two spaces, may results in semantically richer monitoring dashboards. QuantifiedCity consists of three layers for data acquisition, data analysis, and data visualization. Its main contribution is the use of the spatio-temporal space as a blocking function to fuse data that is generated in different spaces (e.g., physical, social). This is inspired from the work of Jaro [5] in which he introduced the use of blocking methods for record linkage. The intu- ition behind our adaptation is to say that two pieces of information are related if they are generated in same spatial location and within the same time interval. The second big contribution of QuantifiedCity is the use of cross-modal data mining and machine learning tech- niques in order to better combine features from the dif- ferent spaces of data in the knowledge discovery pro- cesses [8]. We describe in the following sections the global archi- tecture of QuantifiedCity and provide some details about each of its layers. We also include a demonstra- tion of its capabilities in a real-life scenario. Figure 1: Overview of data extraction and data fusion Approach 2. THE QUANTIFIED CITY The general architecture of the QuantifiedCity is composed of three major components: (1) data acqui- sition, (2) data analysis, and (3) data visualization. The data acquisition component contains the needed modules to scrap and aggregate streams of data, as well as static geographical data sets such as census. Data streams are processed concurrently in real-time and stored in one database. The data analysis component contains modules that analyze the data. The results of the analysis are visual- ized through in a map-centric dashboard. The system is designed as a suite of independently, small, scalable modules to make it easier and simpler for development. In what follows, we provide the details of each com- ponent. 2.1 Data Acquisition This module allows the collection and fusion of sev- eral data sets. As a first step, we acquire static geo- graphical and census data 1, this later is used to spa- tially index the city's data. In order to fuse the knowl- edge from different data sources, we use spatio-temporal units to conduct our studies. We divide the city into high resolution spatial units, and the day on fine-grained time units to make it easier to study an urban phenom- ena from different angels. As an example: when we study the effects of on-road traffic on urban air pollu- tion, it's important to collect data on air quality from 1http://www.mdps.gov.qa/ar/Pages/default.aspx Social DataInternet of Things Human mobilityEvent DetectionTrafficInferenceLocation identificationContent AnalysisSQLMongoStream acquisitionInsightsSparkStatisticsRecomm-nedationsRheemHTML5 CSS3 NodeJSData AcquisitionData AnalysisData VisualizationFunctionalitiesTechnologiesMappingGiraphFigure 2: Main GUI of QuantifiedCity . This showcases some statistics around "Qatar ExxonMobil Open" tennis tournament such as overall sentiment per location (map) and time (time line), density of interactions on the map, traffic status, etc. environmental sensors, number of cars from proximity sensors, and events happening in the same neighbor- hood from social streams. The cross-referenced datasets can provide insights on relationship between the choice of local events' places and long-term air quality level in the city. This later can lead urban planners to conclu- sions that can not be identified easily. Our system captures, concurrently and continuously, data streams on the city's infrastructure and on the city's social interactions. Once captured, data streams are indexed based on time, and geographical informa- tion. Then, data is loaded to centralized geo-database. 2.2 Data analysis This module consists of a set of data mining and machine learning algorithms to enable knowledge dis- covery from cross-modal fused data. This set of intel- ligence techniques vary from content-based techniques (sentiment, NER, topic modeling, event detection, etc.), to structure-based techniques such as network analysis (community detection, etc. ). As an example, the system is able to capture and mine social content, detect events, and sense the sentiment of people. By mining social networks, and photo-sharing platforms, the system can reveal places that attract peo- ple most. Using content-analysis, we can identify places from text, link activity to places and verify whether people complain about traffic using sentiment analysis. Our results can optimize traffic light system in the city to effectively manage big events. 2.3 Data visualization As we have different mediums of information to show- case, a highly interactive system is needed to allow ef- ficient and smooth interactions with complex intercon- nected data. The visualization explorer enables users to explore the data easily and intuitively. It is mainly composed of two sub-views, namely, (1) the dashboard, (2) map view. Dashboard View: The dashboard view shows the re- sults of the stats information. The view shows different pages, categorized in main categories and can be ac- cessed by scrolling from top to bottom. Every page has different panels, and every panel contain a chart to visually communicate results of the different analysis. Map view: Using Map-centred exploratory approach allows users to easily select spatial choices, as it intu- itively provides contextual insights. In addition to the map layer, we use colored zones shapes to showcase the different results. The zones layer enables users to spec- ify the targeted zone. 3. DEMONSTRATION In this section, we showcase a detailed demonstra- tion plan and describe how different users (monitors, urban planners, decision makers, police, etc.) can in- teract with QuantifiedCity . Current EventQatar ExxonMobil OpenThe system provides contextual insights based on three dimensions: i. time, ii. location, and iii. event. As data in our system is indexed to a particular time span, users can select the beginning and the end of the time range, and the granularity of the time analysis (e.g., group data by 20 secs or 5 minutes.) Users are able to nar- row focus to a subset of data based on the selected time bounds. Similarly, users can select specific regions or a particular location to filter the data and have limited views to their selections. More interestingly, users can tap into the wealth of data around one event. Subse- quently, users can combine different geographical and temporal criteria to gain access to intuitive, yet insight- ful reports. As an example to illustrate the results, we will fo- cus on some big Tennis tournaments that took place in Qatar this January and show how QuantifiedCity captures the urban dynamics (social activity, sentiment analysis, fans mobility, traffic congestion, etc.) around these events. With the use of content-analysis meth- ods, QuantifiedCity can identify and distinguish be- tween events in the same context (i.e., "Tennis tour- nament" will be automatically associated with QATAR EXXONMOBIL OPEN for men during the first week of January, and will be linked to Qatar Total Tourna- ment for women in the last week, thanks to the spatio- temporal blocking function). The system is able to fuse the knowledge from different contextual data to estimate crowd movement patterns (from sensor data about cars moving around the stadiums, and social me- dia activity and engagement), and evaluate the com- plaints to the city managers during the event (using topic modeling and sentiment analysis). The informa- tion can be used to fuel a recommendation engine to help re-routing people from-to the game venues, it can also notify the authorities about the need of an evacu- ation plan if needed. The system can be connected to the road traffic light system to adjust crowd movement. Figure 2 shows QuantifiedCity system visualizing results of Doha city. On the right side of the dash- board, various panels allow users to see the different results of data analysis modules. On the left side, a map with zones polygons are visualized. By clicking on every zone, the system will generate statistics that will take the same form, but projected on one zone. 4. CONCLUSION In this paper, we presented QuantifiedCity , a sys- tem that captures the complex dynamics taking place in fast growing cities. The system relies on a compre- hensive set of data sources coming from different lay- ers: physical overlay through IoT-based sensing and hu- man overlay through social-based sensing which enables a holistic understanding of how cities work and oper- ate. QuantifiedCity brings a novel perspective into data integration paradigm by using the spatio-temporal space as a blocking function to link parcels of data com- ing from different sources. The intuition being that data generated in the same geographical space and time win- dow can be grouped together and used for knowledge discovery related to what is happening in that location. By bridging the gap between the physical and the social sides of cities, and through an interactive and intuitive map-centric interface, our system allows urban planners and decision makers to visually inspect problems, and interactively analyze and track in real-time the devel- opment of events. 5. REFERENCES [1] L. Atzori, A. Iera, and G. Morabito. The internet of things: A survey. Computer networks, 54(15):2787 -- 2805, 2010. [2] M. Batty, K. W. Axhausen, F. Giannotti, A. Pozdnoukhov, A. Bazzani, M. Wachowicz, G. Ouzounis, and Y. Portugali. Smart cities of the future. The European Physical Journal Special Topics, 214(1):481 -- 518, 2012. [3] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi. Understanding individual human mobility patterns. Nature, 453(7196):779 -- 782, 2008. [4] S. Grauwin, S. Sobolevsky, S. Moritz, I. G´odor, and C. Ratti. Towards a comparative science of cities: Using mobile traffic records in new york, london, and hong kong. In Computational approaches for urban environments, pages 363 -- 387. Springer, 2015. [5] M. A. Jaro. Advances in record-linkage methodology as applied to matching the 1985 census of tampa, florida. Journal of the American Statistical Association, 84(406):414 -- 420, 1989. [6] J. Shang, Y. Zheng, W. Tong, E. Chang, and Y. Yu. Inferring gas consumption and pollution emission of vehicles throughout a city. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '14, pages 1027 -- 1036, New York, NY, USA, 2014. ACM. [7] S. Wang, L. He, L. Stenneth, P. S. Yu, and Z. Li. Citywide traffic congestion estimation with social media. In Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems, SIGSPATIAL '15, pages 34:1 -- 34:10, New York, NY, USA, 2015. ACM. [8] Y. Zheng, L. Capra, O. Wolfson, and H. Yang. Urban computing: concepts, methodologies, and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 5(3):38, 2014.
1009.0842	2	1009	2010-09-26T10:55:01	Empirical study and modeling of human behaviour dynamics of comments on Blog posts	[ "cs.SI", "cs.HC", "physics.soc-ph" ]	On-line communities offer a great opportunity to investigate human dynamics, because much information about individuals is registered in databases. In this paper, based on data statistics of online comments on Blog posts, we first present an empirical study of a comment arrival-time interval distribution. We find that people interested in some subjects gradually disappear and the interval distribution is a power law. According to this feature, we propose a model with gradually decaying interest. We give a rigorous analysis on the model by non-homogeneous Poisson processes and obtain an analytic expression of the interval distribution. Our analysis indicates that the time interval between two consecutive events follows the power-law distribution with a tunable exponent, which can be controlled by the model parameters and is in interval (1,+{\infty}). The analytical result agrees with the empirical results well, obeying an approximately power-law form. Our model provides a theoretical basis for human behaviour dynamics of comments on Blog posts.	cs.SI	cs	arXiv:1009.0842v2 [physics.soc-ph] Empirical study and modeling of human behaviour dynamics of comments on Blog posts * (Business School, University of Shanghai for Science and Technology, Shanghai 200093, China) Jin-Li Guo Abstract On-line communities offer a great opportunity to investigate human dynamics, because much information about individuals is registered in databases. In this paper, based on data statistics of online comments on Blog posts, we first present an empirical study of a comment arrival-time interval distribution. We find that people interested in some subjects gradually disappear and the interval distribution is a power law. According to this feature, we propose a model with gradually decaying interest. We give a rigorous analysis on the model by non-homogeneous Poisson processes and obtain an analytic expression of the interval distribution. Our analysis indicates that the time interval between two consecutive events follows the power-law distribution with a tunable exponent, which can be controlled by the model parameters and is in interval (1,+∞). The analytical result agrees with the empirical results well, obeying an approximately power-law form. Our model provides a theoretical basis for human behaviour dynamics of comments on Blog posts. Key words: human dynamics; Poisson process; power-law distribution; comments on Blog posts. ** Project 70871082 supported by the National Natural Science Foundation of China (Grant No 70871082) and Project S30504 supported by the Shanghai Leading Academic Discipline Project (Grant No S30504). Corresponding author. E-mail address: [email protected]. 1 arXiv:1009.0842v2 [physics.soc-ph] 博客评论的人类行为动力学实证研究和建模* 郭进利（上海理工大学管理学院工业工程研究所，上海，200093）摘要：由于 Internet 上的在线通信信息可以存储在数据库中，它为研究人类动力学提供了机遇。本文基于网上博客评论的数据统计，给出评论时间间隔分布的实证研究。发现人们对于某个话题的兴趣逐渐消失，评论的时间间隔服从幂律分布。根据这个特征，我们提出一个兴趣逐渐消失的人类行为动力学模型。用非齐次 Poisson 过程对模型进行严格分析，获得事件的时间间隔分布的解析表达式。结果表明，这个模型的时间间隔服从幂律分布，幂律指数由两个参数调节，其范围在区间（1, ∞）中. 理论预测和实证结果相吻合。关键词：人类动力学；Poisson 过程；幂律分布，博客帖子评论。 1．引言人们每天参与大量不同类型的活动, 如，发电子邮件, 处理销售订单到参加娱乐活动等等。上世纪人们试图用计数过程和排队模型刻画人类行为。在这方面最简单也是应用最广的是 Poisson 过程。20 世纪初， Erlang 在研究电话交换台呼叫过程中进一步讨论了 Poisson 分布, 并将 Poisson 分布推广得到了爱尔朗分布。此后, Poisson 过程在物理学、通讯技术、交通运输和管理科学等领域得到了广泛的应用。[1-2] Poisson 过程特征为事件的到达时间间隔相互独立而且服从同一个指数分布。然而，一般人类活动的时间间隔不一定相互独立，也不一定有相同的分布。2005 年，Barabási 通过对 E-mail 和爱因斯坦等人信件的数据实证发现，人类的许多活动的时间间隔服从幂律分布，这个发现对传统人类活动模型中任务按照 Poisson 过程描述提出了挑战。 Barabási 的实证表明 *国家自然科学基金项目(批准号：70871082) 和上海市重点学科建设项目（批准号：S30504）资助 E-mail: [email protected]. 2 arXiv:1009.0842v2 [physics.soc-ph] 人类活动方式是相当不均匀的，短期高活跃状态分隔长期的休止状态[3,4]。这种不均匀是以给定任务的两次连续执行时间间隔的重尾分布为特征。引起这个特征的现象是多种多样的，人们按照优先级执行任务[3-8]、基于记忆自适应兴趣的驱动[9,10]和对网页浏览的兴趣均都有可能引起重尾分布。Grabowski 在一类社会网络中研究了人们对一件事件感兴趣的时间长度[7]，也发现了人类动力学中有趣的标度律。很多实证结果表明人类行为动力学表现出指数为 1 和 1.5 的幂律分布，以至于 Vázquez 等人在文献[11]中将人类动力学划分为两个适普类。对这两个适普类划分的质疑是周涛等人关于钱学森信件的实证 [12]。文献[13]中意识到非齐次 Poisson 过程在人类行为动力学研究中的重要性，形成了用非齐次 Poisson 过程研究人类动力学的思路。最近，文献[14]用级联式的非齐次 Poisson 过程研究人类通信的适普性。其实，人类行为是非常复杂的。人们可能对某件事情随时间推移不断的表现出浓厚兴趣，比如，吸毒和吸烟随着毒瘾越来越大吸食的次数越来越频繁。人们也可能对某件事情随时间推移兴趣越来越淡漠，甚至兴趣消失，比如，央视的综艺大观和实话实说栏目最初收视率非常高，随着时间的推移人们慢慢的对节目内容失去兴趣，现在它们的收视率已经不是很高了。有些事情和个体的差异也有很大的关系，有人对上网聊天兴趣越来越高，而有人对其越来越淡漠。再例如，如果有人在网上发一个关于学术不端的帖子。开始很多人跟贴评论，随着时间推移慢慢没有人关心此事。由此可见，人们对某事件兴趣逐渐消失的变化规律是值得研究的。兴趣消失事件的时间间隔分布是否具有幂律特征？如果具有幂律特征是否属于 Vázquez 等人的适普类？本文的目的是试图回答这些问题。 Internet 的高速发展，为人们相互交流提供了方便平台，也为研究人类行为动力学提供了数据源。在第二节中，我们分别收集网上四个敏感话题的评论数据。通过统计分析发现，虽然这四个博客话题互不相干，但是它们的共同特征是：人们开始对其兴趣极浓，时间长了兴趣消失。实证表明，这些评论的时间间隔分布具有幂律分布特征。在第三节中，根据上一节的实证研究，提出一个兴趣逐渐消失的人类行为动力学模型。在第四节中对我们提出的模型进行分析。分析表明理论预测和实证结果相吻合。第五节是结论和讨论。 2．博客评论时间间隔分布实证研究我们收集了新浪网博客和中国科学网博客上四个关于科研、教学和研究生培养话题的讨论。通过模拟和利用回归方法对主体数据进行拟合（见图 1）。新浪网一个名人博客 2008 年 3 arXiv:1009.0842v2 [physics.soc-ph] 10 月 9 日发了一个与诺贝尔奖有关的帖子 A，到 2010 年 2 月 5 日收到 721 条评论。其中， 2008 年 10 月份 99.2%条评论，2008 年 11 月份 0.4%条评论，2008 年 12 月份 0.4%条评论， 2009 年以后没有评论。由此可见，人们开始对这个话题的兴趣极浓，现在已经对它没有什么兴趣。图 1(a)是在双对数坐标中评论的时间间隔分布图，图中的直线是利用回归方法得到主体数据的拟合直线，相关系数 R=0.9644。从图 1(a)可见这个帖子的评论时间间隔分布是指数近似为 1.9777 的幂律分布。 (a) empirical data slope=-1.9777 ) t ( p 1.000 0.100 0.010 0.001 1 10 100 1000 t 10000 100000 1.000 0.100 0.010 ) t ( p 0.001 1 (c) empirical data slope=-1.6104 10 100 t 1000 1.000 0.100 0.010 ) t ( p 0.001 1 1.000 0.100 0.010 ) t ( p 0.001 1 (b) empirical data slope=-1.0448 10 1000 100 t 10000 (d) empirical data slope=-1.47 10 t 100 图 1. 双对数坐标系中评论时间间隔分布. (a) 帖子 A 的评论时间间隔分布. (b) 帖子 B 的评论时间间隔分布. (c) 帖子 C 的评论时间间隔分布. (d) 帖子 D 的评论时间间隔分布. 中国科学网博客上三个博客分别发表了帖子 B、C 和 D。2009 年 5 月 17 日发了关于科研办学敏感话题的帖子 B，到 2010 年 2 月 5 日收到 320 条评论。2009 年 5 月份占 55.4%， 2009 年 6 月份占 21%，2009 年 7 月份占 14%, 2009 年 8 月份占 9%, 2009 年 9 月份占 0%, 2009 年 10 月份占 0.3%, 11 月份占 0.3%，2009 年 12 月份以后没有评论。可见，人们对这个话题开始非常热衷，慢慢淡漠，最后完全没有了兴趣。图 1(b)是在双对数坐标中评论时间间隔分布图，利用回归方法得到主体数据的拟合直线，相关系数 R=0.8564。从图 1(b)可见这个帖子的评论时间间隔分布是指数近似为 1.0448 的幂律分布。 2009 年 6 月 4 日发了帖子 C。到 2010 年 2 月收到了 189 条评论. 其中，2009 年 6 月占 93.66%, 2009 年 7 月占 3.17%, 2009 年 10 月占 0.53%, 2009 年 12 月占 2.11%, 2010 年 1 月占 0.53%, 2010 年 2 月以后没有评论. 图 1(c)是在双对数坐标中评论时间间隔分布图，利用回归方法得到主体数据的拟合直线，相关系数 R=0.9324。从图 1(c)可见这个帖子的评论时间间隔分布是指 4 arXiv:1009.0842v2 [physics.soc-ph] 数近似为 1.6104 的幂律分布。 2010 年 1 月 6 日发了贴子 D。到 2010 年 2 月 10 日收到 171 条评论。1 月 6 日至 1 月 10 日占 63.53%, 1 月 11 日至 1 月 20 日占 16.47%, 1 月 21 日至 1 月 31 日占 13.53%, 2 月 1 日至 2 月 10 日占 6.47%.图 1(d)是在双对数坐标中评论时间间隔分布图，利用回归方法得到主体数据的拟合直线，相关系数 R=0.911。从图 1(d)可见这个帖子的评论时间间隔分布是指数近似为 1.47 的幂律分布。 3．模型 Barabási 在文献[3]中指出：当个体依据优先级参数选择任务执行时, 任务的等待时间分布服从幂律分布。Barabási 模型定义如下：任务列表有 L 项必须做的任务, 当一项新任务加入到列表中时, 以概率密度函数 )( xρ 分配给他一个优先权 0≥x 。初始状态 t=0 时, 有 L 项新任务加入列表。在每个离散时间步 t>0, 以概率 p 选择列表中有最高优先权的任务接受服务, 以概率 1-p 随机选择一项任务接受服务。选择的任务从列表中立即移出, 同时有一项新任务加入列表。文献[3]中，Barabási 在 1→p 条件下给出了任务在第τ时间步被执行的概率为: 1−≈ττP (1) 从上式可见的, 对于最高优先权首选规则的模型任务等待时间呈现指数为 1 的幂律分布。 Barabási 模型刻画了一类人们对任务选择的模型。但是，不是每类事情都可以被看作这类任务。我们有这样的经验，当人们对初次接触的新鲜事物兴趣往往很大，随着时间推移人们对其慢慢失去兴趣。上一节中关于博客评论的实证就是这类兴趣消失的例子。从实证可见，评论时间间隔服从幂律分布。我们对这类兴趣消失的人类行为动力学现象刻画为：(a) 在不同的时间间隔事件发生的次数是相互独立的；(b) 事件发生率随时间线性递减，即在时刻 t 事件的发生率为 tλ )( = b /( at + )1 ，其中， a ≥ ,0 b > 0 ；(c) 在 t 与 dt t + 之间事件发生一次的概率为 t )(λ ，短时间 dt 内事件几乎不会发生两次以上。 dt 事件的发生率为 tλ )( = b /( at + )1 体现出人们对这个事件的兴趣随时间不断衰减，长时间后兴趣完全消失。 a 越大兴趣衰减越快， b 越大兴趣衰减越慢。 5 arXiv:1009.0842v2 [physics.soc-ph] 4．模型分析随机过程 tN ({ ), ≥t }0 称为一个计数过程, 若 )(tN 表示到时刻 t 为止已发生的“事件”的总数 , 因此 , 一个计数过程满足，(ⅰ) )( ≥tN 0 ； (ⅱ) )(tN 是整数值； (ⅲ) 若 t s < , 则 sN )( ≤ tN )( ； (ⅳ) 若 t s < 时, tN − )( sN )( ts 中发生的事件个数。如果在不等于区间 ],( 相交的时间区间中发生的事件个数是独立的, 则称计数过程有独立增量。若在任意时间区间中发生的事件个数的分布只依赖于时间的区间长度, 则称计数过程有平稳增量。假设 nT 表示计数过程 tN ({ ), ≥t }0 第 n 个事件与第 1−n 个事件之间的时间间隔, 则过程的点发生时间序列为 0 =S 0 , S n n = ∑ i 1 = T i , n = L,2,1 (2) 计数过程 tN ({ ), ≥t }0 的强度(若定义的极限存在) tNt }0)( ) −Δ+ > t Δ 是一个能很好地刻画过程计数性质的量, 可以通过它来描述一个计数过程。 dt t )(λ 是在时间 (3) tNP ({ t )( λ = lim t 0 →Δ t 和 dt t + 之间事件出现的概率。齐次 Poisson 过程作为一类重要的计数过程, 它的一个主要特征是具有平稳独立增量。我们将齐次 Poisson 过程的增量平稳性移去, 得到非齐次 Poisson 过程[15]。由文献[15]可见, 我们的模型为强度为 )(tλ 非齐次 Poisson 过程。引理 1[2]. 设 tN ({ ), ≥t }0 为具有强度函数 )(tλ 的非齐次 Poisson 过程, 对于任意给定的正整数 n 和 r ，过程的 n 个点发生时间 S , S r 1 + r + 2 , L , S nr + 的联合分布密度函数是 f S r 1 + , S r + 2 , , L S nr + t ( r 1 + , t r + 2 , L , t nr + ) = n ⎧ ∏ ( ⎪ ⎪⎪ i 1 = ⎨ ⎪ ⎪ 0 ⎪ ⎩ t ( λ ir + )) exp{ − nr + t ∫ 0 1})( dxx λ r ! ⎛ ⎜ ⎝ t ∫ 0 1 + r dxx )( λ r ⎞ ⎟ ⎠ 0 < t r 1 + ≤ L ≤ t nr + 其它情形引理 2[2]. 设 tN ({ ), ≥t }0 为具有强度函数 )(tλ 的非齐次 Poisson 过程, 对于任意给定的正整数 n ，过程的 n 个点发生时间为 SS nS 1 L ,第 1+n 个事件与第 n 个事件之间的时 , , , 2 间间隔 1+nT 的条件分布函数是 6 arXiv:1009.0842v2 [physics.soc-ph] TP { n 1 + ≤ St n = x n , L , S 1 = x 1 1} −= exp{ − t x ∫ + n x n duu )( } λ = x n , L , S 1 = x 2 exp{ − S 1 , , L fx } 1 ( S n n ∏ duu )( } λ i 1 = x ( λ x n 1 − n − 2 ∫ 0 − 1 + t x n ∫ 0 dx ) x 1 , L , x n ) dx 1 L dx n ( λ x i dx ) 1 ) dx n − 2 L x 3 ∫ 0 ( λ x 2 ) dx 2 x 2 ∫ 0 ( λ x 1 ) dx 1 TP { n > t } 由引理 1 和引理 2，我们有 ∫ ∫ L x ∫ n 1 − 0 TP { n St dx dx L > = x n x 3 − + + n 1 1 1 n ∞ dx ∫ 0 ∫ 0 ∞ ( n n ∫ 0 exp{ − b n )!1 − a + t x n ∫ 0 ∫ 0 n − 1 b n )!1 − n a ∞ ∫ 0 ( n 2 n − dx 2 ∫ 0 duu x )( (} λ λ n ln ( 1 ∫ 0 x ∫ n n 0 )1 + b t ()1) ++ a 1 dx ) n ax ax xa (( y n − ∞ − dy y ( e + at ) b a = = = = ( λ x n − 1 dx + )1 n 因此，时间间隔 1+nT 的分布函数为 F Tn t 1)(1 −= + b n )!1 − n a ∫∞ 0 ( n n − 1 y y ( e + at ) b a dy ， (4) 当 0=a 时，时间间隔 1+nT 的分布函数为 tF )( = lim a 0 → F T n 1 + t lim1)( −= a 0 → b n )!1 − n a ∞ ∫ 0 ( n n − 1 y y ( e + at ) b a dy 1 −= e − bt ， (5) 式(5)表明，当 0=a 时，时间间隔 1+nT 的分布函数为指数分布，模型退化为齐次 Poisson 过程。当 0>a 时，由式(4)，我们知道时间间隔 1+nT 分布函数的密度函数为当 T 充分大时， f t )(1 + Tn = b − n 1 + )!1 n a ∫∞ 0 ( n − ( b a e + )1 y n − 1 y 1( + ae − y t ) b a + 1 dy ， (6) f n T t )(1 + ≈ b − n 1 + )!1 n a T ∫ 0 ( n − ( b a e + )1 y n − 1 y 1( + ae − y t ) b a + 1 dy ， (7) 由积分中值定理，存在 ,0[ T∈ξ ] ，使得， 7 arXiv:1009.0842v2 [physics.soc-ph] f n T t )(1 + ≈ ( b − n 1 + )!1 n a T ∫ 0 e ( n − ( b a + )1 y y n − 1 dy ) 1 1( + ae t ξ − ) = n b + ( a + 1 n ) α 1( − 1 − ( n Γ bn (,( a )!1 + T )1 )) b a + 1 1 ， (8) 1( + ae t ξ − ) b a + 1 1 = n 1 − dy 由于这里， ye y − Γ xn ,( ) 是不完全的 Gamma 函数。 ∫∞ x tF− 为时间间隔的累积概率分布，因此，从(8)可见，时间间隔 1+nT 的分布函数 )( b a 到的结果： =γ 1.997, 1.0448, 1.6104, 1.47. 因此，分析表明理论预测和实证结果相吻合。 ) ,1( ∞ 内。实证得的幂律分布。由于，故指数范围区间是幂律指数为 γ = a > ,0 b > 0 1+ 5．结论与讨论我们通过对网上贴子评论时间间隔的实证研究，发现人们对新帖子兴趣较大，随着时间的推移人们对其兴趣衰减，以至于可能失去兴趣。根据这个特征建立兴趣消失的人类行为动力学模型。通过对模型的严格分析，发现兴趣消失的人类行为动力学的事件间隔时间服从幂律分布，幂律指数可以是任何大于 1 的实数。我们的模型为实证结果： =γ 1.997, 1.0448, 1.6104, 1.47 提供了理论依据。体现出了人类动力学中幂律指数的多样性，说明人类行为动力学机制不像文献[11]所描述的那样简单，预示着人类行为动力学机制的复杂性。文献[9]中研究了兴趣驱动的人类动力学模型，这是一个时间离散模型。作者在模型中假设衰减函数为 1 + 1 ia , i = L,2,1 。文献[9]中指出：若事件在时刻 t 发生，则在时刻 1+t 发 1 生事件的概率为 a+1 . 从文献[9]中的式(1)可知，若事件在时刻 t 没有发生，则在时刻 1+t 发生事件的概率不小于 1 + ta ( 1 + )1 . 则 P{在时刻 1+t 事件发生} = P{在时刻 1+t 事件发生╱在时刻 t 事件发生}+ P{在时刻 1+t 事件发生╱在时刻 t 事件不发生} ≥ 1 a+ 1 + 1 + ta ( 1 + )1 8 arXiv:1009.0842v2 [physics.soc-ph] 如果将文献[9]中模型的衰减函数扩展为 b + 1 ia , i = L,2,1 ，则要求满足条件 b + a 1 < 1 . (9) 文献[9]中提出的模型是时间离散模型。从文献[9]中的分析可知事件是等时间间隔发生，这与博客评论的时间间隔不相符。本文提出的是时间连续模型，假设发生率（数学上也称强度函数）为 tλ )( = b + 1 ta . 对于充分小的 dt ，在时间 t 和 dt t + 之间事件出现的概率是 dt t )(λ 。因此，发生率 )(tλ 与文献[9]中的衰减函数不完全相同，也不需要条件（9）的约束。更重要的是，本文对这类问题提供了一个更严密的研究思路和估计事件发生时间间隔分布的途径。致谢与周涛教授进行了有益讨论，特此感谢。感谢国家自然科学基金项目(批准号：70871082) 和上海市重点学科建设项目（批准号：S30504）资助。参考文献： [1] Ross S M. Stochastic Processes. New York: John Wiley & Sons, inc, 1983. [2] 邓永录、梁之瞬. 随机点过程及其应用. 北京: 科学出版社, 1992. [3] Barabási A L. The origin of bursts and heavy tails in human dynamics. Nature (London), 2005,435:207-211 [4] Vázquez A. Exact Results for the Barabási Model of Human Dynamics. Phys. Rev. Lett. 2005, 95: 248701. [5] Vázquez A. Impact of memory on human dynamics,. Physica A, 2007, 373: 747-752. [6] Blanchard P and Hongle M O. Modeling human activity in the spirit of Barabasi’s queueing systems. Phys. Rev. E, 2007, 75:026102. [7] Grabowski A. Interpersonal interactions and human dynamics in a large social network. 2007 Physica A, 2007, 385: 363-369. 9 arXiv:1009.0842v2 [physics.soc-ph] [8] 郭进利. 一个人类行为动力学模型及其精确解. 物理学报. 2010, 59(6) 3851—3855. [9] Shang M S, Chen G X, Dai S X, Wang B H and Zhou T. Interest-Driven Model for Human Dynamics. Chinese Physics Letters, 2010, 27(4): 048701. [10] Han X P, Zhou T and Wang B H. Modeling human dynamics with adaptive interest. New Journal of Physics, 2008, 10: 073010 [11] Vázquez A, Oliveira J G and Dezsö Z et al. Modeling bursts and heavy tails in human dynamics. Phys. Rev. E, 2006, 73:036127. [12] Li N N, Zhang N and Zhou T. Empirical analysis on temporal statistics of human correspondence patterns. Physica A, 2008, 387:6391—6394 [13] 郭进利. 人类动力学模型与计数过程简介. 人类行为动力学模型（郭进利、周涛等主编）, 香港：上海系统科学出版社，2008, 1—10. [14] Malmgren R D, Stouffer D B, and Campanharo A S L O et al. On Universality in Human Correspondence Activity. Science, 2009, 325: 1696-1700. [15] 郭进利、史定华. 带连续强度函数的非齐次 Poisson 过程的公理化定义. 数理统计与应用概率, 1993, 8(4) : 52-57 10
1507.05224	5	1507	2017-09-20T14:26:57	Quantifying Controversy in Social Media	[ "cs.SI" ]	Which topics spark the most heated debates on social media? Identifying those topics is not only interesting from a societal point of view, but also allows the filtering and aggregation of social media content for disseminating news stories. In this paper, we perform a systematic methodological study of controversy detection by using the content and the network structure of social media. Unlike previous work, rather than study controversy in a single hand-picked topic and use domain specific knowledge, we take a general approach to study topics in any domain. Our approach to quantifying controversy is based on a graph-based three-stage pipeline, which involves (i) building a conversation graph about a topic; (ii) partitioning the conversation graph to identify potential sides of the controversy; and (iii) measuring the amount of controversy from characteristics of the graph. We perform an extensive comparison of controversy measures, different graph-building approaches, and data sources. We use both controversial and non-controversial topics on Twitter, as well as other external datasets. We find that our new random-walk-based measure outperforms existing ones in capturing the intuitive notion of controversy, and show that content features are vastly less helpful in this task.	cs.SI	cs	1 (cid:30)antifying Controversy on Social Media KIRAN GARIMELLA, Aalto University GIANMARCO DE FRANCISCI MORALES, Qatar Computing Research Institute ARISTIDES GIONIS, Aalto University MICHAEL MATHIOUDAKIS, CNRS LIRIS & INSA Lyon Which topics spark the most heated debates on social media? Identifying those topics is not only interesting from a societal point of view, but also allows the (cid:27)ltering and aggregation of social media content for disseminating news stories. In this paper, we perform a systematic methodological study of controversy detection by using the content and the network structure of social media. Unlike previous work, rather than study controversy in a single hand-picked topic and use domain-speci(cid:27)c knowledge, we take a general approach to study topics in any domain. Our approach to quantifying controversy is based on a graph-based three-stage pipeline, which involves (i) building a conversation graph about a topic; (ii) partitioning the conversation graph to identify potential sides of the controversy; and (iii) measuring the amount of controversy from characteristics of the graph. We perform an extensive comparison of controversy measures, di(cid:29)erent graph-building approaches, and data sources. We use both controversial and non-controversial topics on Twitter, as well as other external datasets. We (cid:27)nd that our new random-walk-based measure outperforms existing ones in capturing the intuitive notion of controversy, and show that content features are vastly less helpful in this task. CCS Concepts: • Information systems → Social networking sites; Social networks; • Human-centered computing → Social networks; • Applied computing → Sociology; Additional Key Words and Phrases: controversy, polarization, echo chambers, (cid:27)lter bubble, twitter ACM Reference format: Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2017. Quanti- fying Controversy on Social Media. ACM Trans. Soc. Comput. 1, 1, Article 1 (January 2017), 26 pages. DOI: 10.1145/3140565 1 INTRODUCTION Given their widespread di(cid:29)usion, online social media have become increasingly important in the study of social phenomena such as peer in(cid:30)uence, framing, bias, and controversy. Ultimately, we would like to understand how users perceive the world through the lens of their social media feed. However, before addressing these advanced application scenarios, we (cid:27)rst need to focus on the fundamental yet challenging task of distinguishing whether a topic of discussion is controversial. Our work is motivated by interest in observing controversies at societal level, monitoring their evolution, and possibly understanding which issues become controversial and why. The study of controversy in social media is not new; there are many previous studies aimed at identifying and characterizing controversial issues, mostly around political debates [1, 10, 39, 40] but also for other topics [27]. And while most recent papers have focused on Twitter [10, 27, 39, 40], controversy in other platforms, such as blogs [1] and opinion fora [2], has also been analyzed. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro(cid:27)t or commercial advantage and that copies bear this notice and the full citation on the (cid:27)rst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speci(cid:27)c permission and/or a fee. Request permissions from [email protected]. © 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM. 2469-7818/2017/1-ART1 $$15.00 DOI: 10.1145/3140565 ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:2 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis However, most previous papers have severe limitations. First, the majority of previous studies focus on controversy regarding political issues, and, in particular, are centered around long-lasting major events, such as elections [1, 10]. More crucially, most previous works can be characterized as case studies, where controversy is identi(cid:27)ed in a single carefully-curated dataset, collected using ample domain knowledge and auxiliary domain-speci(cid:27)c sources (e.g., an extensive list of hashtags regarding a major political event, or a list of left-leaning and right-leaning blogs). We aim to overcome these limitations. We develop a framework to identify controversy regarding topics in any domain (e.g., political, economical, or cultural), and without prior domain-speci(cid:27)c knowledge about the topics in question. Within the framework, we quantify the controversy associated with each topic, and thus compare di(cid:29)erent topics in order to (cid:27)nd the most controversial ones. Having a framework with these properties allows us to deploy a system in-the-wild, and is valuable for building real-world applications. In order to enable such a versatile framework, we work with topics that are de(cid:27)ned in a light- weight and domain-agnostic manner. Speci(cid:27)cally, when focusing on Twitter, a topic is speci(cid:27)ed as a text query. For example, "#beefban" is a special keyword (a "hashtag") that Twitter users employed in March 2015 to signal that their posts referred to a decision by the Indian government about the consumption of beef meat in India. In this case, the query "#beefban" de(cid:27)nes a topic of discussion, and the related activity consists of all posts that contain the query, or other closely related terms and hashtags, as explained in Section 4.1. We represent a topic of discussion with a conversation graph. In such a graph, vertices represent users, and edges represent conversation activity and interactions, such as posts, comments, mentions, or endorsements. Our working hypothesis is that it is possible to analyze the conversation graph of a topic to reveal how controversial the topic is. In particular, we expect the conversation graph of a controversial topic to have a clustered structure. This hypothesis is based on the fact that a controversial topic entails di(cid:29)erent sides with opposing points of view, and individuals on the same side tend to endorse and amplify each other's arguments [1, 2, 10]. Our main contribution is to test this hypothesis. We achieve this result by studying a large number of candidate features, based on the following aspects of activity: (i) structure of endorsements, i.e., who agrees with whom on the topic, (ii) structure of the social network, i.e., who is connected with whom among the participants in the conversation, (iii) content, i.e., the keywords used in the topic, (iv) sentiment, i.e., the tone (positive or negative) used to discuss the topic. Our study shows that all features, except from content-based ones, are useful in detecting controversial topics, to di(cid:29)erent extents. Particularly for Twitter, we (cid:27)nd endorsement features (i.e., retweets) to be the most useful. The extracted features are then used to compute the controversy score of a topic. We o(cid:29)er a systematic de(cid:27)nition and provide a thorough evaluation of measures to quantify controversy. We employ a broad range of topics, both controversial and non-controversial ones, on which we evaluate several measures, either de(cid:27)ned in this paper or coming from the literature [27, 40]. We (cid:27)nd that one of our newly-proposed measures, based on random walks, is able to discriminate controversial topics with great accuracy. In addition, it also generalizes well as it agrees with previously-de(cid:27)ned measures when tested on datasets from existing work. We also (cid:27)nd that the variance of the sentiment expressed on a topic is a reliable indication of controversy. The approach to quantifying controversy presented in this paper can be condensed into a three- stage pipeline: (i) building a conversation graph among the users who contribute to a topic, where edges signify that two users are in some form of agreement, (ii) identifying the potential sides of the controversy from the graph structure or the textual content, and (iii) quantifying the amount of controversy in the graph. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:3 The rest of this paper is organized as follows. Section 2 discusses how this work (cid:27)lls gaps in the existing literature. Subsequently, Section 3 provides a high level description of the pipeline for quantifying controversy of a topic, while Sections 4, 5, and 6 detail each stage. Section 7 shows how to extend the controversy measures from topics to users who participate in the discussion. We report the results of an extensive empirical evaluation of the proposed measures of controversy in Section 8. Section 9 extends the evaluation to a few measures that do not (cid:27)t the pipeline. We conclude in Section 10 with a discussion on possible improvements and directions for future work, as well as lessons learned from carrying out this study. 2 RELATED WORK Analysis of controversy in online news and social media has attracted considerable attention, and a number of papers have provided very interesting case studies. In one of the (cid:27)rst papers, Adamic and Glance [1] study the link patterns and discussion topics of political bloggers, focusing on blog posts about the U.S. presidential election of 2004. They measure the degree of interaction between liberal and conservative blogs, and provide evidence that conservative blogs are linking to each other more frequently and in a denser pattern. These (cid:27)ndings are con(cid:27)rmed by the more recent study of Conover et al. [10], who also study controversy in political communication regarding congressional midterm elections. Using data from Twitter, Conover et al. [10] identify a highly segregated partisan structure (present in the retweet graph, but not in the mention graph), with limited connectivity between left- and right-leaning users. In another recent work related to controversy analysis in political discussion, Mejova et al. [39] identify a signi(cid:27)cant correlation between controversial issues and the use of negative a(cid:29)ect and biased language. The papers mentioned so far study controversy in the political domain, and provide case studies centered around long-lasting major events, such as presidential elections. In this paper, we aim to identify and quantify controversy for any topic discussed in social media, including short-lived and ad-hoc ones (for example, see the topics in Table 2). The problem we study has been considered by previous work, but the methods proposed so far are, to a large degree, domain-speci(cid:27)c. The work of Conover et al. [10], discussed above, employs the concept of modularity and graph partitioning in order to verify (but not quantify) controversy structure of graphs extracted from discussion of political issues on Twitter. In a similar setting, Guerra et al. [27] propose an alternative graph-structure measure. Their measure relies on the analysis of the boundary between two (potentially) polarized communities, and performs better than modularity. Di(cid:29)erently from these studies, our contribution consists in providing an extensive study of a large number of measures, including the ones proposed earlier, and demonstrating clear improvement over those. We also aim at quantifying controversy in diverse and in-the-wild settings, rather than carefully-curated domain-speci(cid:27)c datasets. In a recent study, Morales et al. [40] quantify polarity via the propagation of opinions of in(cid:30)uential users on Twitter. They validate their measure with a case study from Venezuelan politics. Again, our methods are not only more general and domain agnostic, but they provide more intuitive results. In a di(cid:29)erent approach, Akoglu [2] proposes a polarization metric that uses signed bipartite opinion graphs. The approach di(cid:29)ers from ours as it relies on the availability of this particular type of data, which is not as readily available as social-interaction graphs. Similarly to the papers discussed above, in our work we quantify controversy based on the graph structure of social interactions. In particular, we assume that controversial and polarized topics induce graphs with clustered structure, representing di(cid:29)erent opinions and points of view. This assumption relies on the concept of "echo chambers," which states that opinions or beliefs stay ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:4 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis Table 1. Summary of related work for identfying/quantifying controversial topics. Paper Choi et al. [8] Popescu and Pennacchiotti [45] Mejova et al. [39] Klenner et al. [33] Tsytsarau et al. [49] Dori-Hacohen and Allan [14] Jang et al. [29] Conover et al. [10] Coletto et al. [9] Akoglu [2] Amin et al. [3] Guerra et al. [27] Morales et al. [40] Garimella et al. [20] Identifying Quantifying Content Network inside communities created by like-minded people, who reinforce and endorse the opinions of each other. This phenomenon has been quanti(cid:27)ed in many recent studies [4, 18, 26]. A di(cid:29)erent direction for quantifying controversy, followed by Choi et al. [8] and Mejova et al. [39], relies on text and sentiment analysis. Both studies focus on language found in news articles. In our case, since we are mainly working with Twitter, where text is short and noisy, and since we are aiming at quantifying controversy in a domain-agnostic manner, text analysis has its limitations. Nevertheless, we experiment with incorporating content features in our approach. A summary of related work categorized along di(cid:29)erent dimensions is presented in Table 1. As we mention above, most existing work to date tries to identify controversial topics as case studies on a particular topic, either using content or networks of interactions. Our work is one of the few that quanti(cid:27)es the degree of controversy using language and domain independent methods. Section 8 shows that our method outperforms existing ones [27, 40]. Finally, our (cid:27)ndings on controversy have several potential applications on news-reading and public-debate scenarios. Quantifying controversy can provide a basis for analyzing the "news diet" of readers [34, 35], o(cid:29)ering the chance of better information by providing recommendations of contrarian views [41], deliberating debates [16], and connecting people with opposing views [15, 25]. 3 PIPELINE Our approach to measuring controversy is based on a systematic way of characterizing social media activity. We employ a pipeline with three stages, namely graph building, graph partitioning, and controversy measure, as shown in Figure 1. The (cid:27)nal output of the pipeline is a value that measures Fig. 1. Block diagram of the pipeline for computing controversy scores. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. Graph BuildingGraph PartitioningControversy Measure(cid:30)antifying Controversy on Social Media 1:5 how controversial a topic is, with higher values corresponding to higher degree of controversy. We provide a high-level description of each stage here and more details in the sections that follow. 3.1 Building the Graph The purpose of this stage is to build a conversation graph that represents activity related to a single topic of discussion. In our pipeline, a topic is operationalized as a set of related hashtags (details in §4.1), and the social media activity related to the topic consists of those items (e.g., posts) that match this set of hashtags. For example, the query might consist simply of a keyword, such as "#ukraine", in which case the related activity consists of all tweets that contain that keyword, or related tags such as #kyiv and #stoprussianaggression. Even though we describe textual queries in standard document-retrieval form, in principle queries can take other forms, as long as they are able to induce a graph from the social media activity (e.g., RDF queries, or topic models). Each item related to a topic is associated with one user who generated it, and we build a graph where each user who contributed to the topic is assigned to one vertex. In this graph, an edge between two vertices represents endorsment, agreement, or shared point of view between the corresponding users. Section 4 details several ways to build such a graph. 3.2 Partitioning the Graph In the second stage, the resulting conversation graph is fed into a graph partitioning algorithm to extract two partitions (we defer considering multi-sided controversies to further study). Intuitively, the two partitions correspond to two disjoint sets of users who possibly belong to di(cid:29)erent sides in the discussion. In other words, the output of this stage answers the following question: "assuming that users are split into two sides according to their point of view on the topic, which are these two sides?" Section 5 describes this stage in further detail. If indeed there are two sides which do not agree with each other –a controversy– then the two partitions should be loosely connected to each other, given the semantic of the edges. This property is captured by a measure computed in the third and (cid:27)nal stage of the pipeline. 3.3 Measuring Controversy The third and last stage takes as input the graph built by the (cid:27)rst stage and partitioned by the second stage, and computes the value of a controversy measure that characterizes how controversial the topic is. Intuitively, a controversy measure aims to capture how sell separated the two partitions are. We test several such measures, including ones based on random walks, betweenness centrality, and low-dimensional embeddings. Details are provided in Section 6. 4 GRAPH BUILDING This section provides details about di(cid:29)erent approaches to build graphs from raw data. We use posts on Twitter to create our datasets.1 Twitter is a natural choice for the problem at hand, as it represents one of the main fora for public debate in online social media, and is often used to report news about current events. Following the procedure described in Section 3.1, we specify a set of queries (indicating topics), and build one graph for each query. We choose a set of topics which is balanced between controversial and non-controversial ones, so as to test for both false positives and false negatives. We use Twitter hashtags as queries. Users commonly employ hashtags to indicate the topic of discussion their posts pertain to. Then, we de(cid:27)ne a topic as the set of hashtags related to the given query. Among the large number of hashtags that appear in the Twitter stream, we consider those 1From the full Twitter (cid:27)rehose stream. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:6 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis Table 2. Datasets statistics: hashtag, sizes of the follow and retweet graphs, and description of the event. The top group represent controversial topics, while the bo(cid:29)om one represent non-controversial ones. Description and collection period (2015) E 204 332 Government of India bans beef, Mar 2–5 155 904 Death of Boris Nemtsov, Feb 28–Mar 2 2 009 277 Netanyahu's speech at U.S. Congress, Mar 3–5 Protests after death of Boris Nemtsov ("march"), Mar 1–2 42 553 131 566 Controversial Indian documentary, Mar 1–5 690 944 Riots in Baltimore after police kills a black man, Apr 28–30 880 814 286 603 Ukraine con(cid:30)ict, Feb 27–Mar 2 841 466 Gun violence in U.S., Jun 1–30 1 211 956 Debate during the U.K. national elections, May 3 Indiana pizzeria refuses to cater gay wedding, Apr 2–5 Hashtag # Tweets #beefban #nemtsov #netanyahuspeech #russia_march #indiasdaughter #baltimoreriots #indiana #ukraine #gunsense #leadersdebate #sxsw #1dfamheretostay #germanwings #mothersday #nepal #ultralive #FF #jurassicworld #wcw #nationalkissingday 422 908 371 732 1 196 215 317 885 776 109 1 989 360 972 585 514 074 1 022 541 2 099 478 343 652 501 960 907 510 1 798 018 1 297 995 364 236 408 326 724 782 156 243 165 172 Retweet graph E 30 180 77 330 280 375 17 662 144 935 432 621 74 214 91 764 58 514 136 290 11 003 26 819 39 075 176 915 57 544 15 544 7646 32 515 11 809 4816 V 21 590 43 114 122 884 10 883 68 608 289 483 43 252 50 191 30 096 54 102 9304 15 292 29 763 155 599 40 579 9261 5401 26 407 10 674 4638 Follow graph V 9525 17 717 49 081 4844 38 302 214 552 21 909 31 225 17 335 22 498 4558 3151 2111 2225 4242 2113 3899 4395 3264 790 SXSW conference, Mar 13–22 91 356 20 275 Last OneDirection concert, Mar 27–29 7329 Germanwings (cid:30)ight crash, Mar 24–26 14 160 Mother's day, May 8 42 833 Nepal earthquake, Apr 26–29 16 070 Ultra Music Festival, Mar 18–20 63 672 31 802 23 414 Women crush Wednesdays, Jun 17 5927 National kissing day, Jun 19 Follow Friday, Jun 19 Jurassic World movie, Jun 12-15 that were trending during the period from Feb 27 to Jun 15, 2015. By manual inspection we (cid:27)nd that most trending hashtags are not related to controversial discussions [19]. We (cid:27)rst manually pick a set of 10 hashtags that we know represent controversial topics of discussion. All hashtags in this set have been widely covered by mainstream media, and have generated ample discussion, both online and o(cid:31)ine. Moreover, to have a dataset that is balanced between controversial and non-controversial topics, we sample another set of 10 hashtags that represent non-controversial topics of discussion. These hashtags are related mostly to "soft news" and entertainment, but also to events that, while being impactful and dramatic, did not generate large controversies (e.g., #nepal and #germanwings). In addition to our intuition that these topics are non-controversial, we manually check a sample of tweets, and we are unable to identify any clear instance of controversy.2 As a (cid:27)rst step, we now describe the process of expanding a single hashtag into a set of related hashtags which de(cid:27)ne the topic. The goal of this process is to broaden the de(cid:27)nition of a topic, and ultimately improve the coverage of the topic itself. 4.1 From hashtags to topics In the literature, a topic is often de(cid:27)ned by a single hashtag. However, this choice might be too restrictive in many cases. For instance, the opposing sides of a controversy might use di(cid:29)erent hashtags, as the hashtag itself is loaded with meaning and used as a means to express their opinion. Using a single hashtag may thus miss part of the relevant posts. To address this limitation, we extend the de(cid:27)nition of topic to be more encompassing. Given a seed hashtag, we de(cid:27)ne a topic as a set of related hashtags, which co-occur with the seed hashtag. To (cid:27)nd related hashtags, we employ (and improve upon) a recent clustering algorithm tailored for the purpose [17]. 2Code and networks used in this work are available at http://github.com/gvrkiran/controversy-detection. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:7 Feng et al. [17] develop a simple measure to compute the similarity between two hashtags, which relies on co-occurring words and hashtags. The authors then use this similarity measure to (cid:27)nd closely related hashtags and de(cid:27)ne clusters. However, this simple approach presents one drawback, in that very popular hashtags such as #(cid:29) or #follow co-occur with a large number of hashtags. Hence, directly applying the original approach results in extremely noisy clusters. Since the quality of the topic a(cid:29)ects critically the entire pipeline, we want to avert this issue and ensure minimal noise is introduced in the expanded set of hashtags. Therefore, we improve the basic approach by taking into account and normalizing for the popularity of the hashtags. Speci(cid:27)cally, we compute the document frequency of all hashtags on a random 1% sample of the Twitter stream,3 and normalize the original similarity score between two hashtags by the inverse document frequency. The similarity score is formally de(cid:27)ned as sim(hs , ht) = 1 1 + log(d f (ht)) (α cos(Ws ,Wt) + (1 − α) cos(Hs , Ht)) , (1) where hs is the seed tag, ht is the candidate tag, Wx and Hx are the sets of words and hashtags that co-occur with hashtag hx, respectively, cos is the cosine similarity between two vectors, d f is the document frequency of a tag, and α is a parameter that balances the importance of words compared to hashtags in a post. By using the similarity function in Equation 1, we retrieve the top-k most similar hashtags to a given seed. The set of these hashtags along with the initial seed de(cid:27)nes the topic for the given seed hashtag. The topic is used as a (cid:27)lter to get all tweets which contain at least one of the hashtags in the topic. In our experiments we use α = 0.3 (as proposed by Feng et al. [17]) and k = 20. Figure 2 shows the top-20 most similar hashtags for two di(cid:29)erent seeds: (a) #baltimoreriots, which identi(cid:27)es the discussion around the Baltimore riots against police violence in April 2015 and (b) #netanyahuspeech, which identi(cid:27)es the discussion around Netanyahu's speech at the US congress in March 2015. By inspecting the sets of hashtags, it is possible to infer the nature of the controversy for the given topic, as both sides are represented. For instance, the hashtags #istandwithisrael and #shutupbibi represent opposing sides in the dicussion raised by Netanyahu's speech. Both hashtags are recovered by our approach when #netanyahuspeech is provided as the seed hashtag. It is also clear why using a single hashtag is not su(cid:28)cient to de(cid:27)ne a topic: the same user is not likely to use both #safespacetoriot and #segregatenow, even though the two hashtags refer to the same event (#baltimoreriots). 4.2 Data aspects For each topic, we retrieve all tweets that contain one of its hashtags and that are generated during the observation window. We also ensure that the selected hashtags are associated with a large enough volume of activity. Table 2 presents the (cid:27)nal set of seed hashtags, along with their description and the number of related tweets.4 For each topic, we build a graph G where we assign a vertex to each user who contributes to it, and generate edges according to one of the following four approaches, which capture di(cid:29)erent aspects of the data source. 1. Retweet graph. Retweets typically indicate endorsement.5 Users who retweet signal endorse- ment of the opinion expressed in the original tweet by propagating it further. Retweets are not constrained to occur only between users who are connected in Twitter's social network, but users are allowed to retweet posts generated by any other user. 3From the Twitter Streaming API https://dev.twitter.com/streaming/reference/get/statuses/sample. 4We use a hashtag in Russian, 5We do not consider 'quote retweets' (retweet with a comment added) in our analysis. , which we refer to as #russia_march henceforth, for convenience. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. #марш1:8 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis (a) (b) Fig. 2. Sets of related hashtags for the topics (a) #baltimoreriots and (b) #netanyahuspeech. We select the edges for graph G based on the retweet activity in the topic: an edge exists between two users u and v if there are at least two (τ = 2) retweets between them that use the hashtag, irrespective of direction. We remark that, in preliminary experimentation with this approach, building the retweet graph with a threshold τ = 1 did not produce reliable results. We presume that a single retweet on a topic is not enough of a signal to infer endorsement. Using τ = 2 retweets as threshold proves to be a good trade-o(cid:29) between high selectivity (which hinders analysis) and noise reduction. The resulting size for each retweet graph is listed in Table 2. In an earlier version of this work [20], when building a conversation graph for a single hashtag, we created an edge between two vertices only if there were "at least two retweets per edge" (in either direction) between the corresponding pair of users. When de(cid:27)ning topics as sets of hashtags, there are several ways to generalize this (cid:27)ltering step. The simplest approach considers "two of any" in the set of hashtags that de(cid:27)nes the topic. However, this approach is too permissive, and results in an overly-inclusive graph, with spurious relationships and a high level of noise. Instead, we opt to create an edge between two nodes only if there are at least two retweets for any given hashtag between the corresponding pair of users. In other words, the resulting conversation graph for the topic is the union of the retweet graphs for each hashtag in the topic, considered (and (cid:27)ltered) separately. 2. Follow graph. In this approach, we build the follow graph induced by a given hashtag. We select the edges for graph G based on the social connections between Twitter users who employ the given hashtag: an edge exists between users u and v if u follows v or vice-versa. We stress that the graph G built with this approach is topic-speci(cid:27)c, as the edges in G are constrained to connections between users who discuss the topic that is speci(cid:27)ed as input to the pipeline. The rationale for using this graph is based on an assumption of the presence of homophily in the social network, which is a common trait in this setting. To be more precise, we expect that on a given topic people will agree more often than not with people they follow, and that for a controversial topic this phenomenon will be re(cid:30)ected in well-separated partitions of the resulting graph. Note that using the entire social graph would not necessarily produce well-separated partitions that correspond to single topics of discussion, as those partitions would be "blurred" by the existence of additional edges that are due to other reasons (e.g., o(cid:31)ine social connections). On the practical side, while the retweet information is readily available in the stream of tweets, the social network of Twitter is not. Collecting the follower graph thus requires an expensive crawling phase. The resulting graph size for each follow graph is listed in Table 2. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:9 (a) (b) (c) (d) (e) (f) (g) (h) Fig. 3. Sample conversation graphs with retweet (top) and follow (bo(cid:29)om) aspects (visualized using the force- directed layout algorithm in Gephi). The le(cid:28) side is controversial, (a,e) #beefban, (b,f) #russia_march, while the right side is non-controversial, (c,g) #sxsw, (d,h) #germanwings. Only the largest connected component is shown. 3. Content graph. We create the edges of graph G based on whether users post instances of the same content. Speci(cid:27)cally, we experiment with the following three variants: create an edge between two vertices if the users (i) use the same hashtag, other than the ones that de(cid:27)nes the topic, (ii) share a link to the same URL, or (iii) share a link with the same URL domain (e.g., cnn.com is the domain for all pages on the website of CNN). 4. Hybrid content & retweet graph. We create edges for graph G according to a state-of-the-art process that blends content and graph information [46]. Concretely, we associate each user with a vector of frequencies of mentions for di(cid:29)erent hashtags. Subsequently, we create edges between pairs of users whose corresponding vectors have high cosine similarity, and combine them with edges from the retweet graph, built as described above. For details, we refer the interested reader to the original publication [46]. 5 GRAPH PARTITIONING As previously explained, we use a graph partitioning algorithm to produce two partitions on the conversation graph. To do so, we rely on a state-of-the-art o(cid:29)-the-shelf algorithm, METIS [32]. Figure 3 displays the two partitions returned for some of the topics on their corresponding retweet and follow graphs (Figures 3(a)-(d) and Figures 3(e)-(h), respectively).6 The partitions are depicted in blue or red. The graph layout is produced by Gephi's ForceAtlas2 algorithm [28], and is based solely on the structure of the graph, not on the partitioning computed by METIS. Only the largest 6Other topics show similar trends. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:10 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis (a) (b) Fig. 4. Partitions obtained for (a) #beefban, (b) #russia_march by using the hybrid graph building approach. The partitions are more noisy than those in Figures 3(a,b). connected component is shown in the visualization, though in all the cases the largest connected component contains more than 90% of nodes. From an initial visual inspection of the partitions identi(cid:27)ed on retweet and follow graphs, we (cid:27)nd that the partitions match well with our intuition of which topics are controversial (the partitions returned by METIS are well separated for controversial topics). To make sure that this initial assessment of the partitions is not an artifact of the visualization algorithm we use, we try other layouts o(cid:29)ered by Gephi. In all cases we observe similar patterns. We also manually sample and check tweets from the partitions, to verify the presence of controversy. While this anecdotal evidence is hard to report, indeed the partitions seem to capture the spirit of the controversy.7 On the contrary, the partitions identi(cid:27)ed on content graphs fail to match our intuition. All three variants of the content-based approach lead to sparse graphs and highly overlapping partitions, even in cases of highly controversial issues. The same pattern applies for the hybrid approach, as shown in Figure 4. We also try a variant of the hybrid graph approach with vectors that represent the frequency of di(cid:29)erent URL domains mentioned by a user, with no better results. We thus do not consider these approaches to graph building any further in the remainder of this paper. Finally, we try graph partitioning algorithms of other types. Besides METIS (cut based), we test spectral clustering, label propagation, and a(cid:28)liation-graph-based models. The di(cid:29)erence among these methods is not signi(cid:27)cant, however from visual inspection METIS generates the cleanest partitions. 6 CONTROVERSY MEASURES This section describes the controversy measures used in this work. For completeness, we describe both those measures proposed by us (§6.1, 6.3, 6.4) as well as the ones from the literature that we use as baselines (§6.5, 6.6). 7For instance, of these two tweets for #netanyahuspeech from two users on opposing sides, one is clearly supporting the speech https://t.co/OVeWB4XqIg, while the other highlights the negative reactions to it https://t.co/v9RdPudrrC. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:11 6.1 Random walk This measure uses the notion of random walks on graphs. It is based on the rationale that, in a controversial discussion, there are authoritative users on both sides, as evidenced by a large degree in the graph. The measure captures the intuition of how likely a random user on either side is to be exposed to authoritative content from the opposing side. Let G(V , E) be the graph built by the (cid:27)rst stage and its two partitions X and Y, (X ∪ Y = V , X ∩ Y = ∅) identi(cid:27)ed by the second stage of the pipeline. We (cid:27)rst distinguish the k highest-degree vertices from each partition. High degree is a proxy for authoritativeness, as it means that a user has received a large number of endorsements on the speci(cid:27)c topic. Subsequently, we select one partition at random (each with probability 0.5) and consider a random walk that starts from a random vertex in that partition. The walk terminates when it visits any high-degree vertex (from either side). We de(cid:27)ne the Random Walk Controversy (RWC) measure as follows. "Consider two random walks, one ending in partition X and one ending in partition Y , RWC is the di(cid:29)erence of the probabilities of two events: (i) both random walks started from the partition they ended in and (ii) both random walks started in a partition other than the one they ended in." The measure is quanti(cid:27)ed as RWC = PX X PY Y − PY X PX Y , (2) where PAB, A, B ∈ {X , Y} is the conditional probability PAB = Pr[start in partition A end in partition B]. (3) The aforementioned probabilities have the following desirable properties: (i) they are not skewed by the size of each partition, as the random walk starts with equal probability from each partition, and (ii) they are not skewed by the total degree of vertices in each partition, as the probabilities are conditional on ending in either partition (i.e., the fraction of random walks ending in each partition is irrelevant). RWC is close to one when the probability of crossing sides is low, and close to zero when the probability of crossing sides is comparable to that of staying on the same side. 6.2 An e(cid:27)icient variant of the random walk controversy score The most straightforward way to compute RWC is via Monte Carlo sampling. We use this approach in an earlier version of this work [20], with samples of 10 000 random walks. Nevertheless, collecting a large number of samples is computationally intensive, and leads to slow evaluation of RWC. In this section, we propose a variant of RWC de(cid:27)ned as a special case of a random walk with restart – thus leading to a much more e(cid:28)cient computation. This variant can handle cases where the random walker gets stuck (i.e., dangling vertices), by using restarts. This feature is important for two reasons: (i) retweet graphs (one of our main considerations in this paper) are inherently directed, hence the direction of endorsement should be taken into account, and (ii) since these directed graphs are very often star-like, there are a few authoritative users who generate information that spreads through the graph. Our previous Monte Carlo sampling does not take into consideration such graph structure, and the direction of information propagation, as the random walk process needs to be made ergodic for the sampling process to function. To de(cid:27)ne the proposed variant of RWC, we assume there are two sides for a controversy, de(cid:27)ned as two disjoint sets of vertices X and Y. In the original de(cid:27)nition of the measure, we start multiple random walks from random vertices on either side, which terminate once they reach a high-degree vertex. For this variant of RWC, random walks do not terminate, rather they restart once they reach a high-degree vertex. More formally, we consider two instances of a random walk with restart (RWR), based on whether they start (and restart) from X (start = X) or Y (start = Y). When start = X, the RWR has a restart ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:12 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis vector uniformly distributed over X, and zero for vertices in Y (the situation is symmetric for start = Y). Moreover, the random walk runs on a modi(cid:27)ed graph with all outgoing edges from high-degree vertices removed. This modi(cid:27)cation transforms the high-degree vertices into dangling vertices, hence forcing the random walk to restart once it reaches one of these vertices.8 To formally de(cid:27)ne this variant of RWC, let P1 and P2 be the stationary distributions of the RWR obtained for start = X and start = Y, respectively. We consider the conditional probability Pr[start = A end = B+] that the random walk had started on side A ∈ {X , Y}, given that at some step at steady-state it is found in one of the high-degree vertices of side B ∈ {X , Y} (denoted as B+). We thus consider the following four probabilities: PX ,X + = Pr[start = X end = X +] = PX ,Y + = Pr[start = X end = Y +] = PY ,Y + = Pr[start = Y end = Y +] = PY ,X + = Pr[start = Y end = X +] = Notice that for the probabilities above we have X X V V V V X X v ∈X + P1(v) v ∈X + P1(v) + Y v ∈X + P2(v) , v ∈Y + P1(v) v ∈Y + P1(v) + Y v ∈Y + P2(v) , X X V V V V Y Y V V V V v ∈Y + P2(v) v ∈X + P2(v) v ∈Y + P1(v) + Y v ∈Y + P2(v) , v ∈X + P1(v) + Y v ∈X + P2(v) . (4) (5) (6) (7) (8) and Pr[start = X end = X +] + Pr[start = Y end = X +] = 1 Pr[start = X end = Y +] + Pr[start = Y end = Y +] = 1 as we ought to. The variant of the RWC score can be now de(cid:27)ned as RWC = PX X + PY Y + − PX Y + PY X + , which, like the original version, intuitively captures the di(cid:29)erence in the probability of staying on the same side and crossing the boundary. To verify that the new variant of the score works as expected, we compare it to the original version of the score (obtained via Monte Carlo sampling). The results are shown in Figure 5, from which it can be clearly seen that the new variant is almost identical to the original one. However, for the datasets considered in this work, we found empirically that this algorithm based on random walk with restart is up to 200 times faster compared to the original Monte Carlo algorithm. 8To compute the stationary distribution of the random walks, we use the implementation of Personalized PageRank from NetworkX https://networkx.github.io/documentation/latest/reference/generated/networkx.algorithms.link_analysis. pagerank_alg.pagerank.html. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:13 Fig. 5. Comparison between RWC scores computed via Monte Carlo sampling and those computed via RWR. Pearson's r = 0.96. 6.3 Betweenness Let us consider the set of edges C ⊆ E in the cut de(cid:27)ned by the two partitions X , Y. This measure uses the notion of edge betweenness and how the betweenness of the cut di(cid:29)ers from that of the other edges. Note that the cut here refers to the partioning obtained using METIS, as described in Section 3. Recall that the betweenness centrality bc(e) of an edge e is de(cid:27)ned as bc(e) = σs,t(e) σs,t , s(cid:44)t ∈V (9) where σs,t is the total number of shortest paths between vertices s, t in the graph and σs,t(e) is the number of those shortest paths that include edge e. The intuition here is that, if the two partitions are well-separated, then the cut will consist of edges that bridge structural holes [7]. In this case, the shortest paths that connect vertices of the two partitions will pass through the edges in the cut, leading to high betweenness values for edges in C. On the other hand, if the two partitions are not well separated, then the cut will consist of strong ties. In this case, the paths that connect vertices across the two partitions will pass through one of the many edges in the cut, leading to betweenness values for C similar to the rest of the graph. Given the distributions of edge betweenness on the cut and the rest of the graph, we compute the KL divergence dK L of the two distributions by using kernel density estimation to compute the PDF and sampling 10 000 points from each of these distributions (with replacement). We de(cid:27)ne the Betweenness Centrality Controversy (BCC) measure as (10) which assumes values close to zero when the divergence is small, and close to one when the divergence is large. BCC = 1 − e −dKL , 6.4 Embedding This measure is based on a low-dimensional embedding of graph G produced by Gephi's ForceAtlas2 algorithm [28] (the same algorithm used to produce the plots in Figures 3 and 4). According to Noack [43], a force-directed embedding also maximizes modularity. Based on this observation, the two- dimensional layouts produced by this algorithm indicate a layout with maximum modularity. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. -0.20.00.20.40.60.8RWC score - Monte Carlo Sampling-0.20.00.20.40.60.81.0RWC score - Random Walk with Restart1:14 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis Let us consider the two-dimensional embedding ϕ(v) of vertices v ∈ V produced by ForceAtlas2. Given the partition X, Y produced by the second stage of the pipeline, we calculate the following quantities: • dX and dY , the average embedded distance among pairs of vertices in the same partition, X and Y respectively; • dX Y , the average embedded distance among pairs of vertices across the two partitions X and Y. Inpsired by the Davies-Bouldin (DB) index [12], we de(cid:27)ne the Embedding Controversy measure EC as (11) EC is close to one for controversial topics, corresponding to better-separated graphs and thus to . EC = 1 − dX + dY 2dX Y higher degree of controversy, and close to zero for non-controversial topics. 6.5 Boundary Connectivity This controversy measure was proposed by Guerra et al. [27], and is based on the notion of boundary and internal vertices. Let u ∈ X be a vertex in partition X; u belongs to the boundary of X i(cid:29) it is connected to at least one vertex of the other partition Y, and it is connected to at least one vertex in partition X that is not connected to any vertex of partition Y. Following this de(cid:27)nition, let BX , BY be the set of boundary vertices for each partition, and B = BX ∪ BY the set of all boundary vertices. = X − BX are said to be the internal vertices of partition X (similarly for By contrast, vertices IX IY ). Let I = IX ∪ IY be all internal vertices in either partition. The reasoning for this measure is that, if the two partitions represent two sides of a controversy, then boundary vertices will be more strongly connected to internal vertices than to other boundary vertices of either partition. This intuition is captured in the formula GMCK = 1 B (12) where di(u) is the number of edges between vertex u and internal vertices I, while db(u) is the number of edges between vertex u and boundary vertices B. Higher values of the measure correspond to higher degrees of controversy. db(u) + di(u) − 0.5, u∈B di(u) 6.6 Dipole Moment This controversy measure was presented by Morales et al. [40], and is based on the notion of dipole moment that has its origin in physics. Let R(u) ∈ [−1, 1] be a polarization value assigned to vertex u ∈ V . Intuitively, extreme values of R (close to −1 or 1) correspond to users who belong most clearly to either side of the controversy. To set the values R(u) we follow the process described in the original paper [40]: we set R = ±1 for the top-5% highest-degree vertices in each partition X and Y, and set the values for the rest of the vertices by label-propagation. Let n+ and n− be the number of vertices V with positive and negative polarization values, respectively, and ∆A the absolute di(cid:29)erence of their normalized size ∆A = polarization value among vertices n+ (n−) and set d as half their absolute di(cid:29)erence, d = ❕c +−❕c− The dipole moment controversy measure is de(cid:27)ned as (cid:12)(cid:12)(cid:12) . Moreover, let ❕c + (❕c−) be the average (cid:12)(cid:12)(cid:12) n+−n− V 2 . (13) The rationale for this measure is that, if the two partitions X and Y are well separated, then label propagation will assign di(cid:29)erent extreme (±1) R-values to the two partitions, leading to higher values of the MBLB measure. Note also that larger di(cid:29)erences in the size of the two partitions MBLB = (1 − ∆A)d. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:15 (re(cid:30)ected in the value of ∆A) lead to decreased values for the measure, which takes values between zero and one. 7 CONTROVERSY SCORES FOR USERS The previous sections present measures to quantify the controversy of a conversation graph. In this section, we propose two measures to quantify the controversy of a single user in the graph. We denote this score as a real number that takes values in [−1, 1], with 0 representing a neutral score, and ±1 representing the extremes for each side. Intuitively, the controversy score of a user indicates how 'biased' the user is towards a particular side on a topic. For instance, for the topic 'abortion', pro-choice/pro-life activist groups tweeting consistently about abortion would get a score close to -1/+1 while average users who interact with both sides would get a score close to zero. In terms of the positions of users on the retweet graph, a neutral user would lie in the 'middle', retweeting both sides, where as a user with a high controversy score lies exclusively on one side of the graph. RW Cuser : The (cid:27)rst proposed measure is an adaptation of RWC. As input, we are given a user u ∈ V in the graph and a partitioning of the graph into two sides, de(cid:27)ned as disjoint sets of vertices X and Y. We then consider a random walk that starts – and restarts – at the given user u. Moreover, as with RWC, the high-degree vertices on each side (X + and Y +) are treated as dangling vertices – whenever the random walk reaches these vertices, it teleports to vertex u with probability 1 in the next step. To quantify the controversy of u, we ask how often the random walk is found on vertices that belong to either side of the controversy. Speci(cid:27)cally, for each user u, we consider the conditional probabilities Pr[start = u end = X +] and Pr[start = u end = Y +] , and estimate them by using the power iteration method. Assuming that user u belongs to side X of the controversy (i.e., u ∈ X), their controversy is de(cid:27)ned as: (14) RW Cuser(u, X) = Pr[start = u end = X +] Pr[start = u end = X +] + Pr[start = u end = Y +] . Expected hitting time: The second proposed measure is also random-walk-based, but de(cid:27)ned on the expected number of steps to hit the high-degree vertices on either side. Intuitively, a vertex is assigned a score of higher absolute value (closer to 1 or −1), if, compared to other vertices in the graph, it takes a very di(cid:29)erent time to reach a high-degree vertex on either side (X + or Y +). Speci(cid:27)cally, for each vertex u ∈ V in the graph, we consider a random walk that starts at u, and u before the random walk reaches any high-degree vertex estimate the expected number of steps, l X u across all vertices u ∈ V , we de(cid:27)ne ρX (u) in X +. Considering the distribution of values of l X as the fraction of vertices v ∈ V with l X u . We de(cid:27)ne ρY (u) similarly. Obviously, we have ρX (u), ρY (u) ∈ [0, 1). The controversy score of a user is then de(cid:27)ned as v < l X ρ(u) = ρX (u) − ρY (u) ∈ (−1, 1). (15) Following this de(cid:27)nition, a vertex that, compared to most other vertices, is very close to high-degree vertices X + will have ρX (u) ≈ 1; and if the same vertex is very far from high-degree vertices Y +, it will have ρY (u) ≈ 0 – leading to a controversy score ρ(u) ≈ 1 − 0 = 1. The opposite is true for vertices that are far from X + but close to Y +, leading to a controversy score ρ(u) ≈ −1. 7.1 Comparison with BiasWatch BiasWatch [37] is a recently-proposed, light-weight approach to compute controversy scores for users on Twitter. At a high level, the BiasWatch approach consists of the following steps: (1) Hand pick a small set of seed hashtags to characterize the two sides of a controversy (e.g., #prochoice vs. #prolife); ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:16 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis (2) Expand the seed set of hashtags based on co-occurrence; (3) Use the two sets of hashtags, identify strong partisans in the graph (users with high controversy score); (4) Assign controversy scores to other users via a simple label propagation approach. We compare the controversy scores obtained by our approaches to the ones obtained by Bi- asWatach9 on two sets of datasets: tweets matching the hashtags (i) #obamacare, #guncontrol, and #abortion, provided by Lu et al. [37] and (ii) the datasets in Table 2. We compute the Pearson correlation between our measure based on Expected hitting time and BiasWatch; the results are shown in Figure 6. We omit the comparison with RW Cuser scores as they are almost identical to the ones by BiasWatch. The authors also provide datasets which contain human annotations for controversy score (in the range [-2,2]) for 500 randomly selected users. We discretize our controversy scores to the same range, and compute the 5-category Fleiss' κ value. The κ value is 0.35, which represents a 'fair' level of agreement, according to Landis and Koch [36]. Topic #abortion #obamacare #guncontrol #beefban #baltimoreriots #netanyahuspeech #nemtsov #indiana #indiasdaughter #ukraine Pearson correlation 0.51 0.48 0.42 0.41 0.41 0.41 0.38 0.40 0.39 0.40 Fig. 6. (le(cid:28)) Pearson's r between the scores obtained by our algorithm (Expected hi(cid:29)ing time) and BiasWatch. (right) Sample sca(cid:29)er plot for #abortion. Our approach thus provides results that are similar to the state-of-the-art approach. Our method also has two advantages over the BiasWatch measure: (i) Even though we do not make use of any content information in our measure, we perform at par; and (ii) RW Cuser provides an intuitive extension to our RWC measure. Given this uni(cid:27)ed framework, it is possible to design ways to reduce controversy, e.g., by connecting opposing views [21, 22], and such a uni(cid:27)ed formulation can help us de(cid:27)ne principled objective functions to approach these tasks. 8 EXPERIMENTS In this section we report the results of the various con(cid:27)gurations of the pipeline proposed in this paper. As previously stated, we omit results for the content and hybrid graph building approaches presented in Section 4, as they do not perform well. We instead focus on the retweet and follow graphs, and test all the measures presented in Section 6 on the topics described in Table 2. In addition, we test all the measures on a set of external datasets used in previous studies [1, 10, 27] to validate the measures against a known ground truth. Finally, we use an evolving dataset from 9For BiasWatch we use parameters µ1 = 0.1, µ2 = 0.4, optimization method 'COBYLA', cosine similarity threshold 0.4, and 10 nearest neighbors for hashtag extension. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. -1.5-1.0-0.50.00.5RWC-1.5-1.0-0.50.00.5BiasWatch#abortion (r = 0.5)(cid:30)antifying Controversy on Social Media 1:17 Fig. 7. Controversy scores on retweet graphs of various controversial and non-controversial datasets. Fig. 8. Controversy scores on follow graphs of various controversial and non-controversial datasets. Table 3. Results on external datasets. The 'C?' column indicates whether the previous study considered the dataset controversial (ground truth). Dataset Political blogs Twitter politics Gun control Brazil soccer Karate club Facebook university NYC teams V 1222 18 470 33 254 20 594 34 281 95 924 E C? 16 714 48 053 349 782 82 421 78 4389 176 249 RWC BCC 0.53 0.42 0.77 0.79 0.68 0.70 0.48 0.67 0.64 0.11 0.35 0.26 0.24 0.34 EC GMCK MBLB 0.45 0.49 0.62 0.34 0.81 0.55 0.75 0.68 0.11 0.51 0.38 0.27 0.19 0.17 0.18 0.28 0.24 0.17 0.17 0.01 0.01 Twitter collected around the death of Venezuelan president Hugo Chavez [40] to show the evolution of the controversy measures in response to high-impact events. To avoid potential over(cid:27)tting, we use only eight graphs as testbed during the development of the measures, half of them controversial (beefban, nemtsov, netanyahu, russia_march) and half non-controversial (sxsw, germanwings, onedirection, ultralive). This procedure resembles a 40/60% train/test split in traditional machine learning applications.10 8.1 Twi(cid:29)er hashtags Figure 7 and Figure 8 report the scores computed by each measure for each of the 20 hashtags, on the retweet and follow graph, respectively. Each (cid:27)gure shows a set of beanplots,11 one for each measure. Each beanplot shows the estimated probability density function for a measure computed on the topics, the individual observations are shown as small white lines in a one-dimensional scatter plot, and the median as a longer black line. The beanplot is divided into two groups, one for controversial topics (left/dark) and one for non-controversial ones (right/light). A larger separation of the two distributions indicates that the measure is better at capturing the characteristics of 10A demo of our controversy measures can be found at https://users.ics.aalto.(cid:27)/kiran/controversy. 11A beanplot is an alternative to the boxplot for visual comparison of univariate data among groups. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. −0.20.00.20.40.60.81.0BCCECGMCKMBLBRWCControversy scoreControversialNon−controversial−0.20.00.20.40.60.81.0BCCECGMCKMBLBRWCControversy scoreControversialNon−controversial1:18 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis controversial topics. For instance, this separation is fundamental when using the controversy score as a feature in a classi(cid:27)cation algorithm. Figures 7 and 8 clearly show that RWC is the best measure on our datasets. BCC and EC show varying degrees of separation and overlap, although EC performs slightly better as the distributions are more concentrated, while BCC has a very wide distribution. The two baselines GMCK and MBLB instead fail to separate the two groups. Especially on the retweet graph, the two groups are almost indistinguishable. For all measures the median score of controversial topics is higher than for non-controversial ones. This result suggests that both graph building methods, retweet and follow, are able to capture the di(cid:29)erence between controversial and non-controversial topics. Given the broad range of provenience of the topics covered by the dataset, and their di(cid:29)erent characteristics, the consistency of the results is very encouraging. 8.2 External datasets We have shown that our approach works well on a number of datasets extracted in-the-wild from Twitter. But, how well does it generalize to datasets from di(cid:29)erent domains? We obtain a comprehensive group of datasets kindly shared by authors of previous works: Political blogs, links between blogs discussing politics in the US [1]; Twitter politics, Twitter messages pertaining to the 2010 midterm election in US [10]; and the following (cid:27)ve graphs used in the study that introduced GMCK [27], (a) Gun control, retweets about gun control after the shooting at the Sandy Hook school; (b) Brazil soccer, retweets about to two popular soccer teams in Brazil; (c) Karate club, the well-known social network by [50]; (d) Facebook university, a social graph among students and professors at a Brazilian university; (e) NYC teams, retweets about two New York City sports teams. Table 3 shows a comparison of the controversy measures under study on the aforementioned datasets.12 For each dataset we also report whether it was considered controversial in the original paper, which provides a sort of "ground truth" to evaluate the measures against. All the measures are able to distinguish controversial graphs to some extent, in the sense that they return higher values for the controversial cases. The only exception is Karate club. Both RWC and MBLB report low controversy scores for this graph. It is possible that the graph is too small for such random-walk-based measures to function properly. Conversely, BCC is able to capture the desired behavior, which suggests that shortest-path and random-walk based measures might have a complementary function. Interestingly, while the Political blogs datasets is often considered a gold standard for polarization and division in online political discussions, all the measures agree that it presents only a moderate level of controversy. Conversely, the Twitter politics dataset is clearly one of the most controversial one across all measures. This di(cid:29)erence suggests that the measures are more geared towards capturing the dynamics of controversy as it unfolds on social media, which might di(cid:29)er from more traditional blogs. For instance, one such di(cid:29)erence is the cost of an endorsement: placing a link on a blog post arguably consumes more mental resources than clicking on the retweet button. For the 'Gun control' dataset, Guerra et al. need to manually distinguish three di(cid:29)erent partitions in the graph: gun rights advocates, gun control supporters, and moderates. Our pipeline is able to (cid:27)nd the two communities with opposing views (grouping together gun control supporters and moderates, as suggested in the original study) without any external help. All measures agree with the conclusions drawn in the original paper that this topic is highly controversial. 12The datasets provided by Guerra et al. [27] are slightly di(cid:29)erent from the ones used in the original paper because of some irreproducible (cid:27)ltering used by the authors. We use the datasets provided to us verbatim. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:19 Fig. 9. Controversy scores on 56 retweet graphs from Morales et al. Day 'D' (indicated by the blue vertical line) indicates the announcement of the death of president Hugo Chavez. Note that even though from the results in Table 3, RWC, BCC and EC appear to outperform each other, it is not the case. These methods are not comparable, meaning, a score of 0.5 for RWC is not the same as a 0.5 for BCC. The insight we can draw from these results is that our methods are able to discern a controversial topic from a non-controversial one consistently, irrespective of the domain, and are able to do so more reliably than existing methods (GMCK and MBLB). 8.3 Evolving controversy We have shown that our approach also generalizes well to datasets from di(cid:29)erent domains. But in a real deployment the measures need to be computed continuously, as new data arrives. How well does our method work in such a setting? And how do the controversy measures evolve in response to high-impact events? To answer these questions, we use a dataset from the study that introduced MBLB [40]. The dataset comprises Twitter messages pertaining to political events in Venezuela around the time of the death of Hugo Chavez (Feb-May 2013). The authors built a retweet graph for each of the 56 days around the day of the death (one graph per day). Figure 9 shows how the intensity of controversy evolves according to the measures under study (which occurs on day 'D'). The measure proposed in the original paper, MBLB, which we use as 'ground truth', shows a clear decrease of controversy on the day of the death, followed by a progressive increase in the controversy of the conversation. The original interpretation states that on the day of the death a large amount of people, also from other countries, retweeted news of the event, creating a single global community that got together at the shock of the news. After the death, the ruling and opposition party entered in a (cid:27)ery discussion over the next elections, which increased the controversy. All the measures proposed in this work show the same trend as MBLB. Both RWC and EC follow very closely the original measure (Pearson correlation coe(cid:28)cients r of 0.944 and 0.949, respectively), while BCC shows a more jagged behavior in the (cid:27)rst half of the plot (r = 0.743), due to the discrete nature of shortest paths. All measures however present a dip on day 'D', an increase in controversy in the second half, and another dip on day 'D+20'. Conversely, GMCK reports an almost constant moderate value of controversy during the whole period (r = 0.542), with barely noticeable peaks and dips. We conclude that our measures generalize well also to the case of evolving graphs, and behave as expected in response to high-impact events. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. DaysD-29D-27D-25D-23D-21D-19D-17D-15D-13D-11D-9D-7D-5D-3D-1D+1D+3D+5D+7D+9D+11D+13D+15D+16D+18D+20D+22D+24Controversy score00.10.20.30.40.50.60.70.80.91MBLBRWCECGMCKBCC1:20 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis (a) (b) Fig. 10. RWC scores for synthetic Erdös-Rényi graphs planted with two communities. p1 is the intra- community edge probability, while p2 is the inter-community edge probability. 8.4 Simulations Given that RWC is the best-performing score among the ones in this study, we focus our attention solely on it henceforth. To measure the robustness of the RWC score, we generate random Erdös- Rényi graphs with varying community structure, and compute the RWC score on them. Speci(cid:27)cally, to mimic community structure, we plant two separate communities with intra-community edge probability p1. That is, p1 de(cid:27)nes how dense these communities are within themselves. We then add random edges between these two communities with probability p2. Therefore, p2 de(cid:27)nes how connected the two communities are. A higher value of p1 and a lower value of p2 create a clearer two-community structure. Figure 10 shows the RWC score for random graphs of 2000 vertices for two di(cid:29)erent settings: plotting the score as a function of p1 while (cid:27)xing p2 (Figure 10a), and vice-versa (Figure 10b). The RWC score reported is the average over ten runs. We observe a clear pattern: the RWC score increases as we increase the density within the communities, and decreases as we add noise to the community structure. The e(cid:29)ects of the parameters is also expected, for a given value of p1, a smaller value of p2 generates a larger RWC score, as the communities are more well separated. Conversely, for a given value of p2, a larger value of p1 generates a larger RWC scores, as the communities are denser. 8.5 Controversy detection in the wild In most of the experiments presented so far, we hand-picked known topics which are controversial and showed that our method is able to separate them from the non-controversial topics. To check whether our system works in a real-world setting, we deploy it in the wild to explore actual topics of discussion on Twitter and detect the ones that are controversial. More speci(cid:27)cally, we obtain daily trending hashtags (both US and worldwide) on the platform for a period of three months (June 25 – September 19, 2015). Then, we obtain all tweets that use these hashtags, and create retweet graphs (as described in Section 4). Finally, we apply the RWC measure on these conversation graphs to identify controversial hashtags. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 0.00.20.40.60.81.0p10.10.00.10.20.30.40.50.60.7RWC scorep2=0.01p2=0.02p2=0.03p2=0.04p2=0.050.000.020.040.060.080.10p20.10.00.10.20.30.40.50.60.7RWC scorep1=0.1p1=0.2p1=0.3p1=0.4p1=0.5(cid:30)antifying Controversy on Social Media 1:21 Fig. 11. Frequency of RWC scores for hashtags trending from June to September 2015. The results can be explored in our online demo [19].13 To mention a few examples, our system was able to identify the following controversial hashtags: • #whosiburningblackchurches (score 0.332): A hashtag about the burning of predominantly • #communityshield (score 0.314): Discussion between the fans of two sides of a soccer game.15 • #nationalfriedchickenday (score 0.393): A debate between meat lovers and vegetarians about black churches.14 the ethics of eating meat. Moreover, based on our experience with our system, most hashtags that are reported as trending on Twitter concern topics that are not controversial. Figure 11 shows the histogram of the RWC score over the 924 trending hashtags we collected. A majority of these hashtags have an RWC score around zero. 9 CONTENT In this section we explore alternative approaches to measuring controversy that use only the content of the discussion rather than the structure of user interactions. As such, these methods do not (cid:27)t in the pipeline described in Section 3. The question we address is "does content help in measuring the controversy of a topic?" In particular, we test two types of features extracted from the content. The (cid:27)rst is a typical IR-inspired bag-of-words representation. The second involves sentiment-related features, extracted with NLP tools. 9.1 Bag of words We take as input the raw content of the social media posts – in our case, the tweets pertaining to a speci(cid:27)c topic. We represent each tweet as a vector in a high-dimensional space composed of the words used in the whole topic, after standard preprocessing used in IR (lowercasing, stopword 13https://users.ics.aalto.(cid:27)/kiran/controversy/table.php. 14https://erlc.com/article/explainer-whoisburningblackchurches. 15https://en.wikipedia.org/wiki/2015_FA_Community_Shield. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 0.10.00.10.20.30.40.50.6RWC score051015202530354045Frequency1:22 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis Fig. 12. Sentiment variance controversy score for controversial and non-controversial topics. removal, stemming). Following the lines of our main pipeline, we group these vectors in two clusters by using CLUTO [31] with cosine distance. The underlying assumption is that the two sides, while sharing the use of the hashtag for the topic, use di(cid:29)erent vocabularies in reference to the issue at hand. For example, for #beefban a side may be calling for "freedom" while the opposing one for "respect." We use KL divergence as a measure of distance between the vocabularies of the two clusters, and the I2 measure [38] of clustering heterogeneity. We use an unpaired Wilcoxon rank-sum test at the p = 0.05 signi(cid:27)cance level, but we are unable to reject the null hypothesis that there is no di(cid:29)erence in these measures between the controversial and non-controversial topics. Therefore, there is not enough signal in the content representation to discern between controversial and non-controversial topics with con(cid:27)dence. This result suggests that the bag-of-words representation of content is not a good basis for our task. It also agrees with our earlier attempts to use content to build the graph used in the pipeline (see Section 4) – which suggests that using content for the task of quantifying controversy might not be straightforward. 9.2 Sentiment analysis Next, we resort to NLP techniques for sentiment analysis to analyze the content of the discussion. We use SentiStrength [48] trained on tweets to give a sentiment score in [−4, 4] to each tweet for a given topic. In this case we do not try to cluster tweets by their sentiment. Rather, we analyze the di(cid:29)erence in distribution of sentiment between controversial and non-controversial topics. While it is not possible to say that controversial topics are more positive or negative than non- controversial ones, we can detect a di(cid:29)erence in their variance. Indeed, controversial topics have a higher variance than non-controversial ones, as shown in Figure 12. Controversial ones have a variance of at least 2, while non-controversial ones have a variance of at most 1.5. In practice, the "tones" with which controversial topics are debated are stronger, and sentiment analysis is able to detect this. While this signal is clear, it is not straightforward to incorporate it into the measures based on graph structure. Moreover, this feature relies on technologies that do not work reliably for languages other than English and hence cannot be applied for topics such as #russia_march. 10 DISCUSSION The task we tackle in this work is certainly not an easy one, and this study has some limitations, which we discuss in this section. We also report a set of negative results that we produced while coming up with the measures presented. We believe these results will be very useful in steering this research topic towards a fruitful direction. Table 4 provides a summary of the various graph building strategies and controversy measures we tried for quantifying controversy. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 0.00.51.01.52.02.53.0Sentiment VarianceControversialNon−controversial(cid:30)antifying Controversy on Social Media 1:23 Table 4. Summary of various graph building and controversy measures tried. Methods that worked reliably are marked in bold. Graphs Measures Retweet Follow Content Mention Hybrid (content + retweet, mention + retweet) Random Walk Edge Betweenness Embedding Boundary Connectivity Dipole Moment Cut-based Measures (conductance, cut ratio) Sentiment Analysis Modularity SPID 10.1 Limitations Twitter only. We present our (cid:27)ndings mostly on datasets coming from Twitter. While this is certainly a limitation, Twitter is one of the main venues for online public discussion, and one of the few for which data is available. Hence, Twitter is a natural choice. In addition, our measures generalize well to datasets from other social media and the Web. Choice of data. We manually pick the controversial topics in our dataset, which might introduce bias. In our choice we represent a broad set of typical controversial issues coming from religious, societal, racial, and political domains. Unfortunately, ground truths for controversial topics are hard to (cid:27)nd, especially for ephemeral issues. However, the topics are unanimously judged controversial by the authors. Moreover, the hashtags represent the intuitive notion of controversy that we strive to capture, so human judgement is an important ingredient we want to use. Over(cid:27)tting. While this work presents the largest systematic study on controversy in social media so far, we use only 20 topics for our main experiment. Given the small number of examples, the risk of over(cid:27)tting our measures to the dataset is real. We reduce this risk by using only 40% of the topics during the development of the measures. Additionally, our measures agree with previous independent results on external datasets, which further decreases the likelihood of over(cid:27)tting. Reliance on graph partitioning. Our pipeline relies on a graph partitioning stage, whose quality is fundamental for the proper functioning of the controversy measures. Given that graph partitioning is a hard but well studied problem, we rely on o(cid:29)-the-shelf techniques for this step. A measure that bypasses this step entirely is highly desirable, and we report a few unsuccessful attempts in the next subsection. Multisided controversies. Not all controversies involve only two sides with opposing views. Some times discussions are multifaceted, or there are three or more competing views on the (cid:27)eld. The principles behind our measures neatly generalize to multisided controversies. However, in this case the graph partitioning component needs to automatically (cid:27)nd the optimal number of partitions. We defer experimental study of such cases to an extended version of this paper. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:24 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis Evaluation. De(cid:27)ning what is controversial/polarized can be subjective. There are many ways to de(cid:27)ne what is controversial, depending on the context, subject and (cid:27)eld of study, e.g. See [6] for around a dozen ways to de(cid:27)ne polarization. Our evaluation is based on our intuitive labelling that a topic is controversial/polarized. This might not always be true, but given that the alternative is to hand-label/survey the thousands of users, we presume that this assumption is reasonable for developing methods that can be adapted to large scale systems. 10.2 Negative results We brie(cid:30)y review a list of methods that failed to produce reliable results and were discarded early in the process of re(cid:27)ning our controversy measures. Mentions graph. Conover et al. [10] rely on the mention graph in Twitter to detect controversies. However, in our dataset the mention graphs are extremely sparse given that we focus on short-lived events. Merging the mentions into the retweet graph does not provide any noticeable improvement. Previous studies have also shown that people retweet similar ideologies but mention across ideologies [5]. We exploit this intuition by using correlation clustering for graph partitioning, with negative edges for mentions. Alas, the results are qualitatively worse than those obtained by METIS. Cuts. Simple measures such as size of the cut of the partitions do not generalize across di(cid:29)erent graphs. Conductance (in all its variants) also yields poor results. Prior work identi(cid:27)es controversies by comparing the structure of the graph with randomly permuted ones [10]. Unfortunately, we obtain equally poor results by using the di(cid:29)erence in conductance with cuts obtained by METIS and by random partitions. Community structure. Good community structure in the conversation graph is often understood as a sign that the graph is polarized or controversial. However, this is not always the case. We (cid:27)nd that both assortativity and modularity (which have been previously used to identify controversy) do not correlate with the controversy scores, and are not good predictors for how controversial a topic is. The work by Guerra et al. [27] presents clear arguments and examples of why modularity should be avoided. Partitioning. As already mentioned, bypassing the graph partitioning to compute the measure is desirable. We explore the use of the all pairs expected hitting time computed by using SimRank [30]. We compute the SPID (ratio of variance to mean) of this distribution, however results are mixed. 10.3 Conclusions In this paper, we performed the (cid:27)rst large-scale systematic study for quantifying controversy in social media. We have shown that previously-used measures are not reliable and demonstrated that controversy can be identi(cid:27)ed both in the retweet and topic-induced follow graph. We have also shown that simple content-based representations do not work in general, while sentiment analysis o(cid:29)ers promising results. Among the measures we studied, the random-walk-based RWC most neatly separates controver- sial topics from non-controversial ones. Besides, our measures gracefully generalize to datasets from other domains and previous studies. This work opens several avenues for future research. First, it is worth exploring alternative approaches and testing additional features, such as, following a generative-model-based approach, or exploiting the temporal evolution of the discussion of a topic [23]. From the application point of view, the controversy score can be used to generate recommen- dations that foster a healthier "news diet" on social media. Given the ever increasing impact of polarizing (cid:27)gures in our daily politics and the rise in polarization in the society [13, 24], it is ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. (cid:30)antifying Controversy on Social Media 1:25 important to not restrict ourselves to our own 'bubbles' or 'echo chambers' [44, 47]. Our methods for identifying controversial topics can be used as building blocks for designing such systems to reduce controversy on social media [21, 22] by connecting social media users with content outside their own bubbles. Finally, polarization by itself may not be a wholly negative phenomenon. Several studies [11, 42] argue that a democracy needs deliberation, and polarization enable such a deliberation to happen in the public, to a certain extent, thus informing people about the issues and arguments from di(cid:29)erent sides. Given such a setting, it is of paramount importance to understand to what extent a discussion is polarized, so that things do not spiral out of control, or create isolated echo chambers. Our paper contributes methods that are useful in this setting, and enable measuring the degree of polarization of a topic in a domain-agnostic fashion. Acknowledgements. This work has been supported by the Academy of Finland project "Nestor" (286211) and the EC H2020 RIA project "SoBigData" (654024). REFERENCES [1] Lada A Adamic and Natalie Glance. 2005. The political blogosphere and the 2004 US election: divided they blog. In LinkKDD. 36–43. In COSN. 13–24. [2] Leman Akoglu. 2014. Quantifying Political Polarity Based on Bipartite Opinion Networks. In ICWSM. [3] Md Tanvir Al Amin, Charu Aggarwal, Shuochao Yao, Tarek Abdelzaher, and Lance Kaplan. 2017. Unveiling Polarization in Social Networks: A Matrix Factorization Approach. Technical Report. IEEE. [4] Jisun An, Daniele Quercia, and Jon Crowcroft. 2014. Partisan sharing: Facebook evidence and societal consequences. [5] Alessandro Bessi, Guido Caldarelli, Michela Del Vicario, Antonio Scala, and Walter Quattrociocchi. 2014. Social Determinants of Content Selection in the Age of (Mis)Information. In Social Informatics. 259–268. [6] Aaron Bramson, Patrick Grim, Daniel J Singer, Steven Fisher, William Berger, Graham Sack, and Carissa Flocken. 2016. Disambiguation of social polarization concepts and measures. The Journal of Mathematical Sociology 40, 2 (2016), 80–111. [7] Ronald S Burt. 2009. Structural holes: The social structure of competition. Harvard university press. [8] Yoonjung Choi, Yuchul Jung, and Sung-Hyon Myaeng. 2010. Identifying controversial issues and their sub-topics in news articles. In Paci(cid:27)c-Asia Workshop on Intelligence and Security Informatics. Springer, 140–153. [9] Mauro Coletto, Kiran Garimella, Aristides Gionis, and Claudio Lucchese. 2017. A Motif-based Approach for Identifying Controversy. In Proceedings of the 10th International on Conference on Web and Social Media. AAAI. [10] Michael Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Gonçalves, Filippo Menczer, and Alessandro Flammini. 2011. Political Polarization on Twitter. In ICWSM. [11] Lincoln Dahlberg. 2007. Rethinking the fragmentation of the cyberpublic: from consensus to contestation. New media & society 9, 5 (2007), 827–847. [12] David L. Davies and Donald W. Bouldin. 1979. A Cluster Separation Measure. IEEE TPAMI 1, 2 (1979), 224–227. [13] Michael Dimock, Carroll Doherty, Jocelyn Kiley, and Russ Oates. 2014. Political polarization in the American public: How increasing ideological uniformity and partisan antipathy a(cid:29)ect politics, compromise and everyday life. Washington, DC: Pew Research Center (2014). [14] Shiri Dori-Hacohen and James Allan. 2015. Automated controversy detection on the web. In European Conference on Information Retrieval. Springer, 423–434. [15] Abraham Doris-Down, Husayn Versee, and Eric Gilbert. 2013. Political blend: an application designed to bring people together based on political di(cid:29)erences. In C&T. 120–130. [16] Kevin M Esterling, Archon Fung, and Taeku Lee. 2010. How Much Disagreement is Good for Democratic Deliberation? The CaliforniaSpeaks Health Care Reform Experiment. SSRN (2010). [17] Wei Feng, Jiawei Han, Jianyong Wang, Charu Aggarwal, and Jianbin Huang. 2015. STREAMCUBE: Hierarchical Spatio-temporal Hashtag Clustering for Event Exploration over the Twitter Stream. In ICDE. [18] Seth R Flaxman, Sharad Goel, and Justin M Rao. 2015. Filter Bubbles, Echo Chambers, and Online News Consumption. [19] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2016a. Exploring Controversy in Twitter. In CSCW [demo]. 33–36. [20] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2016b. Quantifying (2015). ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017. 1:26 K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis 1400–1408. Controversy in Social Media. In WSDM. 33–42. [21] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2017a. Factors in Recommending Contrarian Content on Social Media. In WebSci. [22] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2017b. Reducing Controversy by Connecting Opposing Views. In WSDM. 81–90. [23] Kiran Garimella, Gianmarco De Francisci Morales, Aristides Gionis, and Michael Mathioudakis. 2017c. The E(cid:29)ect of Collective Attention on Controversial Debates on Social Media. In WebSci. 43–52. [24] Kiran Garimella and Ingmar Weber. 2017. A Long-Term Analysis of Polarization on Twitter. In ICWSM. [25] Eduardo Graells-Garrido, Mounia Lalmas, and Daniele Quercia. 2013. Data portraits: Connecting people of opposing views. arXiv preprint arXiv:1311.4658 (2013). [26] Catherine Grevet, Loren G Terveen, and Eric Gilbert. 2014. Managing political di(cid:29)erences in social media. In CSCW. [27] Pedro Henrique Calais Guerra, Wagner Meira Jr, Claire Cardie, and Robert Kleinberg. 2013. A Measure of Polarization on Social Media Networks Based on Community Boundaries. In ICWSM. [28] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. 2014. ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. (2014). [29] Myungha Jang, John Foley, Shiri Dori-Hacohen, and James Allan. 2016. Probabilistic Approaches to Controversy Detection. In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. ACM, 2069–2072. [30] Glen Jeh and Jennifer Widom. 2002. SimRank: A measure of structural-context similarity. In KDD. 538–543. [31] George Karypis. 2002. CLUTO - A clustering toolkit. (2002). [32] George Karypis and Vipin Kumar. 1995. METIS - Unstructured Graph Partitioning and Sparse Matrix Ordering System. [33] Manfred Klenner, Michael Amsler, Nora Hollenstein, and Gertrud Faass. 2014. Verb polarity frames: a new resource and its application in target-speci(cid:27)c polarity classi(cid:27)cation. In KONVENS. 106–115. [34] Juhi Kulshrestha, Muhammad Bilal Zafar, Lisette Espin Noboa, Krishna P Gummadi, and Saptarshi Ghosh. 2015. Characterizing Information Diets of Social Media Users. In ICWSM. [35] Michael LaCour. 2012. A balanced news diet, not selective exposure: Evidence from a direct measure of media exposure. [36] J Richard Landis and Gary G Koch. 1977. The measurement of observer agreement for categorical data. biometrics [37] Haokai Lu, James Caverlee, and Wei Niu. 2015. BiasWatch: A Lightweight System for Discovering and Tracking Topic-Sensitive Opinion Bias in Social Media. In CIKM. 213–222. [38] Ujjwal Maulik and Sanghamitra Bandyopadhyay. 2002. Performance evaluation of some clustering algorithms and [39] Yelena Mejova, Amy X Zhang, Nicholas Diakopoulos, and Carlos Castillo. 2014. Controversy and Sentiment in Online validity indices. IEEE TPAMI 24, 12 (2002), 1650–1654. News. arXiv preprint arXiv:1409.8152 (2014). [40] AJ Morales, J Borondo, JC Losada, and RM Benito. 2015. Measuring political polarization: Twitter shows the two sides [41] Sean A Munson, Stephanie Y Lee, and Paul Resnick. 2013. Encouraging Reading of Diverse Political Viewpoints with a [42] Diana C Mutz. 2002. The consequences of cross-cutting networks for political participation. American Journal of of Venezuela. Chaos 25, 3 (2015). Browser Widget.. In ICWSM. Political Science (2002), 838–855. [43] Andreas Noack. 2009. Modularity clustering is force-directed layout. Physical Review E 79, 2 (2009), 026102. [44] Eli Pariser. 2011. The (cid:27)lter bubble: What the Internet is hiding from you. Penguin UK. [45] Ana-Maria Popescu and Marco Pennacchiotti. 2010. Detecting controversial events from twitter. In Proceedings of the 19th ACM international conference on Information and knowledge management. ACM, 1873–1876. [46] Yiye Ruan, David Fuhry, and Srinivasan Parthasarathy. 2013. E(cid:28)cient community detection in large networks using content and links. In WWW. 1089–1098. [47] Cass R Sunstein. 2009. Republic. com 2.0. Princeton University Press. [48] Mike Thelwall. 2013. Heart and soul: Sentiment strength detection in the social web with SentiStrength. In CyberEmo- (1995). SSRN (2012). (1977), 159–174. [49] Mikalai Tsytsarau, Themis Palpanas, and Kerstin Denecke. 2011. Scalable detection of sentiment-based contradictions. [50] Wayne Zachary. 1977. An Information Flow Model for Con(cid:30)ict and Fission in Small Groups. J. of Anthropological tions. 1–14. DiversiWeb, WWW 2011 (2011). Research 33 (1977), 452–473. ACM Transactions on Social Computing, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1308.2565	1	1308	2013-08-12T13:58:15	A place-focused model for social networks in cities	[ "cs.SI", "physics.soc-ph" ]	The focused organization theory of social ties proposes that the structure of human social networks can be arranged around extra-network foci, which can include shared physical spaces such as homes, workplaces, restaurants, and so on. Until now, this has been difficult to investigate on a large scale, but the huge volume of data available from online location-based social services now makes it possible to examine the friendships and mobility of many thousands of people, and to investigate the relationship between meetings at places and the structure of the social network. In this paper, we analyze a large dataset from Foursquare, the most popular online location-based social network. We examine the properties of city-based social networks, finding that they have common structural properties, and that the category of place where two people meet has very strong influence on the likelihood of their being friends. Inspired by these observations in combination with the focused organization theory, we then present a model to generate city-level social networks, and show that it produces networks with the structural properties seen in empirical data.	cs.SI	cs	A place-focused model for social networks in cities Chloe Brown,1 Anastasios Noulas,1 Cecilia Mascolo,1 and Vincent Blondel2 1Computer Laboratory, University of Cambridge, Cambridge (UK) 2Universit´e Catholique de Louvain, Louvain-La-Neuve (Belgium) (Dated: October 4, 2018) The focused organization theory of social ties proposes that the structure of human social networks can be arranged around extra-network foci, which can include shared physical spaces such as homes, workplaces, restaurants, and so on. Until now, this has been difficult to investigate on a large scale, but the huge volume of data available from online location-based social services now makes it possible to examine the friendships and mobility of many thousands of people, and to investigate the relationship between meetings at places and the structure of the social network. In this paper, we analyze a large dataset from Foursquare, the most popular online location-based social network. We examine the properties of city-based social networks, finding that they have common structural properties, and that the category of place where two people meet has very strong influence on the likelihood of their being friends. Inspired by these observations in combination with the focused organization theory, we then present a model to generate city-level social networks, and show that it produces networks with the structural properties seen in empirical data. I. INTRODUCTION It is an intuitive idea that social relationships between people arise out of, and are reflected by, meetings and shared activities in common spaces. Scott Feld's the- ory of the focused organization of social ties posits that friendships form between individuals whose interactions are organized around extra-network foci, which can in- clude physical places. In his 1981 paper [1], Feld presents this theory and discusses how commonly observed struc- tural properties of social networks could result. Empirical investigation of such theories has tradition- ally been difficult and time-consuming, requiring inter- views with, and observation of, necessarily small groups of people. Large-scale analysis has therefore been im- possible. However, the recent widespread adoption of location-based online social services has provided us with a huge volume of data both about the structure of peo- ple's social networks, as described by social ties explic- itly declared by users of these services, and about their activities and meetings at places in their local environ- ment, thanks to the location-sharing dimension. We now therefore have an unprecedented opportunity to investi- gate the role that places may play in the structure of social networks, on a scale not previously feasible. In addition, the semantic information about places avail- able in location-based online social services allows us to investigate the relationship between the types of places where people meet and the likelihood that those people are friends. In this work, we study a large dataset from Foursquare, which is the most popular location-based online social network, used by over 35 million people worldwide [2]. We analyze the social and spatial properties of social networks in cities, and present a model for a place-based social network based on our observations from the em- pirical data, in combination with Feld's focused organi- zation theory. We then show that the model produces networks with the structural properties expected of so- cial networks, as well as preserving the popularity distri- bution of places in the city and the spans of the sets of places that people visit. In more detail, our work makes the following contribu- tions: • We first define and analyze place-based social net- works at the city scale, to answer the question: what do intra-city social networks look like, and do they have common structural characteristics? We show that the city-based social networks in the Foursquare dataset have the structural prop- erties observed by computational social scientists studying real-world social networks, namely: a power-law degree distribution, small-world prop- erties (high clustering and small diameter), and strong community structure. While the global properties of online and offline social networks have been analyzed previously, our work is the first to examine and compare the structures of place-based social networks within different cities and to show these common structural properties. • We then address the question: is this large location- based social dataset consistent with the theory of fo- cused tie organization? Exploiting the combination of social information and specific semantic informa- tion available in the Foursquare dataset, we are able to examine the relationship between the category of a place where people meet and the probability of friendship. The type of a place where people meet has a strong influence on the likelihood that they are friends, and that there is evidence for the ex- istence of intra-place triads, resulting in clustering of social ties around places in the city. • Inspired by these observations in combination with the focused organization theory, we present a model for a city-level social network, based around meet- ings at places, and show that when simulated this model is able to produce networks with the structural properties observed in real social net- works, while preserving the popularity distribution of places in the city and the span distribution of sets of places visited by individuals. The fact that this model is able to reproduce empirically observed social network features is consistent with the idea that places can act as foci for friendships. Our work has intrinsic interest, as we investigate an area largely unexplored, namely, that of the structural similarities between social networks at the city scale within different cities, and demonstrate that the networks in different cities show striking similarities. Furthermore, our model demonstrates that the focused organization of social ties, with places as foci, results in networks with the structural features commonly observed in social net- works. From a practical perspective, the observation that that the type of a place where people meet strongly affects the probability of friendship could be useful to online location-based social services such as Foursquare. For ex- ample, one important application in location-based social networks is the recommendation to users of venues they might want to visit [3 -- 5]. The fact that that the type of place where people meet has a strong influence on friend- ship suggests that different recommendations would be appropriate depending on the other users present. There are also potential applications in the develop- ment of smarter privacy controls in location-based online social networks: Page et al. found that people's concerns about privacy in location-sharing services center around the desire to preserve one's existing offline relationship boundaries [6], and our observations suggest that these boundaries might be reflected in the types of places where friends meet (for example, closer friends at homes and nightlife spots, less close acquaintances only at profes- sional venues or transport spots). Use of this informa- tion could enable services such as Foursquare to adjust the default audience of a check-in, for example, based on relationship semantics inferred using meeting places. II. RELATED WORK The structure of social networks has been well-studied by sociologists and by computational social scientists, and such networks are known commonly to exhibit some particular structural properties. For example, social net- works usually have small-world properties, that is, small diameter and high clustering [7, 8], and show a power- law degree distribution [9]. Stanley Milgram showed, in his famous study, that social networks tend to have short average path lengths between individuals, finding an av- erage distance of about six hops in the network he stud- ied [10]. Social networks also tend to show strong com- munity structure: the nodes of the network are arranged into tightly-knit groups with their members densely con- 2 nected, and fewer or looser connections between these groups of friends [11]. In the past decade, the growth in popularity of online social networks has enabled the examination of social net- work structure at a previously impossible scale, and these networks have been shown generally to have the same structural properties as their offline counterparts [12 -- 14]. Location-based social networks in particular have been the subject of much recent research, and many studies have shown that being close geographically and visiting the same places is a strong indicator for being friends in the online location-based social network [15 -- 18]. The significance of different kinds of places and their associated probability that two people meeting there are friends was discussed by Cranshaw et al. [19], who de- veloped a metric based on entropy to characterize places according to the diversity of the population of visitors. However, they had a much smaller sample of 489 users than we use here, and they identified locations as cells in a grid, rather than by using specific venues with their as- sociated information available in Foursquare. Other work that has discussed similar concepts includes the studies of urban Bluetooth encounters by Kostakos et al. [20, 21], which observed the differences between people's mobility in different places around the city, for example, in pubs and in the street. However, this work only examined the city of Bath, UK, while we have shown similarities across different cities. Modeling network structure has also long been a highly active research topic, with a huge variety of models be- ing presented over the years to reproduce various em- pirically observed properties through mechanisms such as preferential attachment, triadic closure, and commu- nity association [9, 22, 23]. The work closest to that we present here is the body of research concerning affilia- tion networks, that is, social networks where agents are associated with societies. In this case, the type of place- based city network that we examine can be seen as one where users are associated with places. Many models have been presented that aim to achieve explicit group formation alongside the social network structure [24 -- 26]. Our work differs in that we are not here trying to model well-defined groups in parallel with the social network, but rather we simply use the association of users with places in network generation, in a way inspired by the focused organization theory of social ties. Most recently, Allamanis et al. [27] presented a model for the evolution of specifically location-based social net- works. While this model is similar to that we present here in that it takes into account place information in the placement of social ties, they seek to model some- thing fundamentally different. Our work is concerned with creating a place-based network where social ties are reflected in actual meetings between people, not the on- line social network that commonly contains some ties be- tween people who have never met. 3 III. EMPIRICAL ANALYSIS B. City social networks In this section, we will describe the Foursquare dataset and explain how we define the city-level social networks. We will then address the question: what do intra-city so- cial networks look like, and do they have common struc- tural characteristics? We show that the city-level social networks for different cities not only have similar struc- ture, but also that they have the known properties of gen- eral social networks, namely: a power-law degree distri- bution, small-world properties (high clustering and short paths), and strong community structure. Furthermore, since the city-level networks have the structure expected of social networks in the real world, we can use the place information from the Foursquare dataset to address the question: is this large location- based social dataset consistent with the theory of focused tie organization? The idea is that we can investigate on a large scale whether the idea of focused tie organiza- tion, with places acting as the foci, could play a part in the structure of these networks. To this end, we analyze the spatial properties of the social networks, in order to inform our definition of a model for place-focused social networks. A. Dataset description The dataset was collected from Foursquare, which is an online location-based social service where users check in to their current location using a mobile application, and share these check-ins with their friends. We collected all of the check-ins from Foursquare posted on Twitter dur- ing the 10 months between November 2010 and Septem- ber 2011; it is estimated that this is 20-25% of Foursquare check-ins made during this period. The dataset therefore consists of check-ins having the form (userID, placeID, timestamp). We also have the Twitter friend lists of the users concerned, which we use to define a social network, and venue information down- loaded directly from Foursquare. For every venue, we have the venue ID, venue name, latitude, longitude, and a category giving some indica- tion of the semantics of the place: this is one of the top- level categories provided by Foursquare, namely: Arts and Entertainment, College and University, Food and Drink, Nightlife Spot, Outdoors and Recreation, Profes- sional and Other Places, Residence, Shop and Service, and Travel and Transport. In the following, we present the data from five large US cities for which we have a large number of users and check-ins: Atlanta, Boston, Chicago, Minneapolis and Seattle. The numbers of users, venues and check-ins in the dataset for each city are shown in Table I. We define a place-based social network for each city, using the Twitter friend lists of the users in the dataset. Since we want to study specifically the social networks within cities grounded in physical space, we constructed a social network for each city by considering users to be friends when each is in the friend list of the other, and the connected users have checked in to at least one of the same places on Foursquare. We require the tie to be reciprocated in Twitter's directed social graph to remove ties where, for example, many users may fol- low a celebrity without actually knowing them, but the celebrity does not follow all of their followers in return. A reciprocal tie must be approved by both users, and so better represents some definition of friendship. The place requirement is imposed in order to ground the network in physical space. Formally, for each city, given: 1. The set of n users: U = {u1, u2, u3, . . . , un} 2. The Twitter friend lists for each user ui ∈ U : Fi = {uj ui follows uj on Twitter} 3. The sets of venues for each user ui ∈ U : Vi = {v user i has checked in to venue v in the city} we construct the social graph G(U, E) where the n nodes U of the graph represent the users and the graph G has an undirected edge (ui, uj)in the set of edges E whenever: 1. ui ∈ Fj and uj ∈ Fi, i.e. the users both follow one another on Twitter, and 2. Vi ∩ Vj > 0, i.e. the users have both checked in to at least one of the same places. C. Structural properties We first analyze the structures of the city-level social networks and confirm that they exhibit well-known char- acteristics of social networks: a power-law degree dis- tribution, small-world properties (high clustering with respect to a random network, and small shortest path lengths) and strong community structure. 1. Degree distribution Many networks, from those in biological systems such as metabolic networks, to technological structures such as the topology of the Internet and the page structure of the World Wide Web, exhibit a power-law degree distri- bution, and social networks are known also to have this property [9, 11, 28, 29]. Essentially, this means that the number of others to whom an individual in the network City Atlanta Boston Chicago Users Venues Check-ins 368,608 28,275 18,270 23,579 13,243 296,150 715,652 42,791 33,261 235,793 Minneapolis 13,396 12,696 16,205 15,051 260,023 Seattle 4 TABLE I. Number of users, number of venues, and number of check-ins in the dataset, for each of the five cities. City N K NGC C Atlanta Boston Chicago d dr Q Qr 13,011 46,756 11,476 0.16 0.0006 4.6 4.6 0.53 0.17 10,478 41,505 8,816 0.17 0.0010 4.3 4.0 0.45 0.15 19,931 84,778 17,287 0.16 0.0004 4.6 4.9 0.47 0.14 Minneapolis 6,499 30,640 5,914 0.18 0.0016 4.4 4.2 0.41 0.12 7,445 28,466 6,392 0.18 0.0008 4.4 4.6 0.46 0.16 Seattle Cr TABLE II. Numbers of nodes N and edges K, nodes NGC in the giant connected component, average clustering coefficient C, average shortest path length d, and modularity Q, average clustering coefficient Cr, shortest path length dr, and modularity Qr in a random network with the same number of nodes and edges, for each of the five cities. The high clustering coefficient and similar shortest path length with respect to the random network shows that these social networks have small-world properties. is connected is distributed as a power-law, so that while most have a small number of ties, there are a few hub nodes with a large number of connections. Figure 1 shows the degree distributions of the social networks for the cities we observe in our dataset. We can see that indeed, the degree distributions of the city-level networks in our dataset resemble power-laws. While the exact numbers of people having each degree differ accord- ing to the number of people in the dataset for each city, the exponents of the distributions, reflected in the slopes of the graphs, are remarkably similar across the cities. Using the methodology presented in [30], we have con- firmed the power-law distribution and measured an aver- age exponent of 2.76 across the 50 cities in the Foursquare dataset with the most check-ins. 2. Clustering Another commonly observed property of social net- works is a relatively high level of clustering, compared to the level seen in a random graph [8]. In terms of so- cial relationships, this corresponds to the fact that many of an individual's friends are likely to be friends with one another. The level of clustering in a graph can be measured by the clustering coefficient. The clustering co- efficient C of a node with N neighbors is defined as the number of links between these N neighbors, divided by the number of possible links that could exist between the neighbors, i.e. 2N N (N − 1) (1) The clustering coefficient C of a graph is then defined to be the mean clustering coefficient of all its nodes. Table II shows the clustering coefficients observed in the networks for our five cities under analysis. We can see that the values are consistent across all the five networks, being between 0.1 and 0.2. This is much higher, on the order of thousands of times, than the level observed in random graphs with the same numbers of nodes and edges, as shown in the table. 3. Average shortest path length Social networks are known commonly to have a low average shortest path length. A shortest path from one node m to another n is defined to be the smallest number of steps that are needed to reach n, starting at m and travelling along edges in the graph. The average shortest path length is then defined to be the mean value over all pairs in the graph. This quantity is defined only for connected graphs, and so for our networks we consider only the giant component that always contains at least 80% of the nodes in the graph, the presence of which is indeed another characteristic commonly seen in social networks [12]. The average shortest path lengths d in the giant components of the social networks in the dataset can be seen in Table II, and can be seen to be comparable to those in random graphs, which together with the high clustering denotes so-called small-world networks [8]. 4. Community structure The final prominent structural feature of social net- works we consider here is community structure. So- cial networks tend to exhibit strong community struc- ture, that is, the nodes are arranged into groups tightly connected by many social ties between members of the group, and these groups are more loosely interconnected to make up the entire network [11, 31, 32]. One measure of the strength of community structure in a network is 5 FIG. 1. Degree distributions in the city-level social graphs; each distribution plotted on a log-log scale presents a linear functional form similar to a power-law, with many people having few friends and a few people having very many friends. While the exact numbers of people having each degree differ according to the number of people in the dataset for each city, the exponents of the distributions, reflected in the slopes of the graphs, are remarkably similar across the cities. The average exponent across the 50 cities in the Foursquare dataset with the most check-ins is 2.76. the modularity Q of a partition of a network into its com- munities. Values of 0.3 or above are generally considered high, and indicative of the kind of community structure commonly seen in various biological, technological and social networks [33]. We partition our social networks into communities us- ing the Louvain algorithm [34], a popular community de- tection method that scales well for large networks, and measure the modularity of the partition in each case. The values of Q shown in Table II show that the place-based social networks in the five cities do indeed have strong community structure. D. Spatial properties We now investigate in more detail the spatial proper- ties of the social networks in the dataset, to try to gain in- sight into the nature of place-based friendships. We study the popularity distribution of places in the dataset and find that it resembles a power-law, similarly to the degree distribution of the social network. We further analyze tri- angles in the network and find that most triads have at least one common place, suggesting that some intra-place triadic closure mechanism could be at work. Finally we study the likelihood of friendship between colocated peo- ple given the category of that place, and show that the type of place where people meet has a strong influence on the probability that they are friends. 1. Place popularity Figure 2 shows the popularity of places in each city, de- fined to be the number of unique people who have checked in at that place in the dataset. The place popularity dis- tribution on a log-log scale presents a linear functional form similar to a power-law, with many places having few visitors and yet a few places being extremely popu- lar. The mean exponent for the power-law distributions we measured across the 50 cities with the most check-ins was 1.87. 2. Proportions of triads having one common place We consider the groups of three friends u1, u2 and u3 where the ties (u1, u2), (u2, u3) and (u3, u1) all exist in the social graph, and examine the proportion of such triangles where at least one place has been visited by all three people. Table III shows the proportions for each of the five cities. The vast majority (at least 70%) of triangles are such that each pair of friends making up a triad shares at least one place, indicating the presence of clustering around common places. It would therefore seem promising to consider including triangles made up of people visiting the same places in our model. 110100Number of friends110100100010000Number of peopleAtlantaBostonChicagoMinneapolisSeattle6 FIG. 2. Numbers of people checking in at a place in the dataset. The place popularity distribution on a log-log scale presents a linear functional form similar to a power-law, with many places having few visitors and yet a few places being extremely popular. The mean exponent for the power-law distributions we measured across the 50 cities with the most check-ins was 1.87. City Proportion of triangles Atlanta Boston Chicago Minneapolis Seattle 0.90 0.81 0.80 0.71 0.84 TABLE III. Proportions of social triangles where all three friends have at least one common place. In each city, more than 70% of triangles in the place-based social network are such that the same place is shared between each pair of friends in the triangle. 3. Types of places and likelihood of friendship for a city, given that they have been colocated at a place. We now investigate the potential for meetings at vari- ous places around the city to foster or to reinforce friend- ships. To this end, we make use of the categories pro- vided by Foursquare for each venue in their database. Each place has one category assigned by Foursquare, and the possible categories for a places are: Arts and Enter- tainment, College and University, Food, Nightlife Spot, Outdoors and Recreation, Professional and Other Places, Residence, Shop and Service, and Travel and Transport. We consider pairs of users who have been colocated at venues, that is, they have not only visited the same place but they have also had the chance to meet there through being there at the same time. Since check-ins are identified by a single timestamp, we consider users to have been colocated at a venue if they both checked in there within a 1 hour time window. We then compute the probability that a pair have a tie in the social network Figure 3 shows these probabilities for each city. We can see that some categories (Food, Nightlife and Resi- dence) have a far greater likelihood that colocated pairs are friends than the others do, which suggests that some places have a greater potential for fostering and for re- inforcing social ties, or acting as a focus for friendships, than others. Specifically, across the cities, the probability that a pair colocated at a Food, Nightlife Spot, or Resi- dence venue are friends is between 0.1 and 0.2. The prob- ability for the categories Professional and Other Places, and Shop and Service, is lower, being between 0.05 and 0.1, and the other categories generally have a very low probability. This divides places into three categories: 'social' places, where people tend to meet with their friends, 'semi-social' categories, where some meetings between friends take place but where there are also com- 110100Number of visitors110100100010000Number of placesAtlantaBostonChicagoMinneapolisSeattle7 FIG. 3. Probability of friendship between pairs colocated at a place, by Foursquare category. Categories of places have one of three types: 'social' places, where people tend to meet with their friends, 'semi-social' categories, where some meetings between friends take place but where there are also commonly meetings with strangers, and generally 'non-social' places, where people tend to meet with strangers over friends to a greater degree than in other kinds of places. monly meetings with strangers, and generally 'non-social' places, where people tend to meet with strangers over friends to a greater degree than in other kinds of places. Again, this is consistent with the focused tie organi- zation theory as it suggests that some places have high potential for fostering or for reinforcing friendship, acting as the foci described by the theory. In the next section, we use our observations from the Foursquare dataset to define a model for this type of place-focused network, in order to test computationally whether it results in net- works with the known structural characteristics of social networks. IV. A MODEL FOR PLACE-FOCUSED SOCIAL NETWORKS Having analyzed the Foursquare dataset, we now use our observations to define a model for place-focused net- works, based around people meeting at places in the city. This represents a large-scale investigation of how the fo- cused organization theory fits with empirical data. More- over, it provides an opportunity to investigate how the way in which people meet one another at places around the city could affect the structure of the resulting so- cial network, by running the same model without various components, which we will show later. A. Model description Our model uses three main kinds of information about the places in the city: 1. The popularity of a place, i.e. the number of users who have checked in there. 2. The geographic (latitude and longitude) coordi- nates of a place. 3. The semantics of a place and the activities that take place there, as indicated by the Foursquare categories described in the previous section. These first two pieces of information are relevant to ensure that the mobility of people described by our model is consistent with the mobility patterns that we see in the real world. The information about place categories, on the other hand, is used to define the potential for a place to act as a focus for friendship, given our observations in the previous section. The procedure to generate a social network in the city is defined as follows: 1. Begin with the set of N people and V venues in the dataset for the city. 2. For each person u, consider them to have visited m places, where m is sampled from the distribution of places per user in the dataset for the city. Assign to u an initial place from the set V , chosen with ArtsCollegeFoodNightlifeOutdoorsProfessionalResidenceShopTransportCategory00.050.10.150.2Probability of friendshipAtlantaBostonChicagoMinneapolisSeattleprobability proportional to the popularity of that place. 3. For each of the m − 1 additional places that u has visited in the dataset, assign to u another place v from V , with probability qrα, where q is propor- tional to the popularity of v in the dataset, and r is inversely proportional to the rank distance of the place from the first place of u. Given a set of places V in a city, the probability of a person visit- ing place v1 ∈ V given that they have visited place v2 ∈ V is defined to be: where r ∝ 1 rankv1(v2) rankv1(v2) = n + 1 (2) (3) when n is the number of places in the city closer to v1 than v2. This method of place assignment is de- signed to reproduce observed patterns of intra-city human mobility. We choose the exponent α = 0.84 in accordance with the study of human mobility in cities [35], in which the authors found that people's movements within a wide range of cities of varying sizes and densities tend to follow this distribution. 4. For each place v in V , for each pair of people (u1, u2) who have been assigned place v: (a) Place a social tie between u1 and u2 with prob- ability p = (cid:26) pcat(v) if pop(v) ≤ 30 0.001 otherwise 8 We show that the model produces networks with the structural properties observed in the real data, and demonstrate how the model is also able to preserve im- portant spatial properties such as the popularity of places and the dispersion of places visited by mobile users in real traces. The results given for each city are the aver- age of 10 runs of the model, apart from degree distribu- tion, place popularity, and geographic span, for which we choose one distribution randomly from those generated by the 10 runs. The distributions across all of the 10 runs were similar. We then study the effect of each piece of information by removing each one in turn, running the model, and examining the properties of the resulting networks: first, we analyze the effect of intra-place triadic closure, and then examine the effect of the place information. This in- formation includes distance (place location), place cate- gory, and place popularity. Our analysis shows that these place-based features of our model are important for the generated networks to show the properties observed in real social networks. A. Full model The results of running the full model, incorporating in- formation about distance, place semantics, and including the triangles between people meeting friends of friends, are shown in Figure 4 and Table IV. according to the category of v: pcat(v) = 0.15 for 'social' places (Food, Nightlife Spot, and Residence), pcat(v) = 0.08 for 'semi-social' places (Professional and Other Places, and Shop and Service), and pcat(v) = 0.01 for all other places, in the light of the analysis in the previous section. pop(v) is the popularity of v; we discuss the choice of the threshold 30 in the following section. (b) For each of u1's existing friends f in the social network who have visited v, place a link be- tween f and u2 with probability 0.15, in line with the 'social' probability. This implements an intra-place triangle closing mechanism as suggested by the observation in the previous section that most triangles in the networks have one common place shared by all three members. (4) 1. Structural properties Figure 4 shows that the full model produces networks with a power-law degree distribution, as seen in the ac- tual social networks. As seen in Table IV, the synthetic graphs also display the structural properties of the em- pirically observed networks: high clustering (between 0.1 and 0.2), short path lengths (around 4 hops), and strong community structure (as indicated by a modularity value Q higher than 0.3). B. Spatial properties Besides the known structural properties of social net- works that are present in the generated networks, the im- portant spatial properties of user visits to places within the system are also preserved: the popularity of places, and the distribution of geographic distances between places visited by a user. V. PROPERTIES OF SYNTHETIC NETWORKS We now present the properties of the synthetic net- works generated using the procedure described above. We consider two measurements of venue popularity in cities. First, in Figure 5 we present the CCDF of place popularity distribution as seen in the real data and the 1. Place popularity 9 FIG. 4. Degree distributions of synthetic graphs. The full model produces networks with a power-law degree distribution, as seen in the empirical data for each of the five cities. Model C d Q Full 0.14 4.0 0.40 0.13 3.8 0.33 0.01 3.8 0.29 No triadic closure 0.05 4.0 0.38 No distance No categories TABLE IV. Properties of the synthetic place-based social networks. The figures shown are averages over the cities. The individual values differ between cities but are close to the average in each case. The full model produces networks with the correct properties, namely high clustering coefficient C, low average shortest path length d comparable with that seen in random graphs, and strong community structure as indicated by high modularity Q > 0.3. Removing individual components from the model affects these properties of the generated networks. assignment of places to users performed by the model, using rank distance and popularity as described in the previous section. For all cities the model approximates the empirical distribution well. The skewed distribution on a log-log scale implies that the original hierarchy of places, where a few places are very popular and many are visited by only a few people, is preserved. From a social network modeling perspective, we desire this property because the category of a place, used to determine the probability of an intra-place tie, can be re- lated to its popularity (e.g. an airport or railway station, both in the Transport category, may be very popular but also unlikely to be a focus for friendships). Therefore, we do not wish places that are not very popular in the real data to be extremely popular in the model's assignment of people to places. Figure 6 shows a different perspective of place popular- ity: for every place in the city visited by a given number of users, we plot the corresponding number of users as- signed to that place by the model. The red line shows the average modeled popularity of a place with a given empirical popularity; the value for the model assignment closely matches that in the empirical data. 2. Geographic span of individual users The model also preserves the distribution of geographic span of places visited by individual users. This is a mea- sure of the spread over space of the set of places P , with each p ∈ P having coordinates (px, py), that a user visits, and is computed as: (cid:80) (cid:80) cx = p∈P px P span(P ) = p∈P dist(p, c) P where c is defined to have coordinates (cx, cy) where 110100Number of friends110100100010000Number of peopleAtlantaBostonChicagoMinneapolisSeattle10 FIG. 5. Complementary Cumulative Distribution Function (CCDF) of place popularity, in the model (red squares), and in the real data (black circles), for each of the five cities. The method of assigning places to users according to a combination of place popularity and rank distance is effective to preserve the distribution of popularity of places. FIG. 6. Popularity of individual places in the model and in the real data, for each of the five cities. The popularity of a place is preserved by the model, so places that are unpopular in the empirical data are not disproportionately popular in the assignment of places to users generated by the model. and (cid:80) p∈P py P cy = and the function dist(p1, p2) returns the great-circle dis- tance between places p1 and p2. The distributions of users' geographic span in each of the five cities are shown in Figure 7. The model gives the same distribution of spans as in the empirical data, with users being more likely to visit other venues nearby than places further away. This is in accordance with empirical observations made in studies of human mobility in cities [35], confirming that the mechanism we have used for assigning places to users does not produce an unrealistic pattern of user visits to places. C. Effect of individual components of the model We now investigate the effect of each kind of informa- tion used by our model on the structure of the generated social networks, by running the model using all of the information except that in question, and examining the properties of the networks that result. 1. Model without triadic closure We first examine the effect of the triadic closure mech- anism on the networks, where when a person has a social tie to another at a common place, they also have a so- cial tie to that other person's friends within that place. Table IV shows that in the absence of the triadic closure mechanism, when social ties are placed only according to 1101001000Popularity0.00010.0010.010.11CCDFDataModel1101001000Popularity0.00010.0010.010.11CCDFDataModel1101001000Popularity0.00010.0010.010.11CCDFDataModel1101001000Popularity0.00010.0010.010.11CCDFDataModelSeattleMinneapolisChicagoBoston1101001000Popularity0.00010.0010.010.11CCDFDataModelAtlanta110100Data popularity110100Model popularityPointsMean110100Data popularity110100Model popularityPointsMean110100Data popularity110100Model popularityPointsMean110100Data popularity110100Model popularityPointsMean110100Data popularity110100Model popularityPointsMeanChicagoSeattleMinneapolisBostonAtlanta11 FIG. 7. Probability Density Function of the geographic span of the places visited by individual users in the model and in the real data, for each of the five cities. The span distribution is generally preserved by the model, meaning that there is the same chance that a user will visit places spread over a certain distance in the city. People are generally more likely to visit places within a small geographic range, with occasional longer trips. the probabilities of connection at given places according to their category, clustering in the graph is unsurpris- ingly much lower than in the actual networks. Triadic closure is an idea often explained in terms of social bal- ance theory [36], in that if an individual has two friends who themselves are not friends, more psychological stress will result than if those two friends are also friends and the relevant triangle exists in the social graph. One can imagine that this effect would be particularly strong if the individuals concerned were actually meeting face to face, such as would be likely when those three individuals are connected by the sharing of a common place. 2. Model without distance 3. Model without categories We study the effect of the different probabilities of ties between pairs at places of different categories by running the model without the category-based probabilities and instead using in each case a uniform probability that will produce the same number of ties in the resulting network. The table shows that this dramatically reduces the clus- tering coefficient C of the resulting network, even when the triadic closure mechanism is present. This is because the clusters of social ties tend to be around the more so- ciable places where friends tend to meet, that have higher probabilities for intra-place ties in the full model. Even when the same number of ties are placed, if these are not focused around 'social' places but spread over all the venues with equal probability, the network does not dis- play the high clustering characteristic of a social network. We see also that the community structure of the network is weaker than when the full model is used. To study the effect of the distance between places, and people's tendency to visit places closer to others that they visit, on the networks, we run the model assign- ing places to users using only the popularity of the place (instead of using the popularity combined with the rank distance as outlined in the previous section). The ta- ble shows that this reduces the modularity value Q of the resulting networks, which implies weaker community structure. This can be understood in terms of people be- ing more likely to visit, and therefore to meet with friends at, places nearby to locations they already visit. When distance is not taken into account in the model, this ad- ditional clustering effect based on geography disappears and leaves only clustering at single places produced by triadic closure, which weakens the community structure of the network. VI. DISCUSSION The mechanism of assigning people to places based on popularity and forming intra-place ties has much in com- mon with the preferential attachment model [9], which was proposed to produce scale-free networks by means of having new nodes connect to existing nodes with prob- ability proportional to their degree. That is, the graphs grow in a 'rich-get-richer' fashion; the more neighbors a node already has, the more likely it is that it acquires more neighbors, which produces graphs with a power-law degree distribution. Our model can be seen as employing a version of preferential attachment to places: more pop- ular places are more likely to attract more people, and therefore more potential pairs of friends. 0.010.1110100Geographic span (km)1e-060.00010.011PDFDataModel0.010.1110100Geographic span (km)1e-060.00010.011PDFDataModel0.010.1110100Geographic span (km)1e-060.00010.011PDFDataModel0.010.1110100Geographic span (km)1e-060.00010.011PDFDataModel0.010.1110100Geographic span (km)1e-060.00010.011PDFDataModelMinneapolisBostonChicagoSeattleAtlantaHowever, we have not used only place popularity, but have taken into account the categories of places to deter- mine the likelihood of intra-place ties. We have specified that the probabilities of social tie formation based on the categories of places are used when the places in question have 30 or fewer visitors, and a lower probability for all more popular places. In our experiments, we found that this was necessary in order to be able to obtain networks with the correct community structure, resembling what we see in the real network, where most members of a community share one common place, due to the prop- erty that most triangles have one place shared between all three friends. This might intuitively be considered to be a manifestation of the idea of the constraint of a place outlined in the 1981 paper proposing the focused orga- nization theory for social tie formation [1]. A place with high constraint forces individuals there to interact much and often (for example, a family home) resulting in more or stronger social ties than a place with low constraint where many individuals may go, but will probably not encounter one another, or where visitors to the place do not typically interact with many others (for example, a park). Highly-constraining foci that, as Feld writes, "cre- ate close-knit clusters of various sizes", correspond to the places with a fairly small number of visitors in the city. Interestingly, this threshold is the same as that found in a study of the geographic span of communities in a mo- bile phone communication network [37]. The authors ob- served that communities of 30 or fewer members tended to be geographically tight, with a 100% increase in the geographic spread of communities occurring as the num- ber of members increased from 30 to 40. This might suggest that communities at the intra-city level, such as those we observe in the place-based social networks un- der study, are intrinsically constrained in size in order to be sustainable within the city. This idea is related to previous studies that have found an apparent upper limit of around 30 on the maximum number of more intense relationships that a person may be able to maintain, with contacts who are not part of this 'inner circle' being com- municated with less frequently [38, 39]. In the city-level social networks, one way in which this threshold could manifest is as a limit on the maximum number of peo- ple with whom one is able to maintain regular face-to- face contact at a place, reflected in the size of close-knit groups based around places. As we see here, this affects 12 the community structure of the resulting social network. VII. CONCLUSIONS AND FUTURE WORK In this work, we performed a large-scale investigation into whether places could act as foci for friendships in a social network, in a way consistent with the focused orga- nization theory of social ties. We used a dataset from the online location-based social network Foursquare to ana- lyze intra-city social networks, and we confirmed that they have the structural properties observed by sociol- ogists and computational social scientists studying real- world social networks. Looking deeper into the spatial characteristics of the networks, we showed that triangles in the social network often have a single common place shared between all three members, and that certain types of places seem more likely to indicate friendship between individuals who meet at those places than others, which fits with the focused organization theory. We then used our observations to propose a model for social networks in cities based on focused tie organization with places as foci, and showed that the model generates networks with the well-known structural properties of so- cial networks. We demonstrated the importance of the places as foci in our model by analyzing the effect of each individual piece of information on the resulting networks, as well as showing that the model preserves characteristic spatial properties of empirical data. We intend as future work to investigate how this model of the formation of friendships around certain kinds of places could be exploited in these ways by online location- based social services, for example, in smarter privacy con- trols based on the semantics of a relationship inferred from types of meeting places, and providing better rec- ommendations for places to visit for groups of friends going out together. VIII. ACKNOWLEDGEMENTS Chloe Brown is a recipient of the Google Europe Fellowship in Mobile Computing, and this research is supported in part by this Google Fellowship. We ac- knowledge the support of EPSRC through grant GALE (EP/K019392). [1] S. L. Feld, American Journal of Sociology 86 (1981). [2] Foursquare, Foursquare Blog (2013). [3] B. Berjani and T. Strufe, in Proc. of the 4th Workshop on Social Network Systems (ACM, 2011). [4] X. Long, L. Jin, and J. Joshi, in LBSN'12 (2012). [5] A. Noulas, S. Scellato, N. Lathia, and C. Mascolo, in SocialCom '12 (2012). [6] X. Page, A. Kobsa, and B. Knijnenburg, in ICWSM '12 (2012). [7] I. de Sola Pool and M. Kochen, Social Networks 1 (1979). [8] D. Watts, American Journal of Sociology 105 (1999). [9] A. Barab´asi and R. Albert, Science 286 (1999). [10] J. Travers and S. Milgram, Sociometry (1969). [11] M. Girvan and M. E. J. Newman, Proc. of the National Academy of Sciences 99 (2002). [12] R. Kumar, J. Novak, and A. Tomkins, in Link Mining: Models, Algorithms, and Applications, edited by P. S. S. Yu, J. Han, and C. Faloutsos (Springer New York, 2010). [13] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee, in IMC '07 (New York, NY, USA, 2007). [14] X. Zhao, A. Sala, C. Wilson, X. Wang, S. Gaito, H. Zheng, and B. Y. Zhao, in IMC '12 (2012). [15] J. Chang and E. Sun, in ICWSM '11 (2011). [16] E. Cho, S. A. Myers, and J. Leskovec, in KDD '11 (2011). [17] D. J. Crandall, L. Backstrom, D. Cosley, S. Suri, D. Hut- and J. Kleinberg, Proc. of the National tenlocher, Academy of Sciences 107 (2010). [18] S. Scellato, A. Noulas, R. Lambiotte, and C. Mascolo, in ICWSM '11 (2011). [19] J. Cranshaw, E. Toch, J. Hong, A. Kittur, and N. Sadeh, in Proceedings of the 12th ACM international conference on Ubiquitous computing, Ubicomp '10 (ACM, New York, NY, USA, 2010) pp. 119 -- 128. [20] V. Kostakos, T. Nicolai, E. Yoneki, E. ONeill, H. Kenn, and J. Crowcroft, Personal and Ubiquitous Computing 13, 355 (2009). [21] V. Kostakos, E. O'Neill, A. Penn, G. Roussos, and D. Pa- padongonas, ACM Transactions on Computer-Human Interaction (TOCHI) 17, 2 (2010). [22] D. Romero and J. Kleinberg, in ICWSM '10 (2010). [23] J. Leskovec, J. Kleinberg, and C. Faloutsos, in KDD '05 (2005). [24] S. Lattanzi and D. Sivakumar, in STOC '09 (2009). [25] J. Yang and J. Leskovec, in ICDM '12 (2012). [26] E. Zheleva, H. Sharara, and L. Getoor, in KDD '09 (2009). 13 [27] M. Allamanis, S. Scellato, and C. Mascolo, in IMC '12 (2012). [28] L. A. N. Amaral, A. Scala, M. Barth´el´emy, and H. E. Stanley, Proc. of the National Academy of Sciences 97 (2000). [29] A. Wagner and D. A. Fell, Proc. Royal Society of London. Series B: Biological Sciences 268 (2001). [30] A. Clauset, C. Shalizi, and M. Newman, SIAM review 51 (2009). [31] S. Fortunato, Physics Reports 486 (2010). [32] J. Onnela, D. Fenn, S. Reid, M. Porter, P. Mucha, M. Fricker, and N. Jones, Physical Review E 86 (2012). [33] M. E. J. Newman, Proc. of the National Academy of Sciences 103 (2006). [34] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, Journal of Statistical Mechanics: Theory and Experiment 2008 (2008). [35] A. Noulas, S. Scellato, R. Lambiotte, M. Pontil, and C. Mascolo, PloS One 7 (2012). [36] D. Cartwright and F. Harary, Psychological Review 63 (1956). [37] J.-P. Onnela, S. Arbesman, M. C. Gonz´alez, A.-L. Barab´asi, and N. A. Christakis, PLoS ONE 6 (2011). [38] R. Hill and R. Dunbar, Human Nature 14, 53 (2003). [39] A. Sutcliffe, R. Dunbar, J. Binder, and H. Arrow, British Journal of Psychology 103, 149 (2012).
1803.05068	2	1803	2018-11-26T06:56:05	AURORA: Auditing PageRank on Large Graphs	[ "cs.SI", "physics.soc-ph" ]	Ranking on large-scale graphs plays a fundamental role in many high-impact application domains, ranging from information retrieval, recommender systems, sports team management, biology to neuroscience and many more. PageRank, together with many of its random walk based variants, has become one of the most well-known and widely used algorithms, due to its mathematical elegance and the superior performance across a variety of application domains. Important as it might be, state-of-the-art lacks an intuitive way to explain the ranking results by PageRank (or its variants), e.g., why it thinks the returned top-k webpages are most important ones in the entire graph; why it gives a higher rank to actor John than actor Smith in terms of their relevance w.r.t. a particular movie? In order to answer these questions, this paper proposes a paradigm shift for PageRank, from identifying which nodes are most important to understanding why the ranking algorithm gives a particular ranking result. We formally define the PageRank auditing problem, whose central idea is to identify a set of key graph elements (e.g., edges, nodes, subgraphs) with the highest influence on the ranking results. We formulate it as an optimization problem and propose a family of effective and scalable algorithms (AURORA) to solve it. Our algorithms measure the influence of graph elements and incrementally select influential elements w.r.t. their gradients over the ranking results. We perform extensive empirical evaluations on real-world datasets, which demonstrate that the proposed methods (AURORA) provide intuitive explanations with a linear scalability.	cs.SI	cs	AURORA: Auditing PageRank on Large Graphs Jian Kang Arizona State University [email protected] Yinglong Xia Huawei Research America [email protected] Meijia Wang Arizona State University [email protected] Wei Fan Tencent [email protected] Nan Cao Tongji University [email protected] Hanghang Tong Arizona State University [email protected] 8 1 0 2 v o N 6 2 ] I S . s c [ 2 v 8 6 0 5 0 . 3 0 8 1 : v i X r a Abstract -- Ranking on large-scale graphs plays a fundamental role in many high-impact application domains, ranging from information retrieval, recommender systems, sports team man- agement, biology to neuroscience and many more. PageRank, together with many of its random walk based variants, has become one of the most well-known and widely used algorithms, due to its mathematical elegance and the superior performance across a variety of application domains. Important as it might be, state-of-the-art lacks an intuitive way to explain the ranking results by PageRank (or its variants), e.g., why it thinks the returned top-k webpages are the most important ones in the entire graph; why it gives a higher rank to actor John than actor Smith in terms of their relevance w.r.t. a particular movie? In order to answer these questions, this paper proposes a paradigm shift for PageRank, from identifying which nodes are most important to understanding why the ranking algorithm gives a particular ranking result. We formally define the PageRank auditing problem, whose central idea is to identify a set of key graph elements (e.g., edges, nodes, subgraphs) with the highest influence on the ranking results. We formulate it as an opti- mization problem and propose a family of effective and scalable algorithms (AURORA) to solve it. Our algorithms measure the influence of graph elements and incrementally select influential elements w.r.t. their gradients over the ranking results. We perform extensive empirical evaluations on real-world datasets, which demonstrate that the proposed methods (AURORA) provide intuitive explanations with a linear scalability. Index Terms -- Graph mining, PageRank, explainability I. INTRODUCTION Ranking on graph data is a way to measure node importance and plays a fundamental role in many real-world applica- tions, ranging from information retrieval [1], recommender systems [2], social networks [3], sports team management [4] to biology [5] and neuroscience [6]. Among others, PageR- ank [1], together with many of its random walk based variants, is one of the most well-known and widely used ones. Its mathematical elegance lies in that it only requires the topo- logical structure and the associated edge weights as the input. Such a generality makes it applicable to networks1 from many different application domains. PageRank has a strong ability answering questions like what is the most important page in the World Wide Web; 1In this paper, we use 'graph' and 'network' interchangably. (a) Potential influential edges. (b) Potential influential node. Fig. 1: Example of potential influential edges and node. who is the most influential person in a collaboration network. Despite its superior performance on graph ranking, PageRank lacks intuitive ways to give answers to questions like why the top-k returned webpages are the most important ones; why actor John ranks higher than actor Smith in terms of their relevance w.r.t. a particular movie. How the ranking results are derived from the underlying graph structure has largely remained opaque to the end users, who are often not experts in data mining and mathematics. To address this challenge, we aim to explain PageRank by finding the influential graph elements (e.g., edges, nodes and subgraphs), which we formulate as PageRank auditing problem. The key idea is to quantitatively understand how the ranking results would change if we perturb a specific graph element. To be specific, we measure the influence of each graph element by the rate of change in a certain loss function (e.g., Lp norm, etc.) defined over the ranking vector. We believe that auditing graph ranking can benefit many real- world applications. First, it can render the crucial explain- ability of such ranking algorithms, by identifying valuable information of influential graph elements. Thus, it can help answer questions like why a given node ranks on the top of the ranking list, which link leads to the ranking vector in a certain way. Furthermore, users can use the auditing results to optimize the network topology. In addition, it may help identify the vulnerabilities in the network (e.g. links between two clusters, cutpoints of clusters as shown in Figure 1a and Figure 1b). With the auditing result, users may find several links or nodes that have the greatest influence on the ranking results. It can help users identify if there exist suspicious individuals that manipulate the ranking results by linking heavily with unrelated, off-query topics. Finally, with TABLE I: Table of Symbols Symbols G = (V, E) (i, j) A A(i, j) A(i, :) A(:, j) A(cid:48) A−1 e r r (i) Tr (A) f (r) n m c Definitions the input network edge from node i to node j adjacency matrix of the input network the element at ith row and jth column ith row of matrix A jth column of matrix A transpose of the matrix A inverse of the matrix A the teleportation vector in PageRank PageRank of the input network ranking score of i Trace of the matrix A a loss function over ranking vector r number of nodes in the input network number of edges in the input network damping factor in PageRank the knowledge of graph ranking auditing, we may design a more robust ranking algorithm that is hard to be manipulated by users with strategies like linking heavily [7]. The main contributions of the paper are summarized as follows. • Problem Definition. We formally define PageRank audit- ing problem, and formulate it as an optimization problem, whose key idea is to measure the influence of different graph elements as the rate of change in the PageRank results. • Algorithms and Analysis. We propose fast approxima- tion algorithms to solve the PageRank auditing problem. The algorithms achieve a (1 − 1/e) approximation ratio with a linear complexity. • Empirical Evaluations. We perform extensive experi- ments on diverse, real-world datasets. The experimental results demonstrate that our proposed methods (a) provide reasonable and intuitive information to find influential graph elements, and (b) scale linearly w.r.t. the graph size. The rest of the paper is organized as follows. Section II formally defines the PageRank auditing problem. Section III introduces our proposed algorithms. Then we provide experi- mental evaluations in Section IV. After reviewing related work in Section V, we conclude the paper in Section VI. II. PROBLEM DEFINITION In this section, we first present a table of symbols (Table 1) that contains the main notations used throughout the paper, and then review PageRank for ranking nodes on graphs, and finally give a formal definition of the PageRank auditing problem. We use bold upper-case letters for matrices (e.g., A), bold lower-case letters for vectors (e.g., a), calligraphic fonts for sets (e.g., S), and lower-case letters for scalars (e.g., c). For matrix indexing conventions, we use the rules similar to Matlab that are shown as follows. We use A(i, j) to denote the entry of matrix A at ith row and jth column, A(i, :) to denote the ith row of matrix A, and A(:, j) to denote the jth column of matrix A. We use prime to denote the transpose of matrix (i.e., A(cid:48) is the transpose of matrix A). Given a graph G with n nodes and m edges, PageRank essentially solves the following linear system, r = cAr + (1 − c)e (1) where e is the teleportation vector with length n and A is the normalized adjacency matrix of the input graph. In n 1; in PageRank, e is chosen as the uniform distribution 1 personalized PageRank, e is a biased vector which reflects user's preference (i.e., 'personalization') [8]; in random walk with restart [9], all the probabilities are concentrated on a single node. A default choice for the normalized adjacency matrix A is the column-normalized matrix (i.e., the stochastic matrix). A popular alternative choice is the normalized graph Laplacian matrix. In fact, as long as the largest eigenvalue of A is less than 1/c, a fix-point solution of the above linear system will converge to r = (1 − c)(I − cA)−1e [10]. In this paper, we use this general form for the adjacency matrix A. For the ease of description, we also define Q = (I − cA)−1, and r = pg(A, e, c) as the resulting PageRank vector with A, e, and c as the corresponding inputs. Regarding explainable learning and mining techniques, Pang et al. [11] propose a novel notation of influence functions to quantify the impact of each training example on the underlying learning system (e.g., a classifier). The key idea is to trace the model's prediction back to its training data, where the model parameters were derived. In this way, it learns how a perturbation of a single training data will affect the resulting model parameters, and then identifies the training examples that are most responsible for model's predictions. Our proposed method to explain the PageRank results is built upon the principle outlined in [11]. To be specific, we aim to find a set of graph elements (e.g., edges, nodes, a subgraph) such that, when we perturb/remove them, the ranking vector will have the greatest change. Formally, we define PageRank auditing problem as follows: Problem 1: PageRank Auditing Problem. Given: a graph with adjacency matrix A, PageRank vector r, a loss function f over its PageRank vector, user-specific element type (e.g. edges vs. nodes vs. subgraphs), and an integer budget k; Find: a set of k influential graph elements that has the largest impact on the loss function over its PageRank vector f (r). III. PROPOSED ALGORITHM In this section, we propose a family of algorithms (AURORA), to solve PageRank auditing problem (Problem 1), together with some analysis in terms of effectiveness as well as efficiency. A. Formulation The intuition behind the proposed methods is to find a set of key graph elements (e.g., edges, nodes, subgraphs) whose perturbation/removal from the graph would affect the PageR- ank results most. To be specific, let r = pg(A, e, c) be the PageRank vector of the input graph A, and rS = pg(AS , e, c) be the new PageRank vector after removing the graph elements in set S. We formulate the PageRank auditing problem as the following optimization problem. maxS ∆f = (f (r) − f (rS ))2 s.t. S = k Descriptions Lp norm (2) In order to solve the above optimization problem, we need to answer two key questions including (Q1) how to quantitatively measure the influence of an individual graph element w.r.t. the objective function; and (Q2) how to collectively find a set of k graph elements with the maximal influence. We first present our proposed solution for Q1 in Subsection III-B, and then propose three different algorithms for Q2 in Subsections III-C, III-D and III-E, depending on the specific type of graph elements (i.e., edges vs. nodes vs. subgraphs). B. Measuring Graph Element Influence To measure how f (r) will change if we perturb/remove a specific graph element, we define its influence as the rate of change in f (r). Definition 1: (Graph Element Influence). The influence of an edge (i, j) is defined as the derivative of the loss function f (r) with respect to the edge, i.e., I(i, j) = df (r) dA(i,j). The influence of a node is defined as the aggregation of all inbound and outbound edges that connect to the node., i.e., I(i) = I(i, j). And the influence of a subgraph is defined as the aggregation of all edges in the subgraph S, I(S) = (cid:80) [I(i, j) + I(j, i)] + I(i, j). n(cid:80) n(cid:80) j=1,j(cid:54)=i j=1,j=i We can see that the influence for both nodes and subgraphs can be naturally computed based on the edge influence. There- fore, we will focus on how to measure the edge influence. By the property of the derivative of matrices, we first rewrite the influence df (r) i,j∈S dA(i,j) as (cid:40) ∂f (r) df (r) dA = ∂A + ( ∂f (r) ∂f (r) ∂A , Directly calculating ∂f (r) rule: ∂A )(cid:48) − diag( ∂f (r) ∂A ), if undirected if directed (3) ∂A(i,j) is hard, and we resort to the chain ∂f (r) ∂A(i, j) = ∂f (r) ∂r ∂r ∂A(i, j) (4) TABLE II: Choices of f (·) functions and their derivatives Functions f (r) = rp n(cid:80) i=1 ∂f Derivatives ∂r = r◦rp−2 rp−1 n(cid:80) ∂r = [ exp(r(i)) ∂r = (M + M(cid:48))r exp(r(i) i=1 ∂f p ∂f ] Soft maximum f (r) = log( exp(r(i))) Energy norm (M in Energy Norm is a Hermitian positive definite matrix.) f (r) = r(cid:48)Mr ∂r b) 2. Computing ∂A(i,j) : Recall that A is the adjacency matrix of the input graph, and the PageRank solution can be written as solving the linear system in Eq. (1). Then we have that ∂r ∂A(i, j) = c ∂A ∂A(i, j) r + cA ∂r ∂A(i, j) Move the second term in Eq. (5) to the left, we have that (5) (6) (7) (I − cA) ∂r = c ∂A(i, j) = c(I − cA)−1 ∂A r ∂A(i, j) ∂A ∂A(i, j) r ∂r ∂A(i, j) By the property of the first order derivative of matrix, we have that ∂A ∂A(i, j) = Sij + Sji, Sij, if undirected if directed (8) (cid:40) where Sij is the single-entry matrix with 1 at ith row and jth column and 0 elsewhere. Recall that Q = (I− cA)−1, we can re-write Eq. (7) as the following equation. ∂r ∂A(i, j) = cr(j)Q(:, i) (9) Combine everything together, we get the closed-form solution for calculating the influence of an edge (i, j) as follows: ∂f (r) ∂A(i, j) = 2cr(j)Tr(r(cid:48)Q(:, i)) (10) Following Eq. (10), we get the matrix of gradients for all edges as ∂f (r) ∂A = 2c(1 − c)Q(cid:48)re(cid:48)Q(cid:48) (11) Next, we present the details on how to solve each partial a) Computing ∂f (r) derivative in Eq. (4) one by one. ∂r : Here we discuss the choices of f (·) function. In this paper, we choose f (·) to be squared L2 norm for simplicity. However, it is worth mentioning that the proposed methods are applicable to a variety of other loss functions as well. We list some alternative choices and their corresponding derivatives in Table II. In the table, Lp norm is the most commonly-used vector norm that measures the overall sizes of the vector. Some commonly-used Lp norm includes L1 norm and L2 norm (also known as the Euclidean norm); soft maximum is used to approximate the maximum value of elements in the vector; energy norm is a measurement often used in system and control theory to measure the internal energy of vector. Two major computational challenges in calculating ∂f (r) ∂A lie in (1) calculating Q with O(n3) time complexity and (2) O(n2) space complexity to save the matrix of gradients. We address both challenges by exploring the low-rank structure of ∂f (r) ∂A . From Eq. (11), we can show that it can be re-written as the following low-rank form ∂f (r) ∂A = 2c(1 − c)Q(cid:48)re(cid:48)Q(cid:48) = 2c(Q(cid:48)r)r(cid:48) (12) where Q(cid:48)r is an n × 1 vector and r is an n × 1 PageRank vector. Since r = (1 − c)Qe, we have that Q(cid:48)r is a person- alized PageRank on the reverse of the input graph with r as teleportation vector with a constant scaling. With this in mind, we do not need to calculate Q explicitly, or to save the entire matrix directly. Instead, we can use power method to calculate r and Q(cid:48)r, each with O(m) time and save these two vectors ∂A at the ith with O(n) space. To extract the element of ∂f (r) row and the jth column, we simply calculate the product of the ith element in Q(cid:48)r and the jth element in r, and scale it by 2c, which takes O(1) time. C. Auditing by Edges: AURORA-E Due to its combinatorial nature, straight-forward methods for solving the optimization problem in Eq. (2) are not feasible. The key behind the proposed family of algorithms is due to the diminishing returns property of Problem 1, which is summarized in Theorem 1 Theorem 1: (Diminishing Returns Property of Problem 1). For any loss function listed in Table 2, and for any set of graph elements S, which could be either a set of edges, nodes or subgraphs, in the given graph, its influence measure I(S) defined in Definition 1 is (a) normalized; (b) monotonically non-decreasing; (c) submodular, where S ⊆ E. Proof: We only prove the diminishing returns property in the edge case. The proofs for nodes and subgraphs are similar and thus is omitted due to the space limitation. Let I(S) = (cid:80) I(i, j). It is trivial that if there is no edge selected, the influence is 0. Thus it is normalized. Let I,J ,K be three sets and I ⊆ J . We further define three sets S,T ,K as follows: S = I ∪ K, T = J ∪ K and R = J \ I, then we have (i,j)∈S I(i, j) − (cid:88) (i,j)∈I I(i, j) (cid:88) (cid:88) (cid:88) (i,j)∈J (i,j)∈R (i,j)∈J \I I(i, j) I(i, j) I(J ) − I(I) = = = ≥ 0 which proves that I(S) is monotonically non-decreasing. Finally, we prove that it is submodular. Define P = T \ S. We have that P = (J ∪ K) \ (I ∪ K) = R \ (R ∩ K) ⊆ R = J \ I. Then we have I(T ) − I(S) = I(i, j) ≤ I(J ) − I(I) (cid:88) (i,j)∈P : The adjacency matrix A, integer budget k Algorithm 1: AURORA-E Input Output: A set of k edges S with the highest influence 1 initialize S = ∅; 2 initialize c (e.g., c = 1/2 max eigenvalue(A)); 3 calculate PageRank r = pg(A, e, c); 4 calculate partial gradients ∂f (r) 5 calculate gradients df (r) 6 while S (cid:54)= k do 7 I(i, j) with Eq. (3); ∂A by Eq. (12); dA by Eq. (3); (i,j) find (i, j) = argmax add edge (i, j) to S; remove (i, j), and remove (j, i) if undirected; re-calculate r, ∂f (r) dA by Eq. (3); ∂A by Eq. (12), and df (r) 8 9 10 11 return S; We can see that AURORA-E finds a (1 − 1/e) near-optimal solution with a linear complexity. Lemma 1: (Approximation Ratio of AURORA-E). Let Sk = {s1, s2, ..., sk} represents the set formed by AURORA-E, O is the optimal solution of Problem 1, I(S) is the influence defined in Definition 1. I(Sk) ≥ (1 − 1/e)I(O) Proof: Omitted for space. Lemma 2: (Time and Space Complexities of AURORA-E). Algorithm 1 is O(mk) in time and O(m + n) in space, where m and n are the numbers of edges and nodes in the input graph; and k is the budget. Proof: It takes O(m) time complexity to calculate r and ∂f (r) ∂A by applying power iterations. In the while-loop, we find the edge with the greatest influence by traversing all edges, which takes O(m) time. Time spent to re-calculate r and ∂f (r) ∂A remains the same as O(m). Since the body inside the loop will run k times, the overall time complexity is O(mk). In Algorithm 1, it takes O(m) space to save the sparse adjacency matrix A and O(n) space to save the PageRank vector r and column vector Q(cid:48)r in Eq. (12). Therefore it has O(m + n) space complexity. which proves the submodularity of the edge influence. The diminishing returns property naturally leads to a greedy algorithm to obtain a near-optimal solution for solving Prob- lem 1. We first present the algorithm for auditing by edges in this subsection. The algorithms for auditing by nodes and by subgraphs will be presented in Subsections III-D and III-E, respectively. With the diminishing returns property, we propose AU- RORA-E (Algorithm 1) algorithm to find top-k influential edges. The key idea of AURORA-E is to select one edge and update the gradient matrix at each of the k iterations. The effectiveness and efficiency of the proposed AURORA- E are summarized in Lemma 1 and Lemma 2, respectively. D. Auditing by Nodes: AURORA-N By Theorem 1, the influence of nodes also enjoys the diminishing returns property. Following this, we propose a greedy algorithm AURORA-N (Algorithm 2) to find a set of k influential nodes with (1 − 1/e) approximation ratio with a linear complexity. The efficiency of the proposed AURORA-N is summarized in Lemma 3. Lemma 3: (Time and Space Complexities of AURORA-N). Algorithm 2 is O(mk) in time and O(m + n) in space, where m and n are the numbers of edges and nodes in the input graph; and k is the budget. Proof: Omitted for space. Algorithm 2: AURORA-N : The adjacency matrix A, integer budget k Input Output: A set of k nodes S with highest influence 1 initialize S = ∅; 2 initialize c (e.g., c = 1/2 max eigenvalue(A)); 3 calculate PageRank r = pg(A, e, c); 4 calculate partial gradients ∂f (r) 5 calculate gradients df (r) 6 while S (cid:54)= k do 7 ∂A by Eq. (12); dA by Eq. (3); find vi = argmax add vi to S; remove all inbound and outbound edges of vi; re-calculate r, ∂f (r) dA by Eq. (3); ∂A by Eq. (12), and df (r) I(i); i 8 9 10 11 return S; E. Auditing by Subgraphs: AURORA-S Here, we discuss how to select an influential subgraph with k nodes and we focus on the vertex-induced subgraph. With the diminishing returns property (Theorem 1) in mind, we propose AURORA-S (Algorithm 3) to greedily identify the influential subgraph with (1− 1/e) approximation ratio with a linear complexity. The efficiency of the proposed AURORA-S is summarized in Lemma 4. highest influence : The adjacency matrix A, output size k Algorithm 3: AURORA-S Input Output: A vertex-induced subgraph of k nodes S with 1 initialize S = ∅; 2 initialize c (e.g., c = 1/2 max eigenvalue(A)); 3 calculate PageRank r = pg(A, e, c); 4 calculate partial gradients ∂f (r) 5 calculate gradients df (r) 6 while S (cid:54)= k do 7 ∂A by Eq. (12); dA by Eq. (3); I(i, j); find (i, j) = argmax if S + 2 ≤ k then (i,j) add vi and vj to S; else 8 9 10 11 12 13 14 15 16 17 find the endpoint v with higher gradient; if v (cid:54)∈ S then add v to S; add the other endpoint to S; else remove all edges in S; re-calculate r, ∂f (r) Eq. (3); 18 return S; ∂A by Eq. (12), and df (r) dA by Lemma 4: (Time and Space Complexities of AURORA-S). Algorithm 3 is O(mk) in time and O(m + n) in space, where m and n are the numbers of edges and nodes in the input graph; and k is the budget. Proof: Omitted for space. F. Generalization and Variants The proposed family of AURORA algorithms assume the input graph is a plain network. However, it is worth pointing out that AURORA algorithms also work on different types of networks and other random walk-based techniques. a) AURORA on Normalized PageRank: Recall that in Section II, we remove the constraint of A being a normalized adjacency matrix and use the general form instead. The effect of removing that constraint will cause the L1 norm of PageR- ank to be not equal to 1. We show that AURORA is also able i r(i), we to work on L1 normalized PageRank. Let S(r) =(cid:80) have ∂f (r) ∂A = cQ(cid:48)(− 2f (r) S(r) 1 + r)r(cid:48) 2 S(r) (13) Then we can apply AURORA algorithms by replacing Eq. (12) with Eq. (13) b) AURORA on Network of Networks: Network of Net- works (NoN) is a type of networks first introduced in [12] with the ability to leverage the within-network smoothness in the domain-specific network and the cross-network consistency through the main network. An NoN is usually defined as the triplet R =< G,A, θ >, where G is the main network, A is a set of domain-specific network and θ is a mapping function to map the main node to the corresponding domain-specific network. In [12], CROSSRANK and CROSSQUERY are two ranking algorithms proposed to solve ranking on NoN. The authors have proved that they are actually equivalent to the well-known PageRank and random walk with restart on the integrated graph. Thus, AURORA algorithms also have the ability to audit CROSSRANK and CROSSQUERY on Network of Networks. c) AURORA on Attributed Networks: Given a large at- tributed network, it is important to learn the most influen- tial node-attribute or edge-attribute w.r.t. a query node. We show that AURORA algorithms have the ability to find top- k influential edge-attributes and node-attributes on attributed networks. The central idea is to treat attributes as attribute nodes and form an augmented graph with those attribute nodes. To support node attributes, let A be the a × a node-to-node adjacency matrix, and W be the w × a node-to-attribute matrix, then we can form an augmented graph G =(cid:0) A 0 (cid:1). To support edge attributes, similar to [13], we embed an edge-node for each edge in the input graph and define a mapping function ψ that maps each edge-node to edge attribute in the original graph. We assume A is the a × a node- to-node adjacency matrix, and b is the number of different edge attribute values. By embedding edge-node, it creates a (a + b) × (a + b) augmented graph Y. Then to find the top-k influential node-attributes and edge-attributes, we can easily W 0 Category SOCIAL COLLABORATION PHYSICAL OTHERS Network Karate Dolphins WikiVote Pokec GrQc DBLP NBA cit-DBLP cit-HepTh cit-HepPh Airport Lesmis Amazon U U D D U U U D D D D U D 7,115 1,632,803 103,689 30,622,564 34 62 5,242 42,252 3,923 12,591 27,770 34,546 1,128 77 78 159 14,496 420,640 127,034 49,743 352,807 421,578 18,736 254 TABLE III: Statistics of the datasets Type Nodes Edges 262,111 1,234,877 (In Type, U means undirected graph; D means directed graph.) run AURORA-N on the augmented attributed graphs G and Y, respectively. IV. EXPERIMENTAL EVALUATION In this section, we evaluate the proposed AURORA algo- rithms. All experiments are designed to answer the following two questions: • Effectiveness. How effective are the proposed AURORA algorithms in identifying key graph elements w.r.t. the PageRank results? • Efficiency. How efficient and scalable are the proposed AURORA algorithms? A. Setup Datasets. We test our algorithms on a diverse set of real- world network datasets. All datasets are publicly available. The statistics of these datasets are listed in Table III. • SOCIAL NETWORKS. Here, nodes are users and edges indicate social relationships. Among them, Karate [14] is a well-known network dataset of a university karate club collected by Wayne Zachary in 1977. Dolphins [15] is an undirected social network of frequent associations between dolphins in a community living off Doubtful Sound, New Zealand. WikiVote [16] is generated by Wikipedia voting data from the inception of Wikipedia till January 2008. Pokec [17] is a popular online social network in Slovakia. • COLLABORATION NETWORKS. Here, nodes are individ- uals and two people are connected if they have col- laborated together. We use the collaboration network in the field of General Relativity and Quantum Cosmology (GrQc) in Physics from arXiv preprint archive2. DBLP3 is a co-authorship network from DBLP computer science bibliography. And NBA [18] is a collaboration network of NBA players from 1946 to 2009. cit-DBLP [19] is the citation network of DBLP, a database of scientific publications such as papers and books. Each node in the network is a publication, and each edge represents 2https://arxiv.org/ 3http://dblp.uni-trier.de/ a citation of a publication by another publication. cit- HepTh [20] is an ArXiv HEP-TH (High Energy Physics - Theory) citation network. The data covers papers from January 1993 to April 2003. If a paper i cites paper j, there is a directed edge from i to j. cit-HepPh [20] is an ArXiv HEP-PH (High Energy Physics - Phenomenology) citation network. The data covers papers from January 1993 to April 2003. If a paper i cites paper j, there is a directed edge from i to j. • PHYSICAL INFRASTRUCTURE NETWORKS. This cate- gory refers to the networks of physical infrastructure enti- ties. Nodes in them correspond to physical infrastructure, and edges are connections. Airport4 is a dataset of airline traffic. Each node represents an airport in the United States, an edge (i, j) represents the airline from i to j while the edge weight stands for the normalized number of passengers. • OTHERS. This category contains networks that do not fit into the above categories. Lesmis [21] is a network of co-appearances of characters in Victor Hugo's novel "Les Miserables". A node represents a character and an edge connects a pair of characters if they both appear in the same chapter of the book. Amazon [22] is a co-purchasing network collected by crawling Amazon website. It is based on the Customers Who Bought This Item Also Bought feature. Baseline Methods. We compare our proposed methods with several baseline methods, which are summarized as follows. • Brute force (brute force). Calculate the changes by iter- ating all possible combinations of graph elements. • Random Selection (random). Randomly select k elements and calculate the change by removing them. • Top-k Degrees (degree). We first define the degree of an edge (u, v) as follows, (d(u) + d(v)) × max (d(u) + d(v)) × d(u), i∈{u,v}d(i), d(u, v) = if undirected if directed where d(u) represents the degree of node u. To audit by graph elements, we select k elements with the highest degrees. For edges, we select k edges with the highest edge degrees defined above; for nodes, we select k nodes with the highest node degrees; for subgraphs, we form a vertex-induced subgraph from k nodes with the highest degrees. • PageRank. We first define the PageRank score of an edge (u, v) as follows, r(u, v) = (r(u) + r(v)) × max (r(u) + r(v)) × r(u), i∈{u,v}r(i), if undirected if directed where r(u) is the PageRank score of node u. To audit by graph elements, we select k elements with the highest PageRank scores. That is, for edges, we select 4https://www.transtats.bts.gov/ k edges with the highest PageRank scores defined above; for nodes, we select k nodes with highest PageRank scores; for subgraphs, we form a vertex-induced subgraph from k nodes with the highest PageRank scores. • HITS. We first define HITS score of an edge (u, v) and node u as follows, HIT S(u, v) = hub(u) × hub(v) + auth(u) × auth(v) HIT S(u) = hub(u) + auth(u) where hub(u) and auth(u) represent the hub score and authority score of node u, respectively. To audit by graph elements, we select k elements with the highest HITS scores. That is, for edges, we select k edges with the highest HITS scores defined above; for nodes, we select k nodes with the highest HITS scores; for subgraphs, we form a vertex-induced subgraph from k nodes with the highest HITS scores. Metrics. Here, we choose the loss function to be squared L2 norm. We quantify the performance of auditing by the goodness score ∆f (defined in Eq. (2)) of the graph elements S found by the corresponding algorithms. Repeatability and Machine Configuration. All datasets are publicly available. We will release the code of our proposed algorithms upon the publication of the paper. All experiments are performed on a virtual machine with 4 Intel i7-8700 CPU cores at 3.4GHz and 32GB RAM. The operating system is Windows 10. All codes are written in Python 3.6. B. Effectiveness Results a) Quantitative Comparison: We perform effectiveness experiments with the baseline methods. We set k from 1 to 10 and find k influential edges and nodes, respectively. We set k only from 2 to 10 to find an influential subgraph of size-k, respectively. This is because a vertex-induced subgraph with only 1 node does not contain any edge and therefore is meaningless for the purpose of PageRank auditing. We conduct the experiments with brute force on relatively small datasets to obtain the ground-truth of Eq. (2), including Karate, Dolphins and Lesmis. On other larger datasets, searching a ground-truth with k most influential elements is prohibitively expensive due to its combinatorial nature. For example, even if we use the small Lesmis dataset, it will take over a day to find ground-truth with k = 5. Therefore, the results by brute force are absent on the remaining larger datasets. The results of quantitative comparison across 9 different datasets are shown from Figure 2 to Figure 4. From those figures, we have the following observations: (1) our family of AURORA algorithms consistently outperform other baseline methods (except for brute force) on all datasets; (2) on the three small datasets with ground-truth brute force, the performance of the proposed AURORA algorithms is very close to the ground-truth. Figure 5 and Figure 6 shows the effect of k on auditing results. We can observe the following findings from those figures: (1) our family of AURORA algorithms in- crementally find influential graph elements w.r.t. the budget k; (2) the proposed AURORA algorithms consistently outperform baseline methods on different budgets. b) Case Studies on Airport Dataset: A natural use case of our AURORA algorithms is to find influential edges and nodes in a given graph. To demonstrate that our algorithms are indeed able to provide intuitive information, we test our algorithms on the Airport dataset. This dataset was manually created from commercial airline traffic data in 2017, which is provided by United States Department of Transportation. More detailed description and statistics of this dataset can be found in Subsection IV-A. We perform AURORA-E and AURORA-N to find the most influential airlines (edges in the graph) and airports (nodes in the graph) across United States with k = 7. Edges selected by AURORA-E are ATL-LAX, LAX-ATL, ATL-ORD, ORD-ATL, ATL-DEN, DEN-ATL and LAX-ORD. In contrast, PageRank selects ATL-LAS instead of DEN-ATL and ATL-DFW instead of LAX-ORD. DEN-ATL plays a more important role in determining the centrality (e.g., PageRank) of other airports. This is because DEN serves as one of the busiest hub airports that connects West coast and East coast; while ATL-LAS is less important in that regard, considering the existence of ATL-LAX and ATL-PHX. Comparing LAX-ORD and ATL-DFW, LAX-ORD directly connects Los Angeles and Chicago, both of them are largest cities in the United States. In the scenario of node-auditing, AURORA-N selects ATL, LAX, ORD, DFW, DEN, LAS and CLT. In contrast, PageRank selects SFO instead of CLT. CLT seems to be a more rea- sonable choice because it serves as a major hub airport, the 6th busiest airport by FAA statistics, to connect many regional airports around States like North Carolina, South Carolina, Virginia, West Virginia, etc. Compared with CLT, SFO is less influential in that regard, mainly due to the following two reasons: (1) it ranks after CLT ( 7th vs. 6th) in the list of busiest airports by FAA statistics; (2) due to the location proximity of SFO to LAX and SJC, even if this node is perturbed (i.e., absent), many surrounding airports (especially regional airports in California) could still be connected via LAX and SJC. c) Case Studies on DBLP Dataset: Another interesting use case of AURORA algorithms is sense-making in graph proximity. We construct a co-authorship network from DBLP computer science bibliography to test our algorithms. We perform AURORA-N and PageRank with k = 6. Different from the previous case study, here we use a personalized PageRank with the query node is Christos Faloutsos. In this case, the top-ranked scholars in the resulting rank- ing vector r form the proximity (i.e., 'neighborhood') of the query node (i.e., who are most relevant to Christos Faloutsos). Consequently, the nodes selected by an audit- ing algorithm indicate those important nodes in terms of mak- ing/maintaining the neighborhood of the query node. Com- paring the results by AURORA-N and PageRank, 6 of them are the same while AURORA-N selects Jure Leskovec instead of Yannis Ioannidis. This result is consistent with the intuition, since Jure Leskovec, as the former stu- dent of Christos Faloutsos with lots of joint publica- tions, plays a more prominent role in the neighborhood of Christos Faloutsos by sharing more common collabo- Fig. 2: Auditing results by edges. Budget k = 10. Higher is better. Best viewed in color. Fig. 3: Auditing results by nodes. Budget k = 10. Higher is better. Best viewed in color. Fig. 4: Auditing results by subgraphs. Budget k = 10. Higher is better. Best viewed in color. (a) Auditing by Edges (b) Auditing by Nodes (c) Auditing by Subgraphs Fig. 5: Effect of k on auditing results (cit-HepPh Dataset). Higher is better. Best viewed in color. rators. d) Case Studies on NBA Dataset: In a collaboration network, a subgraph can be naturally viewed as a team (e.g. sports team). From this perspective, AURORA-S has the potential to find teammates of a player. We set the query node as Allen Iverson. Since there are 5 players for each team on the court, we set k = 5. The subgraph selected by AURORA-S consists of Allen Iverson, Larry Hughes, Theo Ratliff, Joe Smith and Tim Thomas. In con- trast, in the resulting subgraph by PageRank, Tim Thomas is replaced by Drew Gooden. Though Drew Gooden and Allen Iverson shares many common neighbors in the network, which might cause its ranking in PageRank to be higher, they have never played in the same team together. Compared with the subgraph by PageRank, all players in the subgraph by AURORA-S have played with Iverson in Philadelphia Sixers during the time period from 1997 to 1999. C. Efficiency Results We show the running time vs. number of edges m and budget size k on Pokec dataset in Figure 7 and Figure 8. We can see that the proposed AURORA algorithms scale linearly with respect to m and k, respectively. This is consistent with our complexity analysis that the family of AURORA algorithms are linear with respect to the number of edges and the budget. D. Visualization To better understand the auditing results, we developed a prototype system with D3.js to represent the influence of graph elements visually. In the system, we use the strength of line to represent the gradient of an edge and use the size and color for the gradient of nodes. An example of visualizing hand- crafted toy graph is shown in Figure 9. It is easy to see in the figure that Node 5 is the most influential node, and edges around Node 5 is more influential than other edges, both of which are consistent with our intuition. (a) Auditing by Edges (b) Auditing by Nodes (c) Auditing by Subgraphs Fig. 6: Effect of k on auditing results (cit-HepTh Dataset). Higher is better. Best viewed in color. scores, and a node has a high authority score if it links to many nodes with high hub scores. Many variants of PageRank and HITS have been developed in the literature. To name a few, in [24], the author studies the stability of PageRank and HITS, based on which they propose two new algorithms (Randomized HITS and Subspace HITS). Ding et al. [25] provide a unified ranking method for HITS and PageRank. In [8], Haveliwala et al. propose the well known personalized PageRank by replacing the uniform teleportation vector with a biased personalized topic-specific vector; while random walk with restart [9] concentrates all teleportation probabilities to a single node. Other random walk based graph ranking methods include [26] and many more. Explainability. How to enhance interpretability of machine learning and data mining models has been attracting a lot of research interests in recent years. Two representative ways to explain such black-box predictions consist of (a) using other interpretable models to provide interpretable representation of model's predictions [27] -- [29], and (b) quantifying the influ- ence by perturbing either features or training data [11], [30] -- [33]. The proposed AURORA algorithms follow the generic principle in [11] in the context of graph ranking, and has been integrated into a prototype system [34]. Finding influential nodes on graphs is a very active research area. The central theme of these research is to find nodes to maximize the spread of influence in social networks (i.e., influence maximiza- tion) [35] -- [37]. The seminal work in influence maximization problem is attributed to Kempe et al. [38], [39], in which they have discovered its diminishing returns property. The focus of this paper is fundamentally different from the classic influence maximization problem, in the sense that we want to find most influential nodes, edges or subgraphs w.r.t. the graph ranking results, as opposed to the size of 'infected' nodes during the influence propagation process. VI. CONCLUSION In this paper, we study the problem of auditing PageRank, where we aim to find the most influential graph elements (e.g., edges, nodes, subgraphs) w.r.t. graph ranking results. We formally define the PageRank auditing problem by measur- ing influence of each graph element as the rate of change in a certain loss function defined over the ranking vector, and formulate it as an optimization problem. We further Fig. 7: running time vs. the number of edges on Pokec dataset Fig. 8: running time vs. number of k on Pokec dataset Fig. 9: Visualization of toy graph on the visualization system. Best viewed in color. V. RELATED WORK In this section, we briefly review the related work, from the following two perspectives, including (1) graph ranking and (2) explainability. Graph ranking. Regarding graph ranking, PageRank [1] and HITS [23] are probably the most well-known and widely used algorithms. PageRank measures the importance of nodes as a stationary distribution of random walks. HITS assumes that each node has two scores: hub and authority. A node has a high hub score if it links to many nodes with high authority propose a family of fast approximation algorithms, named AURORA, with (1 − 1/e) approximation ratio and a linear complexity in both time and space. The extensive experimental evaluations on more than 10 datasets demonstrate that the proposed AURORA algorithms are able to identify influential graph elements, consistently outperform baseline methods, and scale linearly to large graphs. In the future, we would like to generalize this auditing paradigm to other graph ranking methods (e.g., HITS) as well as dynamic networks. ACKNOWLEDGMENT supported by NSF (IIS-1651203, This work is IIS- 1715385 and IIS-1743040), DTRA (HDTRA1-16-0017), ARO (W911NF-16-1-0168), DHS (2017-ST-061-QA0001), NSFC (61602306, Fundamental Research Funds for the Central Uni- versities), and gifts from Huawei and Baidu. REFERENCES [1] L. Page, S. Brin, R. Motwani, and T. Winograd, "The pagerank citation ranking: Bringing order to the web." Stanford InfoLab, Technical Report 1999-66, November 1999. [2] M. Gori and A. Pucci, "Itemrank: A random-walk based scoring algorithm for recommender engines," in Proceedings of the 20th In- ternational Joint Conference on Artifical Intelligence, San Francisco, CA, USA, 2007, pp. 2766 -- 2771. [3] J. Weng, E.-P. Lim, J. Jiang, and Q. He, "Twitterrank: Finding topic- sensitive influential twitterers," in Proceedings of the Third ACM Inter- national Conference on Web Search and Data Mining. New York, NY, USA: ACM, 2010, pp. 261 -- 270. [4] F. Radicchi, "Who is the best player ever? a complex network analysis of the history of professional tennis," PloS one, vol. 6, no. 2, p. e17249, 2011. [5] R. Singh, J. Xu, and B. Berger, "Pairwise global alignment of protein interaction networks by matching neighborhood topology," in Research in computational molecular biology. Springer, 2007, pp. 16 -- 31. [6] J. J. Crofts and D. J. Higham, "Googling the brain: Discovering hierarchical and asymmetric network structures, with applications in neuroscience," Internet Mathematics, vol. 7, no. 4, pp. 233 -- 254, 2011. [7] D. F. Gleich and M. W. Mahoney, "Using local spectral methods to robustify graph-based learning algorithms," in Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York, NY, USA, 2015, pp. 359 -- 368. [8] T. H. Haveliwala, "Topic-sensitive pagerank," in Proceedings of the 11th international conference on World Wide Web. ACM, 2002, pp. 517 -- 526. [9] H. Tong, C. Faloutsos, and J.-y. Pan, "Fast random walk with restart and its applications," in ICDM. IEEE, 2006, pp. 613 -- 622. [10] L. Li, Y. Yao, J. Tang, W. Fan, and H. Tong, "Quint: On query-specific optimal networks," in KDD. ACM, 2016, pp. 985 -- 994. [11] P. W. Koh and P. Liang, "Understanding black-box predictions via influ- ence functions," in Proceedings of the 34th International Conference on Machine Learning, International Convention Centre, Sydney, Australia, 2017, pp. 1885 -- 1894. [12] J. Ni, H. Tong, W. Fan, and X. Zhang, "Inside the atoms: Ranking on a network of networks," in KDD, New York, NY, USA, 2014, pp. 1356 -- 1365. [13] R. Pienta, A. Tamersoy, H. Tong, and D. H. Chau, "Mage: Matching approximate patterns in richly-attributed graphs," in Big Data (Big Data), 2014 IEEE International Conference on. IEEE, 2014, pp. 585 -- 590. [14] W. W. Zachary, "An information flow model for conflict and fission in small groups," Journal of anthropological research, vol. 33, no. 4, pp. 452 -- 473, 1977. [15] D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and S. M. Dawson, "The bottlenose dolphin community of doubtful sound features a large proportion of long-lasting associations," Behavioral Ecology and Sociobiology, vol. 54, no. 4, pp. 396 -- 405, 2003. diffusion model for social networks." Springer. [16] J. Leskovec, D. Huttenlocher, and J. Kleinberg, "Signed networks in social media," in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, New York, NY, USA, 2010, pp. 1361 -- 1370. [17] L. Takac and M. Zabovsky, "Data analysis in public social networks." [18] L. Li, H. Tong, N. Cao, K. Ehrlich, Y.-R. Lin, and N. Buchler, "Replacing the irreplaceable: Fast algorithms for team member recom- mendation," in WWW, 2015, pp. 636 -- 646. [19] M. Ley, "The dblp computer science bibliography: Evolution, research issues, perspectives," in Proceedings of the 9th International Symposium on String Processing and Information Retrieval, London, UK, 2002, pp. 1 -- 10. [20] J. Leskovec, J. Kleinberg, and C. Faloutsos, "Graphs over time: densi- fication laws, shrinking diameters and possible explanations," in KDD. ACM, 2005, pp. 177 -- 187. [21] D. E. Knuth, The Stanford GraphBase: a platform for combinatorial computing. Addison-Wesley Reading, 1993, vol. 37. [22] J. Leskovec, J. Kleinberg, and C. Faloutsos, "Graph evolution: Den- sification and shrinking diameters," ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 1, no. 1, p. 2, 2007. [23] J. M. Kleinberg, "Authoritative sources in a hyperlinked environment," Journal of the ACM (JACM), vol. 46, no. 5, pp. 604 -- 632, 1999. [24] A. Y. Ng, A. X. Zheng, and M. I. Jordan, "Stable algorithms for link analysis," in Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval. ACM, 2001, pp. 258 -- 266. [25] C. Ding, X. He, P. Husbands, H. Zha, and H. Simon, "Pagerank, hits and a unified framework for link analysis," in Proceedings of the 2003 SIAM International Conference on Data Mining. SIAM, 2003, pp. 249 -- 253. [26] G. Jeh and J. Widom, "Simrank: a measure of structural-context similar- ity," in Proceedings of the 8th ACM SIGKDD international conference on Knowledge discovery and data mining, 2002, pp. 538 -- 543. [27] R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad, "Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission," in Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2015, pp. 1721 -- 1730. [28] M. T. Ribeiro, S. Singh, and C. Guestrin, "Why should i trust you?: Explaining the predictions of any classifier," in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2016, pp. 1135 -- 1144. [29] O. Bastani, C. Kim, and H. Bastani, "Interpreting blackbox models via model extraction," arXiv preprint arXiv:1705.08504, 2017. [30] M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkata- subramanian, "Certifying and removing disparate impact," in Proceed- ings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2015, pp. 259 -- 268. [31] P. Adler, C. Falk, S. A. Friedler, G. Rybeck, C. Scheidegger, B. Smith, and S. Venkatasubramanian, "Auditing black-box models for indirect influence," in Data Mining (ICDM), 2016 IEEE 16th International Conference on. IEEE, 2016, pp. 1 -- 10. [32] A. Datta, S. Sen, and Y. Zick, "Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems," in Security and Privacy (SP), 2016 IEEE Symposium on. IEEE, 2016, pp. 598 -- 617. [33] J. Li, W. Monroe, and D. Jurafsky, "Understanding neural networks through representation erasure," arXiv preprint arXiv:1612.08220, 2016. [34] J. Kang, S. Freitas, H. Yu, Y. Xia, N. Cao, and H. Tong, "X-rank: Explainable ranking in complex multi-layered networks," in CIKM. New York, NY, USA: ACM, 2018, pp. 1959 -- 1962. [35] D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins, "Information diffusion through blogspace," in WWW. ACM, 2004, pp. 491 -- 501. [36] W. Chen, Y. Wang, and S. Yang, "Efficient influence maximization in social networks," in KDD. ACM, 2009, pp. 199 -- 208. [37] M. Gomez Rodriguez, J. Leskovec, and A. Krause, "Inferring networks of diffusion and influence," in KDD. ACM, 2010, pp. 1019 -- 1028. [38] D. Kempe, J. Kleinberg, and ´E. Tardos, "Maximizing the spread of influence through a social network," in Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2003, pp. 137 -- 146. [39] D. Kempe, J. M. Kleinberg, and ´E. Tardos, "Influential nodes in a
1705.10442	2	1705	2017-08-04T21:54:08	Influence Maximization Meets Efficiency and Effectiveness: A Hop-Based Approach	[ "cs.SI" ]	Influence Maximization is an extensively-studied problem that targets at selecting a set of initial seed nodes in the Online Social Networks (OSNs) to spread the influence as widely as possible. However, it remains an open challenge to design fast and accurate algorithms to find solutions in large-scale OSNs. Prior Monte-Carlo-simulation-based methods are slow and not scalable, while other heuristic algorithms do not have any theoretical guarantee and they have been shown to produce poor solutions for quite some cases. In this paper, we propose hop-based algorithms that can easily scale to millions of nodes and billions of edges. Unlike previous heuristics, our proposed hop-based approaches can provide certain theoretical guarantees. Experimental evaluations with real OSN datasets demonstrate the efficiency and effectiveness of our algorithms.	cs.SI	cs	Influence Maximization Meets Efficiency and Effectiveness: A Hop-Based Approach Jing Tang Xueyan Tang Interdisciplinary Graduate School Nanyang Technological University Email: [email protected] School of Comp. Sci. & Eng. Nanyang Technological University Email: [email protected] Junsong Yuan School of EEE Nanyang Technological University Email: [email protected] 7 1 0 2 g u A 4 ] I S . s c [ 2 v 2 4 4 0 1 . 5 0 7 1 : v i X r a Abstract-Influence Maximization is an extensively-studied problem that targets at selecting a set of initial seed nodes in the Online Social Networks (OSNs) to spread the influence as widely as possible. However, it remains an open challenge to design fast and accurate algorithms to find solutions in large-scale OSNs. Prior Monte-Carlo-simulation-based methods are slow and not scalable, while other heuristic algorithms do not have any theoretical guarantee and they have been shown to produce poor solutions for quite some cases. In this paper, we propose hop- based algorithms that can easily scale to millions of nodes and billions of edges. Unlike previous heuristics, our proposed hop- based approaches can provide certain theoretical guarantees. Experimental evaluations with real OSN datasets demonstrate the efficiency and effectiveness of our algorithms. Index Terms-Online social networks, influence maximization, hop-based influence estimation. I. INTRODUCTION Information can be disseminated widely and rapidly through Online Social Networks (OSNs) with "word-of-mouth" effects. Viral marketing is such a typical application in which new products or activities are advertised by some influential users in the OSN to other users in a cascading manner [12]. A large amount of recent work [3], [6], [7], [8], [9], [10], [15], [16], [20], [25], [26], [27], [28], [29], [31] has been focusing on influence maximization in viral marketing, which targets at selecting a set of initial seed nodes in the OSN to spread the influence as widely as possible. The influence maximization problem was formulated by [16] with two basic diffusion models, namely the Independent Cascade (IC) and Linear Threshold (LT) models. Although finding the optimal seed set is NP-hard [16], a simple greedy hill-climbing algorithm has a (1− 1/e)-approximation guarantee due to the submodularity and monotone properties of the influence spread under these models [24]. Follow-up studies have mostly concentrated on efficient implementation of the hill-climbing algorithm for large-scale OSNs. The key difference among various methods lies in how to estimate the influence spread of a seed set. Computing the exact influence spread on general graphs is #P-hard for both the IC and LT models [6], [8]. Thus, some Monte-Carlo-simulation-based methods [16], [20], [27], [31] estimate the influence spread by reachability tests, while some reverse influence sampling methods carry out seed selection [3], [25], [26], [28], [29] by leveraging the concept of reverse reachability. These sampling-based methods can provide theo- retical guarantees up to (1−1/e−)-approximation. However, the sampling-based methods can encounter the efficiency prob- lem even for the dramatically improved versions [3], [25], [26], [27], [28], [29] as they may consume a lot of time/memory to obtain/store one sample. Other heuristic methods [6], [7], [8], [9], [15] conduct rough estimation of the influence spread either by exploiting some related features (such as node degrees) or extracting subgraphs where the influence spread is easier to compute. However, these heuristics do not have any theoretical guarantee and they have been shown to suffer from the effectiveness problem. To achieve both efficiency and effectiveness, in this paper, we propose a new hop-based approach for the influence maximization problem. Although it belongs to heuristics, unlike other heuristic methods, our method can provide certain theoretical guarantees. As shall be shown by the experimen- tal results, our hop-based methods outperform the existing heuristics and perform as well as the sampling-based methods in terms of the influence spread produced. Meanwhile, our hop-based methods run much faster than the sampling-based methods. For a large OSN with billions of edges, only our hop- based methods and some ineffective heuristics can work with acceptable time and memory usage for various distributions of propagation probabilities. Our contributions are summarized as follows. 1) We propose hop-based influence estimation algorithms for efficiently selecting seed nodes to maximize the influence spread. 2) We develop an upper bounding approach on the influ- ence generated by a seed node to further speed up seed selection. 3) We carry out theoretical analysis for our hop-based algorithms and derive approximation guarantees. 4) We conduct extensive experiments with several real OSN datasets. The results demonstrate the efficiency and effectiveness of our hop-based algorithms. The rest of this paper is organized as follows. Section II introduces the influence maximization problem and the greedy hill-climbing algorithm. Section III elaborates our algorithm design and analyzes the theoretical guarantees. Section IV presents the experimental study. Section V reviews the related work. Finally, Section VI concludes the paper. 1 II. PRELIMINARIES A. Problem Definition Let G = (V,E) be a directed graph modeling an OSN, where the nodes V represent users and the edges E represent the connections among users (e.g., followships on Twitter). For each directed edge (u, v) ∈ E, we refer to v as a neighbor of u, and refer to u as an inverse neighbor of v. To facilitate the exposition, we shall mainly focus on the Independent Cascade (IC) model - a representative and most widely-studied diffusion model for influence propagation [3], [6], [7], [15], [16], [20], [25], [26], [27], [28], [29], [31]. Extending our hop-based methods to other diffusion models shall be discussed later. In the IC model, a propagation probability pu,v is associated with each edge (u, v), repre- senting the probability for v to be activated by u through the edge. Let Nu denote the set of node u's neighbors, i.e., Nu = {v : v ∈ V, (u, v) ∈ E}. Given a set of seed nodes S, the IC diffusion process proceeds as follows. Initially, the seed nodes S are activated, while all the other nodes are not activated. When a node u first becomes activated, it attempts to further activate its neighbors who are not yet activated. For each such neighbor v ∈ Nu, v would become activated with probability pu,v. This process repeats until no more node can be activated. The influence spread of the seed set S, denoted by σ(S), is the expected number of nodes activated by the above process. The influence maximization problem [16] is to find a set S of k nodes to maximize σ(S), where k is a given parameter. Formally, max S=k σ(S). (1) B. Greedy Heuristic The influence function σ(·) has been proved to be sub- modular and monotone under the IC model [16]. Thus, a simple greedy hill-climbing algorithm that provides (1− 1/e)- approximation [24] was proposed for influence maximization as described in Algorithm 1. It starts with an empty seed set S = ∅. In each iteration, the greedy heuristic chooses a new seed u from the non-seed nodes V \ S with largest marginal influence gain σ(S ∪ {u}) − σ(S) and adds u to S. The algorithm stops after selecting k seeds. Algorithm 1: Greedy(G, σ) 1 initialize S ← ∅; 2 while the size of S is smaller than k do 5 return S; 3 4 find u ← arg maxv∈V\S {σ(S ∪ {v}) − σ(S)}; S ← S ∪ {u}; A CELF technique [20] can be used to enhance the ef- ficiency of the greedy algorithm due to the submodularity of influence spread. Specifically, it may not be necessary to evaluate the marginal influence gain for every node of V \S in each iteration. Due to the submodularity, the marginal gains 2 (a) High degree nodes Influence spread for different hops of propagation on LiveJournal. (b) Top influential nodes Fig. 1. can only decrease over iterations. Thus, the marginal gains obtained in the previous iterations can be used as upper bounds for a new iteration. In the new iteration, the nodes can be evaluated in decreasing order of these upper bounds. Once the largest marginal gain evaluated is greater than the upper bound of the next node to evaluate, the evaluation can stop as none of the remaining nodes would be able to produce larger marginal gain. III. HOP-BASED APPROACHES A. Hop-Based Influence Estimation under IC Model We start by studying some empirical results of influence spread tested on a real OSN graph – LiveJournal (5M nodes and 69M edges) [21]. Fig. 1 plots the influence spread within different numbers of hops of propagation (the data points of ∞ hops represent the actual influence spread without any hop limit, and the seed set size is represented as a percentage of the entire node set in the OSN). We test two typical seed sets for influence maximization: selecting the nodes with highest degrees and selecting the top influential nodes using a greedy hill-climbing heuristic [16] as described in Section II-B. The propagation probability on each edge (u, v) in the OSN is set to the reciprocal of v's in-degree as widely adopted in previous studies [7], [16], [26], [28]. As seen from Fig. 1, the increase in the influence spread for considering each additional hop of propagation generally decreases with increasing number of hops. The majority of influence spread is produced within the first few hops of propagation. Similar trends have also been observed by several measurement-driven studies on real OSNs [4], [13], [19]. For example, Goel et al. [13] showed that less than 10% of the cascades in the diffusion are more than 2 hops away from the seed. These observations motivate us to design hop-based algorithms to efficiently capture the major influence propagation, especially for the first two hops of propagation. A hop-based algorithm focuses on the influence propagation up to a given number of h hops starting from the initial seed set. For h = 1 and h = 2, we can efficiently calculate the exact influence spread within h hops of propagation and maintain it incrementally when the seed set expands. Corresponding to the denotation of a node v's neighbors as Nv, let Iv denote v's inverse neighbors, i.e., Iv = {w : w ∈ V, (w, v) ∈ E}. Let πS h (v) denote the probability for a node v to be activated within h hops of propagation from a seed set S, and let 0123456∞Numberofhops0.00.20.40.60.81.01.2Influencespread1e60.1%0.2%0.4%0.8%0123456∞Numberofhops0.00.20.40.60.81.01.21.4Influencespread1e60.1%0.2%0.4%0.8%Fig. 2. The effect on v ∈ Nu by adding u to the seed set S. Fig. 3. An example of how a non-seed node v is activated. σh(S) denote the influence spread produced within h hops of propagation from S. One Hop of Propagation: We first model one hop of prop- agation. Obviously, for all the seed nodes v ∈ S, πS 1 (v) = 1. With one hop of propagation, for all the non-seed nodes v /∈ S, v can only be activated directly by its inverse neighbors Iv them to fail to activate v is(cid:81) who are seed nodes in S. Since each of such v's inverse (cid:81) neighbors activates v independently, the probability for all of w∈Iv∩S (1−pw,v). Consequently, the probability for v to be activated is 1− w∈Iv∩S (1−pw,v). Thus, for any node v ∈ V, its one-hop activation probability is given by (cid:81) if v ∈ S, w∈Iv∩S (1 − pw,v), otherwise. 1 − (2) (cid:40)1, πS 1 (v) = Now, we show how to maintain πS 1 (v) when the seed set changes. Suppose that πS 1 (v) is known for every node v ∈ V. If a new seed node u is added to a seed set S, it is clear that the activation probability of the new seed becomes 1, i.e., π (u) = 1. In addition to u, only the one-hop activation probabilities of its neighbors are affected. So, we 1 (v) for each can incrementally update π node v ∈ Nu (see Fig. 2). The new activation probability S∪{u} π 1 (v) based on πS (v) is given by S∪{u} 1 S∪{u} 1 S∪{u} 1 π (v) = 1 − (1 − pw,v) (cid:89) (cid:0)1 − πS 1 (v)(cid:1) w∈(Iv∩S)∪{u} = 1 − · (1 − pu,v). (3) Algorithm 2 calculates the increment of one-hop influence spread σ1(S ∪ {u})− σ1(S) efficiently by maintaining πS 1 (v) for every node v based on the above equation. 3 π S∪{u} 1 π 1 (v) ← 1 − v∈{u}∪(Nu\S) (u) ← 1; S∪{u} 1 1 (v)(cid:1) (cid:0)1 − πS 1 (v)(cid:1); (cid:0)π · (1 − pu,v); (v) − πS Algorithm 2: OneHopIncrement(G,S, u) 2 for each node v ∈ Nu \ S do 4 return(cid:80) Two Hops of Propagation: To better approximate σ(S), we next model two hops of propagation. As illustrated in Fig. 3(a), with two hops of propagation, a non-seed node v may be activated directly by a seed node ui or indirectly via a neighbor wj of a seed node ui. In the former case, S∪{u} 1 3 Fig. 4. The effect on v ∈N 2 u by adding u to the seed set S. πS 2 (v) = 1 (ui) since πS the probability for v to be activated by ui is pui,v, which can be rewritten as pui,v · πS 1 (ui) = 1. In the latter case, the probability for v to be activated by wj is pwj ,v · πS 1 (wj). Since the activation probability of each seed node ui ∈ S is 1, Fig. 3(a) is equivalent to Fig. 3(b) in which (cid:81) v is activated independently by all of its inverse neighbors. As a result, the probability for v to be activated is given by 1 (w)). Thus, for any node v ∈ V, its 1 − two-hop activation probability is given by w∈Iv (1 − pw,v · πS (cid:40)1, (cid:81) 1 − w∈Iv (cid:0)1 − pw,v · πS (4) if v ∈ S, 1 (w)(cid:1), otherwise. (cid:1) (cid:0)(cid:83) According to the above equation, we can obtain πS 2 (v) based on the πS 1 (w)'s of its inverse neighbors. When a new seed node u is added, only the nodes within two hops of u are affected. \ {u}. We denote these nodes by N 2 Fig. 4 illustrates the effect of adding u to S on the activation u . Due to the independence among all two- of a node v in N 2 hop activation paths as discussed above, we can incrementally update the two-hop activation probability by considering the outgoing edges from u one at a time. u = Nu ∪ w∈Nu Nw π π (cid:89) Theorem 1: The new two-hop activation probability S∪{u} (v) after adding a seed u to S can be computed by 2 S∪{u} (w) (v) = 1− 2 (cid:0)1− πS 2 (v)(cid:1) S∪{u} 1 − pw,v · π 1 1 − pw,v · πS (5) where Mu,v denotes the set of intermediate nodes connecting u and v, i.e., Mu,v = {w : (u, w) ∈ E and (w, v) ∈ E}. We leave the formal proofs of all theoretical results to the appendix. w∈(Mu,v∪{u}) 1 (w) · Algorithm 3 performs the updates on the activation proba- bilities of the nodes within two hops of u when u is added as a new seed to S. Lines 1–2 set the one-hop and two-hop activation probabilities of the new seed node u to 1. Lines 3– 4 initialize π (v) for all the nodes within two hops of S∪{u} 2 , vpu;vuS¼1(v) S v(a)v(b)w1 w1 w2 w2 w3 w3 w4 w4 u1 u1 u1 u1 u2 u2 u2 u2 Seed nodespw ;v vuFirst HopSecond HopS¼2(v) S ¼1(w2) S w1 1w2 ¼1(w1) S pw ;v 2.........pu;w 2pu;w 1Algorithm 3: T woHopsIncrement(G,S, u) S∪{u} 1 π 1 S∪{u} 2 π 2 (u) ← 1; (u) ← 1; S∪{u} 2 6 7 8 4 3 for each node v ∈ N 2 u \ S do (v) ← πS 2 (v); 5 for each node w ∈ Nu \ S do π π π (w) ← 1 − (w) ← 1− S∪{u} 1 S∪{u} S∪{u} 2 2 for each node v ∈ Nw \ S do S∪{u} 2 S∪{u} 2 1 (w)(cid:1) (cid:0)1 − πS (cid:0)1−π (w)(cid:1) · (1 − pu,w); · 1−pu,w·π (cid:0)1−π (v)(cid:1) · 1−pw,v·π (cid:0)π 2 (v)(cid:1); (v) − πS S∪{u} 2 (v) ← 1− u\S) v∈{u}∪(N 2 9 π 10 return(cid:80) S∪{u} 1 (u) 1−pu,w·πS 1 (u) S∪{u} 1 (w) 1−pw,v·πS 1 (w) ; ; S∪{u} 2 u. Line 6 computes the new one-hop activation probabilities for all of u's neighbors as explained earlier. For each node u \ S, lines 7–9 calculate the new two-hop activation v ∈ N 2 probability π (v) in an iterative manner according to Theorem 1. In this way, we can save a huge amount of space for storing the intermediate nodes Mu,v for every pair of nodes u and v. Finally, the algorithm returns the total increment of two-hop influence spread σ2(S ∪ {u}) − σ2(S). B. Complexity V +E w∈Nu Nw We shall refer to the greedy heuristic (Algorithm 1) as the OneHop and TwoHop algorithms respectively when the influence spread is approximated by one hop and two hops of propagation. Thus, the total time complexity of TwoHop is O plexity of selecting one seed in the TwoHop algorithm is V + E + Time Complexity: The time complexity of Algorithm 2 is O(1 + Nu). Thus, the time complexity of selecting one seed in the OneHop algorithm is O(cid:0)(cid:80) u∈V (1 + Nu)(cid:1) = (cid:1). Therefore, the total time complexity of OneHop O(cid:0) is O(cid:0)k(V + E)(cid:1). The time complexity of Algorithm 3 is O(cid:0)1 + Nu +(cid:80) (cid:1). Therefore, the time com- O(cid:0)(cid:80) w∈Nu Nw)(cid:1) = O(cid:0) u∈V (1 + Nu + (cid:80) w∈V (Iw · Nw)(cid:1). (cid:1) = O(cid:0) (cid:80) u∈V(cid:80) V + E +(cid:80) (cid:16) w∈V (Iw · Nw)(cid:1)(cid:17) (cid:80) Space Complexity: Besides the space used to store the graph, the OneHop algorithm only requires O(V) space to store the one-hop activation probability πS 1 (v) for every v ∈ V before and after a new seed is added. Similarly, the TwoHop algorithm requires O(V) space to store the one-hop and two- hop activation probabilities πS 2 (v) for computing the influence increment. Thus, the space complexities of the OneHop and TwoHop algorithms are both O(V). C. Further Improvement on Efficiency 1 (v) and πS w∈Nu Nw V+E+ k(cid:0) . Note that selecting the first seed in Algorithm 1 requires calculating the influence spread σ({v}) for every node v even when the CELF technique [20] is adopted. To avoid such 4 computation, we develop an upper bound on σ({v}) when hop-based influence estimation is applied. Theorem 2: For each node v ∈ V, the h-hop influence spread σh({v}) satisfies (cid:88) (cid:0)pv,w · σh−1({w})(cid:1). (6) σh({v}) ≤ 1 + w∈Nv w∈Nv w∈Nv Note that when h = 1, σh({v}) ≤ σh({v}). Furthermore, let σ0({v}) = σ0({v}) = 1 and σh({v}) = (cid:0)pv,w · σh−1({w})(cid:1), then 1 +(cid:80) 1+(cid:80) a time complexity of O((cid:80) all nodes v ∈ V is O(V + E) + O((cid:80) Algorithm 3 which is O(cid:0) V + E +(cid:80) (7) the upper bound σ1({v}) = pv,w is the exact 1-hop influence spread of a single seed {v}, i.e., σ1({v}) = σ1({v}). Computing σ1({v}) for a node v has a time complexity of O(1 + Nv). Thus, it takes v∈V (1 + Nv)) = O(V + E) to calculate σ1({v}) for all nodes v ∈ V. The time complexity for computing the upper bound σ2({v}) given in Theorem 2 is O(1 + Nv) after obtaining σ1({w}) for all nodes w ∈ V. Thus, the total time complexity for calculating σ2({v}) for v∈V (1 + Nv)) = O(V + E) + O(V + E) = O(V + E), which is much lower than the time complexity for computing the exact two- hop influence spread σ2({v}) for all nodes v ∈ V using On obtaining the upper bounds σh({v}), the CELF technique described in Section II-B can then be applied to the first iteration of Algorithm 1 so that only the influence spreads of a subset of nodes in V need to be calculated for selecting the first seed. We shall show in Section IV-B that the upper bounding approach can dramatically reduce the running time of the two-hop method. w∈V (Iw · Nw)(cid:1). D. Theoretical Analysis In this section, we carry out theoretical analysis for our hop-based algorithms. We first show that the influence spread within h hops of propagation is submodular and monotone. Theorem 3: For any h ≥ 1, the influence spread produced within h hops of propagation is submodular and monotone under the IC model. ∗ denote the optimal seed set for maximizing the actual ∗) = maxS=k σ(S). influence without any hop limit, i.e., σ(S Then, we can derive the following guarantees for our hop- based methods. Let S Theorem 4: Under the IC model, if σh(S)/σ(S) ≥ α for any seed set S = k, the solution Sh returned by the greedy heuristic (Algorithm 1) with hop-based influence estimation satisfies · σ(S σ(Sh) ≥ (8) Theorem 4 indicates that if the ratio σh(S)/σ(S) is lower bounded by α, the hop-based methods can provide a multi- plicative guarantee of(cid:0)(1 − 1/e)α(cid:1). Next, we derive a lower bound on the ratio σh(S)/σ(S) in the class of scale free random graphs which are commonly used to model OSNs [2], [22]. The degree distribution of a scale-free (undirected) (cid:0)(1 − 1/e)α(cid:1) ∗). (cid:80) influence spread by adding u to S is σ1(S ∪ {u}) − σ1(S) = v∈Nu\S bu,v. For two hops of propagation, all the paths from seed nodes to a non-seed node within two hops are from node u itself; 2)(cid:0)1 − πS 1 (u)(cid:1) (cid:80) acyclic. The increment of influence spread by adding a new seed u to a seed set S consists of three parts: 1) 1 − πS (cid:0)bu,v· (cid:1) non-seed neighbors; 3)(cid:80) (cid:80) 2 (u) v∈Nu\S bu,v from u's w∈Nv\(S∪{u}) bv,w 1 (u)(cid:1) 2 (u)+(cid:0)1−πS from u's neighbors' neighbors which are not seeds. Thus, the (cid:0)bu,v · (cid:1). It is (cid:80) v∈Nu\S bu,v +(cid:80) (cid:80) total increment of influence spread by adding a new seed u to a seed set S is σ2(S∪{u})−σ2(S) = 1−πS · easy to show that the influence spread within a fixed number of hops under the LT model is submodular and monotone as well. Thus, our analysis on the hop-based algorithms can also be applied to the LT model. w∈Nv\(S∪{u}) bv,w v∈Nu\S v∈Nu\S · IV. EVALUATION A. Experimental Setup Datasets. We use several real OSN datasets in our experi- ments [1], [17], [21]. Due to space limitations, we just show the results for three representative datasets: NetHEPT (15K nodes and 32K edges), LiveJournal (5M nodes and 69M edges) and Twitter (42M nodes and 1.5B edges). The results for other datasets are similar. Algorithms. We compare our OneHop, TwoHop and TwoHop-O (without the upper bounding technique described in Section III-C) algorithms with the following state-of-the-art algorithms. • HighDegree: Select the k nodes with highest degrees [16]. • DegreeDiscount: The degree discount heuristic was de- gation probability via the following two models. [6], • TRIVALENCY model • WC model [7], [16], [26], [28]: pu,v of each edge (u, v) is set to the reciprocal of v's in-degree, i.e., pu,v = 1/Iv. [15]: pu,v of each edge (u, v) is set by choosing a probability from the set {0.1, 0.01, 0.001} at random. To evaluate the seed sets returned by different algorithms, we estimate the influence spread of each seed set by taking the average measurement of 10, 000 Monte-Carlo simulations. The algorithms are all implemented in C++ and the experiments are carried out on a machine with an Intel Xeon E5-2695 2.4GHz CPU and 64GB memory. We limit the running time of each algorithm up to 100 hours (3.6 × 105 seconds). 5 veloped by [7]. • IRIE: IRIE [15] is a state-of-the-art heuristic. We set α = 0.7 and θ = 1/320 respectively as suggested in [15]. • IMM: IMM [28] is one of the most advanced sampling- based methods that can provide a (1 − 1/e − )- approximation guarantee with probability at least 1 − δ. • D-SSA: D-SSA [26] aims to further reduce the number of samples generated compared to IMM while providing the same approximation guarantee. We set = 0.1 and δ = 1/V for both IMM and D-SSA according to the default setting in [26]. Parameter Settings. For the IC model, we set the propa- Fig. 5. Lower bound α for different propagation probabilities and seed ratios (γ = 3). −γ i di=1 d d−γ(cid:80)∞ graph follows a power law. That is, the probability of a node having degree d is P0(d) = , where γ is a given power scale parameter whose typical value is in the range of 2 ≤ γ ≤ 3. We first analyze the expected number of nodes activated within one hop of propagation, which gives a lower bound on σh(S) for any h ≥ 1 since σh(S) increases with h. Next, we derive an upper bound on the expected number of nodes activated σ(S). Using the lower bound on σh(S) and upper bound on σ(S), we can derive a lower bound α on the ratio σh(S)/σ(S). Theorem 5: For scale free random graphs with propagation probability pu,v = p for every edge (u, v) ∈ E and any seed set S and any hop of h ≥ 1, we have E[σh(S)] E[σ(S)] ≥ (cid:0)1 − kV (cid:1)P1(1) and P1(d) = 1 − (1 − k/V)(1 − pk/V) 1 − (1 − k/V)P0(1)(1 − pA) (cid:80)∞ d1−γ di=1 d1−γ where A = 1 − Fig. 5 shows the lower bound derived in (9) when varying the propagation probability p from 0 to 0.1 and the seed ratio k/V from 0 to 0.5. We can see that the lower bound generally increases with both the seed ratio and propagation probability. (9) , i . E. Extension to Linear Threshold Model The Linear Threshold (LT) model is another basic diffusion model described in [16], which is also widely used [8], [28], [29]. In the LT model, each directed edge (u, v) is associated with a weight bu,v, and the total weight of all the incoming edges to each node does not exceed 1. In the diffusion process, each node randomly chooses a threshold between 0 and 1. The diffusion starts from an initial seed set S. A non-seed node becomes activated only when the aggregate weight of the incoming edges from its inverse neighbors that have been activated reaches its threshold. 1 (v) = 1; otherwise, πS Our proposed hop-based algorithms can be easily extended to work with the LT model. For one hop of propagation, if v ∈ S, πS u∈Iv∩S bu,v. When adding a new seed u to a seed set S, for each neighbor v of u where v /∈ S, the increment of v's activation probability S∪{u} is π 1 (v) = bu,v. Thus, the total increment of 1 1 (v) =(cid:80) (v) − πS Propagationprob.p0.000.020.040.060.080.10Seedratiok/V0.00.10.20.30.40.5Lowerboundα0.00.20.40.60.81.0(a) NetHEPT (b) LiveJournal (c) Twitter (IRIE cannot run) Fig. 6. Influence spread on various graphs under the WC model. (a) NetHEPT (b) LiveJournal (IMM and D-SSA cannot run) (c) Twitter (IRIE, IMM and D-SSA cannot run) Fig. 7. Influence spread on various graphs under the TRIVALENCY model. B. Results Influence Spread: Figs. 6 and 7 show the influence spread produced by different algorithms on various graphs when the size of seed set is set to k = 1, 10, 100, 1000 under the WC and TRIVALENCY models respectively. Due to out-of- memory reasons and prohibitively long computation times, IRIE failed to produce results on the Twitter dataset under both WC and TRIVALENCY models, while both IMM and D-SSA failed on the LiveJournal and Twitter datasets under the TRIVALENCY model. From the results obtained, we can make the following observations. Our OneHop, TwoHop and TwoHop-O methods usually generate influence spread as high as that by the IMM and D-SSA methods which can provide the state-of-the-art (1 − 1/e − )-approximation guarantee. Our methods remarkably outperform both the HighDegree and DegreeDiscount heuristics (by up to 40%) on the NetHEPT dataset under both WC and TRIVALENCY models (Figs. 6(a), 7(a)) and on the LiveJournal dataset under the WC model (Fig. 6(b)). These observations demonstrate the effectiveness of our hop-based methods. Running Time: Figs. 8 and 9 show the running times of different algorithms. The OneHop and DegreeDiscount methods run almost at the same speed which can find the top 1000 influential nodes on the Twitter dataset (with billions of edges) within 30 seconds. They are just slightly slower than the HighDegree method and run several orders faster than other methods, including the IRIE, IMM, D-SSA, TwoHop and TwoHop-O methods (note that the y-axis is in logscale). This shows a tradeoff between efficiency and effectiveness for the hop-based methods. Estimating the influence spread with a higher hop limit takes more time but can improve the quality of the seed set chosen. For example, on the LiveJournal dataset under the WC model (Fig. 6(b)), the TwoHop method performs notably better than the OneHop method. If the application is highly time-sensitive, the OneHop method could be preferable to the TwoHop method. Otherwise, the TwoHop method is favoured since its running time is quite acceptable even for very large networks. We also observe that while the TwoHop and TwoHop-O methods always produce the same seed set solution, the former runs significantly faster than the latter (by up to 4 orders of magnitude). This is because TwoHop-O consumes too much time on computing the influence spread of all the single seed sets. This demonstrates the efficiency of our upper bounding approach. Moreover, we observe that the OneHop and TwoHop meth- ods generally run much faster than the state-of-the-art IRIE, IMM and D-SSA methods. This demonstrates the efficiency of our hop-based methods. Since IMM and D-SSA are reverse influence sampling methods, their running times heavily de- pend on the sizes of the reverse reachable sets sampled. Under the TRIVALENCY model, all the edges are likely to have substantial propagation probabilities. As a result, high-degree nodes can have many direct reverse reachable neighbors in a 6 HighDegreeDegreeDiscountIRIEIMMD-SSAOneHopTwoHopTwoHop-O100101102103Sizeofseedset0123456Influencespread1e3100101102103Sizeofseedset0.00.51.01.52.02.53.03.54.04.5Influencespread1e5100101102103Sizeofseedset0.00.20.40.60.81.01.21.41.6Influencespread1e7HighDegreeDegreeDiscountIRIEIMMD-SSAOneHopTwoHopTwoHop-O100101102103Sizeofseedset0.00.20.40.60.81.01.21.41.61.8Influencespread1e3100101102103Sizeofseedset6.606.626.646.666.686.706.726.74Influencespread1e5100101102103Sizeofseedset0.981.001.021.041.061.081.10Influencespread1e7(a) NetHEPT (b) LiveJournal (c) Twitter (IRIE cannot run) Fig. 8. Running time on various graphs under the WC model. (a) NetHEPT (b) LiveJournal (IMM and D-SSA cannot run) (c) Twitter (IRIE, IMM and D-SSA cannot run) Fig. 9. Running time on various graphs under the TRIVALENCY model. RUNNING TIME (SECONDS) FOR SELECTING 1000 SEEDS IN THE TWITTER GRAPH UNDER THE WC MODEL WITH DIFFERENT PROPAGATION PROBABILITY SCALE FACTORS f . TABLE I Method f = 1.0 HighDegree DegreeDiscount IRIE IMM D-SSA OneHop TwoHop 0.20 52.23 – 2.13e3 91.21 15.07 7.70e3 1.1 0.20 51.79 – 3.52e4 3.76e3 16.08 7.89e3 1.2 0.21 51.71 – – 2.44e4 17.70 8.11e3 1.3 0.21 52.14 1.4 0.20 52.70 1.5 0.21 51.53 – – – – – – – – – 17.18 8.25e3 17.33 7.94e3 17.85 8.32e3 The field with "–" means that the method cannot run under the setting. based methods even at a minor factor of 1.2. In contrast, our hop-based methods are much less sensitive to the propagation probabilities. Remark on Memory Usage: Recall that our hop-based al- gorithms just need O(V) space which is negligible compared to the space O(V +E) required for storing the OSN graph. On the other hand, the IRIE, IMM and D-SSA methods have significantly higher space complexities, e.g., those of IMM and D-SSA are both O [26], [28]. Thus, they fail to produce results on very large datasets. Our hop-based algorithms never face the out-of-memory problems as long as the memory is large enough to store the OSN graph. (cid:16)(cid:0) ln(1/δ)+ln (V k )(cid:1)·(V+E) (cid:17) 2 sample outcome of influence propagation. Thus, the samples of reverse reachable sets are large and require significant memory and time to compute, which makes IMM and D-SSA intractable on the LiveJournal and Twitter datasets. Under the WC model, for any node, the expected number of its direct reverse reachable neighbors is 1 because the edge propagation probabilities are given by the reciprocal of the node's in- degree. This significantly limits the sizes of reverse reachable sets and favors IMM and D-SSA. However, the running times of IMM and D-SSA are still very sensitive to the propagation probabilities. Table I shows the trends when we scale the propagation probabilities under the WC model by a small factor up to 1.5 (for selecting 1000 seeds in the Twitter graph). It can be seen that the running times of IMM and D-SSA increase rapidly with the scale factor and far exceed our hop- V. RELATED WORK Since the (1 − 1/e − )-approximation greedy algorithm was proposed by [16] for influence maximization, there has been considerable research on improving the efficiency of the greedy algorithm by using heuristics to trade the accuracy of influence estimation for computational efficiency [6], [7], [8], [15], or optimizing the Monte-Carlo simulations for influence estimation [3], [25], [26], [27], [28], [29]. Among them, the DegreeDiscount heuristic [7] roughly estimated influence spread within one-hop neighborhood and the PMIA heuristic [6] used independent propagation paths to construct arbores- cences for rough influence estimation. However, PMIA is very costly in both time and space compared to a follow-up IRIE algorithm [15]. DegreeDiscount and IRIE as well as advanced 7 HighDegreeDegreeDiscountIRIEIMMD-SSAOneHopTwoHopTwoHop-O100101102103Sizeofseedset10−410−310−210−1100101102Runningtime(seconds)100101102103Sizeofseedset10−210−1100101102103104Runningtime(seconds)100101102103Sizeofseedset10−1100101102103104105106Runningtime(seconds)HighDegreeDegreeDiscountIRIEIMMD-SSAOneHopTwoHopTwoHop-O100101102103Sizeofseedset10−410−310−210−1100101102Runningtime(seconds)100101102103Sizeofseedset10−210−1100101102103104Runningtime(seconds)100101102103Sizeofseedset10−1100101102103104105106Runningtime(seconds)sampling-based methods are all included in our experimental comparison. Leveraging the independency among propagation paths, we have developed efficient hop-based methods to compute the exact influence spread within a certain number of hops. Our hop-based influence estimation is in spirit similar to the time-constrained independent cascade model studied in [5], [11], [23] by concentrating on the diffusion within a fixed number of hops. We make new technological advances by inventing algorithms to compute the exact influence spreads of one-hop and two-hop propagations, which could only be approximately estimated in previous work where the approx- imation guarantee is difficult to analyze [18]. Our algorithms enable very efficient evaluation of the change in influence spread when a new seed node is added. We also derive an upper bound on the influence spread to further speed up our hop-based algorithms. Our hop-based approaches can be easily applied to many influence-based applications, such as topic- aware influence maximization [30] and community detection via influence maximization [14]. VI. CONCLUSION In this paper, we have proposed lightweight hop-based methods to address the problem of influence maximization in OSNs. We have also developed an upper bounding technique to further speed up the seed selection algorithm. Through analysis, we show that our methods can provide certain theoretical guarantees. Experiments are conducted with real OSN datasets to compare the efficiency and effectiveness of our algorithms with state-of-the-art ones. In terms of solution quality, our hop-based methods are on par with the most advanced IMM and D-SSA methods which can provide the best (1 − 1/e − )-approximation guarantee and remarkably outperform the HighDegree and DegreeDiscount heuristics for quite some cases. In terms of efficiency, our hop-based methods run much faster than the IRIE, IMM and D-SSA methods for most cases tested. Furthermore, while all these existing methods fail to run on some test cases, our hop-based methods can always execute and find solutions. ACKNOWLEDGMENT This research is supported by the National Research Foun- dation, Prime Minister's Office, Singapore under its IDM Futures Funding Initiative, and by Singapore Ministry of Education Academic Research Fund Tier 1 under Grant 2013- T1-002-123. REFERENCES [1] http://research.microsoft.com/en-us/people/weic/projects.aspx. [2] A.-L. Barab´asi and R. Albert, "Emergence of scaling in random net- works," Science, vol. 286, no. 5439, pp. 509–512, 1999. [3] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier, "Maximizing social influence in nearly optimal time," in Proc. SODA, 2014, pp. 946–957. [4] M. Cha, A. Mislove, and K. P. Gummadi, "A measurement-driven analysis of information propagation in the flickr social network," in Proc. WWW, 2009, pp. 721–730. [5] W. Chen, W. Lu, and N. Zhang, "Time-critical influence maximization in social networks with time-delayed diffusion process," Proc. AAAI, pp. 592–598, 2012. 8 [6] W. Chen, C. Wang, and Y. Wang, "Scalable influence maximization for prevalent viral marketing in large-scale social networks," in Proc. ACM KDD, 2010, pp. 1029–1038. [7] W. Chen, Y. Wang, and S. Yang, "Efficient influence maximization in social networks," in Proc. ACM KDD, 2009, pp. 199–208. [8] W. Chen, Y. Yuan, and L. Zhang, "Scalable influence maximization in social networks under the linear threshold model," in Proc. IEEE ICDM, 2010, pp. 88–97. [9] S. Cheng, H. Shen, J. Huang, W. Chen, and X. Cheng, "IMRank: Influence maximization via finding self-consistent ranking," in Proc. ACM SIGIR, 2014, pp. 475–484. [10] E. Cohen, D. Delling, T. Pajor, and R. F. Werneck, "Sketch-based influence maximization and computation: Scaling up with guarantees," in Proc. ACM CIKM, 2014, pp. 629–638. [11] T. N. Dinh, H. Zhang, D. T. Nguyen, and M. T. Thai, "Cost-effective vi- ral marketing for time-critical campaigns in large-scale social networks," IEEE/ACM Trans. Networking, vol. 22, no. 6, pp. 2001–2011, 2014. [12] P. Domingos and M. Richardson, "Mining the network value of cus- tomers," in Proc. ACM KDD, 2001, pp. 57–66. [13] S. Goel, D. J. Watts, and D. G. Goldstein, "The structure of online diffusion networks," in Proc. ACM EC, 2012, pp. 623–638. [14] F. Jiang, S. Jin, Y. Wu, and J. Xu, "A uniform framework for community detection via influence maximization in social networks," in Proc. IEEE/ACM ASONAM, 2014, pp. 27–32. [15] K. Jung, W. Heo, and W. Chen, "IRIE: Scalable and robust influence maximization in social networks," in Proc. IEEE ICDM, 2012, pp. 918– 923. [16] D. Kempe, J. Kleinberg, and E. Tardos, "Maximizing the spread of influence through a social network," in Proc. ACM KDD, 2003, pp. 137–146. [17] H. Kwak, C. Lee, H. Park, and S. Moon, "What is Twitter, a social network or a news media?" in Proc. WWW, 2010, pp. 591–600. [18] J. R. Lee and C. W. Chung, "A fast approximation for influence maximization in large social networks," in WWW Companion, 2014, pp. 1157–1162. [19] J. Leskovec, L. A. Adamic, and B. A. Huberman, "The dynamics of viral marketing," ACM Trans. Web, vol. 1, no. 1, 2007. [20] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance, "Cost-effective outbreak detection in networks," in Proc. ACM KDD, 2007, pp. 420–429. [21] J. Leskovec and A. Krevl, "SNAP Datasets: Stanford large network dataset collection," http://snap.stanford.edu/data, 2014. [22] Y. Li, B. Q. Zhao, and J. C. S. Lui, "On modeling product advertisement in large-scale online social networks," IEEE/ACM Trans. Networking, vol. 20, no. 5, pp. 1412–1425, 2012. [23] B. Liu, G. Cong, D. Xu, and Y. Zeng, "Time constrained influence maximization in social networks," in Proc. IEEE ICDM, 2012, pp. 439– 448. [24] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, "An analysis of ap- proximations for maximizing submodular set functions-I," Mathematical Programming, vol. 14, no. 1, pp. 265–294, 1978. [25] H. T. Nguyen, T. N. Dinh, and M. T. Thai, "Cost-aware targeted viral marketing in billion-scale networks," in Proc. IEEE INFOCOM, 2016. [26] H. T. Nguyen, M. T. Thai, and T. N. Dinh, "Stop-and-stare: Optimal sampling algorithms for viral marketing in billion-scale networks," in Proc. ACM SIGMOD, 2016, pp. 695–710. [27] N. Ohsaka, T. Akiba, Y. Yoshida, and K. Kawarabayashi, "Fast and accurate influence maximization on large networks with pruned monte- carlo simulations," in Proc. AAAI, 2014, pp. 138–144. [28] Y. Tang, Y. Shi, and X. Xiao, "Influence maximization in near-linear time: A martingale approach," in Proc. ACM SIGMOD, 2015, pp. 1539– 1554. [29] Y. Tang, X. Xiao, and Y. Shi, "Influence maximization: Near-optimal time complexity meets practical efficiency," in Proc. ACM SIGMOD, 2014, pp. 75–86. [30] C. Zhang, J. Sun, and K. Wang, "Information propagation in microblog networks," in Proc. IEEE/ACM ASONAM, 2013, pp. 190–196. [31] C. Zhou, P. Zhang, J. Guo, X. Zhu, and L. Guo, "UBLF: An upper bound based approach to discover influential nodes in social networks," in Proc. IEEE ICDM, 2013, pp. 907–916. APPENDIX = (w1) 1 − π S∪{u} 2 Proof of Theorem 1: To consider the outgoing edges from u one at a time, we first disable all the edges from u to its neighbors except for one edge (u, w1). Then, for each neighbor v of w1, all of v's inverse neighbors other than w1 have their one-hop activation probabilities unchanged by (vw1) denote the new two-hop adding (u, w1). Let π activation probability of v. Then, we have S∪{u} 1 − pw1,v · π 1 1 − pw1,v · πS second edge the Next, we S∪{u} new two-hop π 2 probability of v. Following similar arguments, neighbor v of w2, we have (vw1, w2) (vw1) S∪{u} 1 − pw2,v · π 1 1 − pw2,v · πS S∪{u} 2 1 − πS (vw1, w2) (vw1) 2 (v) enable Let activation for each We continue to enable the outgoing edges of u sequentially. In general, when an edge (u, wi) is enabled after edges (u, w1), (u, w2), ··· , (u, wi−1), for each neighbor v of wi, we have (wi) . 1 − π 1 − π the denote S∪{u} 2 S∪{u} 2 (u, w2). . (11) 1 (w1) 1 (w2) (w2) (10) = = . S∪{u} 1 − pwi,v · π 1 1 − pwi,v · πS (v) with πS S∪{u} 2 1 (wi) (12) 2 (v) and S∪{u} 2 S∪{u} 2 (vw1,··· , wi) (vw1,··· , wi−1) 1 − π 1 − π Therefore, we can initialize π (v) with iteratively update π S∪{u} 2 (cid:0)1 − π 1 − (v)(cid:1) S∪{u} 2 S∪{u} 1 − pw,v · π 1 1 − pw,v · πS 1 (w) · (w) , (13) for all the nodes w ∈ Nu \ S and v ∈ Nw \ S. Moreover, for the direct neighbors of u, their two-hop activation probabilities also need to be adjusted because u's one-hop activation probability has changed from πS 1 (u) to 1. For each neighbor v of u, the adjustment can be made in a similar way by updating S∪{u} π 2 (v) with 1 − S∪{u} 2 S∪{u} 1 − pu,v · π 1 1 − pu,v · πS 1 (u) Then, the final two-hop activation probability π (cid:89) the iterative updates (13) and (14) is (cid:0)1 − π (v)(cid:1) (cid:0)1− πS 2 (v)(cid:1) (v) = 1− S∪{u} 2 π · · w∈(Mu,v∪{u}) (u) . (14) S∪{u} 2 (v) by S∪{u} 1 − pw,v · π 1 1 − pw,v · πS 1 (w) (w) . (15) Hence, the theorem is proven. Proof of Theorem 2: Consider a single seed {u}. Let Au ⊆ Nu denote a subset of a node u's neighbors. Let p(Au) denote the probability that all the nodes in Au are activated directly by u in the IC model while all the nodes in Nu \ Au are not directly activated by u (they may not even be activated eventually). Since each of u's neighbors is activated by u independently, we have · p(Au) =(cid:0) (cid:89) (1 − pu,v)(cid:1). (cid:0) (cid:89) (16) pu,v (cid:1) v∈Au v∈Nu\Au π π w∈V (cid:17) (cid:17) v∈Au = 1 + = 1 + = 1 + Au⊆Nu Au⊆Nu Au⊆Nu Au⊆Nu Au⊆Nu ≤ 1 + ≤ 1 + p(Au) · p(Au) · p(Au) · p(Au) · Au h−1(w) σh({u}) (cid:88) (cid:88) w∈V\{u} Au h−1(w) (cid:17) σh−1({v}) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:88) (cid:16) (cid:88) (cid:16) Furthermore, with h hops of propagation, for each node w ∈ V\{u}, w can only be activated by a propagation path starting from a node v ∈ Au whose path length is no longer than h−1 hops. In other words, the probability for w to be activated by Au h−1(w). Considering all the possible node sets Au Au is π activated directly by u, we have (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:17) (cid:88) p(Au) · σh−1(Au) (cid:0)σh−1({v}) · p(v ∈ Au)(cid:1)(cid:17) (cid:88) (cid:0)p(Au) · σh−1({v}) · p(v ∈ Au)(cid:1)(cid:17) (cid:0)p(Au) · σh−1({v}) · p(v ∈ Au)(cid:1)(cid:17) (cid:0)p(Au) · p(v ∈ Au)(cid:1)(cid:17) (cid:80) (cid:88) (cid:0)p(Au) · p(v ∈ Au)(cid:1) (cid:88) (cid:0)p(Au) · p(v ∈ Au)(cid:1) (cid:88) (cid:0)p(Au ∪ {v}) · p(v ∈ Au ∪ {v})(cid:1) (cid:88) (17) The second "≤" is due to the submodularity of σh(·) (see Theorem 3) such that σh−1(Au) ≤ σh−1({v}). In the third "=", p(v ∈ Au) is such a binary value that p(v ∈ Au) = 1 if and only if v ∈ Au. Meanwhile, we have Au⊆Nu σh−1({v}) · Au⊆Nu\{v} + Au⊆Nu\{v} (cid:88) Au⊆Nu Au⊆Nu Au⊆Nu = 1 + = 1 + = 1 + v∈Nu v∈Nu v∈Nu v∈Nu v∈Au . = = (18) Au⊆Nu\{v} p(Au ∪ {v}). The last "=" follows the fact that p(v ∈ Au) = 0 since v (cid:54)∈ Au ⊆ Nu\{v} and p(v ∈ Au∪{v}) = 1 since v ∈ Au∪{v}. (cid:17) Therefore, from (17) and (18), we have (cid:16) (cid:88) (cid:88) σh({u}) ≤ 1 + v∈Nu σh−1({v}) · Au⊆Nu\{v} . p(Au ∪ {v}) (19) Furthermore, by definition, Au⊆Nu\{v} = Au⊆Nu\{v} p(Au ∪ {v}) (cid:16)(cid:0) (cid:89) w∈Au∪{v} (cid:1) (cid:0) (cid:89) pu,w · w∈Nu\(Au∪{v}) (1 − pu,w)(cid:1)(cid:17) (cid:88) (cid:88) 9 (cid:88) (cid:16) (cid:88) (cid:0) (cid:89) (cid:16)(cid:0) (cid:89) w∈Au pu,w (cid:1) (cid:1) (cid:0) (cid:89) (cid:0) (cid:89) · w∈Nu\(Au∪{v}) (1 − pu,w)(cid:1)(cid:17) (1 − pu,w)(cid:1)(cid:17) pu,w · w∈Nu\(Au∪{v}) pu,v · = Au⊆Nu\{v} = pu,v · Au⊆Nu\{v} w∈Au = pu,v · 1 = pu,v. (cid:17) (cid:16) (cid:80) Thus, by (19) and (20), σh−1({v}) · pu,v (20) it holds that σh({u}) ≤ 1 + . v∈Nu Inequality (7) can be proved by induction. When h = 1, the inequality follows directly from Inequality (6). Suppose (cid:17) that it holds for h− 1 hops of propagation, i.e., σh−1({u}) ≤ σh−1({u}). Then, for h hops of propagation, we have (cid:17) pu,v · σh−1({v}) pu,v · σh−1({v}) σh({u}) ≤ 1 + (cid:88) (cid:88) (cid:16) (cid:16) v∈Nu (21) ≤ 1 + v∈Nu = σh({u}). (cid:88) h (S) = (cid:12)(cid:12)(cid:12). Thus, Therefore, for any h ≥ 0, we have σh({u}) ≤ σh({u}). Proof of Theorem 3: This can be proved using a similar approach to that by Kempe et al. in [16]. For each edge (u, v) ∈ E, we independently flip a coin of bias pu,v to decide whether the edge (u, v) is live or blocked to generate a sample influence propagation outcome X. We use p(X) to denote the probability of a specific outcome X in the sample space. Let h (v) denote the node set that can be reached from a node v V X within h hops in the sample outcome X. Then, the number of nodes that can be reached from a seed set S within h hops in the outcome X is given by σX (cid:12)(cid:12)(cid:12)(cid:83) (cid:0)p(X) · σX h (S)(cid:1), where the monotonicity of σh(S) holds since σX (cid:12)(cid:12)(cid:12)V X as S expands. (cid:91) v∈S V X (cid:83) and T where S ⊆ T , we have(cid:83) v∈T V X is the number of nodes that are reachable from a node u within h hops but are not reachable from any node in a seed set S within h hops in a sample outcome X. For any two node sets S h (v). v∈T V X h (v), which Thus, V X implies that (cid:83) v∈S V X h (u) \ σX h (S ∪ {u}) − σX The marginal influence gain h (S) increases h (v) ⊇ V X σh(S) = h (S) = v∈S V X v∈S V X h (v) ⊆ h (u) \ h (u) \ (cid:83) h (v) h (v) (22) (23) (cid:12)(cid:12)(cid:12) X σX h (S ∪ {u}) − σX (24) Since p(X) ≥ 0 for any X, taking the linear combination, we have h (T ∪ {u}) − σX h (S) ≥ σX h (T ). σh(S ∪ {u}) − σh(S) ≥ σh(T ∪ {u}) − σh(T ). Thus, σh(·) is submodular. ∗ h denote the optimal seed set for maximizing the influence within h hops of propagation, Proof of Theorem 4: Let S (25) 10 (cid:0)(1− 1/e)α(cid:1) that ∗) ≥ ∗ h) ≥ (1− 1/e)σh(S ∗ h) = maxS=k σh(S). We have σ(Sh) ≥ σh(Sh) ≥ i.e., σh(S ∗). · σ(S (1− 1/e)σh(S the exact inequality follows from the fact The first influence spread is equal to the influence spread without any hop limitation of propagation. The second inequality is due to the submodularity and monotonicity of σh(·) (Theorem 3). ∗ h is the optimal solution for The third inequality is because S maximizing σh(·). The last inequality is by the assumption that σh(S)/σ(S) ≥ α for any S = k. We first introduce some lemmas used to prove Theorem 5. Lemma 1: For scale free random graphs with propagation probability pu,v = p for every edge (u, v) ∈ E, the expected influence spread produced within one hop of propagation from a random seed set S satisfies E[σ1(S)] ≥ (p + 1)k − pk2/V. (26) Proof of Lemma 1: With one hop of propagation, for a randomly selected node v, it is not activated if and only if v is not a seed and v is not activated by any of its inverse neighbors. The probability for v to be a non-seed node is 1 − kV. The probability for an inverse neighbor of v to be a seed is kV and thus, the probability for it to activate v is p· kV. Therefore, the probability for all of v's inverse neighbors to fail to activate v is πS 1 (v) = 1 − (cid:89) u∈Iv . · k V pk V pk V k V (cid:1)Iv (cid:0)1 − p · (cid:0)1 − (cid:1) (cid:1) =(cid:0)1 − (cid:1) (cid:1)Iv (cid:0)1 − (cid:1)Iv(cid:105) (cid:0)1 − (cid:16) ∞(cid:88) k (cid:0)1 − V k (cid:0)1 − V k · V pk2 V2 . V − (cid:0)1 − (cid:0)1 − (cid:0)1 − (cid:0)1 − pk V P0(Iv) · pk V pk V ∞(cid:88) (cid:1) (cid:1) (cid:1) (cid:1) (cid:1) Iv=1 Iv=1 · · ≥ 1 − = 1 − (1 + p)k = pk V P0(Iv) 1 (v)] = E(cid:104) As a result, the expectation of the activation probability of a random node v is given by E[πS k V · (cid:1)Iv(cid:17) (cid:0)1 − = 1 − 1 − · (29) Therefore, it holds that E[σ1(S)] = V · E[πS 1 (v)] ≥ (p + 1)k − pk2/V. This completes the proof. Lemma 2 ([22]): For an infinite random power law graph, the expected fraction of nodes activated φ(S) = E[σ(S)]/V can be computed by 1 − ϕ(S) =(cid:0)1 − kV 1 − φ(S) =(cid:0)1 − kV d=0 P1(d + 1)(cid:0)1 − pϕ(S)(cid:1)d (cid:1)(cid:80)∞ d=1 P0(d)(cid:0)1 − pϕ(S)(cid:1)d (cid:1)(cid:80)∞ , , (30) Note that if v is selected as a seed, it must be activated. Hence, the overall activation probability of v is (27) . (28) (cid:16) (cid:0)1 − kV (cid:0)1 − k V k V (cid:0)1 − (cid:1)P1(1). (cid:1)P1(1)(cid:0)1 − pϕ(S)(cid:1)0 where A = 1 − 1 − ϕ(S) ≥ and (cid:0)1 − 1 − φ(S) ≥ (cid:1)P0(1)(cid:0)1 − pϕ(S)(cid:1). k V Hence, by (32) and (33), the lemma follows. Proof of Theorem 5: Lemma 1 indicates that E[σh(S)] ≥ E[σ1(S)] ≥ (p + 1)k − pk2/V. (cid:17) (cid:1)P0(1)(1 − pA) (cid:0)1 − E[σ(S)] ≤ V · 1 − (cid:16) . Lemma 3 indicates that Putting (34) and (35) together, the theorem follows. k V (33) (34) (35) (cid:80)∞ d1−γ di=1 d1−γ i where P1(d) = is the probability of a node connecting to a neighbor whose degree is d, and ϕ(S) is an instrumental variable. Lemma 3: The expected fraction of nodes activated φ(S) is E[σ(S)] ≤ V · (cid:17) (cid:1)P0(1)(1 − pA) bounded by 1 − (31) , Proof of Lemma 3: From (30) in Lemma 2, we have = 1 − A, (32) 11
1809.08368	2	1809	2018-12-31T18:55:09	Attacking Similarity-Based Link Prediction in Social Networks	[ "cs.SI", "cs.CR" ]	Link prediction is one of the fundamental problems in computational social science. A particularly common means to predict existence of unobserved links is via structural similarity metrics, such as the number of common neighbors; node pairs with higher similarity are thus deemed more likely to be linked. However, a number of applications of link prediction, such as predicting links in gang or terrorist networks, are adversarial, with another party incentivized to minimize its effectiveness by manipulating observed information about the network. We offer a comprehensive algorithmic investigation of the problem of attacking similarity-based link prediction through link deletion, focusing on two broad classes of such approaches, one which uses only local information about target links, and another which uses global network information. While we show several variations of the general problem to be NP-Hard for both local and global metrics, we exhibit a number of well-motivated special cases which are tractable. Additionally, we provide principled and empirically effective algorithms for the intractable cases, in some cases proving worst-case approximation guarantees.	cs.SI	cs	Attacking Similarity-Based Link Prediction in Social Networks Kai Zhou1, Tomasz P. Michalak2, Talal Rahwan3, Marcin Waniek2,3, and Yevgeniy Vorobeychik1 1Department of Computer Science and Engineering, Washington University in St. Louis, St. Louis 2Institute of Informatics, University of Warsaw, Warsaw, Poland 3Department of Computer Science, Khalifa University of Science and Technology, Abu Dhabi, UAE 8 1 0 2 c e D 1 3 ] I S . s c [ 2 v 8 6 3 8 0 . 9 0 8 1 : v i X r a ABSTRACT Link prediction is one of the fundamental problems in compu- tational social science. A particularly common means to predict existence of unobserved links is via structural similarity metrics, such as the number of common neighbors; node pairs with higher similarity are thus deemed more likely to be linked. However, a number of applications of link prediction, such as predicting links in gang or terrorist networks, are adversarial, with another party incentivized to minimize its effectiveness by manipulating observed information about the network. We offer a comprehensive algo- rithmic investigation of the problem of attacking similarity-based link prediction through link deletion, focusing on two broad classes of such approaches, one which uses only local information about target links, and another which uses global network information. While we show several variations of the general problem to be NP-Hard for both local and global metrics, we exhibit a number of well-motivated special cases which are tractable. Additionally, we provide principled and empirically effective algorithms for the intractable cases, in some cases proving worst-case approximation guarantees. KEYWORDS Computational social science; link prediction; security and privacy; adversarial attacks 1 INTRODUCTION Link prediction is a fundamental problem in social network analysis. A common approach to predicting a target link (u, v) is to use an observed (sub)network to infer the likelihood of the existence of this link using a measure of similarity, or closeness, of u and v; we call this similarity-based link prediction [1, 12, 20, 26]. For example, if u and v are individuals who have many friends in common, it may be natural to assume that they are themselves friends. Representational power of social networks implies very broad application of link prediction techniques, ranging from friend recommendations to inference of criminal and terrorist ties. A crucial assumption in conventional similarity-based link pre- diction approaches is that the observed (sub)network is measured correctly. However, insofar as link prediction may reveal relation- ships which associated parties prefer to keep hidden -- either for the ACM Conference, July 2017, Washington, DC, USA © 2019 Association for Computing Machinery. ...$ACM ISBN 978-x-xxxx-xxxx-x/YY/MM ...$15.00 https://doi.org/10.1145/nnnnnnn.nnnnnnn sake of privacy, or to avoid being apprehended by law enforcement -- it introduces incentives to manipulate network measurements in order to reduce perceived similarity scores for target links. In order to systematically study the ability of an "adversary" to manipulate link prediction, we formulate attacks on link prediction as an optimization problem in which the adversary aims to mini- mize the total weighted similarity scores of a set of target links by removing a limited subset of edges from the observed subnetwork. We present a comprehensive study of this algorithmic problem, focusing on two important subclasses of similarity metrics: local metrics, which make use of only local information about the target link, and global metrics, which use global network information. We show that the problem is in general NP-Hard even for local metrics, and our hardness results are stronger for the commonly used Katz and ACT global similarity metrics (for example, the problem is hard for these metrics even if there is only a single target link). On the positive side, we exhibit a number of important special cases when the problem is tractable. These include attacks on local metrics when there is a single target link, or a collection of tar- get nodes (such as gang members) with the goal of hiding links among them. Additionally, we present practical algorithms for the intractable cases, including global similarity metrics. In a number of such settings, we are able to provide provable approximation guar- antees. Finally, we demonstrate the effectiveness of the approaches we develop through an extensive experimental evaluation. Related Work. Link prediction has been extensively studied in multiple domains such as social science [12], bioinformatics [2], and security [8]. There are two broad classes of approaches for link prediction: the first based on structural similarity [12, 14] and the second using learning [1, 15, 19, 21]. This work is focused on the former, which commonly use either local information [10, 28], rely on paths between nodes [9, 13], or make use of random walks [5] (we view the latter two categories as examples of global metrics). Our work is connected to several efforts studying vulnerability of social network analysis (SNA). Michalak et al. [16] suggest consid- ering strategic considerations in SNA, but do not offer algorithmic analysis. Waniek et al. study attacks against centrality measures and community detection [22, 23]. There is considerable literature on hiding or anonymizing links on networks (e.g., [24, 25, 27]), but these approaches allow arbitrary graph modifications and are in any case heuristic, often proposing randomly swapping or rerouting edges. In contrast, we provide the first comprehensive algorithmic study of the problem of hiding links by merely deleting observed edges (i.e., preventing them from being observed), and the first strong positive algorithmic results. 2 PROBLEM FORMULATION 2.1 Similarity Metrics One of the major approaches for link prediction both in the network science literature and in practice is via the use of similarity met- rics [12]. Specifically, suppose we wish to know whether a particular link (u, v) connecting nodes u and v exists. A structural similarity metric Sim(u, v) quantifies the extent to which the nodes u and v have shared topological properties, such as shared neighbors, with the idea that higher similarity scores imply greater likelihood that u and v are connected. Below, we will distinguish two types of similarity metrics: local, which only use information about the nodes and their immediate neighbors, and global, which make use of global information about the network. 2.2 Attack Model At the high level, our goal is to remove a subset of observed edges in order to minimize perceived similarity scores of a collection of target (and, presumably, existing) links. This could be viewed both from the perspective of vulnerability analysis, where the goal of link prediction is to identify relationships among malicious parties (such as gang members), or privacy, where the "attacker" is not malicious, but rather aims to preserve privacy of a collection of target relationships. To formalize the problem, consider an underlying graph G = (V , E) representing a social network, where V is the set of nodes and E is the set of edges. This graph is not fully known, and instead an analyst obtains answers for a collection of edge queries Q from the environment, where for each query (u, v) ∈ Q, they observe the associated edge if (u, v) ∈ E, and determine that the edge doesn't exist otherwise. The partially constructed graph GQ = (VQ , EQ) based on the queries Q is then used to compute similarity metrics Sim(u′, v′) for any potential edges (u′, v′) (cid:60) Q. An attacker has a collection of target links H they wish to hide, and can remove a subset of at most k edgs in EQ ≡ E ∩ Q to this end. While there are many ways to express the attacker's objective mathematically, a relatively natural and general approach is to minimize the weighted sum of similarity scores of links in H: ft(Ea) ≡ (u,v)∈H min Ea ⊂EQ wuv Sim(u, v; Ea), s.t. Ea ≤ k, (1) where wuv is the weight representing the relative importance of hiding the link (u, v), and we make explicit the dependence of similarity metrics on the set of removed edges Ea. Henceforth, we simplify notation by keeping this dependence implicit. 3 ATTACKING LOCAL SIMILARITY METRICS Our analysis covers nine representative local similarity metrics (summarized in the supplement) that are commonly used in the state-of-the-art link prediction algorithms. We first systematically divide local metrics into two sub-class: Common Neighbor Degree (CND) and Weighted Common Neighbor (WCN) metrics, depend- ing on their special structures. Next, we show that attacking all local metrics is NP-Hard. We follow this negative result with an approximation algorithm exhibiting a solution-dependent bound. Finally, we present polynomial-time algorithms for well-motivated special cases. We begin by introducing some notation. We denote U = {ui} as the union of end-nodes, termed target nodes, of the target links in H. Assume U = n. Let W = {w1, w2, · · · , wm} be the set of common neighbors of the target nodes, where each wi ∈ W connects to at least two nodes in U . Let N(ui , uj) denote the set of common neighbors of ui and uj. For any node ui ∈ V , let d(ui) be its degree. We use a decision matrix X ∈ {0, 1}m×n to denote the states of edges among the nodes in W and U , where the entry xij in the i-th row and j-th column of X equals 1 if there is an edge between wi and uj; otherwise, xij = 0. We will say the attacker erases xij (when xij = 1) to denote the fact that the attacker deletes the edge between wi and uj (thus setting xij as 0). 3.1 Classification of Local Metrics We now make a useful distinction between two classes of local metrics that use somewhat different local information. i, j(ui ,uj )∈H xr i ·xr j total similarity ft has the formm the sum as ij henceforth. Definition 3.1. A metric Sim is a CND metric if the corresponding , where fr is a metric-dependent increasing function of Sr , the sum of rth row of decision matrix X, and Wr is an associated weight. r =1 Wr fr (Sr ) The metrics Adamic-Adar (AA), Resource Allocation (RA), and Common Neighbors (CN) are CND metrics. We note that the sum i, j (ui,uj)∈H xri ·xr j is over all links in H. For simplicity, we write Definition 3.2. A metric Sim is a WCN metric if N(ui,uj) • it has the form Sim(ui , uj) = д(d(ui),d(uj), N(ui,uj)), where д is strictly increasing in d(ui) and d(uj). That is д(d(ui) − t, d(uj) − s) ≤ д(d(ui), d(uj)) for any valid non-negative integers t and s and any valid value of N(ui , uj). • Sim is strictly increasing in N(ui , uj). That is, Sim(N(ui , uj)− t) ≤ Sim(N(ui , uj)), for any valid non-negative integer t and any valid values of d(ui) and d(uj). The WCN metrics include many common metrics, such as Jac- card, Sørensen, Salton, Hub Promoted, Hub Depressed, and Leicht. By the above definitions, we know a rational attacker will only delete edges between nodes in W and nodes in U , since deleting other types of edges will either decrease d(ui) or d(wi), causing the similarity to increase. Thus, the total similarity ft is fully captured by the decision matrix X. As a result, attacking local similarities is formulated as an optimization problem, termed as Prob-Local: ft(X), min X (2) 0 is the original decision matrix and Sum(·) denotes the 0 − X) ≤ k, s.t. Sum(X where X element-wise summation. 3.2 Hardness Results We start by making no restrictions on the set of target links H. In this general case, we show that attacking all local metrics is NP-hard. Theorem 3.3. Attacking local similarity metrics is NP-Hard. Proof. As attacking local similarity metrics is modelled as an optimization problem, we consider the corresponding decision prob- lem: can an attacker delete up to k edges such that the total similar- ity ft is no greater than a constant θ? We note that the minimum possible ft for all local metrics in a connected graph is 0. Thus, we consider the decision problem PL, which is to decide whether one can we delete k edges such that ft = 0. We use the vertex cover problem for reduction. Let PV C denote the decision version of vertex cover, which is to decide whether there exists a vertex cover of size k given a graph G and an integer k. Given an instance of vertex cover (i.e., a graph G = (V , E) and an integer k), we construct our decision problem PL as follows. We first construct a new graph Q in the following steps: • For each node vi ∈ V , create a node vi for graph Q. • Add another node w to Q and connect w to each vi. • Add n = V nodes u1, · · · , un and add an edge between each pair of nodes (ui , vi). • Add an edge between (ui , ui +1), for i = 1, 2, · · · n − 1. The set H of target links is then H = {(vi , vj)} in Q if and only if (vi , vj) is an edge in G. Our decision problem PL is then constructed regarding this graph Q and target set H. 1: = Example: {v2, v3} is a set cover in G while H Figure G. Vc {(v1, v2),(v1, v3),(v2, v3),(v2, v4),(v2, v5),(v2, v6)} of target links in Q. constructing is = the set graph Q from Now, we show PL and PV C are equivalent. We use CN metrics as an example and show that the same proof can be applied to other local metrics by slightly modifying the constructed graph Q. First, we show if there is a vertex cover of size k in graph G, then we can delete k edges such that ft(H) = 0 in Q. Suppose Vc is a vertex cover with Vc = k. Without loss of generality, let Vc = {v1, · · · , vk}. Then we show that delete k edges(v1, w), · · · ,(vk , w) will make ft(H) = 0. Let (vi , vj) ∈ H be an arbitrary target link. Then (vi , vj) corresponds to an edge in G. By the definition of vertex cover, we have at least one of vi and vj is in Vc. We can assume vi ∈ Vc. Since viandvj has only one common neighbor w in Q, deleting (vi , w) will make CN(vi , vj) = 0. As (vi , vj) is arbitrarily selected, we have CN(vi , vj) = 0 for any target link (vi , vj) ∈ H. Thus, we have found k edges whose deletion will make ft(H) = 0. Second, we show if we can delete k edges to make ft(H) = 0 in Q, the we can find a vertex cover of size k in G. Suppose we found k edges whose deletion will make ft(H) = 0. Then each deleted edge must be (w, vi) for some i = 1, · · · , n, since deleting other types of edges will not decrease ft(H). Without loss of generality, we assume the k deleted edges are (w, v1), · · · ,(w, vk). We then show that Vc = {v1, · · · , vk} forms a vertex cover in G. Since ∀(vi , vj) ∈ H, CN(vi , vj) ≥ 0, ft(H) = 0 means that CN(vi , vj) = 0 for very target link. As each target link (vi , vj) initially has one common neighbor w, we know at least one of vi and vj is in set Vc; otherwise, CN(vi , vj) = 1 making ft(H) > 0. As each (vi , vj) corresponds to an edge in G, we know each edge in G has at least one end node in VC. By definition, Vc is a vertex cover of size k. As a result, PL and PCV is equivalent, proving that minimiz- ing CN metric is NP-hard. The other local metrics are different variations of CN metrics. To make the above proof applicable for other metrics, we need to construct graph Q such that ft(H) = 0 if and only if there is no common neighbors between each pair of target link. To achieve this, we can slightly modify the graph Q constructed previously for CN metric. For CND metrics, we can add some isolated nodes to Q and connect w with each of the isolated nodes. For WCN metrics, we can add some isolated nodes for each node vi and connect each isolated node with vi to make sure that the degree of each vi is always positive. Then the previous proof holds for other local metrics. □ 3.3 Practical Attacks Since in general attacking even local metrics is hard, we have two ways of achieving positive results: approximation algorithms and restricted special cases. We start with the former, and exhibit several tractable special cases thereafter. To obtain an approximation algorithm for the general case, we use submodular relaxation. Specifically, we bound the denominator of each term of ft by constants as if all the budget were assigned to decrease that single term, arriving at an upper bound ftu for the original objective ft . For WCN metrics, let дij be the denominator of Sim(ui , uj). For each дij, we bound it by Lij ≤ дij ≤ Uij, where Lij is obtained when k edges are deleted and Uij is obtained when no edge is deleted. Take 2N(ui,uj) Sørensen metric as an example, where Sim(ui , uj) = d(ui)+d(uj). 0 Then d denote i the original degrees of ui and uj, respectively. In this way, each similarity is bounded as N(ui , uj) j − k ≤ d(ui) +d(uj) ≤ d 0 , where d and d +d +d 0 i 0 i 0 j 0 j = Uij ≤ Sim(ui , uj) ≤ N(ui , uj) = and f W C N Lij ij . t tl tu Ui j . Then N(ui,uj) ≤ f W C N N(ui,uj) Let f W C N ij tu Li j ≤ f W C N . f W C N tl Similarly, for CND metrics, the denominator in each term fr(Sr) r) − k ≤ fr(Sr) ≤ fr(S is bounded by fr(S r), where S 0 0 0 r denotes 0. Then the sum of the rth row of the original decision matrix X ijxr i xr j ≤ f C N D , where f C N D f C N D fr (Sr ) tu tl . Due to the similarity between and f C N D r =1 Wr the structures of f W C N and omit the superscript W CN in the following analysis. The proposed approximation algorithm the associated bound analysis are also applicable for f C N D , we will focus on f W C N ijxr i xr j fr (Sr )−k and f C N D t = m =m ≤ f C N D r =1 Wr tu tl . t t Optimizing Bounding Function. We now consider minimizing ftu. Let S′ be the set of edges that the attacker chooses to delete. Then set S′ is associated with a decision matrix X ′. For any S ⊂ S′, we have X ≥ X ′, where X is the matrix associated with S and ≥ denotes component-wise comparison. Define a set function t t v1v1v2v2v3v3v4v4v5v5v6v6u1u2u3u4u5u6wGQF(S) = ftu(X is equivalent to 0) − ftu(X). Clearly, F(∅) = 0. Then minimizing ftup F(S), s.t. S ≤ k. (3) max S ⊂EQ ij Li j i , C′ wi j ⟨Ci,Cj ⟩ product. Now, ftu(X) = Theorem 3.4. F(S) is a monotone increasing submodular function. Proof. Assume S ⊂ S′, we need to show F(S) ≤ F(S′). It is equivalent to show ftu(X) ≥ ftu(X ′). Let Ci be the ith column of X. Then N(ui , uj) = ⟨Ci , Cj⟩, where ⟨Ci , Cj⟩ denotes their inner , where the weights wij and Lij are constants. Since X ≥ X ′, we have ⟨Ci , Cj⟩ ≥ ⟨C′ j⟩ for every pair of i, j. Thus, ftu(X) ≥ ftu(X ′). That is, F(S) is monotone increasing. Let an edge e (cid:60) S′ be associated with the p-th row and q-th column entry in X. Let e ∪ S be associated with a matrix X e, where pq = 0 while xpq = 1. the only difference between X e and X is that xe Similarly, let e ∪S′ be associated with a matrix X ′e. Define ∆(eS) = F(e ∪ S) − F(S) and ∆(eS′) = F(e ∪ S′) − F(S′). Then we need to show ∆(eS) ≥ ∆(eS′). ∆(eS) = ftu(X) − ftu(X e) = xpj · xpq − = where the sum Similarly, we can obtain ∆(eS′) = ∆(eS′) = wjq Ljq wjq Ljq j is over all pairs of (j, q) such that (uj , uq) ∈ H. The second equality holds as deleting edge e will only affect the q-th column. The last equality holds since xpq = 1 and xe pj). Since (xpj −x′ wjq Ljq ∆(eS′) ≥ 0. By definition, F(S) is submodular. . Then ∆(eS) − pj) ≥ 0, we have ∆(eS)− □ Problem (3) is to maximize a monotone increasing submodular function under cardinality constraint. The typical greedy algorithm for such type of problems achieves a (1− 1/e)-approximation of the maximum. In particular, the greedy algorithm will delete the edge that will cause the largest increase in F(S) step by step until k edges are deleted. Suppose the greedy algorithm outputs a sub-optimal set S∗, which corresponds to a minimizer X∗ u of ftu(X). We then take the value ft(X∗ u) as the approximation of ft(X∗), where X∗ is the optimal minimizer of ft . We term this approximation algorithm as Approx-Local. ⟨Cj , Cq⟩ − pq = ⟨Ce q⟩ j , Ce xpj , wjq Ljq wjq Ljq (xpj −x′ pj · xe xe pq = 0. wjq Ljq x′ pj wjq Ljq j j j j j j j Bound Analysis. We theoretically analyze the performance of our proposed approximation algorithm Approx-Local.1 Let X∗, X∗ u, and X∗ be the minimizers of ft , ftu, and ftl , respectively. Define l the gap between ft and ftu as α(X) = ftu(X) − ft(X), which is a function of the decision matrix X. Theorem 3.5. The gap α(X) is an increasing function of X . Proof. Consider a particular term of α(X), which is denoted as αij(X) = wi j j ⟩, where , CX Li j 1We note that for the CN metric in particular, the set function F(S) is the actual objective. Consequently, the greedy algorithm above yields a (1 − 1/e)-approximation in this case. д(d(ui),d(uj), ⟨C X j ⟩)⟨CX j ⟩ − , CX ⟨CX i i ,C X wi j i + = ). ≥ − wij ⟨CX i ⟨CX i j ⟩ i , CY j ⟩ i ,CY Li j denotes the ith column of X. For simplicity, write д(d(ui), d(uj), CX i ⟨CX j ⟩) as д(X). , CX i Consider an edge connecting to ui is deleted. This corresponds to the case when an entry in CX is erased. Denote the resulting i matrix as Y. Then X ≥ Y. The gap at Y is αij(Y) = wij( ⟨CY − ⟨CY j ⟩ i ,CY д(Y) αij(X) − αij(Y) j ⟩ ⟨CX , CX i д(X) As д is strictly increasing in d(ui) and d(uj), it is increasing in ⟨CX j ⟩ , CX i д(Y) X. Then we have д(X) ≥ д(Y). Thus, j ⟩ − ⟨CY αij(X) − αij(Y) , CX Lij j ⟩ − ⟨CY , CX ⟨CY j ⟩ i , CY д(Y) − 1 д(Y)) ≥ 0. j ⟩ i , CY j ⟩)( 1 j ⟩ − ⟨CY , CX Lij ⟨CY j ⟩ i , CY д(Y) The last inequality holds as Lij is the lower bound (i.e., Lij ≤ д(Y)). As α(X) is the weighted sum over all pair of target links, we have α(X) ≥ α(Y). □ Theorem 3.5 states that the gap between the total similarity and its upper bound function is closing as we delete more edges (i.e., X becomes smaller). We further provide a solution-dependent bound of g = ft(X∗ u) − ft(X∗), which measures the gap between the minimum of ft output by our proposed algorithm and the real minimum. ∗ ∗) ≤ ftu(X ∗ g ≤ ftu(X u) − ftl(X l ). Such a gap depends on the solutions X∗ u and X∗ . We evaluate the l gap through extensive experiments in Section 5. ∗) ≤ ftu(X ∗ u) − ftl(X ∗ u) − ft(X = (⟨CX i i , CY wij Lij − + 3.4 Tractable Special Cases We identify two important special cases for which the attack models are significantly simplified. The first case considers attacking a single target link and optimal attacks can be found in linear time for all local metrics. The second case considers attacking a group of nodes and the goal is to hide all possible links among them. We demonstrate that optimal attacks in this case can be found efficiently for the class of CND metrics. Due to the space limit, we only highlight some key observations and present some important results. The full analysis is in the extended version [3] of the paper. 3.4.1 Attacking a Single Link. When the target is a single link (u, v), the attacker will focus only on the links connecting u or v with their common neighbors, denoted as N(u, v) = {wi}s i =1. Let xiu = 0 denotes that attacker chooses to delete the link between wi and u and xiu = 1 otherwise. Proposition 3.6. For CND metrics, Sim(u, v) =s i =1 xiu xiv д(d(wi)) , where д is a non-decreasing function of d(wi). To minimize a CND, the attacker will remove edges incident to common neighbors w in increasing order of degree d(w). In fact, this algorithm is optimal and has a time complexity O(N(u, v)). For WCN metrics, consider a tuple (u, w, v) where w is a common neighbor of u and v. We divide the links surrounding (u, v) into four sets: E1 = {(u, w)}, E2 = {(v, w)}, E3 = {(u, s)}, and E4 = {(v, s)}, where s denotes a non-common neighbor of u and v. As the attacker deletes links from EQ , there are four possible states of the tuples between u and v. In state 1, both (u, w) and (w, v) are deleted. In state 2, (u, w) is deleted while (w, v) is not. In state 3, (w, v) is deleted while (u, w) is not. In state 4, neither (u, w) not (w, v) is deleted. We use integer variables y1, y2, y3 to denote the number of tuples in state 1, 2, 3, respectively. Furthermore, let y4 and y5 be the number of deleted edges from E3 and E4, respectively. In this way, the vector (y1, y2, y3, y4, y5) fully captures an attacker's strategy. Proposition 3.7. A WCN metric can be written as Sim(u, v) = f (y1, y2, y3, y4, y5) such that f is decreasing in y2 and y3 and f is increasing in y4 and y5. 5 = 0 and y∗ 3 = k. That is, the attacker will always choose k edges from E1 ∪ E2 to delete. The following theorem then specifies how the attacker can optimally choose edges. Our analysis shows that in an optimal attack, y∗ 4 = y∗ 1 = y∗ 2 + y∗ Theorem 3.8. The optimal attack on WCN metrics with a single target link selects arbitrary y∗ 2) links from E2 to delete with the constraint that for any selected links (u, w1) ∈ E1 and (v, w2) ∈ E2, w1 (cid:44) w2. The value of y∗ 2 is the solution of a single-variable integer optimization problem. 2 links from E1 and (k−y∗ The time complexity of solving the single-variable integer opti- mization problem is bounded in O(k). 3.4.2 Attacking A Group of Nodes. We consider the special case where 1) the target is a group of nodes U and the links between each pair of nodes in U consist the target link set H; 2) each link in H has equal weight. In this case, optimal attacks on CND metrics can be found in polynomial time. is a convex increasing function of Si . Proposition 3.9. For CND metrics, the total similarity ft has the i =1 fi(Si), where Si is the sum of the ith row of X and fi(Si) Proposition 3.9 states that ft for CND metrics can be written as a sum of independent functions, where each function fi is a convex increasing function. We then propose a greedy algorithm termed Greedy-CND to minimize f C N D . In essence, Greedy-CND takes as the input S0, which is the row sum of the initial decision matrix X, and decreases an entry in S0 whose decreasing causes the maximum decrease in f C N D step by step until an upper bound of k edges are deleted. This algorithm turns out to be optimal, as we prove in the extended version of the paper. t t formm 4 ATTACKING GLOBAL METRICS In this section, we analyze attacks on two common global similarity metrics: Katz and ACT. We begin with attacks on a single link and show that finding optimal attack strategies is NP-hard even for a single target link. Let A ∈ {0, 1}N×N and D be the adjacency matrix and degree matrix of the graph GQ , respectively. The Laplacian matrix is de- fined as L = D − A. The pseudo-inverse of L is L† = (L − E)−1 + E, where E is an (N × N) matrix with each entry being 1 . We use a bi- N nary vector y ∈ {0, 1}M to denote the states of edges in EQ , where yi = 0 iff the ith edge in EQ is deleted. Finally, y ≤ y′ (A ≤ A′) is a component-wise inequality between vectors (matrices). 4.1 Problem Formulation for Katz Similarity The Katz similarity is a common path-based similarity metric [9]. For a pair of nodes (u, v), Katz similarity is defined as Katz(u, v) = βl pathl u,v = (βA + β 2 2 + β A 3 3 + · · · )uv , A ∞ l =1 u,v denotes the number of walks of length l between where pathl u and v, β > 0 is a parameter and (·)uv denotes the entry in the uth row and vth column of a matrix. By definition, the adjacency matrix A is fully captured by the vector y. Thus, Katz(u, v) is a function of y, written as Katzuv(y). As one would expect, it is an increasing function of y. Lemma 4.1. Katzuv(y) is an increasing function of y. Proof. Let A and A′ be the corresponding adjacency matrices of y and y′. If y ≤ y′, we have A ≤ A′. Now, consider the jth term of the Katz similarity matrix K, which is β jAj. As every entry in A is non-negative and β > 0, we have β jAj ≤ β jA′j, for every j. Thus, Katzuv(y) ≤ Katzuv(y′). □ As a result, deleting a link will always decrease Katzuv(y), and the attacker would therefore always delete k links in EQ (if EQ has at least k links). Thus, minimizing Katz for a particular target link (u, v) is captured by Prob-Katz: Katzuv(y), s.t. min y yi = M − k, y ∈ {0, 1}M . M i =1 4.2 Problem Formulation for ACT The second global similarity metric we consider is based on ACT, which measures a distance between two nodes in terms of random walks. Specifically, for a pair of nodes (u, v), ACT(u, v), is the ex- pected time for a simple random walker to travel from a node u to node v on a graph and return to u. Since ACT(u, v) is a distance metric, the attacker's aim is to maximize ACT(u, v), defined as † uv), ACT(u, v) = VG(L † † vv − 2L uu + L where VG is the volume of the graph [5]. Directly optimizing ACT(u, v) is hard. Indeed, deleting an edge may either increase or decrease ACT(u, v), so that unlike other metrics, ACT is not monotone in y. Fortunately, Ghosh et al. [6] show that when edges are unweighted (as in our setting), ACT(u, v) can be defined in terms of Effective Resistance (ER): ACT(u, v) = VG ER(u, v). It is also not difficult to see that both the volume VG and ER can be represented in terms of y. We begin by investigating the effect of deleting an edge on ER(y). We use a well-known result by Doyle and Snell [4] to this end. Lemma 4.2 ([4]). The effective resistance between two nodes is strictly increasing when an edge is deleted. The following lemma is then an immediate corollary. Lemma 4.3. ER(y) is a decreasing function of y. As a result, maximizing ER(y) would always entail deleting all allowed edges. Let t be the maximum number of edges that can be deleted. Then, maximizing ER(y) can be formulated as Prob-ER: ER(y), s.t. max y yi = M − t, y ∈ {0, 1}M . M i =1 2M However, while ER(y) increases as we delete edges, volume VG = i =1 yi decreases. Fortunately, since volume is linear in the num- ber of deleted edges, we reduce the problem of optimizing ACT to that of solving Prob-ER by solving the latter for t = {0, . . . , k}, and choosing the best of these in terms of ACT. Similarly, hardness of Prob-ER implies hardness of optimizing ACT. Consequently, the rest of this section focuses on solving Prob-ER. 4.3 Hardness Results We prove that minimizing Katz and maximizing ER between a single pair of nodes by deleting edges with budge constraint are both NP-hard. Theorem 4.4. Minimizing Katz similarity and maximizing ACT distance is NP-hard even if H contains a single target link. Proof. We consider the decision version of minimizing Katz, termed as PK , which is to decide whether one can delete k edges to make Katz(u, v) ≤ q given a graph Q and a target node pair (u, v) in Q. Similarly, we consider the decision version of maximizing ER, termed as PE: which is to decide whether one can delete k edges to make ER(u, v) ≥ q given a graph Q and a target node pair (u, v) in Q. We use the Hamiltonian cycle problem, termed PH , for reduction. PH is to decide whether there exists a circle that visits each nodes in a given connected graph G exactly once (thus called Hamiltonian circle). Before reduction, we first consider the minimum of Katz and max- imum of ER between two nodes u and v in all possible connected graphs over a fixed node set. By the definition of Katz similarity, Katz(u, v) is minimized when the graph is a string with u and v as two end nodes and all others as inner nodes in that string; that is the graph over that set of nodes is a Hamiltonian path with u and v as end nodes. We denote the minimum value of Katz(u, v) in this case as minK . Similarity, by the definition of effective resistance, ER(u, v) is maximized when the graph is also a Hamiltonian path over that set of nodes with u and v as the two end nodes. We assume that all edges have equal resistance. We denote the maximum value of ER(u, v) in this case as maxE. We then set q = minK in the decision problem PE and set q = maxE in PE. As a result, the two decision problems PE and PK are then both equivalent to the following decision problem, termed PS : given a graph Q and two nodes u and v in Q, can we delete k edges such that the remaining graph S forms a string (i.e., a Hamiltonian path) with u and v as two end nodes? Now the reduction. Given an instance of Hamiltonian circle (i.e., a graph G = (V , E)), we construct a new graph Q from G in the following steps: • Select an arbitrary node w in G. Let N(w) = {l1, l2, · · · , lW } be the neighbors of w, where W = N(w). • Add two nodes u and v. • Add edge (u, w) and edges (v, li), ∀li ∈ N(w). The resulting graph is then the graph Q in decision problem PS , where the budget k = W + E − V . Below is an example showing the construction of Q and the process of deleting edges. Figure 2: Illustration: construct graph Q from G and delete edges from Q to obtain S. We now show that problem PH and problem PS are equivalent. First, we show if there exists a Hamiltonian circle in G, then we can delete k = W +E−V edges such that the measurement (Katz or ER) between u and v in graph Q is q. Assume the Hamiltonian circle travels to w through edge (li , w) and leaves w through edge (w, lj). We then 1) delete (W − 1) edges (v, lt) for each lt ∈ N(W ) and lt (cid:44) li; 2) delete all E − V edges in G that do not appear in the Hamiltonian circle; 3) delete edge (w, li). Thus, we deleted a total of W + E − V edges. After deleting all these k edges, in the remaining graph S, there exists a Hamiltonian path between w and li. As u only connects to w and v only connects to li, the remaining graph forms a Hamiltonian path between u and v. As a result, the measurement between u and v equals q. Second, we show if we can remove k = W + E − V edges from Q such that the remaining graph S forms a Hamiltonian path between u and v, then we can find a Hamiltonian circle in the graph G. Suppose in the reaming string, v connects to li and w connects to lj. As u connects only to w in graph Q, u must connect to w in S. From the construction of Q, the total number of edges of Q is E + W + 1. After deleting k edges, the remaining number of edges is V + 1. Excluding the two edges (u, w) and (v, li), we know there are V − 1 edges among the node set V of the original graph G. As the remaining graph is connected, there must exist a Hamiltonian path between w and li. As (w, li) is an edge in the graph G, we have found a Hamiltonian circle in G, consisting of the Hamiltonian path between w and li plus the edge (w, li). maximizing ER (ACT) are NP-hard. Thus, decision problem PS is NP-complete; minimizing Katz and □ 4.4 Practical Attack Strategies While computing an optimal attack on Katz and ACT is NP-Hard, we now devise approximate approaches which are highly effective in practice. 4.4.1 Attacking Katz Similarity. To attack Katz similarity, we transform the attacker's optimization problem into that of maximiz- ing a monotone increasing submodular function. We begin with the single-link case (i.e., H is a singleton), and subsequently generalize to an arbitrary H. We define a set function f (Sp) as follows. Let Sp ⊆ EQ be a set of edges that an attacker chooses to delete. Let Ap be the adjacency matrix of the graph GQ after all the edges in Sp are deleted. Define f (Sp) = βAp + β 2 2 p + β A 3 p + · · · 3 A GQSwwwuuvvl1l1l1lilililjljljlWlWlWSince there is a one-to-one mapping between the set Sp and the matrix Ap, the function f (Sp) is well-defined. We note that f (Sp) gives the Katz similarity matrix of the graph G after all the edges in Sp are deleted. We further define a set function дuv(Sp) = (K − f (Sp))uv , where K = f (∅) (the Katz similarity matrix when no edges are deleted) and (·)uv denotes the uth row and vth column of a matrix. Clearly, when Sp = ∅, дuv(Sp) = 0. Then, Prob-Katz is equivalent to дuv(Sp), (4) Theorem 4.5. The set function дuv(Sp) is monotone increasing s.t. Sp = k max Sp ⊂Et and submodular. Proof. To prove that дuv is monotone increasing, we need to show that ∀Sp ⊂ Sq ⊂ Q, дuv(Sp) ≤ дuv(Sq). It is equiva- lent to show (f (Sp))uv ≥ (f (Sq))uv. We note that (f (Sp))uv and (f (Sq))uv are the Katz similarity between u and v after the edges in Sp and Sq are deleted, respectively. Theorem 4.1 states that the Katz similarity will decrease as more edges are deleted. Since Sp ⊂ Sq, we have f (Sp) ≥ f (Sq). Thus, дuv(Sp) ≤ дuv(Sq). Next, we prove дuv is submodular. Let e ∈ Et \ Sq be an edge between node i and node j in the graph. Let G be an n × n matrix where Gij = Gji = 1 and the rest of the entries are 0. Then we have the set Sp ∪ e is associated with Ap − G and Sq ∪ e is associated with Aq − G. For a set S, let ∆(eS) = f (S ∪ e)− f (S). Then we need to show ∆(eSp) ≤ ∆(eSq). Denote the tth item of ∆(eS) as ∆(t)(eS). In the following, we will first prove ∆(t)(eSp) ≤ ∆(t)(eSq) by induction. Assume that the inequality holds for t = s (it's straightforward to verify the case for t = 1 and t = 2). That is βs[(Ap − G)s − (Ap)s − (Aq − G)s + (Aq)s] ≤ 0. (5) When t = s + 1, we have By induction, we have ∆(t)(eSp) ≤ ∆(t)(eSq) for t = 1, 2, 3, · · · . Note that when β is chosen to be less than the reciprocal of the (∆ (s +1)(eSp) − ∆ (s +1)(eSq))/βs +1 = (Ap − G)s +1 − (Ap)s +1 − (Aq − G)s +1 + (Aq)s +1 = (Ap − G)s Ap − (Ap)s +1 − (Aq − G)s Aq + (Aq)s +1 − [(Ap − G)s +1 − (Aq − G)s +1]G ≤ (Ap − G)s Ap − (Ap)s +1 − (Aq − G)s Aq + (Aq)s +1 The inequality comes from the fact that (Ap − G) ≥ (Aq − G) when G ≥ 0. Furthermore, since Sp ⊂ Sq, we have Ap = Aq + F for some F ≥ 0. Thus, (∆ (s +1)(eSp) − ∆ (s +1)(eSq))/βs +1 ≤ (Ap − G)s(Aq + F) − (Ap)s(Aq + F) − (Aq − G)s Aq + (Aq)s +1 = [(Ap − G)s − (Ap)s − (Aq − G)s + (Aq)s]Aq + [(Ap − G)s − (Ap)s]F ≤ 0 maximum of the eigenvalues of Aq − G, the sum will converge. Thus, ∆(eSp) ≤ ∆(eSq). □ i, j wij Kij. Next, for the multi-link case, the total similarity ft = Let F(S) be a function of the set of deleted edges, defined as F(S) = βAS + β 2 2 A S + β 3 3 A S + · · · , where AS denotes the adjacency matrix after all edges in S are deleted. Note that F(S) gives the Katz similarity matrix when edges in S are deleted. Further define дij(S) = (K 0 is i, j wijдij(S). By ij − ft . Thus, minimizing ft 0 the original Katz similarity matrix. Let Gt(S) = definition, we have Gt(S) = 0 − F(S))ij, where K i, j wij K is equivalent to max S ⊂EQ Gt(S), s.t. S ≤ k. The following result is then a direct corollary of Theorem 4.5. Corollary 4.6. Gt(S) is monotone increasing and submodular. Proof. This is an immediate conclusion of two results. First, дij(S) is monotone increasing and submodular in S as proved in Theorem 4.5. Second, a positive linear combination of submodular functions is submodular [17]. As Gt(S) is the sum of дij(S), Gt(S) is monotone increasing and submodular. □ As a result, minimizing the total Katz similarity is equivalent to maximizing a monotone increasing submodular function under cardinality constraint. We can achieve a (1− 1/e) approximation by applying a simple iterative greedy algorithm in which we delete one edge at a time that maximizes the marginal impact on the objective. We call this resulting algorithm Greedy-Katz. d(u) + 1 4.4.2 Attacking ACT. From the analysis of minimizing Katz sim- ilarity, it is natural to investigate submodularity of the effective resistance or ACT as a function of the set of edges. Unfortunately, counter examples show that the effective resistance is neither sub- modular nor supermodular. Consequently, we need to leverage a different kind of structure for ER. Our first step is to approximate the objective function ER(u, v) based on the results by Von Luxburg et al. [18], who show that ER(u, v) can be approximated by 1 d(v) for large geometric graphs as well as random graphs with given expected degrees. Consequently, we use the approximation ER(u, v) ≈ ERap(u, v) = d(u) + 1 1 d(v). Then the total effective resistance is approximated as d(ui), where Wi > 0 is some constant weight associated with each ui. Let Di be the original degree of node ui and zi be an integer variable denoting the number of deleted edges connecting to ui. Then maximizing ERap(H) is equivalent to Wi Di − zi ER(H) ≈ ERap(H) = d(uj)) =n We assume that deleting edges would not make the graph dis- i =1 connected. That is ∀i ∈ [1, n], k < Di. We formulate the above problem as a linear integer program. Specifically, let ∆ij be the decrease in ERap(H) after j edges con- necting to node ui are deleted. As any such j edges will cause the same decrease, the value of each ∆ij for j = 0, 1, · · · , k could be zi ≤ k, zi ∈ [0, k]. 1 d(ui) + 1 n n ij wij( i =1 Wi max i =1 s.t. (6) z , efficiently computed in advance. We use a binary variable hij = 1 to denote that the attacker chooses to delete such j edges; otherwise, hij = 0. Then problem (6) is equivalent to n k k n k i =1 j=0 hij ≤ k,∀i, (Wi Di − ∆ij)hij , s.t. hij ≤ 1, max h The above problem is a linear program of (k + 1) × n binary variables with (n + 1) linear constraints. A numerical solution [7] gives the number of edges incident to each node that needs to be deleted. i =1 j=0 j=0 A Note on a Special Case. We can consider problem (6) for the two special cases introduced in Section 3.4. For both cases (attack- ing a single link and attacking a group of nodes), the weights Wi associated with each node are equal. Under this condition, we can prove that the optimal solution to problem (6) is to delete all k edges connecting the node with the smallest degree. Details are provided in the extended version of the paper. 5 EXPERIMENTS Our experiments use two classes of networks: 1) randomly gen- erated scale-free networks and 2) a Facebook friendship network [11]. In the scale-free networks, the degree distribution satisfies P(k) ∝ k−γ , where γ is a parameter. Baseline algorithms. We compare our algorithms with two base- line algorithms. We term the first one as RandomDel, which ran- domly deletes the edges connected to the target nodes. The second baseline, termed GreedyBase, is a heuristic algorithm proposed in [24]. This algorithm will try to delete the link whose deletion will cause the largest decrease in the number of "closed triads" as de- fined in [24]. Our experiments show that while the performance of GreedyBase varies regarding different metrics, RandomDel performs poorly for all metrics (Fig. 3). Henceforth, we only compare our algorithm with GreedyBase for global metrics (Fig. 4 and Fig. 5). For local metrics, we evaluate Approx-Local in the general case. We consider a target set of size 20. We select RA (CND metric) and Sorensen (WCN metric) as two representatives, for which the results are presented in Fig. 3. All similarity scores are scaled to 1.0 when no edges are deleted. Due to space limit, we only present the results on one scale of the scale-free network (n = 1000, γ = 2.0) and Facebook network(n = 786, m = 12291). A more comprehensive set of experiments is presented in the extended version. We note that deleting a relatively small number of links can significantly decrease the similarities of a set of target links. The gap between the upper and lower bound functions, which reflects the approximation quality of Approx-Local, is within 20% of the original similarity. For global metrics, we evaluate Greedy-Katz and Local-ACT re- garding a set of target links (H = 20) on different scales of net- works. As shown in Fig. 4 and Fig. 5, the performances are sig- nificantly better than those of the baseline algorithm. Additional results for the special cases are provided in the extended version. 6 CONCLUSION We investigate the problem of hiding a set of target links in a network via minimizing the similarities of those links, by deleting (a) scale-free, RA (b) Facebook, RA (c) scale-free, Sørensen (d) Facebook, Sørensen Figure 3: Approx-Local vs. GreedyBase on CND (e.g., RA) and WCN (e.g., Sørensen) metrics in general case. (a) scale-free (b) Facebook Figure 4: Greedy-Katz vs. GreedyBase on Katz similarity. Solid lines: Greedy-Katz. Dotted lines: GreedyBase (a) scale-free (b) Facebook Figure 5: Local-ACT vs. GreedyBase on ACT distance. Solid lines: Local-ACT. Dotted lines: GreedyBase a limited number of edges. We divide similarity metrics associated with potential links into two broad classes: local metrics (CND and WCN) and global metrics (Katz and ACT). We prove that computing optimal attacks on all these metrics is NP-hard. For local metrics, we proposed an algorithm minimizing the upper bounds of local metrics, which corresponds to maximizing submodular functions under cardinality constraints. Furthermore, we identify two special cases, attacking a single link and attacking 02468101214Number of deleted edges0.40.60.81.01.21.4Similarity scoreRA upper boundRA approxRA lower boundRA GreedyBaseRA RandomDel02468101214Number of deleted edges0.30.40.50.60.70.80.91.01.11.2Similarity scoreRA upper boundRA approxRA lower boundRA GreedyBaseRA RandomDel02468101214Number of deleted edges0.40.60.81.01.2Similarity scoreSoren U boundSoren approxSoren GreedyBaseSoren L BoundSoren RandomDel02468101214Number of deleted edges0.50.60.70.80.91.01.11.2Similarity scoreSoren U boundSoren approxSoren GreedyBaseSoren L BoundSoren RandomDel0246810Number of deleted edges0.00.20.40.60.81.0Katz similarity(n=400,=2.0)(n=400,=2.3)(n=800,=2.0)(n=800,=2.3)0246810Number of deleted edges0.00.20.40.60.81.0Katz similarity(n=168,m=1656)(n=333,m=2519)(n=150,m=1693)(n=61,m=270)02468101214Number of deleted edges1.01.52.02.53.0ACT distance(n=500,=2.0)(n=500,=2.3)(n=1000,=2.0)(n=1000,=2.3)02468101214Number of deleted edges1.001.051.101.151.201.251.30ACT distance(n=168,m=1656)(n=333,m=2519)(n=786,m=12291)(n=1034,m=26749)Autonomous Agents and Multiagent Systems, 1341 -- 1349. [23] Marcin Waniek, Tomasz P Michalak, Michael J Wooldridge, and Talal Rahwan. 2018. Hiding individuals and communities in a social network. Nature Human Behaviour 2, 2 (2018), 139. [24] Marcin Waniek, Kai Zhou, Yevgeniy Vorobeychik, Esteban Moro, Tomasz P Micha- lak, and Talal Rahwan. 2018. Attack Tolerance of Link Prediction Algorithms: How to Hide Your Relations in a Social Network. arXiv preprint (2018). [25] Shanqing Yu, Minghao Zhao, Chenbo Fu, Huimin Huang, Xincheng Shu, Qi Xuan, and Guanrong Chen. 2018. Target Defense Against Link-Prediction-Based Attacks via Evolutionary Perturbations. (2018). arxiv preprint. [26] Muhan Zhang and Yixin Chen. 2018. Link Prediction Based on Graph Neural Networks. arXiv preprint arXiv:1802.09691 (2018). [27] Peng Zhang, Xiang Wang, Futian Wang, An Zeng, and Jinghua Xiao. 2016. Mea- suring the robustness of link prediction algorithms under noisy environment. Scientific reports 6 (2016), 18881. local information. The European Physical Journal B 71, 4 (2009), 623 -- 630. [28] Tao Zhou, Linyuan Lü, and Yi-Cheng Zhang. 2009. Predicting missing links via a group of nodes, where the first case ensures optimal attacks for all local metrics and the latter ensures optimal attacks for CND metrics. For global metrics, we prove that even when attacking a single link, both the problem of minimizing Katz and that of maximizing ACT are NP-Hard. We then propose an efficient greedy algorithm (Greedy-Katz) and a principled heuristic algorithm (Local- ACT) for the two problems, respectively. Our experiments show that our algorithms are highly effective in practice and, in particular, significantly outperform a recently proposed heuristic. Overall, the results in this paper greatly advance the algorithmic understanding of attacking similarity-based link prediction. REFERENCES [1] Mohammad Al Hasan, Vineet Chaoji, Saeed Salem, and Mohammed Zaki. 2006. Link prediction using supervised learning. In SDM06: workshop on link analysis, counter-terrorism and security. [2] Wadhah Almansoori, Shang Gao, Tamer N Jarada, Abdallah M Elsheikh, Ayman N Murshed, Jamal Jida, Reda Alhajj, and Jon Rokne. 2012. Link prediction and classification in social networks and its application in healthcare and systems biology. Network Modeling Analysis in Health Informatics and Bioinformatics 1, 1-2 (2012), 27 -- 36. [3] Anonymous Authors. 2018. Attacking Similarity-Based Link Prediction in So- cial Networks (Extended Version). (2018). https://drive.google.com/open?id= 13I1wE8ZgdkW4EiZT_V_BtEVi48JbxOzS [4] Peter G Doyle and J Laurie Snell. 2000. Random walks and electric networks. arXiv preprint math/0001057 (2000). [5] Francois Fouss, Alain Pirotte, Jean-Michel Renders, and Marco Saerens. 2007. Random-walk computation of similarities between nodes of a graph with appli- cation to collaborative recommendation. IEEE Transactions on knowledge and data engineering 19, 3 (2007), 355 -- 369. [6] Arpita Ghosh, Stephen Boyd, and Amin Saberi. 2008. Minimizing effective resistance of a graph. SIAM review 50, 1 (2008), 37 -- 66. http://www.gurobi.com [7] LLC Gurobi Optimization. 2018. Gurobi Optimizer Reference Manual. (2018). [8] Zan Huang and Dennis KJ Lin. 2009. The time-series link prediction problem with applications in communication surveillance. INFORMS Journal on Computing 21, 2 (2009), 286 -- 303. [9] Leo Katz. 1953. A new status index derived from sociometric analysis. Psychome- [10] Elizabeth A Leicht, Petter Holme, and Mark EJ Newman. 2006. Vertex similarity trika 18, 1 (1953), 39 -- 43. in networks. Physical Review E 73, 2 (2006), 026120. Dataset Collection. http://snap.stanford.edu/data. (June 2014). [11] Jure Leskovec and Andrej Krevl. 2014. SNAP Datasets: Stanford Large Network [12] David Liben-Nowell and Jon Kleinberg. 2007. The link-prediction problem for social networks. Journal of the American society for information science and technology 58, 7 (2007), 1019 -- 1031. [13] Linyuan Lü, Ci-Hang Jin, and Tao Zhou. 2009. Similarity index based on local paths for link prediction of complex networks. Physical Review E 80, 4 (2009), 046122. [14] Linyuan Lü and Tao Zhou. 2011. Link prediction in complex networks: A survey. Physica A: statistical mechanics and its applications 390, 6 (2011), 1150 -- 1170. [15] Aditya Krishna Menon and Charles Elkan. 2011. Link prediction via matrix factorization. In Joint european conference on machine learning and knowledge discovery in databases. Springer, 437 -- 452. [16] Tomasz P Michalak, Talal Rahwan, and Michael Wooldridge. 2017. Strategic Social Network Analysis. In AAAI. 4841 -- 4845. [17] George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. 1978. An analysis of approximations for maximizing submodular set functionsâĂŤI. Mathematical programming 14, 1 (1978), 265 -- 294. [18] Ulrike Von Luxburg, Agnes Radl, and Matthias Hein. 2014. Hitting and commute times in large random neighborhood graphs. The Journal of Machine Learning Research 15, 1 (2014), 1751 -- 1798. [19] Hao Wang, Xingjian Shi, and Dit-Yan Yeung. 2018. Relational Deep Learning: A Deep Latent Variable Model for Link Prediction. In AAAI Conference on Artificial Intelligence. 2688 -- 2694. [20] Peng Wang, BaoWen Xu, YuRong Wu, and XiaoYu Zhou. 2015. Link prediction in social networks: the state-of-the-art. Science China Information Sciences 58, 1 (2015), 1 -- 38. [21] Xu-Wen Wang, Yize Chen, and Yang-Yu Liu. 2018. Link Prediction through Deep Learning. (2018). arxiv preprint. [22] Marcin Waniek, Tomasz P Michalak, Talal Rahwan, and Michael Wooldridge. 2017. On the construction of covert networks. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. International Foundation for
1612.08629	4	1612	2018-11-30T16:23:38	Simulating SIR processes on networks using weighted shortest paths	[ "cs.SI", "physics.soc-ph" ]	We present a framework to simulate SIR processes on networks using weighted shortest paths. Our framework maps the SIR dynamics to weights assigned to the edges of the network, which can be done for Markovian and non-Markovian processes alike. The weights represent the propagation time between the adjacent nodes for a particular realization. We simulate the dynamics by constructing an ensemble of such realizations, which can be done by using a Markov Chain Monte Carlo method or by direct sampling. The former provides a runtime advantage when realizations from all possible sources are computed as the weighted shortest paths can be re-calculated more efficiently. We apply our framework to three empirical networks and analyze the expected propagation time between all pairs of nodes. Furthermore, we have employed our framework to perform efficient source detection and to improve strategies for time-critical vaccination.	cs.SI	cs	Simulating SIR processes on networks using weighted shortest paths Dijana Toli´c1, Kaj-Kolja Kleineberg2, and Nino Antulov-Fantulin2,* 1Laboratory for Machine Learning and Knowledge Representations, Rudjer Boskovi´c Institute, Zagreb, Croatia 2Computational Social Science, ETH Zurich, Clausiusstrasse 50, 8092 Zurich, Switzerland *[email protected] ABSTRACT We present a framework to simulate SIR processes on networks using weighted shortest paths. Our framework maps the SIR dynamics to weights assigned to the edges of the network, which can be done for Markovian and non-Markovian processes alike. The weights represent the propagation time between the adjacent nodes for a particular realization. We simulate the dynamics by constructing an ensemble of such realizations, which can be done by using a Markov Chain Monte Carlo method or by direct sampling. The former provides a runtime advantage when realizations from all possible sources are computed as the weighted shortest paths can be re-calculated more efficiently. We apply our framework to three empirical networks and analyze the expected propagation time between all pairs of nodes. Furthermore, we have employed our framework to perform efficient source detection and to improve strategies for time-critical vaccination. Introduction Modern socio-technological systems can be described as networks with naturally occurring spreading phenomena such as information propagation in social networks1, 2, or the epidemic spread among individuals through their contact network 3 -- 6. In order to exactly solve the spreading process one must be able to evaluate the probabilistic evolution of the process configurations in time i.e. solve a set of coupled differential equations called master equation7 -- 10 or Chapman-Kolmogorov equation. However, due to the exponential growth of possible configurations only in small number of special cases9, 10 it is possible to solve the master equation directly. Therefore, different approximations are used: (i) neglecting dynamical correlations by assuming statistical independence at first neighbourhood, pair or node level in mean field approximations11 -- 16, (ii) neglecting loops in the network structure with message passing or belief propagation techniques17 -- 19, or (iii) neglecting the temporal evolution by mapping to the percolation process20 -- 22. Alternative way is the simulation or sampling the statically correct process configurations or realizations with the dynamic Monte Carlo techniques23 -- 28. In this paper, we propose how to exactly transform the SIR spreading dynamics to a shortest path problem on an ensemble of weighted networks using a Monte Carlo approach. In this framework, weights represent interaction time delays along edges. We show the equivalence between the time respecting paths (shortest paths) on the constructed weighted networks and the propagation times of spreading dynamics. The mapping is applicable to the generalized Susceptible Infected Recovered (SIR) model5 without memory (exponential inter-event distribution/Poisson process) and with memory (arbitrary inter-event distributions). Contrary to the standard dynamic Monte Carlo methods23 -- 28, we do not have to specify initial conditions upfront and we can sample new realizations from the previous ones by making local random Markov Chain transitions between the weighted networks. In a limit of infinite process time, we establish the connection with bond percolation theory20, 21, 29, 30. In the special case of independently assigned edge weights, we establish the connection to the disordered in networks31 -- 35. Finally, we show the applications to source detection18, 19, 36 -- 39 and time-critical vaccination40 -- 42. Methods Mapping SIR dynamics to weighted shortest paths Let us start with the Poissonian SIR model, which is memoryless, such that the next state only depends on the present state of the system. This process is defined by two parameters (β ,γ), which describe transitions where susceptible nodes become infected with rate β from infected neighbours and transitions where infected nodes recover with rate γ. This process has exponential inter-event time distributions: ψ(t) = β e−βt (spread) and φ (t) = γe−γt (recovery). However, many realistic processes have memory26 and their inter-event time distributions are non-exponential. Therefore, we allow for arbitrary inter-event density distributions (ψ(t),φ (t)) for generalized SIR processes. 1 Figure 1. The mapping of spreading dynamics to an ensemble of weighted networks. The weights represent the propagation time delays. Each weighted network represents one realization of the stochastic dynamics from an arbitrary source node. The propagation time from the source node to any other node is equal to the shortest path length in weighted network (color coded). The dotted line represents edges with infinite edge weight. For a given network G and class of generalized SIR spreading models, we show how to create weighted networks {Gk} which encode realizations of the stochastic spreading dynamics. Each weighted network Gk (see Fig. 1) encodes one possible outcome of the spreading process from every potential source node in a network. In particular, a time-respecting weighted network instance Gk is created by taking the input network G and assigning weights to the edges of the network instance with the Inverse Smirnov transform43 as (cid:40) ρi, j = Ψ−1(x) : Ψ−1(x) ≤ Φ−1(y) Ψ−1(x) > Φ−1(y) ∞ : , (1) distributions Φ(t) =(cid:82) t 0 dt(cid:48)φ (t(cid:48)) and Ψ(t) =(cid:82) t where x and y are uniform random numbers ∈ [0,1], Φ−1(x) and Ψ−1(y) are inverse functions of the cumulative inter-event 0 dt(cid:48)ψ(t(cid:48)). The quantities Ψ−1(y) and Φ−1(x) respectively represent the samples of the transmission and recovery time obtained with the Inverse Smirnov transform of inter-event distributions. In the special case of the Poissonian SIR process one obtains (cid:40)−ln(x)/β : −ln(x)/β ≤ −ln(y)/γ −ln(x)/β > −ln(y)/γ . ρi, j = ∞ : (2) If nodes recover faster than a transmission occurs, the weight is set to infinity to indicate no transmission through the edge. We denote the distance as shortest paths on weighted networks, dGk (vi,v j) = minχi j ∑(k,l)∈χi j ρk,l, where χi j is the set of all possible paths from node vi to node v j on network Gk and ρk,l denotes the weights defined in Eq. (1). Importantly, this distance is equivalent to the propagation time from node i to j, i.e. t(vi → v j) = dGk (vi,v j) (time respecting path equivalence, see Supplementary information, Sec. 1). The run-time complexity of finding shortest paths44 from specific source node to others is O(E + N logN), where N denotes the number of nodes and E number of edges in a network. The exact mapping is done by generating a random variable y for each node and variable x for each edge. This takes all dynamical correlations into the account and generates the ensemble of directed weighted networks. A simplified mean-field mapping can be obtained by generating random variables x and y per each edge, independently. This simplified mean-field mapping holds when β >> γ (SI case) and generates the ensemble of undirected weighted networks (see Supplementary information, Sec. 5 and 6. for more details). In the section Additional information of this paper, we provide a link to the code that implements the proposed mapping. Sampling methods for simulation The instances of the ensemble of weighted networks can be sampled either (i) independently, or (ii) by traversing elements of the ensemble with Markov Chain. The Markov Chain (rejection-free Gibbs sampler -- see Supplementary information, Sec. 3) consists of transitions Wi, j = w(Gi → G j) over the set of weighted networks {Gk}, where each weighted network corresponds to one realization of the 2/11 Figure 2. Markov Chain Monte Carlo sampling. The transitions Wi, j = w(Gi → G j) between different instances of weighted networks in the ensemble. Randomization of the weight of an edge (denoted by red) corresponds to the transition between the states (realizations of the process). spreading process from every potential source node. This allows us to construct transitions between weighted networks by changing the weights in the first neighbourhood of a randomly selected node for exact mapping or by assigning a new weight to the randomly selected edge for a mean-field mapping, both according to Eq. (1) (see Fig. 2). The existence of a stationary distribution of this Markov Chain is guaranteed by detailed balance property and the uniqueness by ergodicity (see Supplementary information, Sec. 3). In other words, the Gibbs sampling -- as well as independent sampling -- constructs the probability distribution P(G ) of the weighted network ensemble, such that it is a statistically exact representation of the stochastic process. From the ensemble of weighted networks, due to ergodicity, the expectation (cid:104) f (G )(cid:105) can be estimated as the average over n samples of weighted networks: (cid:104) f (G )(cid:105) = ∑ Gk P(Gk) f (Gk) ≈ 1/n n ∑ k=1 f (Gk) (3) By changing functions f (G ), we estimate different properties such as total outbreak size, temporal evolution, expected propagation time or source likelihood (see Supplementary information, Sec. 4). The same estimation formula 3 can be applied to the independent sampling. The convergence rate of independent estimation are bounded with the Berry -- Esseen inequality as O(n−1/2). If all to all shortest paths need to be calculated, after each weight change in Markov Chain transition, shortest paths can be dynamically recalculated45, 46 with computational complexity O(cid:0)N2log3N(cid:1), where N denotes the number of nodes in a network (see Supplementary information, Sec. 8 for more details). In Fig. 3, we demonstrate that the estimations converge to the analytical solution on a toy network model. This confirms that the shortest paths in the ensemble of weighted networks are taking into the account stochastic spreading along all possible paths in the original unweighted network. Relation to bond percolation In the following, we state the connection to percolation theory. From the perspective of source node vi, the shortest path dG (vi,v j) indicates the first infection time of node v j. Therefore all nodes that are not reachable within the temporal distance t stay in the susceptible state. If a node is reachable within the temporal distance t, then it was infected. By letting t go to infinity we obtain asymptotic realization equivalence to bond percolation. In particular, the size of the set of nodes v j that are reachable in any finite time from the source node vi, Sp(vi), is given by Sp(vi) = lim t→∞ {d(vi,v j) < t}. (4) The full proof is given in the Supplementary information, Sec. 2. and holds for simplified mean-field mapping, where edge weights are independent. Hence, Sp(vi) denotes the size of the connected bond percolation component, where the transmissibility parameter29 from source vi is given by (cid:90) ∞ 0 (cid:90) τ 0 p = dτφi(τ) dτ(cid:48)ψi(τ(cid:48)). (5) In Fig. 4, we present the experiments where the equivalence to bond percolation is validated. The transmissibility quantifies the probability that an infected node transmits infection along a link to a susceptible neighbour before it recovers. The second integral accounts for the conditional probability of transmission up to the fixed recovery time τ, and the first integral account for all possible recovery times τ. Furthermore, to take dynamical correlations into the account that were previously missing 3/11 j=0 pn, j ·(cid:16) Figure 3. Toy network model example. (a) Sketch of simple toy network, that consists of nc chains with l nodes. The probability that source node s infects destination node d can be calculated analytically (see supplementary information, Sec. 5. for details) and is given by P(s → d) = 1− ∑nc neighborhood Eq. 7 for Poissionan SIR process and Eq. 6 for generalized non-Markovian process. (b) Simulation results on the toy network with n = 20, l = 3 with Poisson SIR (β = 1,γ = 1). The blue curve represents the estimations with the exact mapping (conditional independence among edge weights), the red curve represents the mean-field mapping (weights are independent) and the black dotted line represents the analytical solution (see supplementary information, Sec. 5. for a million size toy network experiment). . The term pn,k, denotes transmissibility for the first 1− pl 1,1 (cid:17) j in29, we generalize and formalize the transmissibility for the first neighborhood or, more precisely, the probability pn,k that k out of n directed links will be active (6) (7) (cid:19) (cid:18)n k pn,k = (cid:90) ∞ where Ψ(τ) =(cid:82) τ (cid:18)n (cid:19) γ 0 pn,k = k β φ (τ)dτ (1− Ψ(τ))(n−k) Ψ(τ)k, 0 ψ(t)dt. For Poissionan process this becomes Γ(k + 1)Γ( γ+β (n−k) Γ(k + 1 + γ+β (n−k) ) β ) β . The generalized transmissibility for the first neighborhood Eq.7 establishes the connection to semi-directed bond percola- tion30 (see Supplementary information, Sec. 5 and 6). To have exact mapping of the SIR process to percolation, having a single percolation parameter p is no longer accurate when β ≈ γ and generalized transmissibility pn,k for the first neighborhood is needed to capture all dynamical correlations. Relation to disordered networks The question how the average propagation time between two individuals scales with the system size is crucial for many real applications and of special theoretic interest. In particular, if this distance scales logarithmically with the system size, an epidemic can spread nearly instantaneously in arbitrary large networks. However, if it scales like Nσ , in large enough systems the disease will not spread through the network in realistic timescales. We can address this question for the SI-model (γ = 0) in an elegant fashion by establishing an analogy to the shortest path problem in disordered networks31 -- 35, which are generated by assigning weights ρ independently to edges drawn from some distribution f (ρ). In31 the authors have shown that the shortest path length from node vi to node v j in these weighted networks scales differently with the system size depending on f (ρ), and distinguish between strong and weak disorder. The shortest path length is defined as d(vi,v j) = minχi j ∑(k,l)∈χi j ρk,l, where χi j is the set of all possible paths. For the example of Erdos-Renyi networks, in the case of strong disorder the shortest path length scales as (cid:104)d(cid:105) ∝ N1/3, and for weak disorder as (cid:104)d(cid:105) ∝ logN. The temporal distance we have defined earlier is equivalent to the shortest paths on weighted graphs. Therefore, we can use this analogy to study the scaling of the average propagation time with the system size. From31 it is know that if f (ρ) is taken as the exponential distribution, only weak disorder can occur, and the 4/11 Figure 4. Equivalence to bond percolation. Average outbreak size in time estimated with the proposed mapping and comparison with bond percolation late-time outbreak size (dotted lines), see Eq.4. In a limit of infinite time the mapping becomes equivalent to the bond percolation. Results were obtained for the SIR dynamics (continuous time) with different transmission rates β (0.3, 0.03, 0.003, 0.0003) and recovery rate γ = 0.001 and with n = 104 simulations on the 4-connected two dimensional regular lattice (11x11). average distance scales with the logarithm of the systems size for Erdos-Renyi networks. This means that for the memory-less spreading process, where ψ(·) is the exponential distribution, the average propagation time scales with the logarithm of the system size and diseases spread quickly in systems of any size. However, if the inter-event distribution is for instance given by a lognormal distribution (non-Markovian case), we know from 31 that strong disorder can occur such that the average distance and hence the average propagation time scales like (cid:104)d(cid:105) ∝ N1/3. In this case, the finite time properties of the spreading process are drastically different to the non-Markovian case, as the average propagation time becomes very large as the size of the system increases, such that the disease will not spread globally in realistic timescales. To conclude, although it has recently been shown that the infinite time steady states of non-Markovian processes can be mapped to Markovian ones47, their finite time dynamics can be dramatically different. Results Expected propagation time between node pairs Starting from a given node, what is the expected time after which the disease reaches each other node in the network? The answer to this question is encoded in the ensemble of weighted networks. For SI processes (γ = 0), the expected propagation time is always finite and can be estimated as the average over n samples of weighted networks as Di j = 1/n∑n k=1 f (Gk), where f (Gk) = dGk (vi,v j) denotes the length of weighted shortest path between vi and v j. Note, that for the SIR process, the expected propagation time is not finite i.e. it diverges. In Fig. 5(a), we show an explicit example from the perspective of a specific source node. The advantage of the temporal maps we have created is they encode this type of information between all possible node pairs. The result is shown in Fig. 5(b-d) for three empirical networks: the email network48, the airport network39, and the Petser online social network49. The plots show the average temporal distance or expected propagation time Di j for each pair of nodes i, j, and therefore encompass information about the expected spreading dynamics. In particular, we have ordered the nodes by their characteristic spreading timescale, which we define as the expected time after which a given node infects N other nodes, τi, as τi = t∑ j(cid:54)=i Θ(t−Di j)=N, (8) where Θ(·) denotes the Heaviside step function, which is one when t ≥ Di j and zero otherwise. A small value of τi means that node i is of high importance or influence. This ordering of nodes displays the smooth gradient of the average temporal distance between pairs of nodes on all three empirical networks. Source detection For a given snapshot r∗ of a spreading process at time t0 on a network, how does one detect the source node that generated the snapshot? Source detection18, 19, 36 -- 39 is an important theoretical and practical problem in network science. It is at least as hard as any problem in NP-complete class37, 51, because it belongs to #P-complete class52. We approach the problem by 5/11 Figure 5. The expected propagation time Di j on empirical networks. The average propagation time is calculated as the average shortest path distances in the ensemble of weighted networks (mapped dynamics). (a) Nodes in the email network network are placed on a circle, where the angular coordinate represents the average propagation time from the source node (black) with index 500, increasing in the clockwise direction (from blue to red). The average propagation time Di j for: (b) the email network48 (β = 0.01), (c) the airport network (β estimated from flux data see Supplementary information, Sec. 7), (d) the Petster network49 (β = 0.01). Nodes are ordered by their characteristic spreading timescale τi, see Eq. 8) with condition N = N/2. probabilistic kernel estimations36, 53 of the likelihood that source node generated the observed snapshot. Our framework can be used to perform this task, as the ensemble of weighted networks encompasses information about different realizations of the spreading process from all possible sources. For the observed realization r∗, our source probability estimates for every node vi are constructed from ensemble averages si ∝ ∑m j f (G j). From weighted network G j, we extract realization based on temporal distances and calculate the similarity ϕ j ∈ [0,1] to the observed realization r∗. Similarity ϕ j is the normalized number of equal corresponding states between two realizations. Finally, we use the Gaussian kernel function f (G j) = exp(−(ϕ j − 1)2/a2) (9) for probability estimations36, 53 (see Supplementary information, Sec. 4 for more details). We compare our source probability estimations with direct Monte Carlo approach36 and topological distance approach50. The direct Monte Carlo approach generates large number of simulations n = 106 − 108, from each potential source node vi, and source probability estimation is si ∝ ni, where ni is the number of simulations that match the observed realization r∗. The topological distance approach calculates the average topological distance (cid:104)di(cid:105) from each potential source vi to other infected nodes in the observed realization r∗ and source estimation is si ∝ (cid:104)di(cid:105)−1. The result is shown in Fig. 6, where source ranked probability estimates for different methods are shown. Indeed, the method based on the ensemble of weighted networks (n = 105 instances) performs better than a topological distance estimation50 and correctly assigns a high probability to the source candidate identified by the direct Monte Carlo method36. 6/11 Figure 6. Source detection. We compare source detection performance for different methods for the SIR model: (a) β = 0.7,γ = 0.3 and (b) β = 0.7,γ = 0.7 at time T = 5 (discrete time) on the 4-connected lattice (30x30 nodes). For each observed realization, we rank nodes according to their estimated probability score according to the Direct Monte Carlo method36, in case of a tie we prioritize the node with a higher estimator according to the temporal distance method. We compare different methods as explained in the text: Direct Monte Carlo source probability ranking, source detection ranking by topological distance50, and the estimation from our mapping (temporal distance). Results are averaged over 30 observed realizations and error bars show one standard deviation, from top to bottom. Time critical vaccination Now, let us consider the following vaccination scenario. We observe at time t0 the outbreak of a disease and we have a certain amount, m, of vaccines, see Fig. 7(a). We can vaccinate m individuals instantaneously, but the vaccines will only be effective after some time ∆t (a vaccinated individual can get infected before t0 + ∆t, and in this case this vaccine portion is wasted, see Fig. 7(b). Who shall we vaccinate? Our framework can be used to improve the vaccination strategy. The idea is that, at time t0, we vaccinate only individuals who -- with high probability -- will not get infected before t0 + ∆t. Therefore, we estimate the probability that a node i will not be infected before t0 + ∆t as pi = 1/n n ∑ k=1 Θ(dGk (vθ ,vi)−t0 − ∆t), (10) where θ denotes the source node and Θ(·) is one when dGk (vθ ,vi) ≥ t0 +∆t and zero otherwise. We vaccinate nodes proportional to p. In Fig. 7(b-c) we show using empirical Petster network that this procedure performs better than randomly choosing from all susceptible nodes at t0. Interestingly, vaccinating proportional to degree (hubs strategy) performs even worse than the fully random strategy, since the hubs usually are infected earlier, which increases the chance to waste vaccine portions. Discussion We have mapped spreading dynamics on networks to an ensemble of weighted networks. This mapping allows us to define a temporal distance as shortest paths between nodes on an ensemble of weighted networks. We concentrate on the stochastic formulation of the generalized continuous and discrete time Susceptible Infected Recovered (SIR) spreading dynamics without memory (exponential inter-event distribution) and with memory (arbitrary inter-event distributions). The problem of shortest path with the stochastic independent and identically distributed weights (i.i.d.) has been extensively researched in theoretical computer science54 -- 57, probability theory58, 59 and statistical physics community31 -- 35. However, the general SIR dynamics is exactly represented with networks where the edge weights are conditionally independent random variables and only in special case of spreading they become independent variables e.g. SI dynamics (γ = 0) or SIR dynamics with fixed recovery time (ψ(t) = δ (t − T )). Thus, our mapping to the ensemble of weighted networks, provides additional dynamical interpretation of the shortest paths problem in disordered networks. Previously, only the Independent Cascade spreading models (edge weights are i.i.d.) were connected to the problem of shortest paths60 on networks, but with no connections to disordered networks or bond percolation literature. 7/11 Figure 7. Time critical vaccination. (a) Shows an illustration of the time critical vaccination process. Vaccination candidates are highlighted by the diamond symbols. Red note denote infected individuals, and blue nodes are susceptible. In (b) and (c), we show the total number of infected nodes up to time t for the different vaccination strategies described in the text (m = 0.2 of the total number of nodes). The first dashed vertical line denotes T0 = 3, and the second one T0 + τ = 13. In (b) for the discrete time SIR β = 0.03,γ = 0.01 and in (c) for the discrete time SIR β = 0.05,γ = 0.01 with source node 10 in the Petster network. In contrast to message passing, our framework is applicable for arbitrary network structures without neglecting loops in the network structure. Furthermore, our method takes into account dynamical correlations between node states overcoming the assumptions of mean-field like approximations12, 14. In contrast to the well known historical Gillespie or kinetic Monte Carlo methods23 -- 28, 36, we do not have to specify initial conditions upfront and we can sample new realizations from the previous ones by making local random perturbations on the weighted networks. We have shown that for some non-Markovian processes, the average propagation time can scale as a polynomial with the system size, even if the underlying network is small world. In stark contrast to the Markovian case, this means that for large enough systems the disease does not spread globally in a realistic timescale. We have shown that our framework allows for direct applications like ranking of nodes according to their characteristic spreading timescale, source detection and time-critical vaccination. Combining our mapping framework with existing vaccination strategies is likely to further improve the results, especially when time is a crucial factor. Note that previous studies39, 61 -- 63 for estimating the spreading arrival times were concentrated on a global disease spread with different approximations and theoretical understanding of effective distances64 for metapopulation models. However, the exact mathematical equivalences for mapping generalized spreading dynamics at a microscopic level were still lacking. Furthermore, in a limit of infinite process time, we establish the connection with bond percolation theory. Previously the outcome in infinite time of the SIR process have been successfully mapped to the percolation20, 21, 29, 30 but not for the finite time. Contrary to previous state-of-the-art source inference studies18, 19, 36, estimations with our method require same computational effort to sample realizations at an arbitrary point in time t not only for discrete time but also for continuous time dynamics. 8/11 Acknowledgments The authors would like to thank to prof. D. Helbing, A. Lancic, T. Lipic, O. Woolley and K. Caleb for useful comments, constructive feedback and discussions. The work of N.A.F. has been funded by the EU Horizon 2020 SoBigData project under grant agreement No. 654024. and in part by the EU Horizon 2020 CIMPLEX project under grant agreement No. 641191. The work of D.T. has been funded by the Croatian Science Foundation IP-2013-11-9623, "Machine learning algorithms for insightful analysis of complex data structures". K-K. K. acknowledges support by the ERC Grant "Momentum" (324247). Author contributions All authors contributed to the writing and editing of the manuscript, proofreading and analyses. D.T. contributed to defining, formalizing and generalizing the mapping, relation to bond percolation and applications, K.K.K. contributed to defining temporal maps, time critical vaccination and connection to disordered networks and N.A.F. contributed to probabilistic modelling, source inference, algorithmic contributions and lead the research. Additional information Dataset availability: All data that we used in our analysis are freely available48, 49, 65. Supplementary information accompanies this paper. The authors declare no competing interests. Main code available at: https://github.com/ninoaf/SpreadingMapping References 1. Guille, A., Hacid, H., Favre, C. & Zighed, D. A. Information diffusion in online social networks. ACM SIGMOD Rec. 42, 17 (2013). 2. Lerman, K. & Ghosh, R. Information contagion: an empirical study of the spread of news on digg and twitter social networks. In in Proc. 4th Int. Conf. on Weblogs and Social Media (ICWSM), 2010. 3. Anderson, R. M. & May, R. M. Infectious Diseases in Humans (Oxford University Press, 1992). 4. Vespignani, A. Modelling dynamical processes in complex socio-technical systems. Nat. Phys. 8, 32 -- 39 (2011). 5. Pastor-Satorras, R., Castellano, C., Van Mieghem, P. & Vespignani, A. Epidemic processes in complex networks. Rev. Mod. Phys. 87, 925 -- 979 (2015). 6. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M. & Hwang, D.-U. Complex networks: Structure and dynamics. Phys. Reports 424, 175 -- 308 (2006). 7. Barrat, A., Barth´elemy, M. & Vespignani, A. Dynamical Processes on Complex Networks (Cambridge University Press, 2012). 8. Gillespie, D. T. A rigorous derivation of the chemical master equation. Phys. A: Stat. Mech. its Appl. 188, 404 -- 425 (1992). 9. Van Mieghem, P. & Cator, E. Epidemics in networks with nodal self-infection and the epidemic threshold. Phys. Rev. E 86, 016116 (2012). 10. Van Mieghem, P., Omic, J. & Kooij, R. Virus spread in networks. IEEE/ACM Transactions on Netw. 17, 1 -- 14 (2009). 11. Gleeson, J. P. Binary-state dynamics on complex networks: Pair approximation and beyond. Phys. Rev. X 3, 021004 (2013). 12. Castellano, C. & Pastor-Satorras, R. Thresholds for epidemic spreading in networks. Phys. Rev. Lett. 105, 218701 (2010). 13. Pastor-Satorras, R. & Vespignani, A. Epidemic spreading in scale-free networks. Phys. Rev. Lett. 86, 3200 -- 3203 (2001). 14. Sharkey, K. J. Deterministic epidemic models on contact networks: Correlations and unbiological terms. Theor. Popul. Biol. 79, 115 -- 129 (2011). 15. Sharkey, K. J., Kiss, I. Z., Wilkinson, R. R. & Simon, P. L. Exact equations for sir epidemics on tree graphs. Bull. Math. Biol. 77, 614 -- 645 (2015). 16. Kiss, I. Z., Morris, C. G., S´elley, F., Simon, P. L. & Wilkinson, R. R. Exact deterministic representation of markovian sir epidemics on networks with and without loops. J. Math. Biol. 70, 437 -- 464 (2015). 17. Karrer, B. & Newman, M. E. J. Message passing approach for general epidemic models. Phys. Rev. E 82, 016101 (2010). 9/11 18. Lokhov, A. Y., M´ezard, M., Ohta, H. & Zdeborov´a, L. Inferring the origin of an epidemic with a dynamic message-passing algorithm. Phys. Rev. E 90, 012801 (2014). 19. Altarelli, F., Braunstein, A., Dall'Asta, L., Lage-Castellanos, A. & Zecchina, R. Bayesian inference of epidemics on networks via belief propagation. Phys. Rev. Lett. 112, 118701 (2014). 20. Mollison, D. Spatial contact models for ecological and epidemic spread. J R Stat Soc Ser. B 39, 283 -- 326 (1977). 21. Grassberger, P. On the critical behavior of the general epidemic process and dynamical percolation. Math. Biosci. 63, 157 -- 172 (1983). 22. Meyers, L., Newman, M. & Pourbohloul, B. Predicting epidemics on directed contact networks. J. Theor. Biol. 240, 400 -- 418 (2006). 23. Gillespie, D. T. A general method for numerically simulating the stochastic time evolution of coupled chemical reactions. J. Comput. Phys. 22, 403 -- 434 (1976). 24. Fichthorn, K. A. & Weinberg, W. H. Theoretical foundations of dynamical monte carlo simulations. The J. Chem. Phys. 95, 1090 (1991). 25. Bortz, A., Kalos, M. & Lebowitz, J. A new algorithm for monte carlo simulation of ising spin systems. J. Comput. Phys. 17, 10 -- 18 (1975). 26. Bogun´a, M., Lafuerza, L. F., Toral, R. & Serrano, M. A. Simulating non-markovian stochastic processes. Phys. Rev. E 90, 042108 (2014). 27. Vestergaard, C. L. & G´enois, M. Temporal gillespie algorithm: fast simulation of contagion processes on time-varying networks. PLoS Comput. Biol 11, 1 -- 28 (2015). 28. Antulov-Fantulin, N., Lancic, A., Stefancic, H. & Sikic, M. FastSIR algorithm: A fast algorithm for the simulation of the epidemic spread in large networks by using the susceptible -- infected -- recovered compartment model. Inf. Sci. 239, 226 -- 240 (2013). 29. Newman, M. E. J. Spread of epidemic disease on networks. Phys. Rev. E 66, 016128 (2002). 30. Kenah, E. & Robins, J. M. Second look at the spread of epidemics on networks. Phys. Rev. E 76, 036113 (2007). 31. Chen, Y., L´opez, E., Havlin, S. & Stanley, H. E. Universal behavior of optimal paths in weighted networks with general disorder. Phys. Rev. Lett. 96, 068702 (2006). 32. Braunstein, L. A., Buldyrev, S. V., Cohen, R., Havlin, S. & Stanley, H. E. Optimal paths in disordered complex networks. Phys. Rev. Lett. 91, 168701 (2003). 33. Sreenivasan, S. et al. Effect of disorder strength on optimal paths in complex networks. Phys. Rev. E 70, 046133 (2004). 34. Van Mieghem, P. & van Langen, S. Influence of the link weight structure on the shortest path. Phys. Rev. E 71, 056113 (2005). 35. Van Mieghem, P. & Magdalena, S. M. Phase transition in the link weight structure of networks. Phys. Rev. E 72, 056138 (2005). 36. Antulov-Fantulin, N., Lanci´c, A., Smuc, T., Stefanci´c, H. & Siki´c, M. Identification of patient zero in static and temporal networks: Robustness and limitations. Phys. Rev. Lett. 114, 248701 (2015). 37. Shah, D. & Zaman, T. Detecting sources of computer viruses in networks: theory and experiment. In Proceedings of the ACM SIGMETRICS international conference on Measurement and modeling of computer systems, SIGMETRICS '10, 203 -- 214 (ACM, New York, NY, USA, 2010). 38. Pinto, P. C., Thiran, P. & Vetterli, M. Locating the Source of Diffusion in Large-Scale Networks. Phys. Rev. Lett. 109, 068702+ (2012). 39. Brockmann, D. & Helbing, D. The Hidden Geometry of Complex, Network-Driven Contagion Phenomena. Sci. 342, 1337 -- 1342 (2013). 40. Jalili, M. & Perc, M. Information cascades in complex networks. J. Complex Networks 5, 665 -- 693 (2017). 41. Wang, Z. et al. Statistical physics of vaccination. Phys. Reports 664, 1 -- 113 (2016). 42. Helbing, D. et al. Saving human lives: What complexity science and information systems can contribute. J. Stat. Phys. 158, 735 -- 781 (2015). 43. Devroye, L. Sample-based non-uniform random variate generation. In Proceedings of the 18th Conference on Winter Simulation, WSC '86, 260 -- 265 (ACM, New York, NY, USA, 1986). 10/11 44. Dijkstra, E. W. A note on two problems in connexion with graphs. Numer. Math. 1, 269 -- 271 (1959). 45. Demetrescu, C. & Italiano, G. F. Algorithmic techniques for maintaining shortest routes in dynamic networks. Electron. Notes Theor. Comput. Sci. 171, 3 -- 15 (2007). Proceedings of the Second Workshop on Cryptography for Ad-hoc Networks (WCAN 2006). 46. Demetrescu, C. & Italiano, G. F. A new approach to dynamic all pairs shortest paths. J. ACM 51, 968 -- 992 (2004). 47. Starnini, M., Gleeson, J. P. & Boguna, M. Equivalence between non-markovian and markovian dynamics in epidemic spreading processes. Phys. Rev. Lett. 118, 128301 (2017). 48. Guimer`a, R., Danon, L., D´ıaz-Guilera, A., Giralt, F. & Arenas, A. Self-similar community structure in a network of human interactions. Phys. Rev. E 68, 065103 (2003). 49. Rossi, R. A. & Ahmed, N. K. The network data repository with interactive graph analytics and visualization. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence (2015). 50. Comin, C. H. & da Fontoura Costa, L. Identifying the starting point of a spreading process in complex networks. Phys. Rev. E 84, 056105 (2011). 51. Zhai, X., Wu, W. & Xu, W. Cascade source inference in networks: a markov chain monte carlo approach. Comput. Soc. Networks 2, 17 (2015). 52. Provan, J. S. & Ball, M. O. The complexity of counting cuts and of computing the probability that a graph is connected. SIAM J. on Comput. 12, 777 -- 788 (1983). 53. Marron, J. S. & Nolan, D. Canonical kernels for density estimation. Stat. & Probab. Lett. 7, 195 -- 199 (1988). 54. Valiant, L. G. The complexity of enumeration and reliability problems. SIAM J. on Comput. 8, 410 -- 421 (1979). 55. Kulkarni, V. G. Shortest paths in networks with exponentially distributed arc lengths. Networks 16, 255 -- 274 (1986). 56. Corea, G. A. & Kulkarni, V. G. Shortest paths in stochastic networks with ARC lengths having discrete distributions. Networks 23, 175 -- 183 (1993). 57. Peer, S. & Sharma, D. K. Finding the shortest path in stochastic networks. Comput. & Math. with Appl. 53, 729 -- 740 (2007). 58. Hofstad, R. van der, Hooghiemstra, G. & Van Mieghem, P. Size and weight of shortest path trees with exponential link weights. Comb. Probab. Comput. 15, 903 (2006). 59. Bhamidi, S., Goodman, J., Hofstad, R. van der & Komj´athy, J. Degree distribution of shortest path trees and bias of network sampling algorithms. The Annals Appl. Probab. 25, 1780 -- 1826 (2015). 60. Gomez-Rodriguez, M., Song, L., Du, N., Zha, H. & Scholkopf, B. Influence estimation and maximization in continuous- time diffusion networks. ACM Transactions on Inf. Syst. 34, 1 -- 33 (2016). 61. Gautreau, A., Barrat, A. & Barth´elemy, M. Global disease spread: statistics and estimation of arrival times. J. Theor. Biol. 251, 509 -- 522 (2008). 62. Colizza, V. & Vespignani, A. Epidemic modeling in metapopulation systems with heterogeneous coupling pattern: Theory and simulations. J. Theor. Biol. 251, 450 -- 467 (2008). 63. Lawyer, G. Measuring the potential of individual airports for pandemic spread over the world airline network. BMC Infect. Dis. 16, 70 (2016). 64. Iannelli, F., Koher, A., Brockmann, D., Hovel, P. & Sokolov, I. M. Effective distances for epidemics spreading on complex networks. Phys. Rev. E 95, 012313 (2017). 65. Woolley-Meza, O. et al. Complexity in human transportation networks: a comparative analysis of worldwide air transportation and global cargo-ship movements. The Eur. Phys. J. B - Condens. Matter Complex Syst. 84, 589 -- 600 (2011). 11/11
1611.10094	1	1611	2016-11-30T11:21:41	The influence of the network topology on the agility of a supply chain	[ "cs.SI", "physics.soc-ph" ]	The right performance of a supply chain depends on the pattern of relationships among firms. Although there is not a general consensus among researchers yet, many studies point that scale-free topologies, where few highly related firms are combined with many low-related firms, assure the highest efficiency of a supply chain. This paper studies the network topology that leads to the highest agility of the supply chain when sudden demand changes occur. To do this, an agent-based model of a supply chain with restricted relationship between agents is built. The model includes three tiers, where the flow of material is distributed from the bottom supplier to the final customer passing necessarily through firms in every tier. Agility is measured in the model simulations through the order fulfillment rate. Unlike to previous theoretical and lab results, the simulation of the model shows that the highest levels of agility are not obtained with a scale-free topology. Instead, homogeneous distribution of links, such as those induced by regular or Poisson probability laws, shows higher agility values than heterogeneous distributions. Other previous recommendations, such as redundancy or having multiple suppliers, are confirmed by the simulations. The general conclusion is that the most suitable network topology in terms of agility depends on the specific conditions of the supply chain and the aspects of the performance to be analyzed.	cs.SI	cs	The influence of the network topology on the agility of a supply chain Hernández, Juan M.* Department of Quantitative Methods, Institute of Tourism and Sustainable Economic Development (TIDES), University of Las Palmas de Gran Canaria, c/ Saulo Torón, nº17, 35017 Spain. E-mail: [email protected] Corresponding author Pedroza, Carmen Unidad Académica de Estudios Regionales de la Coordinación de Humanidades de la Universidad Nacional Autónoma de México (UNAM), Sede La Ciénega, Av. Lázaro Cárdenas s/n, esq. Felícitas del Río, Colonia, Centro, C.P. 59510, Jiquilpan de Juárez, Michoacán, Mexico. E- mail: [email protected] Abstract. The right performance of a supply chain depends on the pattern of relationships among firms. Although there is not a general consensus among researchers yet, many studies point that scale-free topologies, where few highly related firms are combined with many low-related firms, assure the highest efficiency of a supply chain. This paper studies the network topology that leads to the highest agility of the supply chain when sudden demand changes occur. To do this, an agent-based model of a supply chain with restricted relationship between agents is built. The model includes three tiers, where the flow of material is distributed from the bottom supplier to the final customer passing necessarily through firms in every tier. Agility is measured in the model simulations through the order fulfillment rate. Unlike to previous theoretical and lab results, the simulation of the model shows that the highest levels of agility are not obtained with a scale-free topology. Instead, homogeneous distribution of links, such as those induced by regular or Poisson probability laws, shows higher agility values than heterogeneous distributions. Other previous recommendations, such as redundancy or having multiple suppliers, are confirmed by the simulations. The general conclusion is that the most suitable network topology in terms of agility depends on the specific conditions of the supply chain and the aspects of the performance to be analyzed. Keywords: topology, agility, supply network, scale-free, homogeneous distribution 1. Introduction Supply chain (SC) has been conceptualized as a network of agents (e.g., suppliers, manufacturers, distributors, retailers and consumers) who are interconnected throughout the transference of material or information (Carter et al., 2015). This perspective departs from the dyadic (buyer-supplier) and triadic view which center on the particular relationships between categories (Choi and Wu, 2009; Cox et al., 2001) to a more extended focus where the object of analysis is the supply chain as a whole, represented by a collection of nodes (agents) and links (transactions among them). Assuming this view, the social network analysis is an adequate tool to analyze this kind of systems, making use of quantitative metrics to identify some key nodes or the structure of the network as a whole (Borgatti and Li, 2009). 1 Following the network perspective, a supply chain is also termed supply network (Choi and Hong, 2002; Kim et al., 2011). A supply network is represented by a complex system of interdependence among multiple firms, each one with a number of links to specific partners. In this context, the supply network performing is not just the result of an additive aggregation of the individual firms' decisions, such as it would be the case in a simple and linear SC. Instead of this, a coherent and autonomous global behavior emerges from the non-linear interaction among agents (Surana et al., 2005; Choi et al., 2001). This behavior is dependent on the particular structure of the total relationships among agents in the supply network, what is called the network topology. Although some scholars have made efforts to distinguish supply chains and supply networks recently (Braziotis et al., 2013), in this paper we use both terms interchangeably. Some SC archetypes or topologies have been identified in the scientific literature (Capaldo and Giannoccaro, 2015a; Hearnshaw and Wilson, 2013; Kim et al., 2015). Among them, structures characterized by a heterogeneous distribution of links, such as the scale-free topology, are among the most frequently analyzed. Induced by power-law degree distribution, scale-free supply networks are characterized by the existence of few large hub firms with many connections combined with many small peripheral firms with few connections. This topology is presented in some real SCs and is seen by some authors as the most efficient topology in SC (Hearnshaw and Wilson, 2013). In particular, it has been theoretically shown that scale-free is the most suitable structure to respond against disturbances and disruptions, what determines a resilient SC (Kim et al., 2015; Nair and Vidal, 2011; Zhao et al., 2011). In this paper, we analyze the influence of the network topology on the agility of the supply chain. SC agility can be defined as the ability to respond quickly to sudden changes in supply and demand and is considered as one of the dimensions of a resilient SC (Christopher and Peck, 2004). Agility is usually analyzed from a focal agent's view, assuming the position of a single company in the network and looking the surrounding supply chain in relation to this company (Lee, 2004). Unlikely, our focus adopts a more extended view and analyses agility of the network as a whole, not from the perspective of a single agent (Choi et al., 2001). In supply networks, agility is influenced by the number of intermediate agents through which the flow of material or information needs to pass between the initial supplier and the final customer. In general, the theoretical analyses of resilience in SC so far assume that a link is possible between any two firms in the supply network, independently on the specific tier where the firms belong to (Nair and Vidal, 2011; Zhao et al., 2011). However, in some specific real supply networks the flow needs to pass through firms which belong to stablished tiers and bypass seldom occurs. This is the case for military logistic networks and many food supply chains. The main hypothesis of this paper is that, in these kinds of supply networks, topologies characterized by heterogeneous distributions of links, such as those induced by the power-law degree distribution, present lower values of agility than others induced by homogeneous distributions of links. Therefore, the paper limits to specific classes of SCs the extended belief that scale-free and derivatives are the most efficient SC topologies. In order to check the hypotheses, we build an agent-based model that represents a supply chain with multiple agents and their rules to allocate orders and supplies. The SC in the model includes three tiers (suppliers, wholesalers and distributors) and the pattern of relationships among agents in different tiers follows some probabilistic distributions. Thus, the theoretical SC is built through the realization of random network models in bipartite graphs (Guillaume and Latapy, 2006). Several patterns of relationships between tiers are considered. Following the common recommendation of using agent-based simulation to analyze large, stochastic and non-centralized supply networks (Datta et al., 2007; Kim, 2009), we simulate a sudden demand change and measure the degree of success in adapting to these changes. The results show 2 that, under certain conditions, the percentage of unfilled orders in the SC when assuming power-law degree distributions is substantially higher in value than when assuming other homogeneous distributions. The paper is organized as follows. Next section presents a literature review, starting from which the hypotheses are stated. Section 3 shows the supply chain random model, order and supplies allocation rules, and the algorithm to measure SC agility. Section 4 presents the results of the simulations. Last two sections are devoted for discussion and conclusions. 2. Literature review 2.1. Characteristics of the food supply chain In this subsection, we review some features of the food supply chain (FSC). One of them is the role of intermediaries, what makes that relationships among agents in a FSC are usually more restrictive than in industrial SCs. This paper analyzes agility in SC with restrictive relationships and FSCs are real examples proximate to these types of SCs. FSCs are characterized by the perishability of their products. Additionally, many of their traded goods come from small-scale producers from rural economies in developing countries. Small- scale producers in the agricultural or fisheries sectors have also in common close relations with middlemen or intermediaries. To be able to sell their products, small-scale rural producers participate in a multi-tiered SC, where middlemen are key actors that connect producers with external markets. More specifically, middlemen act as facilitators of trade between producers and wholesalers. To give some examples, this market structure has been identified in many rural economies in agriculture and fishing, such as the potato market in Ethiopia (Abebe et al., 2016), the rice in Philipines (Hayami et al., 1999), or the fisheries Malaysia (Merlijn, 1989), Kenia (Crona et al., 2010) and Mexico (Pedroza, 2013). Small-scale producers in these economies have low production volumes, limited trading skills and lack of conservation or transportation facilities. Furthermore, the lack of trading skills is usually related to low educational levels and isolated geographic conditions. Under these circumstances, the bulking and perishability of products become the main problems (Abebe et al., 2016; Keys, 2005). These factors do not help the small-scale producers to organize and have fluid horizontal relationships among them, which might allow them achieving local, national or international markets. An empirical evidence of the lack of strong horizontal relationships at producer is the malfunctioning of cooperatives. These producer organizations have served to organize production and obtain subsidies from government and foreign aid agencies, but are ineffective to improve commercialization and competitiveness (Abebe et al., 2016; Bernard and Taffesse, 2008; Hayami et al., 1999; Pedroza, 2013; Ruben and Heras, 2012). in this type of economies levels The reasons why small-scale producers trade their products through middlemen and the influence that these intermediaries have in the FSC are from social and economic nature. Middlemen are a social optimal choice to decrease transactions costs, specifically those related to search and matching costs (Abebe et al., 2016; Arya et al., 2015; Gabre‐Madhin, 2001). Middlemen act as a time saving institution by shortening the negotiation time of sellers and buyers in a transaction (Rubinstein and Wolinsky, 1987). Moreover, they are considered important providers of market information (Abebe et al., 2016), match geographically dispersed buyers and sellers (Gabre‐Madhin, 2001), have better information and in general have the ability to contract separately with upstream and downstream parties and achieve supply and delivery coordination, in sum higher expertise (Arya et al., 2015). Middlemen link 3 actors across different social domains and hierarchical levels (Crona et al., 2010), reducing risk and assuring fast distribution of food. Middlemen operate most of the time informally and through cash transactions. All these abilities are a source of flexibility, which enhances SC agility. Middlemen go to rural areas to look for the product by means of a face-to-face relationship with each producer. Each small producer sells only a small amount, which tends to increase the transaction cost per unit of product (Hayami et al., 1999). Thus, middlemen can assure the demanded supply through its own network of producers. This represents a competitive advantage with respect to wholesalers who buy the product directly to few producers, which normally do not guarantee them sufficient supply (Abebe et al., 2016; Keys, 2005). Middlemen income level and relevance in the FSC depends on its capacity to increase, maintain and manage their collection volume, and this is normally done through their linking abilities. Commonly, middlemen in these markets have links with a large network of producers and a small network of wholesalers. Therefore, middlemen play a key role in FSCs, since they can channel market demands and directly influence on the organization of production (Wilson, 1980). In a supply network where the product passes throughout the middlemen' hands, food will arrive on time and with a good quality to the table of consumers. This is the reason why middlemen are a necessary agent to transport the flow of product in a FSC. 2.2. Agility in a resilient SC Agility has been recognized as one of the key qualities for top-performing supply chains (Lee, 2004). Additionally, it has been also identified as a constituent element of a resilient SC (Christopher and Peck, 2004; Jüttner and Maklan, 2011; Ponomarov and Holcomb, 2009; Wieland and Wallenburg, 2013). In this context, the concept of SC resilience embraces the one of agility, which is considered a dimension of it. The definition and practice of supply chain resilience has been largely debated in the scientific literature since the beginning of the century. In general, resilience alludes to the SC capability to respond to unexpected events. Kamalahmadi and Parast (2016), based on a review of the different definitions given in the literature, propose the following extended definition: "The adaptive capability of a supply chain to reduce the probability of facing sudden disturbances, resist the spread of disturbances by maintaining control over structures and functions, and recover and respond by immediate and effective reactive plans to transcend the disturbance and restore the supply chain to a robust state of operations". This definition integrates agility in a resilient SC by mentioning explicitly the need of fast response to disturbances. The different conceptualizations given since the beginning of the century have generally considered agility as a constituent (or formative) element of the SC resilience. Christopher and Peck (2004) considers it as a principle of resilience in combination with other three: supply chain re-engineering or the SC design, collaboration among entities in the SC and supply chain risk management culture, the last one favored by leadership and innovation. Other authors propose new conceptual frameworks to the question of resilience, but maintain a principal role for agility. For example, Pettit et al. (2010) enumerate 14 capabilities, divided in 111 subfactors, to counteract SC vulnerabilities, where agility is not explicitly included but split in several capability factors. Wieland and Wallenburg (2013) assume resilience divided between proactive (agility) and reactive (robustness) responses to disturbances. In empirical studies, agility was found as the major enabler of resilience by a sample of firm managers in Indian firms, followed by collaboration and visibility (Soni et al., 2014). 4 Thus, the definition and role of agility is repeated with few variations in most of the conceptualizations of resilience in the literature. Here we follow Christopher and Peck (2004), which defines agility as the "ability to respond rapidly to unpredictable changes in demand or supply". It is generally accepted that agility includes two dimensions: (a) visibility, or the "knowledge of the status of operating assets and the environment" (Pettit et al., 2010), which refers to the ability to see all entities, production and distribution capacities, together to other restrictions down and upstream in the SC, in order to avoid ineffective actions when a disturbance occurs, and (b) velocity, defined as the "the total time it takes to move product and materials from one end of the supply to the other" (Christopher and Peck, 2004), refers to the speed of the response when the change in demand and supply is produced. Some authors also include flexibility as other ingredient of agility (Scholten et al., 2014; Jüttner and Maklan, 2011). We adopt the definitions of flexibility given by Pettit et al. (2010) as "ability to quickly change inputs or the mode of receiving inputs" (flexibility in sourcing) and the "ability to quickly change outputs or the mode of delivering outputs" (flexibility in order fulfillment). According to Sheffi and Rice (2005), flexibility can be achieved by working with multiple suppliers, having multiple capabilities at each plant location and a "coherent process for setting priorities during the time-sensitive postdisruption period". These strategies are related to the so-called redundancy, which refers to the existence of extra stock in case of disturbances (high inventory levels) or having multiple suppliers as well. Redundancy can be effective to get an agile response to disturbances although it uses to be costly (Sheffi and Rice, 2005). So, agility is a multidimensional concept and influenced by multiple factors. Specifically, collaboration among the different entities in the SC exerts a positive influence on agility. The collaborative partners share information or material and "demonstrates a sense of responsibility towards its supply chain" (Wieland and Wallenburg, 2013). Thus, they can give a faster and more efficient response to disturbances by sharing resources. Collaboration can be vertical, if produced among agents in different tiers of the supply chain, or horizontal, if produced among entities in the same tier. Several empirical tests have proven the essential role of horizontal and vertical collaboration in SCs for an efficient disaster management process (Scholten et al., 2014). Additionally, the positive influence of collaboration on agility has been empirically tested by some recent studies (Scholten and Schilder, 2015; Wieland and Wallenburg, 2013). In this paper we restrict the analysis to the relationship between the supply network topology and agility, leaving aside other factors that influence on it. We also depart from the focal agent's view and analyze agility of the supply network as a whole, not from the perspective of a single agent. In this context, we question the conditions for the supply network to be agile. 2.3. The influence of topology on SC performance As commented in the introduction, we follow the network perspective to analyze SC agility (Borgatti and Li, 2009). Generally speaking, a network is a set of vertices (nodes), edges (links) and the way both elements are connected. In the context of SC, nodes are represented by firms, while edges represent relationships among those agents. This relationship can be formalized through the existence of some flow of material, information or contract among the parts (Beamon, 1998; Kim et al., 2011). For simplicity, in this paper we assume the same network structure for the three types of relationship. So, a link between firms indicates flow of material, information or the presence of a formal (or informal) agreement among them. Supply networks have been also identified as complex adaptive systems (CAS) (Choi et al., 2001; Pathak et al., 2007; Surana et al., 2005). One of the most prominent characteristics of 5 CAS is the presence of self-organization. This phenomenon occurs when a global functioning of the network emerges from the individual decisions of the partners which are combined in a non-linear and decentralized way. In a supply chain like this, the network behaves coherently from the aggregation of every agent's decisions, where everyone influences on the functioning of the network but not anyone controls the global behavior. In brief words, this can be summarized by the sentence "the whole is more than the sum of the parts". The self-organization phenomenon in real networks has been also observed in other fields of science, such as Physics or Biology, and has given origin to the complex network theory (Albert and Barabasi, 2002; M. E. J. Newman, 2003). This theory analyses the global behavior of high- scale networks, such as is the case of Internet, author's citation and on-line recommendations, starting from simple rules of relationships among nodes. Some interdependence patterns or topologies of the network have been identified and their influence on action performance or dynamical models in the network, such as removing nodes or contagion diffusion, has been analyzed (Newman, 2010). In real social networks, three common topological features have been regularly observed (Newman and Watts, 2002): (a) Evidences of "Small World"; (b) High clustering; and (c) Power- law degree distribution. Below we detail an explanation of these evidences and the corresponding facts in SCs. Larger formalization can be found in the review by Newman (2003). The concept of "Small World" property has become popular from the experiment by Travers and Milgram (1969), who found that two strangers can be connected throughout few paths of acquaintances. In social network words, "Small World" indicates that the network diameter, defined as the maximum of the shortest path length among any two nodes in the network, is low as compared with the network size. In a SC context, this means that, although the supply chain includes a lot of firms and relationships, the flow of material or information passes through few firms from the initial supplier to the final customer. Clustering in a network refers to the probability that a triadic relationship forms a closed cycle. In terms of a friendship relationship, this means how often two persons are friend if both are friends of a third one. Hence, the clustering coefficient of the network is the percentage of cycles in all the possible triadic relationships in a network. As expected, high clustering coefficient has been observed in many real social networks (Newman, 2003). In the supply chain context, triads can be formed by buyer-supplier-supplier relationships (Choi and Wu, 2009), but also by buyer-buyer-supplier and three firms in the same tier. High clustering has been observed in some real SCs (Hearnshaw and Wilson, 2013). In network theory, the number of links from a node is called node degree. The network degree distribution (𝑝(𝑘)) indicates the proportion of nodes that have a k degree. In other words, the degree distribution indicates the probability that an arbitrary node includes k edges. Many real networks follow a power-law distribution, 𝑝(𝑘)~𝑘−𝛾, with  a parameter usually between 1 and 3. One of the most outstanding characteristic of the power-law is the existence of fat tails in the probability distribution graph. In other words, in these networks there is few but significant amount of highly connected nodes together with the rest of low connected ones. This phenomenon leads to some relevant properties of the network, such as high diffusion rates of epidemics or attack resilience (Newman, 2010). Some variations of power-law distributions, which also include high clustering, have also been introduced to represent real networks (Dorogovtsev and Mendes, 2013). Empirical analyses have found that some real SCs resemble this topological archetype. This is the case for some automotive SCs, which show a highly centralized network structure (Kim et al., 2011), and the supplier-customer network in the Indian auto-component industry (Parhi, 6 2005). The structure of relationships in these examples is represented by a heterogeneous degree distribution of links, more specifically a power-law degree distribution. This organization has also been identified in certain industrial districts, dominated by few leading firms which outsource production among a big number of smaller firms. This organization favors the adaptation to market changes from the leading to the subcontracted firms helped by production redundancy, proximity and long-term transactional intensity (Lazerson and Lorenzoni, 1999). Nevertheless, other network topologies have been described in industrial districts as well: 1) Many small firms buying and selling material to each other; 2) Satellite platforms, formed by branches of large and external headquarter firms with few trade among them (Markusen, 1996). These patterns may be represented by a homogeneous degree distribution network with high or low clustering coefficient, respectively. Likewise, other possible interaction patterns in real SCs have been studied recently (Capaldo and Giannoccaro, 2015a; Kim et al., 2015). Some of them have been identified in other fields, such as Poisson random networks (Newman, 2003), while other not so common topologies have been described as well. For example, the block-diagonal pattern is characterized by the existence of blocks of firms highly interconnected but with few connections among blocks. Other topologies are hierarchical, where a dominant firm includes many suppliers, or dependent, where few wholesalers include many independent retailers. 2.4. Network topology of a resilient SC The interrelationship pattern of firms influences on the SC resilience. In particular, SC understanding is enhanced through the knowledge of the SC topology, which allows the identification of bottlenecks and critical paths in the supply network. Additionally, diversification of suppliers, which has been recommended as a measure to reduce disruption risks in SC (Sheffi and Rice, 2005), determines the network structure in the SC and their benefits can be also analyzed using topological tools. Starting from findings in the previous section, some theoretical contributions hypothesizes that scale-free topologies characterize the most agile and resilient SC (Hearnshaw and Wilson, 2013). Similarly to the case of the industrial districts, SC resilience is achieved by the coordination role of the leading firm, which enhances communication and organization among suppliers, consequently improving the fast adaptation to changes. Leadership is also necessary to promote risk management culture, which increases resilience in SC (Christopher and Peck, 2004). Thus, hub firms promote horizontal links among suppliers, increasing clustering in the supply network. Therefore, high clustering is also a characteristic of a resilient SC, since it facilitates coordination among firms and thereby agility. Some simulation models have also analyzed the SC resilience obtained with different topologies. For example, Thadakamalla et al., (2004) and Zhao et al. (2011) simulate several random network topologies in a three-tiered SC (supplier, manufacturer and retailer). In these papers, agility is measured through the average supply-path length, which is part of a collection of resilient indicators. The results find that heterogeneous structures, such as those induced by scale-free topology and variations, show the highest values of resilience when a random disruption occurs, but exhibit fragility when targeted attacks are produced. In general, these results agree with previous findings in a more general context (Newman, 2010). Additionally, Nair and Vidal (2011) found that, in terms of inventory levels, scale-free topology performs better against disruptions than other homogeneous topologies but worst from the backorders and total cost perspective. They achieve the results by means of an agent-based 7 model in a simple and reduced SC (only 35 nodes). Recently, Kim et al. (2015) analyze the resilience, measured as the probability of SC disruption in case of a node or link disruption, for a simulated supply network with several topologies, showing that the best figures are obtained with a power-law. Some other simulation models have tested the SC performance with several topologies for other characteristics than resilience. Thus, Li et al. (2013) also found that heterogeneous structures also induce cooperation among firms in a greater extent than homogeneous structures. Additionally, Capaldo and Giannoccaro (2015b) show that a dependence pattern, followed by power-law, obtains the highest level of trust among partners. 2.5. Agility in a SC with restricted relationships In general, the results above point to the prevalence of scale-free structures and derivatives as the most convenient topologies for a resilient and, in particular, agile SC. This is so for the types of SCs considered in the previous contributions. In general, the systems analyzed above are complex supply networks, where links between firms in any tier are possible. This conditions adapt to some real industrial SC (Kim et al., 2011), but not the FSC illustrated in section 2.1. In this kind of SC, a restricted relationship between agents exists, where every agent is supplied by other agents in the precedent tier and supply to agents in the subsequent tier. For example, a wholesaler is supplied exclusively by intermediaries, while an intermediary is only supplied by producers. These restrictions determine a specific supply network structure, which may influence on the SC agility. We call this structure restricted relationship. It is similar to the diagonal structure, where every firm is supplied from firms in upstream tiers (Capaldo and Giannoccaro, 2015a; Kim et al., 2015), but adding that firms are supplied only by firms in the precedent tier. Figure 1 illustrates two examples of this supply network structure in a three-tiered SC. Figure 1. Two examples of a three-tiered supply network with different topologies in the wholesalers and retailers: (a) Homogeneous topology (regular degree distribution with mean degree 𝑘 = 2); (b) Heterogeneous topology (power-law degree distribution with 𝑘 = 2). Thus, the type of SCs analyzed here does not reply those in previous papers, for which more flexible relationships among firms in the supply network are allowed. In a restricted relationship structure, the pattern of links from firms in a tier with firms in the precedent or subsequent tier determines the SC topology. For example, Figure 1a shows a realization of a 8 regular degree distribution of links between supplier-wholesaler and wholesaler-retailer. This distribution assumes a constant degree (𝑘) for any firm in the SC. Figure 1b shows a realization for a scale-free degree distribution of links between the same tiers. The details for the construction of these SCs are given in the next section. In the case of SCs with restricted relationship, scale-free degree distribution may not be the most suitable pattern to get the highest levels of agility. As shown in Figure 1b, the SC topology generated by this distribution presents few big firms that have many relationships (wholesaler 1 and 2) combined with small firms having few connections. Large differences between the number of suppliers and buyers of every firm are more likely presented with heterogeneous than with homogeneous distributions of links. For example, in Figure 1a all wholesalers and retailers have the same number of buyers and sellers, while in Figure 1b, wholesaler 1 has only one seller and three buyers. Consequently, assuming the absence of horizontal links among firms in the same tier, when sudden changes of demand occur, some sellers in a heterogeneous topology may be overloaded of demand (e.g. wholesaler 1 in Figure 1b) while other bigger sellers can fulfill their orders making use of part of their capacity (e.g. wholesaler 2). Therefore, the total orders in the SC would not be fulfilled quickly although it had enough goods from the suppliers. This uneven response to demand is avoided with a homogeneous distribution of links among firms, such as the one illustrated in Figure 1a. In this case, the gap between the number of buyers and sellers in every firm is low or null, so the total unfulfilled orders could be reduced. From these arguments, we state the following hypothesis: H1: In case of capacity constraints and absence of horizontal relationships, agility in a SC with restricted than with heterogeneous-degree topologies, such as those induced by the power-law distribution. is higher with homogeneous-degree relationship topologies Horizontal relationships in the SC imply collaboration among competitors. This practice has been observed in food processing companies and is justified by the incentive of receiving back the help in times of crisis (Scholten and Schilder, 2015). The collaboration between competitors implies the transaction of information and material between partners when needed and enhances flexibility, since it assures new material or service to customers. From the topological point of view, agility is related to path length in a supply network. In SCs with restricted relationship, N tiers and non-horizontal relationships, the path length from any supplier connected to the final demand is constant and equal to N-1, the number of links from the initial supplier to the final customer. If horizontal relationships exist, firms can buy and sell to other firms in the same tier when needed. Thus, other new alternative paths for the flow of material from the bottom supplier to the final customer are included, which help to a quick response when sudden increases in demand happen. We posit the second hypothesis of this paper: H2: The larger the number of horizontal relationships among firms in the same tier, the more agile a SC with restricted relationship is. 2.6. Topological metrics of SC agility Although strongly recommended by researchers (Borgatti and Li, 2009), the use of topological metrics to analyze supply networks is quite limited by now. These metrics are based on graph theory and describe the characteristics of a network exclusively by the disposition of nodes and links in the supply chain. One of the relevant contributions was made by Kim et al. (2011), who apply node-level and network-level metrics to three real automotive supply networks, 9 finding some results who complemented those obtained by Choi and Hong (2002) using a qualitative approach. Among the network-level metrics, the authors use supply network centralization, which measures how a supply network is centralized around a focal firm, and network complexity, which combines the observation of the number of nodes and links. Previously, Falasca et al. (2008) propose the use of spatial density, complexity and node criticality as determinants of SC resilience. To date, a specific topological metric for agility in SC has not been proposed yet. In general, it uses to be embedded in a mix of metrics of resilience in a supply network. For example, Zhao et al. (2011) proposes three indicators of a resilient SC: high percentage of retailers having access to previous tiers, large size of the connected supply network and low average supply path length between any pair of supplier and retailer nodes. Other authors (Mari et al., 2015; Nair and Vidal, 2011) also include high clustering as a characteristic of a resilient SC. Among these topological metrics, agility may be specifically represented by the average supply path length. This measure indicates the average number of agents that the flow of material has to pass through before achieving the final user. Since every firm that manages the product includes some delays associated with receiving, processing and delivering, a low average supply path length means that the material passes through few hands, therefore reducing the total delay and increasing agility. The metrics above do not take into account the positive effect of having alternative paths when needed. Additionally, they are determined exclusively by the static topological structure of the SC and do not consider the operational characteristics of the network. The SC is conceived to transport some flow of material or information from the first tier to the final assembler or consumer. This transportation takes also some time, which is essential for analyzing the response of the SC when a disruption occurs. So, the resilience metrics in general and agility metrics in particular should include the time factor in their specifications. In this line, the percentage of satisfied demand over the total demand in a certain period of time, also called order fulfillment rate (OFR), include these aspects. This metric has been traditionally used as a quantitative SC performance measure (Beamon, 1998; Levy, 1995). More recently, some authors propose OFR as part of other resilience measures of SC as well (Barroso et al., 2015; Datta et al., 2007). In particular, Barroso et al. (2015) uses the concept of resilience triangle, which can be defined as the performance trajectory along time since a disruption occurs until the SC achieves the normal performance levels (Tierney and Bruneau, 2007). Other metrics based on the resilience triangle can be found in the paper of (Carvalho et al., 2012), which propose as part of the resilience metric the ratio of the actual and the promised lead time to deliver an order. Alternatively, Plagányi et al. (2014) proposes the Supply Chain Index (SCI), a quantitative metric to identify resilience and connectedness of a SC. The novelty of this index is that it includes the flow of material throughout the SC in the definition. Low values of SCI indicate that the flow of material is diffused among multiple agents, while large values imply high concentration of flow in few nodes. The SCI was tested by several simple examples of linear SCs. The authors find that the most favorable values of SCI in terms of agility depend on the form of adaptation cost of changes. In general, diffuse flows favor quick adaptation to changes if the cost of changing flows from one agent to another is not scale dependent. This paper adopts the operational perspective of the previous papers and measures agility through the OFR. Instead of analyzing the response of the SC when any class of disruption occurs, only sudden shocks in demand are assumed. Thus, a specific metric for agility is obtained. It is assumed that the lowest suppliers produce a certain amount of material in a specific period of time, which is transported downstream until the final customer. For 10 simplicity, the simulation model does not include delivery times and inventory in every firm. Instead, flow of material throughout the SC is calculated through the demand from the final customer and the processing capacity of every firm. 3. A model to test supply chain agility 3.1. Supply chain random network In this section, we present a stylized model for a supply chain, called supply chain random network (SCRN), where the response to demand shocks will be simulated. Figure 2 shows the general representation of the model. According to a topological viewpoint (e.g., Li et al., 2013; Tang et al., 2016), a supply chain is a directed random graph G(V,E), where V indicates the vector of tiers and E the set of links among nodes (firms) in the graph. In general, a supply chain can include multiple tiers, but for the sake of simplicity we limit to three in this model. So, 𝑉 = (𝑇𝑠, 𝑇𝑤, 𝑇𝑟), where 𝑇ℎ = {ℎ/ℎ = 1,2, … , 𝑁ℎ}, with h={s,w,r}, are the sets of nodes in the tier corresponding to supplier, wholesaler and retailer, respectively. These sets do not include any node in common and the number of nodes in every tier is not necessarily the same. We denote 𝛼: 1: 𝛽, with α, β positive scalars, the ratio of the number of nodes between subsequent tiers (e.g., if 𝑁𝑤 = 100, notation 3:1:5 indicates that the number of suppliers, wholesalers and retailers are 300, 100 and 500, respectively). Figure 2. Representation of the supply chain random network with three tiers: Suppliers (𝑇𝑠), Wholesalers (𝑇𝑤) and Retailers (𝑇𝑟). 𝑁𝑠, 𝑁𝑤 and 𝑁𝑟 represent the number of suppliers, wholesalers and retailers, respectively. Arrows indicate flow and direction of material between firms. The bi-directional arrow indicates horizontal relationship. The links included in E indicate the existence of a buyer-seller relationship between two firms, so a flow of some material (product) is transmitted between them when necessary and available. By assumption, links only connect nodes between subsequent tiers 𝑇𝑠 → 𝑇𝑤 → 𝑇𝑟, 11 so a direct nexus 𝑇𝑠 → 𝑇𝑟 is not considered. Therefore, E includes pairs (𝑠, 𝑤) or (𝑤, 𝑟) if node s transmits the product to w, and node w transmits product to r, with s, w, r in 𝑇𝑠, 𝑇𝑤, 𝑇𝑟, respectively. Additionally, links between nodes in the same tier (horizontal relationships) are allowed only for wholesalers. The horizontal relationships are bi-directional, so the two firms in 𝑇𝑤 that are connected (2 and 3 in Figure 2) can send material to each other in order to satisfy their corresponding demands. In set-builder notation, this means that E also includes pairs (𝑤, 𝑤′) and (𝑤′, 𝑤), with w, w'=1,2,…,𝑁𝑤. For simplicity, the simulation model does not include delivery times and inventory levels in every firm. The analysis of agility is restricted to the SC's capability to fulfill instantaneously the orders coming from retailers. Every node in 𝑇𝑠 and 𝑇𝑤 includes a weight (capacity), which indicates the maximum volume of product that the node can produce or distribute, respectively, in a certain unit of time. For simplicity, we assume that every producer has identical production capacity, cs. By default, we also assume that every wholesaler has the ability to distribute the volume of product coming from his/her relationships with producers. So, the wholesaler's capacity is equal to the number of producers he/she has relations with (i.e. the wholesaler's in-degree) times the producer's capacity, cs. Thus, in the example shown in Figure 2, wholesaler 1's capacity is 2cs, since it is linked with two suppliers, while wholesaler 𝑁𝑤's capacity is cs, since it is linked to only one supplier. 𝑖𝑛 links from the supplier and 𝑘𝑤 𝑜𝑢𝑡 links to the wholesalers and every retailer r has 𝑘𝑟 Finally, the topology of the supply chain is a combination of two bipartite random graphs (Newman and Watts, 2002). They are the supplier/wholesaler and wholesaler/retailer random 𝑖𝑛 links from graphs. Every supplier s has 𝑘𝑠 𝑜𝑢𝑡 links to the the wholesalers. Every wholesaler w has 𝑘𝑤 retailers. We assume that the degree distribution among nodes in every tier (two in-degrees and two out-degrees) follows some prescribed probability distributions. A sample of the random graph is obtained by following the same procedure used for the configuration model (Guillaume and Latapy, 2006): First, we generate the in- and out-degree of every node by simulating the specific distribution and every node is assigned a number of "stubs" equal to its degree; Second, we connect the stubs between subsequent tiers randomly, creating links among the nodes. In order to assure consistence (the sums of the degrees in subsequent tiers must be identical), the mean of the distribution times the number of vertices in subsequent tiers must be identical. Nevertheless, the specific sample may be inconsistent yet. In this case, one node from each one of the two tiers is dropped and their degrees are redrawn until the sample is consistent. 3.2. Model of product distribution We assume a determined system of product distribution and order allocation in the SCRN. This procedure will allow obtaining a measure of the OFR, which depends exclusively on every node's capacity and the network topology. Previously, the total demand in the system is assumed to be identical to the total production capacity of suppliers. Therefore, the material demanded coincides with the disposable supply and, in the naïve supply network with just one supplier, wholesaler and retailer, all demand would be satisfied. However, this may be not the case when several firms operate in every tier, such as it is the case in the SCRN. First, we assign the order allocation rule. It is assumed that every retailer orders one unit of material for every link that the retailer maintains with the upstream tier. In other words, this means that the retailer decides the number of relationships with wholesalers according his/her 𝑖𝑛 units, with r=1,2,…,𝑁𝑟, and necessities. In mathematical terms, the retailer r orders 𝐷𝑟 = 𝑘𝑟 𝑖𝑛 the mean value of the in-degree the total demand from 𝑇𝑟 is 𝐷 = ∑ 𝑖𝑛𝑁𝑟, with 𝑘𝑟 𝑁𝑟 𝑟=1 𝑖𝑛 𝑘𝑟 ≈ 𝑘𝑟 12 distribution 𝑘𝑟 supply is equal to total demand, so 𝑖𝑛. Each supplier's capacity, cs, is determined to assure that, in mean terms, total 𝑐𝑠𝑁𝑠 = 𝑘𝑟 𝑖𝑛𝑁𝑟. (1) Since 𝑁𝑠 = 𝛼𝑁𝑤 and 𝑁𝑟 = 𝛽𝑁𝑤, we have that 𝑐𝑠 = 𝛽 𝛼⁄ 𝑘𝑟 𝑖𝑛. Second, we define the supply allocation rule followed by wholesalers. Let us denote e the accumulated unfulfilled orders, which is initially set as e=0. Starting from the retailer r=1, let ′ is other 𝑤𝑟 be the wholesaler linked to r with the lowest position in 𝑇𝑤. In other words, if 𝑤𝑟 ′. As stated in the previous subsection, 𝑤𝑟's initial capacity wholesaler linked to r, then 𝑤𝑟 < 𝑤𝑟 𝑖𝑛 𝑐𝑠. Starting from (𝑐𝑤𝑟) is equal to 𝑤𝑟's in-degree times each supplier's capacity, so 𝑐𝑤𝑟 = 𝑘𝑤𝑟 𝑐𝑤𝑟 and 𝐷𝑟, two cases are possible: 1. 𝑐𝑤𝑟 ≥ 𝐷𝑟. This means that the wholesaler's capacity exceeds the r's demand, so the quantity 𝐷𝑟 is moved from 𝑤𝑟 to r and 𝑤𝑟's capacity reduces to 𝑐𝑤𝑟 − 𝐷𝑟. 2. 𝑐𝑤𝑟 < 𝐷𝑟. In this case, all the material distributed by 𝑤𝑟 is sold to r. So, an amount 𝑐𝑤𝑟 is ′ (in moved from 𝑤𝑟 to r and 𝑤𝑟's capacity transforms to 𝑐𝑤𝑟 = 0. Then the following 𝑤𝑟 numerical order) that links r is considered and the same procedure is applied previously re- ′ = 𝐷𝑟 − 𝑐𝑤𝑟. The process is repeated recursively until achieving case 1 assigning r's demand 𝐷𝑟 ′⋯′ for the largest w that links r. In the latter case, a quantity 𝑐𝑤 is moved from w to or 𝑐𝑤 < 𝐷𝑟 r and the unfulfilled orders e increases in 𝐷𝑟 ′⋯′ − 𝑐𝑤. Once the orders coming from r are completely managed, the process is repeated for retailer r+1's and so on until reaching r=𝑁𝑟. The model resembles the real transmission of material in a supply network, although it includes several strong restrictions. First, identical volume of transactions is assumed for every buyer-supplier relationship. Second, the supply allocation rule assumes a kind of hierarchy among retailers, since any retailer r has preference to be satisfied over all r'>r. Third, any wholesaler must satisfy all the demand from a particular retailer before attending the next retailer in the hierarchy. Other SC simulation models have included alternative order and shipment allocation rules, such as maintaining inventory levels (Datta et al., 2007; Levy, 1995) or assigning preference based on backorders and trust measures (Kim, 2009). These factors are not considered in the model. Any kind of strategic behavior of agents is not included either. The algorithm initiates with r=1 and covers all retailers subsequently. The final value of e indicates the amount of unfulfilled orders. The ratio OFR=1-e/D gives the percentage of fulfilled orders from the total demand or order fulfillment rate. The larger the value of OFR in [0,1], the higher the agility of the supply network. 4. Simulation setup and results In order to test the influence of the network topology on the SC agility, we simulate the performance of the model above considering several degree distributions of firms in a tier. These degree distributions are the result of the specific conditions in the market and each individual's decisions to have relations with some of the downstream or upstream partners. Up to ten possible supply network structures have been recently identified in the literature (Capaldo and Giannoccaro, 2015a), which have been reduced to four by Kim et al. (2015). Nevertheless, since our main aim is to check whether heterogeneous degree distributions, such as power-law, favor or hinder SC agility to a larger extent than other homogeneous 13 degree distributions, we will limit the analysis to some stochastic distributions which represent the alternatives we look for comparison. Thus, three discrete degree distributions have been selected in the simulations: Regular, Poisson and power-law. A regular degree distribution assumes that every firm in a tier has exactly the same number of relationships with other downstream or upstream firms. A Poisson degree distribution allows the firms to have different number of relationships, but around a mean value. Assuming these two degree distributions, the characteristics of the firms in terms of flow of material and transactions are similar and there is not a leading position of a firm with respect to another. Both are homogeneous distributions, since firms in this kind of SC have proximately similar position in terms of number of relationships. On the contrary, a power-law distribution assumes sharp differences in firms' position, since it includes a short but significant number of firms accumulating many of the total relationships, while the rest has few links with the other tier. All distributions in the simulations are zero-truncated, so every node in the network has at least one link with a firm in the upper/lower tier. By means of this restriction, we set aside the case of suppliers or retailers which are not integrated in the supply chain or wholesalers without relationships with either producers or retailers. In order to compare coherently the performance of the sampled supply networks, identical mean value is assigned for the three distributions. Based on a recent estimation of a real-world food supply chain, in particular the seafood supply chain in Guadalajara, Mexico (Pedroza and Hernández, 2016), we have considered a data set of 1300 firms distributed according the ratio 2:1:10, so 200 suppliers, 100 wholesalers and 1000 retailers. The SCRN and the algorithm to calculate OFR was implemented in Matlab®. 𝑖𝑛 units of material. As described in section 3, we assume that every retailer r demands 𝑘𝑟 Therefore, the total demand is approximately 𝐷 ≈ 1000𝑘𝑟 𝑖𝑛. Using the relationship described in (1), each supplier produces 𝑐𝑠 = 5𝑘𝑟 𝑖𝑛 units. We consider that wholesalers have the same behavior in their relations with suppliers and retailers, so the wholesaler's in- and out-degree distributions are assumed identical. Given these conditions, the possible combination of the three distributions among three tiers is still 33=27. In order to simplify the results, we also assume that every supplier trades with only one wholesaler, so the out-degree distribution 𝑜𝑢𝑡 is regular with mean degree equal 1. By doing this, we limit the scope of the analysis to 𝑘𝑠 the comparison of SC agility obtained with three topological structures in two tiers (wholesalers and retailers), which are 32=9. Table 1 presents the results of the OFR for the three degree distributions included in the analysis. Several values of mean in-degree for retailers 𝑘𝑟 𝑖𝑛 is assumed. As it is shown in the four cases presented, in general the highest agility of the supply network is achieved when degree distributions are homogeneous (regular or Poisson). Lower agility results are obtained when assuming power-law distributions. Nevertheless, the figures also show that SC agility depends on the combination of specific probability laws followed in one or another tier. Thus, given that retailers adopt a heterogeneous distribution of links, the SC agility heavily decreases if the wholesalers follow regular or Poisson distributions, while recovers if they also follow a power-law distribution. However, this is not so if a heterogeneous distribution of links in wholesalers is combined with a homogeneous distribution at the level of retailers. In this case, the agility indicators are not far from those achieved with homogeneous distributions in both tiers. 14 Table 1. Order fulfillment rate (OFR) in the supply chain random network (SCRN). Every supplier includes only one 𝑜𝑢𝑡) link with wholesalers. Three discrete distribution functions for the wholesaler's in- and out-degree (𝑘𝑤 𝑖𝑛) are considered: Regular (Reg.), zero-truncated Poisson (Poiss.) and zero-truncated and the retailers' in-degree (𝑘𝑟 power-law (Pow.). The ratio among the number of firms in tiers is 2:1:10, the number of wholesalers is 𝑁𝑤 = 100 and no horizontal relationships among wholesalers exist. The cases are, from the left to the right: (a) The mean in- degree of retailers is 𝑘𝑟 𝑖𝑛 = 2 and relationships among wholesalers and retailers are randomly assigned, not having into account in and out-degrees; (b) 𝑘𝑟 𝑖𝑛 = 2 and relationships among wholesalers and retailers are ordered according to their degrees, so the retailer with the highest number of relationships trade with the wholesaler with the highest out-degree, and so on; (c) Same as (a) with 𝑘𝑟 𝑖𝑛 = 4; (d) Same as (a) with 𝑘𝑟 𝑖𝑛 and 𝑘𝑤 𝑖𝑛 = 8. 𝑖𝑛 = 2 𝑖𝑛 = 4 𝑖𝑛 = 8 𝑖𝑛 𝑘𝑟 (a) 𝑘𝑟 Reg. 0,90* 0,90 0,87 𝑜𝑢𝑡 * 𝑘𝑤 Reg. Poiss. Pow. Identical probability distribution law for the wholesaler's in- and out-degree is assumed. To assure consistency, the mean degree 𝑘𝑤 𝑜𝑢𝑡 is ordered, so the wholesaler with the highest in-degree has also the highest out-degree, and so on. All distributions are zero-truncated. 𝑖𝑛 = 2 (ordered) Pow. 0,53 0,57 0,72 (b) 𝑘𝑟 Reg. 0,90 0,95 1 (d) 𝑘𝑟 Reg. 0,96 0,96 0,94 (c) 𝑘𝑟 Reg. 0,93 0,93 0,90 𝑜𝑢𝑡 = 2. The sample of 𝑘𝑤 Poiss. 0,93 0,93 0,92 Poiss. 0,85 0,91 0,99 Poiss. 0,89 0,89 0,87 𝑜𝑢𝑡 = 𝛽𝑘𝑟 𝑖𝑛 = 10𝑘𝑟 𝑖𝑛 and 𝑘𝑤 𝑖𝑛 = 𝛼𝑘𝑠 Pow. 0,53 0,56 0,84 Poiss. 0,85 0,85 0,83 Pow. 0,52 0,56 0,71 Pow. 0,53 0,55 0,92 𝑖𝑛 and 𝑘𝑤 The figures are obtained by doing 1000 simulations of the SCRN and taking mean values. In order to avoid high differences between total demand and supply in the simulations, samples with a gap between the simulated and theoretical mean degree higher than 5% are not considered. Therefore, given the conditions of the SCRN, the power-law distributions are not the most efficient degree distribution in the supply chain in terms of fast response of supply changes. This is particularly marked when the bottom tier (retailers) follows the power-law distribution of links. These results confirm H1 above and simulations go in the line of reasons given when stating the hypothesis: The uneven demand from retailers lead that some retailers consume a big part or the total capacity of many small wholesalers, implying that some of the other retailers also linked to those wholesalers cannot satisfy their necessities. This effect is not so marked if there are not such big leading retailers, such as it is the case when retailers follow homogeneous degree distributions. The depletion of many wholesalers' capacity by the leading retailers could be partially avoided if these retailers trade also with leading wholesalers. This possibility is captured in Table1b, which assumes a positive correlation of the wholesaler and retailers' degree, what is also called positive assortativity. This condition means that retailers with high demand tend to relate with leading wholesalers with high number of relationships. The results for agility are something higher but do not differ very much from those obtained in the non-correlated case (Table 1a). The effect of positive assortativity between wholesalers and retailers is only significant for the SC agility when wholesalers follow a power-law degree distribution and retailers adopt homogenous degree distributions. Tables 1c and 1d show the results of OFR assuming higher mean in-degree of retailers (𝑘𝑟 𝑖𝑛 = 4 and 𝑘𝑟 𝑖𝑛 = 8, respectively). In order to assure consistency, the mean out-degree of wholesalers and production of suppliers increase consequently. By increasing the number of relationships among wholesalers and retailers, we analyze the scale effect on agility. The results show again that the SCRN performs better when assuming homogeneous distributions of links between all tiers than mixing homogeneous and heterogeneous distributions. However, when both tiers follow power-law distributions, the agility is practically identical to consider Poisson and regular degree distributions. Thus, power-law distributions can be efficient when high number redundancy of result agrees with is presented. This the common relationships 15 recommendation of having multiple suppliers to increase SC agility (Sheffi and Rice, 2005), but restricting it to the case of supply networks with heterogeneous distribution of links. Horizontal links among wholesalers may enhance agility, since the related firms interchange production between them when needed to fulfill their own orders. Figure 3 shows the OFR evolution when increasing the percentage of the total wholesalers with horizontal links (ρ). In the simulations we consider that two linked wholesalers function as one, adding their capacities and orders. The case ρ=0 presents the same conditions and gives the same OFR values shown in Table 1a, while ρ=1 results in the maximum agility (OFR=1) for any degree distribution since, in this case, all wholesalers share their product and therefore the supply and demand are perfectly connected. In between, OFR increases between the two extreme values, but not regularly. It is not very much affected when less than half of wholesalers interchange production and the steeper increase is produced when horizontal links are higher than 80%. The effect of increasing horizontal links is negligible up to point when homogeneous in-degree distributions of retailers are assumed. In the case of power-law in-degree distribution, the positive effect on agility starts to be intense when more than 50% of wholesalers are linked. These results confirm H2, but provide additional information of the benefits of having horizontal links. The marginal positive effect on agility of adopting new horizontal links in the SCRN is low when starting from low interrelationships between wholesalers but high when the number horizontal links exceed a certain threshold. 16 Figure 3. OFR with respect to the percentage of wholesalers with horizontal links (ρ). The initial conditions are identical to case (a) in Table 1. The three trajectories in every graph correspond to three in-degree distributions for 𝑖𝑛: Regular (Reg.), zero-truncated Poisson (Poiss.) and zero-truncated power-law (Pow.). Every graph retailers 𝑘𝑟 𝑜𝑢𝑡): (a) Regular; (b) Zero-truncated Poisson; (c) assumes the following out-degree distributions of wholesalers (𝑘𝑤 Zero-truncated power-law. Every OFR value is obtained by doing 1000 simulations of the SCRN and taking mean values. 5. Discussion This paper aims to analyze the influence of the topology on the SC agility. This is done by building an agent-based model of a SC including three tiers and some other specifications, such as restricted relationship between firms. The latter condition does not adjust to some industrial supply chains, such as automotive (Choi and Hong, 2002), but does to food supply chains which are compounded by small-scale producers located in rural development economies and unable to take their products to outside markets (Abebe et al., 2016; Crona et al., 2010; Hayami et al., 1999; Merlijn, 1989; Pedroza, 2013). In these circumstances, direct trading between wholesalers and producers is mostly absent, so middlemen exercise a 17 coordination role facilitating product transmission. Abebe et al. (2016) describes this type of food supply chains as "where market failure is ubiquitous and food chains still consist of many stages" and where middlemen represent a socially-tied relationship. Agility, or quick response to sudden demand changes, is treated as a dimension of the SC resilience, understood here as the ability of the SC to respond to disruptions and disturbances. Contrary to some previous experimental contributions which found that scale-free topologies characterize the most resilient SCs (Kim et al., 2015; Nair and Vidal, 2011; Zhao et al., 2011), the simulation results show that this is not so for any kind of supply networks. Specifically, in the SC conditions analyzed in the paper, homogeneous topologies, such as those which follow regular or Poisson degree distributions, assure in general higher level of agility than heterogeneous topologies, such as those derived from power-law degree distributions. In previous contributions, some of the positive characteristics found in heterogeneous distributions for SC resilience are founded on the translation of common properties of complex networks (small world effect, high clustering) to the context of SC, such as it has been evidenced in some theoretical developments (Hearnshaw and Wilson, 2013). The design of the simulation experiments to analyze SC resilience, such as the presence or random disruptions and targeted attacks, also resembles those applied for complex networks in other fields, such as Epidemiology and Cybernetics (Newman, 2010), and the results obtained are fairly similar. Our results extend these previous findings by designing a simulation experiment conceived for SC and illustrate that the most suitable topology to achieve the highest level of agility, and consequently resilience, is not scale-free in the case of restricted relationship among firms. Therefore, this study finds that a SC with similar number of relationships among buyers and sellers is under certain conditions more agile than a SC where few leading firms with many relationships are combined with many small firms with few relationships. This result does not agree either with previous theoretical developments in the context of SC resilience. Specifically, they argue that the presence of leading firms is beneficial to SC resilience by enhancing risk management culture and horizontal links among suppliers (Christopher and Peck, 2004; Kamalahmadi and Parast, 2016). However, in addition to these effects, the pursuit of leadership causes competition among firms in the same tier and likely the emergence of a topology with an uneven distribution of links. The findings in this paper show that this is not the most favorable structure for the agility of a certain type of SC. Which one of both opposite effects predominates is a matter of further empirical research and theoretical analysis. Looking at our findings exclusively, there is a trade-off between each firm' objectives, which are dominated by the maximization of revenues and the quick response to own orders, and the agility of the SC as a whole. In addition to the role of leadership, redundancy (increase of relationships) has been also recommended to enhance SC agility and resiliency (Jüttner and Maklan, 2011; Sheffi and Rice, 2005). This view has been questioned by the resilience results obtained from a SC simulation model by Kim et al. (2015). In their model, resilience is measured through the prevalence of the SC functioning when node or link disruptions occur. On the contrary, our results agree with the common view in the literature and redundancy shows beneficial for SC agility. This effect is more pronounced when heterogeneous degree distributions, such as power-law, are presented in the SC. From the comparison of our results with Kim et al. (2015)'s, it can be concluded that in some types of SCs redundancy favors some dimensions of resilience, such as fast response to demand changes, while not having a considerable effect in other types of SC and dimensions, such as the adaptation to node or link disruptions. The agility metric used in this paper (OFR) differs from the one used in other contributions (average supply path length). One of the reasons is the specific type of SC analyzed. The 18 findings obtained in this paper are applied to a SC with a restricted structure of relationships, such as food supply chains, where each firm trades exclusively with partners located in the previous or next tier. If firms could relate each other freely, such as is the case in other industrial supply chains and commonly assumed in SC simulation models, closed triads and high clustering arise. Agility can be approached in this kind of SC by the average supply path length (Nair and Vidal, 2011; Zhao et al., 2011). Using this metric, the best agility performance was obtained with scale-free topologies. However, supply path length is not useful for SCs with restricted relationships, since the shortest path length from the first producer to the customer is constant, independently of the specific topology. Order fulfillment rate is a more suitable metric for agility in this kind of structures and the results show that in general SCs with scale- free topology do not reach the highest values of OFR. The different metrics used for each structure do not allow making a rigorous comparison of findings, but illustrates the opposite performance of scale-free topologies with respect to SC agility. According to the previous analysis, it may be guessed that, in terms of agility, homogeneous distribution of links among nodes is preferable for food supply chains, where restricted relationship SCs are common. On the contrary, heterogeneous distributions or scale-free topologies are more recommendable for industrial supply chains, where relationships among firms are less restricted. However, the empirical analyses of real SCs in both food and industrial supply chains are very scarce up to date to classify them strictly in one or another type, so a conclusive assertion on this point cannot be given yet. Moreover, it cannot be deduced from the findings that a resilient SC is not compatible with scale-free topology. The results are limited not only to a specific SC structure, but also to a specific dimension of resilience (agility). As it is has been shown in previous studies, other dimensions of resilience, such as the response to node/link disruption, are more favored with scale-free than with other homogeneous topologies (Kim et al., 2015). Other relevant properties of the SC are also enhanced by power-law distributions. This is the case of cooperation among firms (Li et al., 2013) and trust among partners (Capaldo and Giannoccaro, 2015b). 6. Conclusions, limitations and implications This paper analyzes the influence of the pattern of relationships on the agility of the supply chain. SC agility means quick response to demand changes and is one of the characteristics of an efficient SC. Many previous theoretical and empirical contributions highlight that most efficient SCs are those with heterogeneous distribution of links, such as the one induced by scale-free topologies. By simulating an agent-based model built to represent a supply network and the flow of material, this paper shows that this is not the case for SCs where relationships are restricted to firms located in subsequent tiers. This type of SCs is presented in military logistic networks and some food supply chains. For a SC like this, homogeneous distribution of links results in higher agility than heterogeneous distributions of links, such as that one induced by power-law distribution. Therefore, the results of this paper illustrate that the most efficient topology in a SC is not necessarily scale-free, but depends on the conditions of the specific SC and the specific aspect of efficiency. Nevertheless, the results confirm other common recommendations for resilient SCs, such as redundancy. Higher agility levels are obtained by increasing the number of relationships among firms in two subsequent tiers. Horizontal links between firms in the same tier also favor the SC agility, although their effect is significant when a certain volume of relationships is surpassed. 19 The methodological approach followed in this paper makes that findings must be taken with caution. SC agility is influenced by qualitative factors that are not included in the simulation model. To give some examples, visibility and the existence of leading firms, which favor SC agility, are not analyzed here. Additionally, other limitations are presented. First, the algorithm includes a mechanistic way of trading, such as hierarchy in the order and supply allocation rules. Likely, this may not represent the real behavior of agents, which can follow some kind of strategy in the allocation rules. Second, it is assumed identical amount of product for every link, not allowing the firms to increase orders to the same seller. Third, it is not solved the optimal design of a SC with respect to agility. Power-law has been detected as the most suitable degree distribution among a selection of three, but it cannot be derived from the results that the induced topology from power-law distribution gives the best performance among any possible topology. Fourth, the analysis does not consider other strategies to improve SC agility, such as adding capacity and increasing inventory levels, which are considered constant in this paper. Having in mind the limitations of the analysis, some managerial recommendations can be derived from the findings. They are circumscribed to those SCs in the conditions analyzed in the paper, i.e. restricted relationships among firms, where food supply chains are a suitable example. High levels of agility are specifically desirable in these types of supply chains, since they manage high volumes of perishable products that need fast distribution. According to the results, actions to avoid trade concentration in few agents belonging to the same tier (middlemen or wholesalers) are justified. By doing so, homogeneous distribution of links among agents in different tiers is favored, which enhances SC agility. Promotion of cooperation among agents in the same tier (horizontal links) is also recommended, although it is not expected that the effect on agility is sensible when cooperation is low. This study can be extended in several ways. First, findings on agility here can be compared with new simulation results for general supply chain random network, removing the condition of restricted relationships among firms located in subsequent tiers. Additionally, this paper does not consider differences in the relationships among partners. However, strong and weak ties are present in SC relationships and show the degree of trade frequency or compromise between firms. The influence of the number and disposition of these ties on SC agility (or resilience in general) is still to be analyzed. Acknowledgments The work was financed by project PAPIIT IA300215 from the Universidad Nacional Autónoma de México (UNAM) and project ECO2014-59067-P from the Ministry of Economy and Competitiveness of the Government of Spain. References Abebe, G., Bijman, J., Royer, A., 2016. Are middlemen facilitators or barriers to improve smallholders' welfare in rural economies? Empirical evidence from Ethiopia. J. Rural Stud. 43, 203-213. Albert, R., Barabasi, A.L., 2002. Statistical mechanics of complex networks. Rev. Mod. Phys. 74, 47–97. doi:10.1088/1478-3967/1/3/006 20 Barroso, a P., Machado, V.H., Carvalho, H., Machado, V.C., 2015. Quantifying the Supply Chain Resilience. Appl. Contemp. Manag. Approaches Supply Chain. 13–38. doi:10.5772/59580 Beamon, B., 1998. Supply chain design and analysis:: Models and methods. Int. J. Prod. Econ. 55, 281–294. doi:http://dx.doi.org/10.1016/S0925-5273(98)00079-6 Bernard, T., Taffesse, A., 2008. Impact of cooperatives on smallholders' commercialization behavior: evidence from Ethiopia. Agricultural Economics 39, 147-161. Borgatti, S.P.S., Li, X., 2009. On Social Network Analysis in a Supply Chain Context. J. Supply Chain Manag. 45, 5–22. Braziotis, C., Bourlakis, M., Rogers, H., Tannock, J., 2013. Supply chains and supply networks: distinctions and overlaps. Supply Chain Manag. An Int. J. 18, 644–652. doi:10.1108/SCM- 07-2012-0260 Capaldo, A., Giannoccaro, I., 2015a. Interdependence and network-level trust in supply chain 180–195. Ind. Mark. Manag. study. 44, networks: doi:10.1016/j.indmarman.2014.10.001 computational A Capaldo, A., Giannoccaro, I., 2015b. How does trust affect performance in the supply chain? J. Prod. Econ. 166, 36–49. interdependence. Int. the moderating doi:10.1016/j.ijpe.2015.04.008 role of Carter, C., Rogers, D., Choi, T., 2015. Toward the theory of the supply chain. J. Supply Chain Management 51, 89-97. Carvalho, H., Azevedo, S.G., Cruz-Machado, V., 2012. Agile and resilient approaches to supply chain management: Influence on performance and competitiveness. Logist. Res. 4, 49– 62. doi:10.1007/s12159-012-0064-2 Choi, T.Y., Dooley, K.J., Rungtusanatham, M., 2001. Supply networks and complex adaptive systems: Control versus emergence. J. Oper. Manag. 19, 351–366. doi:10.1016/S0272- 6963(00)00068-1 Choi, T.Y., Hong, Y., 2002. Unveiling the structure of supply networks : case studies in Honda. J. Oper. Manag. 20, 469–493. doi:10.1016/S0019-8501(03)00009-9 Choi, T.Y., Wu, Z., 2009. Triads in Supply Networks: Theorizing Buyer - Supplier - Supplier Relationships. J. Supply Chain Manag. 45, 8–25. Christopher, M., Peck, H., 2004. Building the resilient supply chain. International Journal of Logistics Management , 15, 1–13. doi:10.1080/13675560600717763 Cox, A., Sanderson, J., Watson, G., 2001. Supply chains and power regimes: Toward an analytic framework for managing extended networks of buyer and supplier relationships. J. Supply Chain Manag. 37, 28–35. doi:10.1111/j.1745-493X.2001.tb00097.x Crona, B., Nyström, M., Folke, C., Jiddawi, N., 2010. Middlemen, a critical social-ecological link in coastal communities of Kenya and Zanzibar. Mar. Policy 34, 761-771. Datta, P.P., Christopher, M., Allen, P., 2007. Agent-based modelling of complex production/distribution systems to improve resilience. Int. J. Logist. Res. Appl. 10, 187– 203. doi:10.1080/13675560701467144 Dorogovtsev, S., Mendes, J., 2013. Evolution of networks: From biological nets to the Internet 21 and WWW. Oxford University Press. Falasca, M., Zobel, C., Cook, D., 2008. A decision support framework to assess supply chain resilience. Proc. 5th Int. ISCRAM Conference, 596-605. Gabre‐Madhin, E., 2001. The role of intermediaries in enhancing market efficiency in the Ethiopian grain market. Agric. Econ. 25, 311-320. Guillaume, J.-L., Latapy, M., 2006. Bipartite graphs as models of complex networks. Phys. A Stat. Mech. its Appl. 371, 795–813. doi:10.1016/j.physa.2006.04.047 Hayami, Y., Kikuchi, M., Marciano, E., 1999. Middlemen and peasants in rice marketing in the Philippines. Agric. Econ. 20, 79-93. Hearnshaw, J.S.E., Wilson, M.J.M., 2013. A complex network approach to supply chain network theory. Int. J. Oper. Prod. Manag. 33, 442–469. doi:10.1108/01443571311307343 Jüttner, U., Maklan, S., 2011. Supply chain resilience in the global financial crisis: An empirical study. Supply Chain Manag. 16, 246–259. doi:10.1108/13598541111139062 Kamalahmadi, M., Parast, M.M., 2016. A review of the literature on the principles of enterprise and supply chain resilience: Major findings and directions for future research. Int. J. Prod. Econ. 171, 116–133. doi:10.1016/j.ijpe.2015.10.023 Keys, E., 2005. Exploring market-based development: market intermediaries and farmers in Calakmul, Mexico. Geogr. Rev. 24-46 Kim, W.S., 2009. Effects of a trust mechanism on complex adaptive supply networks: An agent- based social simulation study. Jasss 12. Kim, Y., Chen, Y., Linderman, K., 2015. Supply network disruption and resilience. J. Oper. Manag. 33, 43–59. Kim, Y., Choi, T.Y., Yan, T., Dooley, K., 2011. Structural investigation of supply networks: A 194–211. Oper. Manag. approach. 29, social doi:10.1016/j.jom.2010.11.001 network analysis J. Lazerson, M.H., Lorenzoni, G., 1999. The firms that feed industrial districts: A return to the Italian source. Ind. Corp. Chang. 8, 235–266. doi:10.1093/icc/8.2.235 Lee, H.L., 2004. The triple-A supply chain. Harv. Bus. Rev. 82, 102–112. Levy, D., 1995. International sourcing and supply chain stability. J. Int. Bus. Stud. 343-360. Li, G.. b, Gu, Y.-G.., Song, Z.-H.., 2013. Evolution of cooperation on heterogeneous supply networks. Int. J. Prod. Res. 51, 3894–3902. doi:10.1080/00207543.2012.754968 Mari, S.I., Lee, Y.H., Memon, M.S., 2015. Complex network theory-based approach for designing resilient supply chain networks. Int. J. Logist. Syst. Manag. 21, 365–384. doi:10.1504/IJLSM.2015.069733 Markusen, A., 1996. Sticky Places in Slippery Space : A Typology of Industrial Districts. Economic Geography, 293-313. Merlijn, A., 1989. The Role of Middlemen in Small‐scale Fisheries: A Case Study of Sarawak, Malaysia. Dev. Change 20, 683-700. 22 Nair, A., Vidal, J.M., 2011. Supply network topology and robustness against disruptions - an J. Prod. Res. 49, 1391–1404. Int. investigation using multi-agent model. doi:10.1080/00207543.2010.518744 Newman, M., 2010. Networks: an introduction. OUP Oxford. Newman, M., 2003. The structure and function of complex networks. SIAM Rev. 45, 167-256. Newman, M., Watts, D., 2002. Random graph models of social networks. Proc. Nac. Academ. Sci. 99, 2566-2572. Parhi, M., 2005. Dynamics of Inter-firm Linkages in Indian Auto Component Industry: A Social Network Analysis. Pap. Present. DRUID Winter Conf. 27th .29th January 2005. Pathak, S.D.., Day, J.M.., Nair, A.., Sawaya, W.J.., Kristal, M.M.., 2007. Complexity and adaptivity in supply networks: Building supply network theory using a complex adaptive systems perspective. Decis. Sci. 38, 547–580. doi:10.1111/j.1540-5915.2007.00170.x Pedroza, C., 2013. Middlemen, informal trading and its linkages with IUU fishing activities in the port of Progreso, Mexico. Mar. Policy. 39, 135-143. Pedroza, C., Hernández, J.M., 2016. Reputation as a central resource to create social networks and market transactions. The case of Mercado del Mar in Guadalajara, Mexico. Pap. Present. IIFET Conf. 11th. 29th July 2016. Pettit, T., Fiksel, J., Croxton, K., 2010. Ensuring supply chain resilience: development of a conceptual framework. J. Bus. Logist. 31, 1-21. Plagányi, É.E., Van Putten, I., Thébaud, O., Hobday, A.J., Innes, J., Lim-Camacho, L., Norman- López, A., Bustamante, R.H., Farmery, A., Fleming, A., Frusher, S., Green, B., Hoshino, E., Jennings, S., Pecl, G., Pascoe, S., Schrobback, P., Thomas, L., 2014. A quantitative metric to identify critical elements within seafood supply networks. PLoS One 9. doi:10.1371/journal.pone.0091833 Ponomarov, S.Y., Holcomb, M.C., 2009. Understanding the concept of supply chain resilience. Int. J. Logist. Manag. 20, 124–143. doi:10.1108/09574090910954873 Ruben, R., Heras, J., 2012. Social capital, governance and performance of Ethiopian coffee cooperatives. Ann. Public Coop. Econ. 83, 463-484. Rubinstein, A., Wolinsky, A., 1987. Middlemen. Q. J. Econ. 581-594. Scholten, K., Schilder, S., 2015. The role of collaboration in supply chain resilience. Supply Chain Manag. An Int. J. 20, 471–484. doi:10.1108/SCM-11-2014-0386 Scholten, K., Scott, P.S., Fynes, B., 2014. Mitigation processes – antecedents for building supply chain resilience. Supply Chain Manag. An Int. J. 19, 211–228. doi:10.1108/SCM-06-2013- 0191 Sheffi, Y., Rice Jr., J.B., 2005. A Supply Chain View of the Resilient Enterprise. MIT Sloan Manag. Rev. 47, 41–48. Soni, U., Jain, V., Kumar, S., 2014. Measuring supply chain resilience using a deterministic modeling approach. Comput. Ind. Eng. 74, 11–25. doi:10.1016/j.cie.2014.04.019 Surana, A., Kumara, S., Greaves, M., Raghavan, U.N., 2005. Supply-chain networks: a complex 23 adaptive systems perspective, International Journal of Production Research 43, 4235- 4265. doi:10.1080/00207540500142274 Tang, L., Jing, K., He, J., Stanley, H.E., 2016. Complex interdependent supply chain networks: its Appl. 443, 58–69. Cascading failure and robustness. Phys. A Stat. Mech. doi:10.1016/j.physa.2015.09.082 Thadakamalla, H.P., Raghavan, U.N., Kumara, S., Albert, R., 2004. Survivability of multiagent- based supply networks: A topological perspective. IEEE Intell. Syst. 19, 24–31. doi:10.1109/MIS.2004.49 Tierney, K., Bruneau, M., 2007. Conceptualizing and measuring resilience: A key to disaster loss reduction. TR news. Travers, J., Milgram, S., 1969. An Experimental Study of the Small World Problem. Sociometry 32, 425–443. Wieland, A., Wallenburg, C.M., 2013. The influence of relational competencies on supply chain resilience: a relational view. Int. J. Phys. Distrib. Logist. Manag. 43, 300–320. doi:10.1108/IJPDLM-08-2012-0243 Wilson, J., 1980. Adaptation to uncertainty and small numbers exchange: The New England fresh fish market. Bell J. Econ. 491-504. Zhao, K., Kumar, A., Harrison, T.P., Yen, J., 2011. Analyzing the Resilience of Complex Supply Network Topologies Against Random and Targeted Disruptions. IEEE Syst. J. 5, 28–39. doi:10.1109/JSYST.2010.2100192 24
1508.03607	1	1508	2015-08-14T18:59:01	Is That Twitter Hashtag Worth Reading	[ "cs.SI" ]	Online social media such as Twitter, Facebook, Wikis and Linkedin have made a great impact on the way we consume information in our day to day life. Now it has become increasingly important that we come across appropriate content from the social media to avoid information explosion. In case of Twitter, popular information can be tracked using hashtags. Studying the characteristics of tweets containing hashtags becomes important for a number of tasks, such as breaking news detection, personalized message recommendation, friends recommendation, and sentiment analysis among others. In this paper, we have analyzed Twitter data based on trending hashtags, which is widely used nowadays. We have used event based hashtags to know users' thoughts on those events and to decide whether the rest of the users might find it interesting or not. We have used topic modeling, which reveals the hidden thematic structure of the documents (tweets in this case) in addition to sentiment analysis in exploring and summarizing the content of the documents. A technique to find the interestingness of event based twitter hashtag and the associated sentiment has been proposed. The proposed technique helps twitter follower to read, relevant and interesting hashtag.	cs.SI	cs	Is That Twitter Hashtag Worth Reading Anusha A and Sanjay Singh ∗ 1 Introduction Abstract Online social media such as Twitter, Facebook, Wikis and Linkedin have made a great impact on the way we consume information in our day to day life. Now it has become increasingly important that we come across appropriate content from the social media to avoid information explosion. In case of Twitter, popular information can be tracked using hashtags. Studying the characteristics of tweets containing hashtags becomes important for a number of tasks, such as breaking news detection, personalized message recommendation, friends recommendation, and sentiment analysis among others. In this paper, we have analyzed Twitter data based on trending hashtags, which is widely used nowadays. We have used event based hashtags to know users' thoughts on those events and to decide whether the rest of the users might find it interesting or not. We have used topic modeling, which reveals the hidden thematic structure of the documents (tweets in this case) in addition to sentiment analysis in exploring and summarizing the content of the documents. A technique to find the interestingness of event based twitter hashtag and the associated sentiment has been proposed. The proposed technique helps twitter follower to read, relevant and interesting hashtag. P R E P RIN T The rapid growth in the popularity of social networking and microblogging has lead to a new way of finding and broadcasting information in the past decade. The websites like Facebook, Twitter, Linkedin, Stack Overflow, and Stack Exchange, etc. have become the go to sites when people need any platform to com- municate, broadcast their thoughts, publicity for an upcoming product or even a new app that they have developed. These websites bridge the gap between you and the rest of world in just one click. Twitter is noted to be the third most popular of such social networking sites incorporated on April 19, 2007 [1]. It is one of the social networking platforms which has about 288 million active users generating 500 million tweets per day. Each tweet comprise of text up to 140 characters. Thousands of people advertise their recruiting services, consultancy, retail stores by using Twitter. The Internet users like this because it is less probing and available on mobile platform as well. These websites are increasingly used for communicating breaking news, eyewitness accounts and organizing large groups of people using hashtag [2] features in Twitter. Users of Twitter have become accustomed to getting regular updates on trending events, both of personal and global value. For instance, twitter was used to propagate information in real-time in many crisis situations such as the results of the Iran election [3], the tsunami in Samoa [4] and more recently earthquake in Nepal [5]. Many organizations and celebrities use their Twitter accounts to connect to customers and fans and propagate their thoughts. Twitter is depicted as a blend of instant messaging, microblogging, and knowledge source. Twitter is a way to connect with a person or a topic and decide for yourself if you like it. With such a progress in social networking, by analyzing textual data obtained from any of these sites which represent ideas, thoughts and communication between the users, it is possible to obtain an understanding of needs and concerns of the users that provides valuable information for academic, marketing and policy-making. Since there are about 500 million tweets generated every day and very vast number of users look for interesting tweets among those tweets, there is a need for a mechanism to find such interesting tweets using ∗Sanjay Singh is with the Department of Information and Communication Technology, Manipal Institute of Technology, Manipal University, Manipal-576104, INDIA, E-mail: [email protected] 1 an unbiased method. It should not depend on author popularity and consider only the content of tweets. Every tweet generated by a user can be retweeted by other users those who are following the author of that tweet. Until recently, "retweet" count of a tweet was used as measure of popularity of that tweet regardless of topic and content of the tweet. However, retweet count depends mainly on the author popularity and not the content popularity of the tweet. More the followers an author has, the more publicity the tweet could obtain. In order to find the tweet that is of interest to large audience we need to consider more than just retweet count. Hence, we need to analyze the short text message that usually includes noise. The text messages from twitter needs to identify the interesting tweets among them by using an automatic method, so that it can be of use to wider range of audience. In order to find the interesting tweets, we have used topic modeling [6] based on Latent Dirichlet Allocation (LDA) [7]. This is followed by sentiment analysis [8] which helps to build a more human like system. The aim here is to understand the sentiment of each tweet using natural language processing. We have utilized the 'hashtag' to extract tweets and find their interestingness and sentiment associated with that tweet. 2 Related Work Topic models [9] are powerful tools to identify latent text patterns in the content. It has been applied in a wide range of areas including recent work on Twitter [10]. Social media differs from some standard text domain (e.g., citation network, web pages) where topic models are usually utilized in a number of ways. One important fact is that there exists many dimensions in social media that we usually want to consider them simultaneously. Many studies have provided insights into social media. Kwak et al. [11] studied Twitter's structure by investigating various Twitter features. Recently, many works [12] [13] have focused mainly on analyzing or obtaining valuable information, such as influential users and posts on Twitter, from a large amount of social data. Most of the existing approaches have considered retweet counts as a measure of popularity, influence, and interestingness, and presented classifiers that predicted whether and how often new tweets will be retweeted in the future. They exploited many features of twitter, such as textual data, author's information, and propagation information. Although the overall retweet count indicates a tweet's popularity, this may apply only to the followers of the tweet's author. Twitter not only has textual data but also has related data, such as follower and retweet links, which enable us to construct a network structure. The link-based approaches applied a variant of the link analysis algorithm [14] to a designed link structure in order to find interesting messages. However, the link structure requires a large volume of linking data to be analyzed and constructed and cannot be updated effectively when new tweets are generated. Alonso et al.[15], used crowdsourcing to categorize a set of tweets as interesting or uninteresting and reported that the presence of a URL link is a single, highly effective feature for selecting interesting tweets with more than 80 % accuracy. This simple rule, however, may incorrectly categorize an uninteresting tweet (i.e., an uninteresting tweet contains links to meaningless pictures, videos, and advertisements)as interesting. Lauw et.al [16] suggested several features to identify interesting tweets but did not experimentally validate them. For user recommendation, Armentano et.al [17] examined the topology of followers/followees network and identified the relevant users using social relation factors. They conducted not only topology-based profiling but also content-based profiling to find semantically similar users. In social media, semantic analysis and topic modeling are widely used to understand textual data and can facilitate many applications such as user interest modeling [18], sentiment analysis [19], content filtering [20], and event tracking etc. Zhao et al. [21] analyzed the topical differences between Twitter and traditional media using TwitterLDA for investigating short messages. Wang and McCallum [22] and Kawamae [23] conducted topic modeling of temporally-sequenced documents in Twitter and tried to model the topics continuously over time. However, in our approach LDA considers the mixtures of latent topics as a trend based on hashtags and is designed to learn changes in topic distributions, while other works focus on learning topic shifts based on word distributions. Chen et al.[24] focused on recommending URLs posted in tweets using various combinations of topic 2 relevance and social graph information. The model by Ramage et al.[25] is an unsupervised learning method with relative importance of latent topics. In 2014, Min-Chul Yang and Hae-Chang Rim [10] proposed a better model based on Ramage's work to find the interesting tweets, but it does not include sentiment polarity analysis. Figure 1: Block Diagram of Proposed System 3 Proposed Method We collected tweets based on hashtags related to a specific event, ICC World Cup 2015, along with the tweets that was not generated by any particular event. Yang and Rim [10] used a time stamp approach in their work to find interesting tweets that belong to a specific time period, where as we have used hashtag feature that is already available in twitter to eliminate the need of time stamp and find the interestingness and sentiment of tweets generated using hashtags that are related to a specific event that has occurred or occurring at present. Figure 1 shows the block diagram of our proposed method. Once the tweets are extracted generally as well as based on hashtags, it needs to be preprocessed by removing stop words, discarding tweets with non English words, special characters, etc. Every individual tweet is considered as a document. Then such tweets are analyzed to find their interestingness and sentiment polarity. A higher score to interesting tweet is assigned and their polarity is found. The major concept used by Yang and Rim [10] is interestingness. It indicates the number of audience that might find this tweet interesting to them without the consideration of author and author's popularity. To find the interestingness we need to extract latent topics that tweets belong to. We have used LDA [7] to infer latent topics that the tweets we have collected belongs to. LDA considers a document as a bag of words and not the context of the words that appearing in documents. Since all topics do not contribute to be a part of interesting tweet, so we find the weight as well as spatial entropy for each topic. Tweets that belong to the topics with higher weight and less spatial entropy is considered to be more interesting. LDA generates the latent topics for the given set of documents and two probability distributions. First is the word distribution in each topic, and second is distribution of topics in each document. These two distributions are used to find the spatial entropy and integrity of each topic. The plate notation of LDA is given in Fig. 2. In Fig.2 N represents the number of words belonging to K number of topics, being distributed over M number of documents. Each document can belong to one or more topics and each word will have its own probability in topics. The α and β are parameters that depict the probability of words and topics respectively 3 Figure 2: Plate Notation of LDA [26] that has not yet occurred in the data set, but may appear in future. We have set their value to be 0.01 and it can be set to any value between 0 and 1. Here θ is the probability of topic z in the given document and ϕ is the probability of word w in the given topic. We have used The Stanford Topic Modeling Toolbox (TMT) [27] to perform analysis on our data set. In order to find weight of the topics, we have to find integrity and spatial entropy of each topic using the distribution obtained after applying LDA on our data set. 3.1 Tweet Scoring Measures Each tweet is evaluated for its interestingness based on the topics it belongs to. If a tweet belongs to topics that are having higher weights, it is said to be more interesting. Note that the weight of each topic represents how important that topic is. If the topic contains words which are not standard words or if the topic is found to be present in too many documents then such topic is considered as a noisy topic. However, to find weight of each topic, first we need to calculate its integrity and spatial entropy. 3.1.1 Integrity Integrity of topic considers that every topic may not be useful to analyze the data that is collected. Each topic has its own distribution of words, and these words are used to determine the integrity of the topic. A lexical dictionary that has all English words and popular non-dictionary words is built manually. It contains more than 0.1 million words. The Integrity of Topic (I) is measured as follows [10]: I(t) = p(wt)L(w) (1) (cid:88) w∈W where p(wt) is the probability of word w in the topic t and L(w) is 1 if the word is present in the lexical dictionary else 0. 3.1.2 Spatial Entropy Spatial Entropy depicts the distinction of topics based on topic distribution. Most meaningful topics are considered to be related to smaller number of documents. If a topic is found to be closely related to too many documents, then that topic is considered to be noisy or general. Spatial Entropy is given by [10]: p(dt) log p(dt) (2) S(t) = −(cid:88) d∈D where document d is a single tweet and p(dt) is the probability of document d given topic t. The values of p(dt) is calculated based on Bayesian inference in LDA [7] . The Integrity and Spatial Entropy are normalized using x = z(x) = x − µ σ (3) where x is variable, µ is average of variables and σ is standard deviation of variables. 4 3.1.3 Weight of Topics Weight of topics is computed based on the normalized Integrity and Spatial Entropy. It is given by: W (t) = I(t) − S(t) (4) The entropy value is subtracted from integrity value because lower the entropy higher the importance of the topic. The topic gets a low weight if it is found to be noisy or meaningless. 3.1.4 Scoring Tweets Scoring the interestingness of tweets is done by assigning relative topic score and probability of that topic in a particular tweet. The scoring function for tweet t is given by W (t)p(td) (cid:88) (5) Score(t) = t∈T where W (t) is the calculated weight of topic t and p(td) is the topic distribution for document d. Eventually, the document which covers topics with higher weight gets higher score of interestingness. The tweet with the higher score is considered as the more interesting than the tweets that have lower scores. 3.2 Sentiment Polarity Analysis Sentiment polarity analysis is a natural language processing task which returns the polarity of each documents as a float value within the range [-1.0, 1.0]. The first value represent the positive polarity and second value represents the negative polarity. There are few NLTK corpora [28] which is considered for training data and their analysis is considered in deciding the sentiment polarity of the documents that we provide. If the value of positive polarity is higher than the value of negative polarity, then the tweet is said to convey a positive sentiment and vice-versa. 4 Results and Discussion We have used Twitter REST API [29] to collect tweets. The experimental results of our method to evaluate the interestingness and sentiment of tweets is given in this section. Figure 3: Document-Score Plot Interesting tweet classification task is completely content driven and the author's popularity does not influence this scoring mechanism in any way. For our method as tweets crawled were not more than 5K, we set the number of topics to be 15 which is based on our emprirical experiments and performed 1000 iterations of variational inference in TMT. 5 Table 1: Tweets and their Interestingness Score Tweet Two host nations then just like the last time. Best 2 teams in the tour- nament fighting for d title. No interest in ICCWC. Three years back on this day we won the ICCWC for the second time in ODIs. Cheers to the victory. Australia Cricket Champs Again, Congrats Injured Australian batsman Phillip Hughes dies Miss coming home from work to the World Cup Loosing in cricket always hurts but take a bow, we have played some brilliant cricket, especially bowling @bcci Has Virat Kohli taken the responsi- bility to make them winner? Score 1.83701 -0.65715 2.07052 2.48527 - 0.8024191 0.03527 1.40191 0.043229 For sake of presentation the score of 50 such tweets related to ICC World Cup is shown in Fig.3. Table 1 lists eight tweets and their corresponding interestingness score. Among the eight tweets fourth is the most interesting tweet and fifth is found to be the least interesting tweet. We have considered each tweet as interesting tweet if the score obtained is above 1. Since the number of topics considered for our dataset is low, for a relatively low number of topics present in the dataset, the score threshold is also fixed to a lower value. However, depending upon the number of topics threshold for interestingness may vary accordingly. We have collected tweets based on hashtags 'ICCWC' and 'CWC15' to analyze the sentiment based on the event of ICC World Cup 2015. Their sentiment polarity along with the rest of the tweets were calculated, Table 2 lists few such tweets and their sentiment polarity. Figure 4 and Fig.5 depicts the sentiment variation of few tweets related to the event of ICC World Cup 2015. Figure 4 shows the positive sentiment variation and Fig.5 shows the variation in negative sentiment. Figure 4: Sentiment Variation on ICCWC 6 Table 2: Tweets and their Sentiment Polarity Tweet @skysportnz thank you so much for securing the rights of @IPL. It was so boring after iccwc. Gonna be fascinating 6 weeks again No interest in ICCWC. Australia Cricket Champs Again, Congrats Dhoni's innings was com- pletely bizarre, what was his plan??? AUSvIND IC- CWC Scotland vs Sri Lanka ic- cwc Positive Polarity 0.9522 Negative Polarity 0.04779 0.4573 0.9171 0.5426 0.0828 0.4252 0.5747 0.5 0.5 Figure 5: Sentiment Variation on ICCWC As we mentioned earlier the hashtags related to the event of ICC World Cup was included to extract tweets and the result of this sentiment analysis is shown in Fig.6. From Fig.6 we observed that the most of the tweets were not completely positive or negative. For example, when New Zealand won against South Africa in semifinals, people felt bad for South Africa but appreciated New Zealand's game as well. This generates almost equal values of both positive and negative polarity to such tweets. Every user has their own opinion about an event occurring. The tweets belong to various topics, events may convey different emotions: positive or negative, which needs to be analyzed before deciding if the tweet is interesting or not. The proposed method can handle numerous latent topics that can be found in large set of data. The probability distribution obtained after applying LDA is reliable and gives expected results, which can be followed by sentiment analysis. It would be good if the dataset is sufficiently large to find proper latent topics. Our sentiment analysis is done against a vast annotated corpora and tend to take a longer time to analyze a large dataset. 7 Figure 6: Sentiment Analysis 5 Conclusion In this paper, we have proposed an effective approach to evaluate the interestingness of tweets based only on the textual contents and find what sentiment they convey. Topic modeling using LDA was applied to discover the latent topics and then find the weight of each topic, which depends on their integrity and spatial entropy. Higher integrity and lower spatial entropy contributes to a higher weight for the corresponding topic. The tweets that are distributed over topics with higher weight is considered as more interesting and if the tweet belongs to topics with lower weight it is said to be less interesting. In addition to this, we performed sentiment analysis on tweets to find the sentiment polarity of each tweet and to find the variation of sentiment with time and with respect to an event that is occurring. Our approach can be used to find the interestingness and sentiment polarity of tweets related to an event or in general. In future, we can think of applying this method to a real time system to suggest tweets that belong to specific topic and depicts the desired sentiment. Since the hashtags are commonly used feature in most of the social media and not just in Twitter, our approach can be used to analyze any platform's textual data that is related to an event by hashtag. References [1] S. Jain, "40 most popular social networking sites of the world," Social Media Today-Blog, 2012. [Online]. Available: http://www.socialmediatoday.com/content/40-most-popular-social-networking-sites-world [2] Twitter, "Using hashtags on twitter," https://support.twitter.com/articles/ 49309-using-hashtags-on-twitter, 2015, [Online; accessed 10-April-2015]. [3] Wikipedia, "Twitter revolution," en.wikipedia.org/wiki/Twitter Revolution, 2014. [4] BBC, "Deadly tsunami strkes in pacific," news.bbc.co.uk/2/hi/asia-pacific/8281616.stm, 2009. [5] --, "Nepal earthquake: Eight million people affected," news.bbc.co.uk/news/worls-asia-32492232, 2015. [6] Wikipedia, "Topic model - wikipedia, the free encyclopedia," 2015, [Online; accessed 28-April-2015]. [Online]. Available: \url{http://en.wikipedia.org/w/index.php?title=Topic model&oldid=652899497} 8 [7] D. M. Blei, A. Y. Ng, and M. I. Jordan, "Latent dirichlet allocation," J. Mach. Learn. Res., vol. 3, pp. 993–1022, Mar. 2003. [Online]. Available: http://dl.acm.org/citation.cfm?id=944919.944937 [8] Wikipedia, "Sentiment analysis - wikipedia, the free encyclopedia," http://en.wikipedia.org/w/index. php?title=Sentiment analysis&oldid=654779025, 2015, [Online; accessed 15-April-2015]. [9] D. M. Blei, "Probabilistic topic models," Commun. ACM, vol. 55, no. 4, pp. 77–84, April 2012. [Online]. Available: http://doi.acm.org/10.1145/2133806.2133826 [10] M.-C. Yang and H.-C. Rim, "Identifying interesting twitter contents using topical analysis," [Online]. Available: Expert Systems with Applications, vol. 41, no. 9, pp. 4330–4336, 2014. http://www.sciencedirect.com/science/article/pii/S0957417414000141 [11] H. Kwak, C. Lee, H. Park, and S. Moon, "What is twitter, a social network or a news media?" in Proceedings of the 19th International Conference on World Wide Web, ser. WWW '10. New York, USA: ACM, 2010, pp. 591–600. [Online]. Available: http://doi.acm.org/10.1145/1772690.1772751 [12] C. Castillo, M. Mendoza, and B. Poblete, "Information credibility on twitter," in Proceedings of the 20th International Conference on World Wide Web, ser. WWW '11. New York, USA: ACM, 2011, pp. 675–684. [Online]. Available: http://doi.acm.org/10.1145/1963405.1963500 [13] Y. Duan, L. Jiang, T. Qin, M. Zhou, and H.-Y. Shum, "An empirical study on learning to rank of tweets," in Proceedings of the 23rd International Conference on Computational Linguistics, ser. COLING '10. Stroudsburg, PA, USA: Association for Computational Linguistics, 2010, pp. 295–303. [Online]. Available: http://dl.acm.org/citation.cfm?id=1873781.1873815 [14] D. M. Romero, W. Galuba, S. Asur, and B. A. Huberman, "Influence and passivity in social media," in Proceedings of the 20th International Conference Companion on World Wide Web, [Online]. Available: http://doi.acm.org/10.1145/1963192.1963250 ser. WWW '11. New York, NY, USA: ACM, 2011, pp. 113–114. [15] O. Alonso, C. Carson, D. Gerster, X. Ji, content Search Evaluation (CSE 2010), Geneva, Switzerland, July 2010, pp. 39–42. http://ir.ischool.utexas.edu/cse2010/materials/alonsoetal.pdf "Detecting uninteresting streams," in Proceedings of SIGIR 2010 Workshop on Crowdsourcing for [Online]. Available: and S. U. Nabar, in text [16] K. K. H. W. Lauw, A. Ntoulas, "Estimating the quality of postings in the real-time web," in Proceedings of WSDM 2010 Workshop on search in social media SSM'10, Feb 2010, pp. 1–4. [Online]. Available: http://ink.library.smu.edu.sg/sis research/1556 [17] M. G. Armentano, D. Godoy, and A. A. Amandi, "Followee recommendation based on text analysis of micro-blogging activity," Inf. Syst., vol. 38, no. 8, pp. 1116–1127, Nov 2013. [Online]. Available: http://dx.doi.org/10.1016/j.is.2013.05.009 [18] M. Pennacchiotti and S. Gurumurthy, social media user recommendation," in Proceedings of the 20th International Conference Companion on World Wide Web, [Online]. Available: http://doi.acm.org/10.1145/1963192.1963244 ser. WWW '11. New York, NY, USA: ACM, 2011, pp. 101–102. "Investigating topic models for [19] C. Lin and Y. He, "Joint sentiment/topic model for sentiment analysis," in Proceedings of the 18th ACM Conference on Information and Knowledge Management, ser. CIKM '09. New York, NY, USA: ACM, 2009, pp. 375–384. [Online]. Available: http://doi.acm.org/10.1145/1645953.1646003 [20] J. Duan and J. Zeng, "Web objectionable text technique," Expert Syst. Appl., vol. 40, no. 15, pp. 6094–6104, Nov 2013. http://dx.doi.org/10.1016/j.eswa.2013.05.032 content detection using topic modeling [Online]. Available: 9 [21] W. X. Zhao, J. Jiang, J. Weng, J. He, E.-P. Lim, H. Yan, and X. Li, "Comparing twitter and traditional media using topic models," in Proceedings of the 33rd European Conference on Advances in Information Retrieval, ser. ECIR'11. Berlin, Heidelberg: Springer-Verlag, 2011, pp. 338–349. [Online]. Available: http://dl.acm.org/citation.cfm?id=1996889.1996934 [22] X. Wang and A. McCallum, "Topics over time: A non-markov continuous-time model of topical trends," in Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD '06. New York, NY, USA: ACM, 2006, pp. 424–433. [Online]. Available: http://doi.acm.org/10.1145/1150402.1150450 [23] N. Kawamae, "Trend analysis model: Trend consists of temporal words, topics, and timestamps," in Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, ser. WSDM '11. New York, NY, USA: ACM, 2011, pp. 317–326. [Online]. Available: http://doi.acm.org/10.1145/1935826.1935880 [24] J. Chen, R. Nairn, L. Nelson, M. Bernstein, and E. Chi, "Short and tweet: Experiments on recommending content from information streams," in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, ser. CHI '10. New York, NY, USA: ACM, 2010, pp. 1185–1194. [Online]. Available: http://doi.acm.org/10.1145/1753326.1753503 [25] D. Ramage, S. T. Dumais, and D. J. Liebling, "Characterizing microblogs with topic models." [Online]. Available: in ICWSM, W. W. Cohen and S. Gosling, Eds. The AAAI Press, 2010. http://dblp.uni-trier.de/db/conf/icwsm/icwsm2010.html#RamageDL10 [26] Wikipedia, "Latent dirichlet allocation - wikipedia, the free encyclopedia," 2015, [Online; accessed 28-April-2015]. [Online]. Available: \url{http://en.wikipedia.org/w/index.php?title=Latent Dirichlet allocation&oldid=650781080} [27] T. S. N. L. P. Group, "Standford topic modeling toolbox," [Available Online] nlp.stanford.edu/software/ tmt/tmt-0.4/, 2010. [28] N. Project, "Nltk 3.0 documentation," [Available Online] www.nltk.org/data.html, 2014. [29] Twitter, "Rest apis," [Available Online] https://dev.twitter.com/rest/public, 2014. 10
1901.01354	1	1901	2019-01-05T02:22:05	Metrics matter in community detection	[ "cs.SI", "physics.soc-ph" ]	We present a critical evaluation of normalized mutual information (NMI) as an evaluation metric for community detection. NMI exaggerates the leximin method's performance on weak communities: Does leximin, in finding the trivial singletons clustering, truly outperform eight other community detection methods? Three NMI improvements from the literature are AMI, rrNMI, and cNMI. We show equivalences under relevant random models, and for evaluating community detection, we advise one-sided AMI under the $\mathbb{M}_{\mathrm{all}}$ model (all partitions of $n$ nodes). This work seeks (1) to start a conversation on robust measurements, and (2) to advocate evaluations which do not give "free lunch".	cs.SI	cs	Preprint Metrics matter in community detection Arya D. McCarthy,∗ Tongfei Chen, and Rachel Rudinger Department of Computer Science Johns Hopkins University Department of Computer Science and Engineering David W. Matula Southern Methodist University (Dated: January 8, 2019) We present a critical evaluation of normalized mutual information (NMI) as an evaluation metric for community detection. NMI exaggerates the leximin method's performance on weak communities: Does leximin, in finding the trivial singletons clustering, truly outperform eight other community detection methods? Three NMI improvements from the literature are AMI, rrNMI, and cNMI. We show equivalences under relevant random models, and for evaluating community detection, we advise one-sided AMI under the Mall model (all partitions of n nodes). This work seeks (1) to start a conversation on robust measurements, and (2) to advocate evaluations which do not give "free lunch". PACS numbers: 89.75.Hc Networks and genealogical trees -- 05.10.-a Computational methods in statistical physics and nonlinear dynamics -- 02.10.Ox Combinatorics; graph theory -- 87.23.Ge Dynamics of social systems I. INTRODUCTION Unsupervised algorithms -- like those for community de- tection (CD), investigated in many discipline for half a century [1] -- present a challenge for appraisal. In CD, we circumvent the problems of intrinsic measures by using ex- ternal evaluation tasks. Practitioners apply CD methods to benchmark graphs containing "ground truth" commu- nities, then compute an agreement measure to determine how well those communities are recovered. These mea- sures differ from typical classification accuracy because there are no specific labels (e.g. no notion of "Cluster 2") -- only groups of similar entities. The popular measure in CD is normalized mutual infor- mation (NMI). Its theoretical flaws have been noted [2 -- 5]. Particularly relevant is the non-homogeneity of the mea- sure: NMI awards credit for low-information guessing [3]. This deficiency has demonstrable implications for method selection, which we later show using the leximin method as an example. A sequence of proposed improvements [5, 6] in the CD community led to the recent corrected NMI (cNMI) [2]. A common, older measure, adjusted mutual information (AMI) [4], has garnered recent attention in CD [3, 7]. AMI augments NMI's consistent upper bound (1.0) with a consistent zero expectation to adjust for chance cluster- ings:1 AMI(C,T ) = I(C,T ) − B(C,T ) M (C,T ) − B(C,T ) , (1) ∗ [email protected]; Partially conducted at Southern Methodist Uni- versity. 1 Negative AMI indicates worse-than-chance clusterings. where B(C,T ) is a baseline function that is used to adjust the metric to zero expectation, and M (C,T ) guarantees consistent upper bound of the metric. In previous liter- ature, one common incarnation of these these functions could be (2) (3) B(C,T ) = EC(cid:48),T (cid:48) [I(C(cid:48),T (cid:48))] ; M (C,T ) =(cid:112)H(C) · H(T ) . Many variants exist, by changing the definitions of the B and M functions. Note the expectation operator in Equation 2. Over what distribution is the expected value computed? This is called random models in recent work [8]. The literature has implicitly computed expectations over Mperm: all partitions of the same class (or cluster-size pattern) as the observation [2, 4 -- 6, 9]. We argue that more appropriate for CD is an expectation over Mall instead of Mperm: all partitions of n nodes. We also advise one-sided random models for comparing against a fixed ground truth. The main contributions of this work are: • With the leximin method as an example, we show the need for an improved evaluation metric in CD. • We identify an evaluation function which better matches the community detection problem domain. • We advocate for the use of the adjusted metric AMI, with slight modifications from its proposed form. • We provide thorough analysis of the relationships between AMI and other evaluation functions. 9 1 0 2 n a J 5 ] I S . s c [ 1 v 4 5 3 1 0 . 1 0 9 1 : v i X r a II. COMMUNITY DETECTION B. Random models 2 A number of tasks on graphs ask that you partition the graph's nodes to maximize a score function. Situated between the microscopic, node level and the macroscopic, whole-graph level, these partitions form a mesoscopic structure -- be it a core -- periphery separation, a graph coloring, or our focus: community detection. Community detection has been historically ill-defined, though the intuition is to collect nodes with high interconnectivity (or edge density) into communities with low edge density between them. The task is analogous to clustering, which groups points that are near one another in some fashioned metric space. To bring some rigor to the task, we often choose modu- larity [10] as the objective function. Modularity balances the objectives of many edges within clusters (the first term) and avoiding giant, all-encompassing communities (the second term). Directly maximizing modularity is NP- hard [11], as is approximating it to within any constant factor [12]. For this reason, most community detection methods are heuristics that approximate modularity. Still, modularity isn't the be-all, end-all in community detection. In real-world data, we may have a notion of the communities already, and in these cases we only value modularity insofar as it may guide us toward these communities. To assess whether our formulation of com- munity detection matches our needs, we perform extrinsic evaluation against a known ground truth. While our focus is community detection, our results are sufficiently general for any mesoscopic pattern discovery on graphs. III. PRELIMINARIES A. Partition structures is a subset of X, their union equals X, i.e. (cid:83)k A partition or clustering C of set X can be written as C = {C1,··· , Ck}, where each community or cluster Ci i=1 Ci = X, and every pair is mutually disjoint, i.e., ∀ i, j, Ci∩Cj = ∅. We reuse some terminology on partition structures from Kingman [13]. A partition Λ of integer N is a multiset of integers {λ1,··· , λk}, whose sum is N , commonly arranging the elements in descending order. A partition C = {C1,··· , Ck} of set X is said to (cid:80)k have shape(C) (cid:44) {C1,··· ,Ck} (a multiset), where i=1 Ci = X = N .2 That is, the shape is an inte- ger partition whose elements map to cluster sizes. And what, in fact, is the domain of C, which we call M? For community detection, it will be the set of all partitions on N nodes. After all, any of these partitions represents a valid community structure, even if it is a poor community structure by intrinsic metrics like modularity. Nevertheless, other random models are relevant to our discussion, restricting the set to those with a fixed number of communities or with a fixed sequence of sizes. Previous work argued that clustering similarity should be computed in the context of a random ensemble of clusterings [9, 15]. What context of clusterings should be chosen? Gates and Ahn [8] argue that the question above is usually ignored in CD research and more broadly. To remedy this, they proposed three random models: Mperm, Mall and Mnum. Given a clustering C over set X whose size is N , these random models are: terings of N elements: Mall(C) = {C(cid:48) (cid:80) 1. Mall: This is the random model that spans all clus- C∈C(cid:48) C = N}. This is the random model that we advocate in- stead of the more common Mperm; our justifications will be elaborated below. 2. Mperm (permutation model ): The shape of the clus- terings are fixed, and all random clusterings are generated by shuffling the elements between the fixed-size clusters. Formally, Mperm(C) = {C(cid:48) shape(C(cid:48)) = shape(C)}. However, despite being widely used for evaluation [2, 4 -- 6, 9], the premises of the permutation model are frequently violated; in many clustering scenarios, either the number of clusters or the size distribution vary drastically [8, 15]. 3. Mnum: The random model that contains all cluster- ings of the same number of clusters: Mnum(C) = {C(cid:48) ∈ Mall(C) C(cid:48) = C}. These sets of clusterings satisfy the following contain- ment order: Mperm(C) ⊆ Mnum(C) ⊆ Mall(C).3 C. Information theory After a method has followed clues through the space of partitions, we want to pull off its blindfold to see how it did. Classification accuracy or F1 score won't cut it: When comparing to the ground truth, there are no specific labels (e.g. no notion of "Cluster 2") -- only groups of like entities. We settle for a measure of similarity in the 2 This shape has alternatively been called a class or decomposition pattern [14]. 3 This abuse of notation blurs the distinction between random models and the spaces over which they define their probabilities. We have restricted ourselves to uniform distributions over the spaces, so we comfortably believe that the abuse will not confuse. groupings, quantifying how much the computed partition tells us about the ground truth. A popular choice is the normalized mutual information (NMI) between the prediction and the ground truth. To understand it (and its flaws), we must review basic information theory concepts. Mutual information depends on an understanding of entropy, which captures the uncertainty in a random variable -- in this case, the category labels. For a clustering C = {C1, C2, . . . , Ck}, the entropy is Pr(C) log Pr(C), (4) H(C) = −(cid:88) C∈C For community detection, we compute it with maximum likelihood estimation. The probability of membership in a given community is proportional to its size, so the clustering's entropy -- a measure of its uncertainty -- is H(C) = −(cid:88) C∈C C N C N log . (5) Mutual information (MI) measures how well knowing one distribution shrinks our uncertainty about another. Again using maximum likelihood estimation, the MI be- tween a clustering C and its "ground truth" clustering T . is I(C,T ) = C ∩ T N C ∩ T CT log C∈C T∈T N (cid:88) (cid:88) It can be understood as the Kullback -- Leibler divergence from the joint distribution Pr(Cc, Tt) to the product of its marginal distributions Pr(Cc) Pr(Tt). Like entropy, the value is nonnegative. It can be normalized by dividing by an upper bound. We will contrast the choices for this bound in section V. IV. WHAT WE'VE BEEN DOING WRONG Community detection is historically evaluated with NMI: NMI(C,T ) = I(C,T ) M (C,T ) . 3 V. RECIPE FOR PROPER EVALUATION Our recommended solution incorporates both practi- cal scoring concerns and an improvement in probability notions over the de facto preference. Desires for an ad- justment for chance to give a constant baseline and a constant top score are longstanding, as shown by the pop- ular adjusted Rand index [15]. The final recommendation is the adjusted mutual information (AMI), computing its expectation with the one-sided random model M1 all. To be useful to practitioners, an evaluation measure for community detection should have the properties of a constant baseline and a constant top score. a. Constant baseline In the statistical sense, we would like a consistent baseline: A random guess should merit no credit. This differs from the notion of a baseline in NMI, which is a lower bound. With this consistent baseline, we eliminate the "free lunch". The proposal of Zhang [6], relative normalized mutual information (rNMI), improves NMI by subtracting the expected NMI for this ground truth, such that a random guess garners a score of 0. In rNMI the expected NMI is computed under the permutation model (C(cid:48) ∼ Mperm(C)): b. Constant top score The flaw of rNMI is that now only one reference point (the expectation) is fixed across clusters, not two like in NMI. This means that we can't compare performance across clusterings of different sizes, and we don't know whether we've succeeded by attain- ing the maximum value. Zhang et al. [5] renormalize rNMI to create renormalized relative normalized mutual information (rrNMI): rrNMI(C,T ) = rNMI(C,T ) rNMI(T ,T ) . (9) . (6) rNMI(C,T ) = NMI(C,T ) − EC(cid:48) [NMI(C(cid:48),T )] . (8) (7) Now, we have both a constant expectation and a constant ceiling, remedying the flaw we've noted. The quantity is unitless: by normalizing, we divide nats by nats or bits by bits, based on our choice of logarithm. Vinh et al. [9] and Zhang [6] both note a finite size effect: the average score creeps upward with the number of pre- dicted clusters, regardless of the true number. This biases results toward the prediction of a large number of clusters, which is a danger to adequate CD evaluation. A related flaw noted by Peel et al. [3] is that the measure is not homogeneous. Much the center of a circle will be closest on average to each point in it, the trivial partition into singleton communities scores highest under NMI, when averaged over all possible ground truths. As a homoge- neous measure is a precondition of the No Free Lunch theorem [16], NMI in fact awards "free lunch" when guess- ing the singletons partition. We will later show practical consequences of this deficiency. c. Symmetry While we take the controversial stance that symmetry in the measure is undesirable, it is nec- essary to contextualize additional work, and for showing the connections between methods in section VI. Note that rrNMI's denominator depends only on T . The in- formation, though, of one distribution about another is not asymmetric. Lai and Nardini [2] symmetrize rrNMI to create the corrected normalized mutual information (cNMI): cNMI(C,T ) = rNMI(C,T ) + rNMI(T ,C) rNMI(C,C) + rNMI(T ,T ) . (10) Another measure, adjusted mutual information (AMI) [4, 9], was devised years earlier and incorporates all of these fixes in the style of the adjusted Rand index [15]: 4 AMI(C,T ) = I(C,T ) − EC(cid:48),T (cid:48) [I(C(cid:48),T (cid:48))] maxC(cid:48),T (cid:48) I(C(cid:48),T (cid:48)) − EC(cid:48),T (cid:48) [I(C(cid:48),T (cid:48))] , (11) where the variables C(cid:48) ∼ Mperm(C) and T (cid:48) ∼ Mperm(T ) under the max and E operators above. A. Probability improvements a. Adjustment considers all relevant partitions All of NMI's successors share a fault. Their expectations are computed by incorrectly considering only partitions with the same decomposition pattern, or bag of cluster sizes. These aren't the only partitions that could exist, though; the space of options is larger. A community detection algorithm could partition two nodes into either {1},{2} or {1, 2}, and ignoring that fact makes assumptions about the structure of the possibility space. We should prefer Mall. Fortunately, simple closed-form approximations of its expectation are known [8]. b. Adjustment recognizes that the ground truth is fixed. Fixing the first fault in our symmetric measures leaves a second fault. We compute the expectation over all clusterings and all truths in the baseline B(C,T ). But the truth for our problem is fixed. A different truth means we're solving a different problem -- core -- periphery, etc. Our expectations should only ever be over the values of C: all(C,T ) = EC(cid:48)∼Mall(T )[I(C(cid:48),T )] . B1 (12) This all too is easy to plug into our formulas, as straight- forward closed-form approximations again exist [8]. The textbook AMI (before these probability improve- ments) is known to be asymptotically homogeneous, im- parting negligible advantage for weak guesses in the limit as N approaches infinity [3]. By correcting for chance properly, we can make this asymptotically diminishing advantage exactly zero [17]. B. Bonus: Generalized mean There's a final component to our recommendation: How do we compute the upper bound function M (C,T )? Tra- ditionally, different generalized means of the cluster en- tropies have been used (e.g., Equation 3 uses the geo- metric mean) [4]. Yang et al. [18] have discussed that the particular choice of generalized mean is unimportant. 4 The bound function in this AMI definition is M (C, T ) = maxC(cid:48),T (cid:48) I(C(cid:48), T (cid:48)). In practice we could use any of the upper bounds as described in [8], for example Equation 3, as long as it is a upper bound consistent with the chosen random model (here Mperm). 4 FIG. 1. Heatmap of mutual information between two parti- tions. AMI computes expectation over all (X(cid:48), Y (cid:48)) pairs in the joint distribution. cNMI examines only the observation's row and column. In Mperm (shown), exchangeability under MI makes these expectations equal. Fortunately, the problem is simpler in our case: Gates and Ahn [8] discuss that in Mall, the bounding entropy for either cluster is log N , and any generalized mean of log N and log N is log N .5 Hence we define our upper bound function in Mall to be Mall(C,T ) = log N . (13) VI. RELATIONSHIPS BETWEEN MEASURES Having waded through the alphabet soup of information-theoretic measures, we now present our major theoretical result: explorations of interrelations between measures. While we began this with our discussion of generalized means, these relationships exploit the random models and sidedness from section V. The relationships are summarized in Figure 2. To characterize these relationships, we rely on the notion of function specialization: limiting a function's expres- siveness by fixing some parameters, altering its behavior [19, 20]. As the simplest example, one-sided measures in general specialize their two-sided counterparts by fixing the ground truth T , rather than taking an expectation over T 's universe. Our first finding is that in Mperm, AMI specializes itself: One-sided and two-sided AMI are equivalent. We demonstrate this using the exchangeability of partitions under permutation in section A. For a graphical intuition, we redirect the viewer to Figure 1. 5 By the generalized mean inequality. 5 FIG. 2. Relationship between various information-based CD measures. Building on this, we also show that rrNMI (an inher- ently one-sided, asymmetric measure) specializes AMI in Mperm. We exploit a clever form of 1: Dividing the numerator and denominator of rrNMI both by the max- imum MI renders the expression identical to one-sided AMI. This does not generalize to other random models because it relies on the exchangeability of partitions. Oddly, rrNMI also specializes cNMI into one-sided cNMI in all random models. This is a straightforward algebraic result. It becomes clear by expressing both measures in terms only of mutual information and its expectation. This may lead the reader to suspect that cNMI and AMI are equivalent, at least under some random model. Unfortunately, under all three of the Gates and Ahn [8] models, there are irremediable differences. In Mperm, the denominators compute expectations over different sets. In Mall and Mnum, the lack of exchangeability renders the numerators distinct. As a final conceptual introduction, we review the me- diant, or "freshman sum", of two fractions. It is given by separately adding the numerators and denominators of two fractions. It has the property that it lies between its two arguments. The cNMI symmetrizes rrNMI as the mediant of each one-sided measure, so we know that the cNMI is always bounded between the one-sided values. Finally, because the numerators are identical in each one- sided variant, the mediant is also the harmonic mean; this can be shown through simple algebraic manipulation. Our recommended measure, one-sided AMI in Mall, is identical to rrNMI in Mall. Though rrNMI is already closer to an appropriate community detection evaluation, we follow parsimony and historical precedent to use the name AMI. Therefore our recommended measure, the one-sided AMI in Mall, by combining our derived baseline function B1 all in Equation 12 and upper bound function Mall in Equation 13, could be formally written as AMI1 all(C,T ) = I(C,T ) − EC(cid:48) [I(C(cid:48),T )] log N − EC(cid:48) [I(C(cid:48),T )] , (14) where C(cid:48) ∼ Mall(T ) = Mall(C). FIG. 3. The rate of gridlock for various mixing parameters µ and network sizes. 25 LFR graphs were sampled for each combination of parameters. VII. A VIEW TO A TRAP Having given the design of a fitting extrinsic evaluation, we now extol its need. Community detection methods can be viewed, at their core, simply as efficient heuristics for maximizing modularity while remaining tractable. In choosing a method, we first select graphs which we be- lieve represent the distribution of graphs in our use-case. We then apply our methods to the graphs, scoring their predictions against the ground truth and choosing the method with best performance. (We may also factor run- ning time into our decision -- a cost-aware objective.) But NMI's scores are misleading, making it a danger to our method selection process. To showcase the danger, we set and spring a trap for NMI. We use a community detection method with a unique architecture, making it susceptible to choosing trivial clusterings. This elicits NMI's pathology: NMI exaggerates the method's performance compared to other methods, awarding credit to the trivial clustering. AMI corrects this exaggeration. VIII. THE TRAP: LEXIMIN, AN ADVERSARIAL METHOD The leximin method is a divisive clustering method, motivated by congestion in traffic networks. It formu- lates community detection as a hierarchical version of the sparsest cut problem, an objective related to modu- larity. (Notably, both reduce to MAX-CUT [21], making them NP-hard. Veldt et al. [20] show that both objectives specialize an underlying function.) The sequence of cuts comes from a linear programming (LP) relaxation of the sparsest cut problem: the maximum concurrent flow problem [22]. The model routes traffic flow between all pairs of nodes, with an objective that minimizes congestion while fairly satisfying demand [22]. (In applications, the traffic could be goods in a commodity distribution network, gossip in a social network, or signals in a neural network.) When rerouting cannot avoid con- NMIAMIrNMIrrNMIcNMIRelative, Normalized, SymmetricRelativeOne-sidedNormalizedSymmetricOne-sided(Adjusted Rand index)gestion, the saturated edges define a sparse "bottleneck cut". Continuing to increase the allocation of flow gives a sequence of bottleneck cuts which dissect the graph [21]. Or at least, that would be nice. Lamentably, the tractability comes from weak duality between hierarchical sparsest cut and the hierarchical maximum concurrent flow problem. The relaxation will either find the spars- est cut or a multipart cut (a grid ) when the maximum throughput is less than the density of any cut. This is consistent with LPs being in the complexity class P while sparsest cut is NP-hard. The competing forces of cut density and the gridlock bound mean that the graph may splinter into single-node communities without any mesoscopic structure in between. We exploit the fact that complete gridlock becomes reliable for graphs of cer- tain structural properties examined below, as shown in Figure 3. IX. SPRINGING THE TRAP: EXPERIMENT Following Yang et al. With the trap in place, we now spring it with an ex- periment of the fashion described above. We benchmark the NMI and AMI of eight popular CD methods, plus the adversarial case: the leximin method. As our graph dis- tribution, we chose the Lancichinetti -- Fortunato -- Radicchi (LFR) benchmark graphs, which manifest a ground truth and obey properties of real-world networks [24]. [18], we test 25 LFR re- alizations for each combination of parameters N ∈ {80, 100, . . . , 240} and µ ∈ {0.03, 0.09, . . . , 0.75}. Here, µ is the mixing parameter. It controls the fraction of each node's edges that connect outside of its community, and it can be thought of as a knob to increase the difficulty of the community detection task. All other parameters were set as in Yang et al. [18]. We report the mean and standard deviation of each score for each combination of parameters. The eight methods we compare against are: Fast- greedy [25], Infomap [26, 27], label propagation [28], lead- ing eigenvector [29], multilevel [30], spinglass [31], and walktrap [32]. For a concise description of each method, the reader is directed to Yang et al. [18]. Implementation details The leximin method is imple- mented in AMPL [33]. While the hierarchical MCFP could be expressed as a single LP, we exploit the lexicographic problem structure to decompose the problem into a se- quence of N − 1 smaller, subordinate LPs [34]. Using AMPL's CPLEX backend, the method solves the LP of Dong et al.'s triples formulation of the MCFP [35]. The eight other algorithms are implemented in the igraph package, accessed through igraph's R interface. Determining opti- mal modularity is built-in for the igraph algorithms; it is done using the networkx Python library [36] for leximin. Cluster evaluation is done for all using the NMI and AMI functions of the scikit-learn Python package [37]. 6 X. POSTMORTEM OF A TRAPPED MEASURE: DISCUSSION AND RESULTS The most useful graph characteristic to capture the performance of a community detection algorithm is the strength of its communities [18]. One proposed distinction of communities is into "strong" and "weak" [23]. Strong communities are tight-knit: their nodes all have more edges to each other (intra-community) than to other com- munities' nodes (inter-community). In the LFR generator, the strength of communities is controlled en masse by the mixing coefficient, µ. Values closer to 1 indicate a harder problem, edging us toward a detectability limit be- yond which no method (real or hypothetical) can identify communities. We present the mean and standard deviation of our scores in Figure 4. Those graphs which took longer than 3 hours to process are excluded, though all graphs with n ≥ 180 and µ ≥ 0.5 were processed within this bound. The surprising fact that larger graphs were processed faster comes from our column generation scheme: When gridlock splinters the graph into singletons at an early stage, we've reached our optimum, and the remaining O(N ) LPs need not be solved -- a tremendous reduction in computational labor. Cranking the mixing coefficient µ higher makes commu- nities weaker. Our tests revealed, as expected, that the leximin method's NMI begins to surpass the others, stay- ing high when they fall. Infomap and Label Propagation plunge down to 0, while Leading Eigenvector and Fast- greedy have gradual downward slopes, never performing nearly as well as the others. More typically, Multilevel, Spinglass, and Walktrap remain high while communities are strong, then fall near the boundary of µ = 0.5. These results, though, are very misleading. Remember from Figure 3 that gridlock occurs almost consistently when we cross the threshold into weak communities. This means that no mesoscopic structure is detected. Why should this be scored higher than methods that extract some of the embedded patterns? When partitions are scored using AMI, most methods' performance profiles are unchanged. The Louvian (mul- tilevel) method and Infomap, for instance, continue to perform strongly on strong communities and drop off at the same values of µ. The edge betweenness algorithm's performance drops faster as µ → 1, but its trend is similar to the trend under NMI. The same is true for Fastgreedy method. By contrast, AMI gives a completely different characterization of the leximin method's performance. Against the backdrop of this relative consistency, we see a dramatic drop in the performance of the leximin method, lining up with the increased rate of predicting singleton clusterings. Leximin's score for high µ is zero. This is a fairer assessment of the methods' ability, matching our intuition about how good one clustering should be com- pared to another -- particularly, the fact that the singleton clustering finds no community structure, so we should regard this as a failure when community structure does 7 (a) NMI, Multilevel (Louvian) (b) NMI, Edge betweenness (c) NMI, Leximin (d) AMI, Multilevel (Louvian) (e) AMI, Edge betweenness (f) AMI, Leximin FIG. 4. The mean and st. dev. of NMI (top) and AMI (bottom) as a function of the mixing parameter, µ. The colors represent different network sizes. The vertical line at µ = 0.5 divides strong and weak communities [23], and the dotted line shows the maximum score, 1.0. exist. After a broad study on factors affecting community detection, Yang et al. [18] proposed a method for choosing a community detection algorithm. Based on runtime and NMI, the authors recommended certain community detection algorithms for particular combinations of N and µ. If we followed their method precisely, then the leximin method would be recommended for small graphs with weak communities. With the truth quantified by AMI, we see that leximin should only be used for networks with strong communities, and its runtime limits its application to small to medium-sized graphs. XI. RELATED WORK Over the years, numerous evaluation methods for com- munity detection have been suggested. While NMI has historically been the most popular (first proposed for community detection by Danon et al. [38]), others include the variation of information [39], the V-measure [40] (a rediscovery of NMI), the adjusted Rand index (ARI) [15], F-score [41], and Cohen's κ [42]. The standardized mutual information [43] was proposed as another adjustment for chance. Using instead the variance as the denominator, it no longer has the constant top score we seek. Romano et al. [44] identified both AMI and ARI as specializations of a general function. Finally, Decelle et al. [45] proposed a measure called overlap that is adjusted for chance. We favor AMI because it is most similar to the conventional measure and also meets our desiderata. XII. CONCLUSION AND FUTURE WORK We see that the leximin method, like many, is suc- cessful on strong communities and is incorrectly ap- praised by NMI. While NMI was the best known measure when Danon et al. first used it for CD [38], we see that NMI can exaggerate community detection perfor- mance. Inertia in the CD community is likely why a departure from NMI has not occurred. Nevertheless, the measure's flaws demand a move to a more robust mea- sure. The choice of AMI corrects this exaggeration, so we encourage AMI's use in the CD community moving forward. The curious reader may note that mutual information is already a symmetric measure. An asymmetric mea- sure like relative entropy makes a promising avenue for future work. Another important avenue is finding fair and appropriate measures for hierarchical or overlapping clus- tering [46]. Finally, future work will assess CD methods and the related k-way partition problem using one-sided AMI under Mall and Mnum respectively, as well as varying the generalized mean parameter p. 8 [1] D. W. Matula, in Classification and clustering, edited by Rev. E 78, 046110 (2008). J. van Ryzin (Elsevier, 1977) pp. 95 -- 129. [25] A. Clauset, M. E. J. Newman, and C. Moore, Phys. Rev. [2] D. Lai and C. Nardini, Journal of Statistical Mechanics: E 70, 066111 (2004). Theory and Experiment 2016, 093403 (2016). [3] L. Peel, D. B. Larremore, and A. Clauset, Science Ad- vances 3 (2017), 10.1126/sciadv.1602548. [4] N. X. Vinh, J. Epps, and J. Bailey, Journal of Machine Learning Research 11, 2837 (2010). [5] J. Zhang, T. Chen, and J. Hu, Journal of Statistical Mechanics: Theory and Experiment 2015, P03009 (2015). [6] P. Zhang, Journal of Statistical Mechanics: Theory and Experiment 2015, P11006 (2015). [7] A. D. McCarthy, Gridlock in Networks: The Leximin Method for Hierarchical Community Detection, Master's thesis, Southern Methodist University (2017). [26] M. Rosvall and C. T. Bergstrom, Proceedings of the National Academy of Sciences 104, 7327 (2007), https://www.pnas.org/content/104/18/7327.full.pdf. [27] M. Rosvall, D. Axelsson, and C. T. Bergstrom, The European Physical Journal Special Topics 178, 13 (2009). [28] U. N. Raghavan, R. Albert, and S. Kumara, Phys. Rev. E 76, 036106 (2007). [29] M. E. J. Newman, Phys. Rev. E 74, 036104 (2006). [30] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, Journal of Statistical Mechanics: Theory and Experiment 2008, P10008 (2008). [31] J. Reichardt and S. Bornholdt, Phys. Rev. E 74, 016110 [8] A. J. Gates and Y.-Y. Ahn, Journal of Machine Learning (2006). Research 18, 1 (2017). [9] N. X. Vinh, J. Epps, and J. Bailey, in Proceedings of the 26th Annual International Conference on Machine Learning, ICML '09 (ACM, New York, NY, USA, 2009) pp. 1073 -- 1080. [10] M. E. J. Newman, Proceedings of Sciences 103, tional Academy https://www.pnas.org/content/103/23/8577.full.pdf. of 8577 the Na- (2006), [11] U. Brandes, D. Delling, M. Gaertler, R. Gorke, M. Hoe- and D. Wagner, arXiv preprint fer, Z. Nikoloski, physics/0608255 (2006). [12] T. N. Dinh, X. Li, and M. T. Thai, in Proceedings of the 2015 IEEE International Conference on Data Mining (ICDM), ICDM '15 (IEEE Computer Society, Washington, DC, USA, 2015) pp. 101 -- 110. [13] J. F. C. Kingman, Journal of the London Mathematical Society s2-18, 374 (1978). [14] B. Hauer and G. Kondrak, Transactions of the Association for Computational Linguistics 4, 75 (2016). [15] L. Hubert and P. Arabie, Journal of Classification 2, 193 (1985). [16] D. H. Wolpert, Neural Computation 8, 1341 (1996), https://doi.org/10.1162/neco.1996.8.7.1341. [17] A. D. McCarthy, in The 7th International Conference on Complex Networks & Their Applications: Book of Abstracts, edited by H. Cherifi, L. M. Aiello, C. Cherifi, P. Lio, R. Lambiotte, and L. M. Rocha (Cambridge, UK, 2018) pp. 166 -- 168. [18] Z. Yang, R. Algesheimer, and C. J. Tessone, Scientific Reports 6, 30750 (2016). [32] P. Pons and M. Latapy, in Computer and Information Sciences - ISCIS 2005, edited by p. Yolum, T. Gungor, F. Gurgen, and C. Ozturan (Springer Berlin Heidelberg, Berlin, Heidelberg, 2005) pp. 284 -- 293. [33] R. Fourer, B. W. Kernighan, Management Science 36, 519 (1990), https://doi.org/10.1287/mnsc.36.5.519. D. M. Gay, and [34] V. Podinovskii, USSR Computational Mathematics and Mathematical Physics 12, 249 (1972). T. E. V. Olinick, [35] Y. Dong, and D. W. Matula, Networks 65, https://onlinelibrary.wiley.com/doi/pdf/10.1002/net.21583. 68 Jason Kratz, (2015), [36] A. A. Hagberg, D. A. Schult, and P. J. Swart, in Proceed- ings of the 7th Python in Science Conference (SciPy2008) (Pasadena, CA USA, 2008) pp. 11 -- 15. [37] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Journal of Machine Learning Research 12, 2825 (2011). [38] L. Danon, A. Daz-Guilera, J. Duch, and A. Arenas, Journal of Statistical Mechanics: Theory and Experiment 2005, P09008 (2005). [39] M. Meila, in Learning Theory and Kernel Machines, edited by B. Scholkopf and M. K. Warmuth (Springer Berlin Heidelberg, Berlin, Heidelberg, 2003) pp. 173 -- 187. [40] A. Rosenberg and J. Hirschberg, in Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Lan- guage Processing and Computational Natural Language Learning (EMNLP-CoNLL) (2007). [41] I. S. Dhillon, S. Mallela, and R. Kumar, Journal of [19] J.-G. Young, G. St-Onge, P. Desrosiers, and L. J. Dub´e, machine learning research 3, 1265 (2003). Phys. Rev. E 98, 032309 (2018). [42] X. Liu, H.-M. Cheng, and Z.-Y. Zhang, arXiv preprint [20] N. Veldt, D. F. Gleich, and A. Wirth, in Proceedings of the 2018 World Wide Web Conference, WWW '18 (Interna- tional World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, Switzerland, 2018) pp. 439 -- 448. [21] D. W. Matula and F. Shahrokhi, Discrete Applied Math- ematics 27, 113 (1990). Radicchi, [22] F. Shahrokhi and D. W. Matula, J. ACM 37, 318 (1990). Cecconi, [23] F. V. Loreto, the National Academy of Sciences 101, 2658 (2004), https://www.pnas.org/content/101/9/2658.full.pdf. and D. Parisi, Proceedings of Castellano, C. F. arXiv:1807.01130 (2018). [43] S. Romano, J. Bailey, V. Nguyen, and K. Verspoor, in Proceedings of the 31st International Conference on Machine Learning, Proceedings of Machine Learning Re- search, Vol. 32, edited by E. P. Xing and T. Jebara (PMLR, Bejing, China, 2014) pp. 1143 -- 1151. [44] S. Romano, N. X. Vinh, J. Bailey, and K. Verspoor, J. Mach. Learn. Res. 17, 4635 (2016). [45] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, Phys. Rev. E 84, 066106 (2011). [46] D. Horta and R. J. G. B. Campello, J. Mach. Learn. Res. 16, 2949 (2015). [24] A. Lancichinetti, S. Fortunato, and F. Radicchi, Phys. [47] C. Schumacher, M. D. Vose, and L. D. Whitley, in Pro- ceedings of the 3rd Annual Conference on Genetic and Evolutionary Computation, GECCO'01 (Morgan Kauf- mann Publishers Inc., San Francisco, CA, USA, 2001) pp. 565 -- 570. This value is a row average in Figure 1; more broadly, it is a one-sided expectation for the randomness model Mperm. What remains is to show that this average matches the two-sided average, which relies on simple algebra. 9 EC(cid:48),T (cid:48) [I(C(cid:48),T (cid:48))] = ET (cid:48) [EC(cid:48) [I(C(cid:48),T (cid:48))]] = ET (cid:48)[K] = K . Thus, the two-sided expectation matches the one-sided expectation. Appendix A: Self-specialization of AMI We show that the one-sided AMI and two-sided AMI are identical under Mperm -- that AMI specializes itself. Superficially, the differences between the two equations are in the expectation and the max-term -- both are over different distributions. By showing that each of these are in fact identical between the equations, we will show that AMI self-specializes. For this, we will discuss clusterings C and T on set X. The upper bound function is straightforward; it is some generalized mean Mp of the two clusterings' entropies H(C) and H(T ). This bound is unchanged whether using one-sided or two-sided AMI, because the entropy of a clustering depends only on its decomposition pattern -- not the actual elements. So much for the upper bound function. Now, we must show that the expectation is the same regardless of whether one or both variables is bound.6 Formally, EC(cid:48),T (cid:48)[I(C(cid:48),T (cid:48))] = EC(cid:48)[I(C(cid:48),T )] given C(cid:48) ∈ Mperm(C) and T (cid:48) ∈ Mperm(T ). We note that either variable may be fixed due to the symmetry of mutual information, and the claim will still hold. We will first show that the expectation is impervious to permutations of the input, then we will use this resilience to show that the expectation is the same for one- or two-sided expectations. It is clear from the definition of mutual information (Equation 6) that it is permutation-invariant: it relies on intersections and sizes of sets, and in aggregate these are unaltered by permuting the labels. Now, we rely on the following theorem: Theorem 1 (Schumacher et al. [47]). For any set X, given partition T1 and T2 on X with the same shape, another partition C on X (not necessarily the same shape as T1), and any permutation-invariant function F (·,·) on pairs of partitions, we have EC(cid:48)∼Mperm(C)[F (C(cid:48),T1)] = EC(cid:48)∼Mperm(C)[F (C(cid:48),T2)] . This amounts to saying that the expectation is inde- pendent of the particular fixed variable (C or T ) so long as the fixed variable is drawn from Mperm. That is: EC(cid:48)[I(C(cid:48),T1)] = EC(cid:48)[I(C(cid:48),T2)] = K. (A1) 6 The reader may recall Figure 1 if a concrete version of the problem is helpful.
1310.3138	1	1310	2013-10-11T14:15:41	Dynamiques globales et locales dans un r\'eseau de t\'el\'ecommunications	[ "cs.SI" ]	Traditional network generation models attempt to replicate global structural properties (degree distribution, average distance, clustering coefficient, communities, etc.) through synthetic link formation mechanisms such as triadic closure or preferential attachment. In this work, we study the evolution of a very big communication network coming from mobile telephony and we analyse the link formation process. A first study conducted on the standard mechanisms allows observing that several mechanisms are responsible for the properties observed in this network. In a second study, we characterize more precisely the link formation process by searching for correlations between the probability of creating a new link and some individual properties such as the degree, the clustering coefficient and the age of the nodes.	cs.SI	cs	Dynamiques globales et locales dans un ré- seau de télécommunications Erick Stattner et Martine Collard Laboratoire LAMIA Université des Antilles et de la Guyane FRANCE {estattne,mcollard}@univ-ag.fr RÉSUMÉ. Les modèles classiques de génération de réseaux tentent de reproduire des proprié- tés structurelles globales (distribution des degrés, distance moyenne, coefficient de clustering, communautés, etc.) observées sur les réseaux du monde réel, par le biais de mécanismes de for- mation de liens synthétiques tels que la fermeture triadique ou l'attachement préférentiel. Dans ce travail, nous étudions l'évolution d'un très grand réseau de communications en télépho- nie mobile et nous analysons le processus de formation des liens. Une première étude permet d'observer que plusieurs mécanismes sont à l'origine des propriétés globales observées sur ce réseau et qu'ils ne correspondent pas tous au schéma classique. Dans une seconde étude, nous caractérisons plus finement le processus de formation des liens en recherchant des corrélations entre la probabilité de créer un nouveau lien et certaines propriétés individuelles telles que le degré, le coefficient de clustering et l'âge des noeuds. ABSTRACT. Traditional network generation models attempt to replicate global structural prop- erties (degree distribution, average distance, clustering coefficient, communities, etc.) through synthetic link formation mechanisms such as triadic closure or preferential attachment. In this work, we study the evolution of a very big communication network coming from mobile tele- phony and we analyse the link formation process. A first study conducted on the standard mechanisms allows observing that several mechanisms are responsible for the properties ob- served in this network. In a second study, we characterize more precisely the link formation process by searching for correlations between the probability of creating a new link and some individual properties such as the degree, the clustering coefficient and the age of the nodes. MOTS-CLÉS : Réseaux sociaux, réseaux complexes, évolution des réseaux, dynamique des ré- seaux, propriétés macroscopiques, réseaux de télécommunications KEYWORDS: Social networks, complex networks, network evolution, network dynamics, macro- scopic properties, telecommunication networks Conférence MARAMI. Volume /2013, pages 1 à 12 2 Conférence MARAMI. Volume /2013 1. Introduction Les réseaux du monde réel ont souvent des structures topologiques non-triviales qui émergent des interactions entre les individus, on parle ainsi de réseaux com- plexes [BOC 06]. Ces dernières années, l'étude de l'évolution de ces réseaux a été un domaine de recherche très actif [BAR 99, DOR 02, BOG 03, STA 12]. En effet, de nombreux modèles de génération ont été proposés pour reproduire les propriétés topologiques particulières observées sur les réseaux du monde réel (scale-free, petit- monde, structure communautaire forte, etc.). La plupart de ces modèles sont basés sur l'hypothèse que ces propriétés globales sont le résultat de mécanismes d'évolution élémentaires gouvernant les changements structurels survenant au sein de l'ensemble du réseau [BAR 02, KUM 07, TOI 09]. Par exemple, des réseaux possédant des coefficients de clustering élevés sont reproduits en appliquant le mécanisme dit de fermeture triadique [OPS 11], c'est-à-dire privilégiant la connexion entre des noeuds possédant des voisins communs. De même, partant d'un réseau régulier, Watts et Strogatz ont proposé un modèle [WAT 98] basé sur une formation aléatoire des liens pour générer des réseaux de type petit-monde. Les réseaux scale-free peuvent, eux, être reproduits en utilisant le modèle de Barabasi et Albert [BAR 99] qui introduit la notion d'attachement préférentiel, c'est-à-dire que le processus de création de liens privilégie la formation avec des individus déjà fortement connectés. Toutefois, bien que ces mécanismes permettent d'obtenir des réseaux possédant des propriétés structurelles globales identiques à celles observées sur les réseaux com- plexes, ils ne reflètent pas les changements microscopiques réels intervenant sur les réseaux du monde. Il est en effet peu probable que tous les individus d'un même ré- seau créent des liens selon le même principe. C'est la raison pour laquelle d'autres travaux se sont intéressés à des modèles de génération faisant intervenir plusieurs mé- canismes d'évolution. Toivonen et al. [TOI 09] ont par exemple comparé plusieurs modèles combinant fermeture triadique, attachement préférentiel et formation aléa- toire, et ont montré qu'ils étaient en mesure de reproduire des réseaux de natures diverses. Ainsi, les études récentes menées sur l'évolution de réseaux réels ont montré que les changements structurels étaient dirigés par une combinaison de mécanismes aussi bien endogènes, qu'exogènes au réseau, liés notamment aux comportements indivi- duels des noeuds [KOS 06, COL 12]. Par exemple, en étudiant plusieurs réseaux issus de sites d'échanges et de partages (Flickr, Delicious, Answers et LinkedIn), Leskovec et al. [LES 08] ont pu observer que les mécanismes traditionnellement utilisés pour la génération de réseaux complexes, tels que la fermeture triadique ou l'attachement préférentiel, s'observaient finalement très peu. Selon eux, "la nature intrinsèquement non-local de l'attachement préférentiel est fondamentalement incapable de capturer les caractéristiques importantes de ces réseaux". Dans ce travail, nous étudions l'évolution d'un très grand réseau de télécommuni- cations issus de la téléphonie mobile et nous analysons le processus de formation de Dynamiques globales et locales 3 liens. Notre objectif est de comprendre et de caractériser, qualitativement et quantita- tivement, les dynamiques impliquées dans la formation. Ainsi, une première étude menée sur l'évolution globale du réseau permet d'effec- tuer deux observations intéressantes : (i) comme de nombreux réseaux du monde réel, ce réseau conserve la propriété "scale-free" durant son évolution, (ii) les mécanismes de formations élémentaires, couramment utilisés pour simuler l'évolution des réseaux complexes ne sont impliqués que dans de faibles proportions dans l'évolution de ce réseau. Dans une deuxième étude, nous cherchons à caractériser plus finement le pro- cessus de formation en recherchant des corrélations entre la formation des nouveaux liens et certaines propriétés individuelles des noeuds telles que le degré, le coefficient de clustering et l'âge. Cet article est organisé comme suit. La Section 2 est dédiée à l'étude menée sur l'évolution globale du réseau. Dans la Section 3, nous caractérisons plus précisément la formation de liens en recherchant les implications de certaines propriétés locales des noeuds. Enfin, la Section 4 conclut cet article et présente nos travaux futurs. 2. Evolution des propriétés globales Dans une première étude, nous nous sommes intéressés à l'évolution du réseau et nous avons cherché à comprendre si la formation des liens au sein de ce réseau suivait les schémas traditionnellement mis en place dans les modèles de génération. Dans la Section 2.1 nous présentons le jeu de données utilisé et nous montrons comment évolue ses principales propriétés structurelles au cours du temps. Dans la Section 2.2, nous nous intéressons plus précisément à l'apparition des liens du réseau, et nous cherchons à déterminer si ces liens correspondent aux motifs traditionnelle- ment utilisés pour la génération de réseaux complexes. 2.1. Jeu de données et évolution Le jeu de données, fourni par un opérateur de téléphonie mobile local, consiste en un ensemble de fichiers représentant les communications sur 7 mois (Juin à Décembre 2009). Chaque fichier regroupe l'ensemble des communications effectuées sur une journée. Enfin dans un fichier, chaque ligne détaille une communication, en précisant notamment le numéro de l'appelant et de l'appelé, le type d'échange (Appel, SMS ou Fax), la durée de la communication, etc. La première étape de notre travail a consisté à traiter ces données pour en ex- traire le réseau de communications sous-jacent entre les abonnés (limité aux appels et SMS). Dans ce réseau, un noeud représente un abonné, et un lien une communication. Ainsi, nous pouvons étudier comment évolue ce réseau jour après jour en analysant les fichiers séquentiellement. En raison de la quantité de données à traiter, et du temps de calculs associé (ce dernier aspect est détaillé dans la section suivante), nous nous sommes limités pour ce travail préliminaire à l'étude de l'évolution du réseau sur 2 se- 4 Conférence MARAMI. Volume /2013 maines. L'approche que nous avons adoptée était la suivante : le réseau était d'abord construit sur 3 jours, puis analysé sur 15, soit 18 fichiers au total. Dans un premier temps, nous nous sommes intéressés à l'évolution des principales propriétés du réseau : (a) nombre de noeuds #V et de liens #E, (b) nombre de nou- veaux noeuds et de nouveaux liens distincts, (c) densité et (d) coefficient de clustering moyen. La Figure 1 montre comment évolue ces caractéristiques sur 15 jours. (a) (c) (b) (d) Figure 1 -- Évolution des principales propriétés du réseau sur 15 jours (a) nombre de noeuds et de liens (b) nombre de nouveaux noeuds et de nouveaux liens (c) densité et (d) coefficient de clustering moyen. Comme attendu, le nombre de noeuds et de liens connaissent tous deux des évo- lutions croissantes sur la période d'étude (cf Figure 1(a)). Cela s'explique par le fait qu'au début de l'analyse, le réseau considéré, bien que résultat d'un cumul sur 3 jours, ne représente pas toutes les communications existantes et est donc peu fourni. Une grande quantité de noeuds et de liens encore non-identifiés est ainsi insérée au ré- seau. On observe tout de même que cette tendance tend a diminuer pour le nombre de noeuds, puisqu'on constate une phase stabilisation de l'évolution qui peut être ap- prochée par une fonction logarithmique. Ces résultats laissent ainsi penser que deux semaines suffisent pour que la grande majorité des abonnés utilisant leur forfait soit impliquée dans au moins une communication, ce qui permet de les ajouter au réseau. Les résultats obtenus pour le nombre de nouveaux noeuds et de liens confirment nos y = 170966ln(x) + 832249y = -10574x2+ 540039x + 2E+060100000020000003000000400000050000006000000700000080000000123456789101112131415Jours#V #E 01000002000003000004000005000006000000123456789101112131415JoursNouveaux#VNouveaux#Ey = -9E-09x2+ 4E-07x + 5E-064,00E-065,00E-066,00E-067,00E-068,00E-069,00E-060123456789101112131415JoursDensitéy = 0,0141ln(x) + 0,0530,0040,0240,0440,0640,0840,1040123456789101112131415Joursmoy.CCDynamiques globales et locales 5 observations précédentes (cf Figure 1(b)). Bien que le réseau ne fasse que croître, les liens n'étant jamais supprimés, on observe que le nombre de nouveaux noeuds et de nouveaux liens, eux, décroissent. La stabilisation du nombre de noeuds #V observée précédemment se traduit ici par une décroissance très forte du nombre de nouveaux noeuds qui semble tendre vers 0. Ce qui confirme qu'une grande partie des abonnés a été insérée au réseau. Enfin, la densité et le coefficient de clustering moyen (cf Figures 1(a) et (d)) sont tous les deux croissants sur la période. On observe d'ailleurs que le coefficient de clus- tering tend à se stabiliser, ce qui suggère que l'ajout de nouveaux liens crée peu de "triangles" dans le réseau et donc ne tend pas à renforcer l'effet communautaire dans ce réseau. Enfin, pour compléter cette étude sur l'évolution des propriétés, nous nous sommes intéressés à l'évolution de plusieurs mesures relatives au degré des noeuds, de façon à vérifier si ce réseau possédait une structure connue. La Figure 2 montre comment évolue (a) la degré moyen et le degré max et (b) la distribution des degrés. (a) (b) Figure 2 -- Évolution des degrés du réseau sur 15 jours : (a) degré moyen et degré max et (b) distribution des degrés Il est intéressant d'observer l'écart considérable qui existe entre le degré moyen et le degré max (cf Figure 2(a)). Cela laisse supposer qu'il existe un très petit nombre d'individus extrêmement connectés dans ce réseau. Les résultats obtenus pour l'évo- lution de la distribution des degrés (cf Figure 2(b)) confirme cette hypothèse et per- mettent d'observer que le réseau conserve une structure de type scale-free tout au long de son évolution. 2.2. Motifs d'évolution traditionnels L'étude précédente nous a permis d'observer que le réseau de télécommunications étudié conservait une structure scale-free lors son évolution. Or, nous savons que cette propriété est souvent expliquée par le mécanisme dit "d'attachement préférentiel". y = 48,378x + 18,725y = 0,4329x + 5,025020040060080010000123456789101112131415JoursmaxDeg moy.Deg 0800,10,20,30,41357911131517192123252729JoursP(k)k6 Conférence MARAMI. Volume /2013 Ainsi, soit G = (V, E) un réseau, avec V l'ensemble des noeuds et E ⊆ V × V l'ensemble des liens. On note kvi le degré du noeud vi. Selon le modèle proposé par Barabasi et Albert, la probabilité pvi de créer un lien avec un noeud vi s'obtient par : kvi(cid:80) j kvj pvi = [1] où(cid:80) j kvj correspond à la somme des degrés des autres noeuds du réseau. Ainsi, la probabilité qu'a le noeud vi d'être impliqué dans la formation d'un nouveau lien croit proportionnellement avec son degré. Dans un second temps, nous nous sommes intéressés au processus de formation des liens au sein de ce réseau de communications entre abonnées, et plus précisément, nous avons cherché à comprendre si la structure scale-free observée s'expliquait uni- quement par le mécanisme d'attachement préférentiel. Ainsi, pour chaque lien créé durant les 15 jours d'étude, nous avons calculé la probabilité pvi qu'il aurait eu dans le modèle de Barabasi et Albert. La Figure 3 montre la distribution des pvi après 15 jours. Figure 3 -- Distribution des pvi lors de la création de liens sur 15 jours Il est particulièrement intéressant d'observer que le nombre de liens à l'origine des pvi élevés, c'est-à-dire assimilables à de l'attachement préférentiel, est très faible. La grande majorité des créations est en effet associée à des pvi relativement faibles, ce qui suggère que le processus d'attachement préférentiel dans ce réseau s'avère finalement être assez marginal. Ainsi, malgré la propriété scale-free que conserve ce réseau durant son évolution, l'attachement préférentiel ne semble pas être l'unique processus à l'origine de la for- mation des liens. Bien qu'on puisse raisonnablement penser que peu de liens de type attachement préférentiel suffisent à maintenir cette structure scale-free, la question centrale concerne les processus impliqués dans la formation des autres liens. Posons β un seuil de probabilité à partir duquel on considère que le lien est de type attachement préférentiel. Pour mieux comprendre comment évolue ce réseau, nous avons tenté de classifier chaque lien selon les mécanismes de formation traditionnel- lement utilisés dans la littérature : 0500100015002000250030001,00E-091,01E-062,02E-063,03E-064,03E-065,04E-066,05E-067,06E-068,07E-069,07E-061,01E-051,11E-051,21E-051,31E-051,41E-051,51E-051,61E-051,71E-05βDynamiques globales et locales 7 r1 : Booléen ← faux r2 : Booléen ← faux si pvj ≥ β alors r1 ← vrai P A ← P A + 1 algorithme 1 Identification du type de formation de liens Précondition : G = (V, E) : Réseau, β : seuil ∈ [0..1] 1. P A : Entier ← 0 2. T C : Entier ← 0 3. R : Entier ← 0 4. pour tout lien e = (vi, vj) lu dans les fichiers faire si e /∈ E alors 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. fin si 20. 21. fin pour fin si si vi et vj ont au moins un voisin commun alors fin si si r1 = faux ET r2 = faux alors r2 ← vrai T C ← T C + 1 R ← R + 1 fin si Ajouter e à E -- Attachement préférentiel (PA), qui fait référence au mécanisme de formation qui privilégie la création d'un lien avec un individu déjà fortement connecté. -- Fermeture triadique (TC), qui correspond au mécanisme à travers lequel un noeud crée un nouveau lien avec les amis de ses amis. -- Aléatoire (R) : concerne tous les liens qui ne peuvent pas être classés dans une des deux classes précédentes et qu'on suppose aléatoires. L'Algorithme 1 détaille le principe d'identification des liens. Nous pouvons notam- ment observer que le type associé à un lien n'est pas exclusif, puisqu'un lien peut être à la fois identifié comme étant de type attachement préférentiel et de type fermeture triadique. En fixant β = 4, 00E-06 (représenté sur la Figure 3), la Figure 4 montre l'évolu- tion du nombre de liens de type PA, TC et R sur les 15 jours (a) en quantité et (b) en proportion. Comme observé précédemment, on constate que le nombre de liens se formant suivant le mécanisme PA reste finalement assez faible dans le réseau. Seul 20% à 30% des liens sont en effet identifiés comme étant de l'attachement préférentiel. Nous pou- vons également observer que les liens de type TC sont présents dans des proportions similaires. Ce qui nous permet d'affirmer que dans l'évolution de ce réseau, il y a glo- balement autant de liens se formant selon l'attachement préférentiel que de liens se 8 Conférence MARAMI. Volume /2013 (a) (b) Figure 4 -- Évolution du nombre de liens de type PA, TC et R (a) en quantité et (b) en proportion formant selon la fermeture triadique, et ce, quelque soit le moment de l'observation. On peut également noter que ces proportions restent relativement constantes dans le temps. Ces résultats viennent ainsi confirmer deux observations également faites par Leskovec et al. [LES 08], qui montraient qu'environ 30% des nouveaux liens formés dans les réseaux qu'ils étudiaient étaient de type fermeture triadique. L'observation la plus intéressante concerne sans doute le fort pourcentage de liens non-identifiés. En effet, 60% des nouveaux liens ne sont identifiés ni comme étant de l'attachement préférentiel, ni comme étant de la fermeture triadique, ce qui nous amène à nous interroger sur la nature de ces liens, en nous intéressant notamment aux implications de certaines propriétés locales dans le processus de formation. 3. Comportements locaux de formation de liens Pour aller plus loin dans l'étude des nouveaux liens apparaissant dans le réseau, nous nous sommes, dans cette seconde étape, intéressés aux corrélations entre les propriétés individuelles des noeuds et le processus de formation des liens. Nous avons ainsi caractérisé les nouveaux liens, selon trois grandes caractéristiques des noeuds : (1) le degré, (2) le coefficient de clustering et (3) l'âge (calculé en fonction du jour d'apparition du noeud dans le réseau). Pour chacune d'elles, nous définissons trois types de formation des liens, comme décrit ci-dessous : 1) Le degré : - DHH correspond à un lien formé entre deux noeuds hautement connectés au regard du degré moyen, c'est-à-dire un lien e = (vi, vj) est de type DHH si kvi ≥ M oy.Deg et kvj ≥ M oy.Deg. - DLL défini un lien formé entre deux noeuds faiblement connectés au regard du degré moyen, c'est-à-dire un lien e = (vi, vj) est de type DLL si kvi < M oy.Deg et kvj < M oy.Deg. 0500000100000015000002000000250000030000003500000123456789101112131415JoursPA TC R 00,10,20,30,40,50,60,7123456789101112131415Jours%PA%TC % R Dynamiques globales et locales 9 - DHL correspond à un lien formé entre des noeuds mixtes au regard du degré moyen, c'est-à-dire si (kvi < M oy.Deg et kvj ≥ M oy.Deg) ou (kvi ≥ M oy.Deg et kvj < M oy.Deg). 2) Le coefficient de clustering : - CHH correspond à un lien formé entre deux noeuds dont les coefficients de clustering sont supérieurs à la moyenne. Plus précisément, soit cc(vi) le coefficient de clustering du noeud vi, un lien e = (vi, vj) est de type CHH si cc(vi) ≥ M oy.CC et cc(vj) ≥ M oy.CC. - CLL est un lien formé entre des noeuds possédant des coefficients de cluste- ring inférieurs à la moyenne, i.e. cc(vi) < M oy.CC et cc(vj) < M oy.CC. - CHL correspond à un lien formé entre des noeuds mixtes au regard du coef- ficient de clustering. 3) L'age du noeud : - AJJ défini un lien formé entre des noeuds identifiés comme "jeune" dans le réseau. Plus précisément, soient t(vi) le jour où le noeud vi a été ajouté au réseau et t(e) le jour où le lien a été ajouté, un lien e = (vi, vj) est de type AJJ si t(vi) ≥ t(e) − 3 jours et t(vj) ≥ t(e) − 3 jours. - AOO défini un lien formé entre deux noeuds moins récents, c'est-à-dire un lien e = (vi, vj) est de type AOO si t(vi) < t(e)−3 jours et t(vj) < t(e)−3 jours. - AJO correspond à un lien formé entre un jeune noeud et un noeud plus agé. La Figure 5 montre comment évolue (1) en quantitié et (2) en proportion, les liens impliquant (a) le degré (DHH, DLL et DHL), (b) le coefficient de clustering (CHH, CLL et CHL) et (c) l'age du noeud (AJJ, AOO, AJO). Si l'expérience précédente a pu montrer que l'attachement préférentiel n'était pas le mécanisme dominant au sein de ce réseau, il est tout de même intéressant d'obser- ver que la majeure partie des liens se crée entre des noeuds dont au moins un des deux possède un degré supérieur à la moyenne (cf Figures 5(a)). En effet, 40% à 45% des liens se créent suivant DHH, alors que 45% à 50% des nouveaux liens s'avèrent être de type DHL. Seuls 10% des liens sont créés entre des noeuds possédant des degrés inférieurs à la moyenne. Comme observé précédemment, ces proportions restent glo- balement stables durant l'évolution du réseau. On peut expliquer ces résultats par une forme "d'attachement préférentiel local", à travers lequel les noeuds créent des liens avec les individus les plus connectés de leur entourage (ce qui explique le fort pourcentage pour DHH et DHL), mais qui ne sont pas nécessairement les plus connectés globalement d'où le faible pourcentage observé précédemment pour PA. On peut également observer que la grande majorité des liens du réseau se crée entre des noeuds dont au moins un possède un coefficient de clustering inférieur à la moyenne (cf Figure 5(b)). Plus précisément, 60% des liens se créent entre deux noeuds possédant des coefficients de clustering inférieurs à la moyenne et 35% se créent suivant le mécanisme CHL. Seuls 5% des liens se forment entre des individus 10 Conférence MARAMI. Volume /2013 (a) (b) (c) Figure 5 -- Évolution des liens impliquant (a) le degré (DHH, DLL et DHL), (b) le coefficient de clustering (CHH, CLL et CHL) et (c) l'age du noeud (AJJ, AOO, AJO) : (1) en quantitié et (2) en proportion possédant des coefficients de clustering supérieurs à la moyenne. Ces tendances restent stables sur toute la durée de l'étude. En ce qui concerne les liens impliquant l'age du noeud, les tendances sont intéres- santes (cf Figures 5(c)). On peut en effet observer que sur les trois premiers jours de l'étude, les liens de type AJO, c'est à dire impliquant des individus vieux et jeunes, 050000010000001500000200000025000003000000123456789101112131415Jours(1)DHH DLL DHL 00,10,20,30,40,50,6123456789101112131415Jours(2)% DHH %DLL % DHL 01000000200000030000004000000123456789101112131415Jours(1)CHH CLL CHL 00,20,40,60,8123456789101112131415Jours(2)CHH CLL CHL 010000002000000300000040000005000000123456789101112131415Jours(1)AJJ AOO AJO00,20,40,60,81123456789101112131415Jours(2)AJJ AOO AJODynamiques globales et locales 11 Figure 6 -- Évolution du temps de calcul en fonction du nombre de jours analysés sont très fréquents. Par exemple, 90% des nouvelles formations sont de type AJO après le premier jour d'étude. Ce résultat peut s'expliquer par le fait qu'au début de l'étude, de nombreux noeuds sont ajoutés au réseau, ce qui explique le taux de for- mation élevé avec des noeuds déjà présents. Cependant, lorsque l'étude est prolongée, la tendance s'inverse et on observe que la plupart des liens se crée entre des noeuds identifiés comme vieux. Par exemple, après le 7e jour, 60% à 70% des nouveaux liens sont de type AOO, alors que 30% à 20% des liens se créent suivant le mécanisme AJO. Seuls 10% des nouveaux liens connectent des individus récemment ajoutés au réseau. 4. Conclusion Dans ce travail, nous avons analysé l'évolution globale et locale d'un réseau de télécommunications. L'étude menée est intéressante à plusieurs égards. En effet, nous avons pu observer que lors de son évolution, le réseau conserve sa propriété scale-free mais n'est pas dirigé par un mécanisme de formation dominant tel que l'attachement préférentiel. Nous avons ainsi pu mettre en évidence certaines corrélations entre les propriétés individuelles des noeuds et le processus de formation, ce qui nous amène à penser qu'un modèle de génération idéal serait un modèle capable de générer des réseaux possédant des propriétés structurelles particulières, tout en reproduisant loca- lement les tendances de formation observées sur des réseaux réels. Cette étude soulève également un certain nombre de questions sur la généralisation des résultats obtenus. Le réseau que nous avons utilisé est en effet un réseau possédant une sémantique bien particulière, puisqu'il s'agit de communications téléphoniques entre des individus. On peut raisonnablement supposer que des études similaires me- nées sur des réseaux d'autres natures donneraient des résultats différents. De telles études pourraient, à termes, permettre de proposer des modèles de génération ciblés, capables de reproduire l'évolution de réseaux de sémantique diverse : amitiés, colla- boration, etc. y = 0,001e0,3393x0h:00min2h:24min4h:48min7h:12min9h:36min12h:00min123456789101112131415161718JoursTemps d'excecution12 Conférence MARAMI. Volume /2013 Enfin, une des difficultés majeures de ce type d'étude concerne la durée d'ob- servation et le passage à l'échelle. Nous nous sommes en effet limité à l'analyse de l'évolution du réseau sur 15 jours en raison de l'énorme quantité de données à traiter et de la croissance exponentielle du temps de calcul avec la durée d'observation (voir Figure 6 ; résultats obtenus avec Core i7, 3,3GHz, 8Go de RAM, Java 7). Ainsi, un des défis majeurs à long terme consistera à mettre en place, pour l'étude de tels volumes de données, des solutions de collecte, de stockage et d'analyse ef- ficientes de type "big data", dans lesquelles la parallélisation des traitements est le facteur d'optimisation. 5. Bibliographie [BAR 99] BARABASI A., ALBERT R., « Emergence of Scaling in Random Networks », Science, vol. 286(5439), 1999, p. 509 - 512. [BAR 02] BARABASI A. L., Linked : The New Science of Networks, Perseus Books, 2002. [BOC 06] BOCCALETTI S., LATORA V., MORENO Y., CHAVEZ M., HWANG D., « Complex networks : Structure and dynamics », Physics reports, vol. 424, no 4, 2006, p. 175 -- 308, Elsevier. [BOG 03] BOGUNA M., PASTOR-SATORRAS R., DIAZ-GUILERA A., ARENAS A., « Emer- gence of clustering, correlations, and communities in a social network model », Arxiv preprint cond-mat 0309263, , 2003. [COL 12] COLLARD M., COLLARD P., STATTNER E., « Mobility and information flow : per- colation in a multi-agent model », 3rd International Conference on Ambient Systems, Net- works and Technologies, , 2012. [DOR 02] DOROGOVTSEV S., MENDES J., « Evolution of networks », Adv. Phys, , 2002. [KOS 06] KOSSINETS G., WATTS D., « Empirical analysis of an evolving social network », Science, vol. 311, no 5757, 2006, p. 88 -- 90, American Association for the Advancement of Science. [KUM 07] KUMPULA J., ONNELA J., SARAMÄKI J., KASKI K., KERTÉSZ J., « Emergence of communities in weighted networks », Physical review letters, vol. 99, no 22, 2007, page 228701, APS. [LES 08] LESKOVEC J., BACKSTROM L., KUMAR R., TOMKINS A., « Microscopic evolution of social networks », Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM, 2008, p. 462 -- 470. [OPS 11] OPSAHL T., « Triadic closure in two-mode networks : Redefining the global and local clustering coefficients », Social Networks, , 2011, Elsevier. [STA 12] STATTNER E., COLLARD M., VIDOT N., « D2SNet : Dynamics of diffusion and dynamic human behaviour in social networks », Computers in Human Behavior, , 2012. [TOI 09] TOIVONEN R., KOVANEN L., KIVELA M., ONNELA J., SARAMAKI J., KASKI K., « A comparative study of social network models : network evolution models and nodal attribute models », Social Networks, vol. 31, 2009. [WAT 98] WATTS D. J., STROGATZ S. H., « Collective dynamics of 'small-world' networks », Nature, vol. 393, 1998, p. 440-42.
1601.02499	1	1601	2016-01-11T16:06:11	Modeling the Dynamics of Discussions in Social Networks	[ "cs.SI" ]	In this paper, a method of modeling the dynamics of electronic discussions is proposed based on the so called FOPDT model (First Order Plus Dead Time) known from the process control. Knowledge of the model points to possibility of estimating dynamic movements in discussions as well as understanding and designing their maintaining and guidance. Real discussions are processed.	cs.SI	cs	Modeling the Dynamics of Discussions in Social Networks Petr Kl´an, Dept. of System Analysis, University of Economics in Prague, Czech Republic, [email protected] Abstract: In this paper, a method of modeling the dynamics of electronic discussions is pro- posed based on the so called FOPDT model (First Order Plus Dead Time) known from the process control. Knowledge of the model points to possibility of estimating dynamic move- ments in discussions as well as understanding and designing their maintaining and guidance. Real discussions are processed. Key Words: Discussion dynamics, FOPDT model, process identification 1 Introduction Social networks are associated with complex systems. In order to investigate properties of such systems, statistical models are widely used. They ignore individual network nodes and prefer average behavior of large number of nodes that carry out communication. Furthermore, average characteristics of communication networks are given by Zipf law [7]. The latter determines, for example, that frequency of Web sites linked from k other sites (in–links) is proportional to the ratio 1/k2 [6]. In other words, pages linked from a huge amount of the other Web pages, are very rare. A similar type of dependence can be measured e.g. in the electronic discussion group Chminf–l [5]. Frequency of discussions including k discussion posts is approximately proportional to 1/k. As shown in Fig. 1, corresponding relationship was measured in the discussion group Chminf–l for year 2010 like in other years. Discussions including more posts than e.g. 15 are rare in this discussion group. In many cases, the patterns of dynamic behavior in complex systems are given by the size changes in some populations n, n > 0, described by differential equations of type n(t) = bn(t)(1 − n(t)). The latter results in well known logistic curves (t ≥ 0) n(t) = n(0) ebt 1 + n(0)(ebt − 1) , where b denotes size change coefficient and n(0) initial condition [1]. Logistic curves with the characteristic shapes of the letter S represents a common pattern appropriate to describe the dynamics of complex systems, regardless of their origin. In some cases, cummulative dynamics of responses in electronic discussions exhibits the characteristic shape of the logistic curve too, as illustrated in Fig. 2 for Chminf–l discussion carried out on July 22, 2011. Advanced methods for PID (Proportional + Integral + Derivative) control use mathematical process models which have a potential to study dynamic events related to communication in social networks as well. Primarily, there are three–parameter models [2] or FOPDT models (First Order Plus Dead Time) with the corresponding differential equation T y(t) + y(t) = Ku(t − L) represented by the transient solution (t ≥ L) y(t) = K  1 − e −(t − L) T   1 120 100 80 60 40 20 e c n e v k e r F 0 0 5 10 15 k 20 25 30 Figure 1: Illustration of Zipf law. Horizontal axis indicates number of discussion posts and vertical axis frequency of these discussions. 18 16 14 12 10 8 6 4 2 n 0 0 2 4 6 8 10 t [h] 12 14 16 18 20 Figure 2: On the S shape of the logistic curve. Horizontal axis represents time, vertical axis shows growth of discussion post population. 2 given by the unit step response for y(0) = 0. An example of such transient response with K = 27, T = 5 and L = 1 is shown in Fig. 3. Variables y, u indicate input and output and non–negative parameters K, T, L gain, time constant and time delay, respectively. Physical units are not specified in this illustrative example. Using Laplace transform, this model corresponds to a transfer function K e−sL T s + 1 . 30 25 20 y 15 10 5 0 0 5 10 15 20 25 t Figure 3: Three–parameter model step response. The paper shows that the dynamics of communication networks, under certain conditions, can be identified by the three–parameter models. The latter is yet suitable for purposes of prediction and management of discussions using principles associated with PID control. Since the three–parameter models are widely used in automatic control, various rules have been developed for their use. Here, knowledge of parameters K, L, T is decisive. Just widely known model, the intuitive clarity of its meaning, identification and use allow to better understand the communication dynamics in social networks, including such elements of communication as posting, liking etc. For example, use of the specific value of 63% of K at t = T + L is available for estimating sizes of discussions. The following section explains motivation for use of three–parameter model for discussion dynamics description. Further sections are concerned with identification of three–parameter models from cummulative data of discussions. Moreover, some real examples of identification in electronic discussion group Chminf–l are introduced. 3 2 Motivation The author was a participant in one of the discussions on social network, which debated the topic of possible Wikipedia decline despite the noble goal of Wikipedia to assemble all human knowledge. Although a unique system in the history of civilization, the number of active Wikipedia contributors constantly decreases significantly. The principal reason is con- sidered in constraints that community of Wikipedia contributors has created to discourage new contributors. Discussion had a classical format. At the beginning it was an initial post about Wikipedia decline giving the link to the paper [8] to which the participants debated. After the end dis- cussion, the author collected the dynamics of discussion, i.e. the cumulative number of posts y to vertical axis versus related times t. It results in Fig. 4 where the circle indicates the cumulative number of discussion posts at the time. The first post on the subject appeared after about an hour from the initial contribution (transport delay). It is followed by the relatively steep growth of posts to gradually slowed after 24 hours to reach the final steady state of 27 posts. Figure 4: Dynamics of discussion on Wikipedia decline. If observing dynamics of this discussion then one can not ignore the similarity with the exponential dependence of transient response in Fig. 3. The same transient response with parameters K = 27 posts/time, T = 5 hours and L = 1 hour is inscribed by continuous line in Fig. 4. It well characterizes and describe the dynamics of discussion and associates with a principal question if is it possible to describe discussions by the related transfer function 4 27 e−s 5s + 1 or whether the three–parameter models describe the dynamics of electronic discussions in social networks and the Internet in general? 3 Discussion dynamics identification At the beginning, it is necessary to answer the question whether the dynamics of discussion behaves as a transient response to step change? Yes, it does. The initial post from which the debate unfolds, appears and lasts. Thus, it fulfills the same ’shock’ role like unit step expressed by 1/s in the Laplace transform. At the time of this step or the initial post, it is possible to set up the start time t = 0 and (with zero initial condition) to start measurement of transient response as the cumulative number of the posts. The dozens of discussions were examined to verify if the discussion dynamics captured in Fig. 4 does not represent a random but repeatable shape. For this purpose, the freely ac- cessible archive of electronic discussion group for chemical informatics Chminf–l (https: // list.indiana.edu) was selected. The author is a member of this group since 1995 [4]. The group has about 1500 participants and well serve as a model paradigm of communication in social networks because indicating the same patterns introduced by the Zipf law. Discussion subjected as ‘What are faculty instead of using chemistry reference titles?’ from February 2010 represents a characteristic example. The initial post (unit step) was sent at 16:35 (t = 0). The total number of the following posts was 23, which immediately represents gain K associated with the unit step. The dynamics of the discussion is captured in Fig. 5 (irregular curve) together with the unit step response of identified three–parameter model by K = 23 posts/time, T = 5.5 hours and L = 0.5 hour. In order to obtain these parameters, the algorithm of identification described in [2] was used. Therefore, discussion is identified by the transfer function 23 e−0.5s 5.5s + 1 . Discussions with a large number of posts (typically > 10) indicate common patterns in the aforementioned public archive. It examines discussions in private social networks, too. Data from large amount of discussions confirm the hypothesis that the dynamics of the discussion can be interpreted (with sufficient accuracy) as the dynamics of the three–parameter model. In other words, dynamics of a social network discussion can be sufficiently represented by three parameters, K, L and T . Moreover, it appears that while the parameters of L and T in the dynamics of the dis- cussions are not too much different (typical delay currently was within one hour and the time constant from 1 to 5 hours), the parameter K representing the cummulative number of posts significantly varies. Thus, the situation relates to dynamic systems with a highly variable gain. On the other hand, the Fig. 1 suggests that discussions with gain of more than K = 15 are significantly rare in the investigated discussions. 5 25 20 15 10 5 y 0 0 5 10 15 t [h] 20 25 30 Figure 5: Dynamics of discussion in Chminf–l and its identification by three–parameter model. 4 Examples of identification This section shows examples of identification and illustrates situations met in the real prac- tice of discussions. Example 1: Chminf–l 10.3. 2011, beginning: time 15:59, subject: ISTL article . . . , ba- sic time unit: hour, Fig. 6. It is a case of the discussion, which had the standard dynamics and that the last two posts were added after many hours. In the case of international discussions due to different time zones (or e.g. working absence), it is a common phenomenon and it currently represents 10% error given by a noise. It is known [3] that, for example, good PID tuning has to deal with 20% changes of the controlled process parameters. Example 2: Chminf–l 6.1. 2012, beginning: time 1:41, subject: Authors listed in search of ACS Pubs, basic time unit: hour, Fig. 7. The example shows the situation with an extremely large time delay, the process by which the first contribution post to the discussion came after about 12 hours. So long delay is ex- amined in discussions rare and when they do, it is usually given by inadequate time posting of the first discussion post. Example 3: Chminf–l 22.7. 2011, beginning: time 17:23, subject: SciFinder Piracy, basic time unit: day, Fig. 8. Chminf–l 7.11. 2011, beginning: time 20:03, subject: Handbook 92d edition . . . some thoughts, basic time unit: day, Fig. 9. It is illustration of discussions with long time delays and time constants (in days). Fur- thermore, Fig. 8 captures a somewhat chaotic shape of the discussion dynamics, in which, 6 16 14 12 10 8 6 4 2 y 0 0 5 10 15 t [h] 20 25 30 Figure 6: 16 e−2.5s 2.5s + 1 . 18 16 14 12 10 8 6 4 2 y 0 0 5 10 15 t [h] 20 25 30 Figure 7: 17 e−13.1s 2.5s + 1 . 7 despite the large variations, it is possible to recognize the exponential character (and finally the logistic curve). 40 35 30 25 y 20 15 10 5 0 0 2 4 6 t [d] 8 10 12 Figure 8: 36 e−1.5s 2.5s + 1 . Example 4: Chminf–l 22.6. 2011, beginning: time 18:13, subject: Call for papers from Journal of Chemistry and Chemical Engineering, basic time unit: hour, Fig. 10. Capturing a break dynamics discussion (e.g. due to the absence in work), which may occur in the discussion, unlike the technological process. It is interesting to see that the dynamics of the discussion is interrupted while the delay continues. Example 5: Chminf–l 2010, first 5 discussions including 8 posts, Fig. 11. Fig. 11 shows examples of the discussion dynamics with a small number of posts. Such discussions are frequently included in the Chminf–l. In one case, the last post came even after more than 30 hours from the initial post. In Fig. 11, a step response without time delay is experienced as well as a delayed response with extremely small time constants as well as response with a time constant of around 10 hours. All of these responses can be identified by three–parameter model similarly as in previous cases. However, using the predictive power of the model, i.e. prediction of the characteristic time L + T is difficult. On the other hand, the gain K for newly starting discussions can be estimated by the Zipf law (well introduced eg. in [7]). With each new discussion will be possible, based on the subject of discussion, to estimate its probable order associated with the number of posts. 5 Conclusion The paper shows an innovative use of three–parameter FOPDT models in identifying dynam- ics of social discussions. By this view, the cummulative dynamics of discussions is described 8 16 14 12 10 8 6 4 2 y 0 0 2 4 6 t [d] 8 10 12 Figure 9: 16 e−0.3s 1.5s + 1 . 14 12 10 8 6 4 2 y 0 0 5 10 15 t [h] 20 25 30 Figure 10: Discussion under work absences. 9 10 9 8 7 6 5 4 3 2 1 y 0 0 5 10 15 t [h] 20 25 30 Figure 11: Discussions with a small number of posts. as e.g. dynamics of mechanical systems. The paper gives examples of identification of real discussions in the electronic discussion group called Chminf–l. It is shown that dynamics of discussions is well described by the three parameter FOPDT model known from the process control. At the same time, the predictive ability of three– parameter models can be used. In the case of a small number of posts, it is possible to use a combination of the model with the effects of Zipf law. In addition to the prediction and estimation of dynamic properties of discussions, the three– parameter model serves in managed discussions when calculating proportional and integral parts. The proportional part can manifest itself by inserting proportionally opposing views or to removing the offending posts, the integral part then historical experience on the subject, experience from previous discussions, etc. References [1] Kahn P.B., Mathematical Methods for Scientists and Engineers, J. Wiley & Sons, 1990. [2] Kl´an P., Gorez R., Process Control, FCC Public, 2011. [3] Kl´an P. , Gorez R., Balanced tuning of PI controllers, European Journal of Control, Vol. 6, No. 6, pp. 541–550, 2000. [4] Kl´an P., Mindl J., Stedr´y A., Rubesov´a E., Chemical Informatics by the Internet, Academy of Sciences of the Czech Republic, Czech Chemical Society, CHEMagaz´ın, 1999. [5] LIST.INDIANA.EDU, Mailing Lists https://list.indiana.edu/sympa/home. [6] Mitchell M., Complexity: A Guided Tour, Oxford University Press, 2011. 10 [7] Pierce J. R., An Introduction to Information Theory: Symbols, Signals and Noise, Dover Books on Mathematics, 1980. [8] Simonite T., The Decline of Wikipedia, MIT Technology Review, http://www. technologyreview.com/featuredstory/520446/the–decline–of–wikipedia/, 2013. 11
1810.09981	1	1810	2018-10-23T17:22:26	A Systematic Framework and Characterization of Influence-Based Network Centrality	[ "cs.SI", "cs.DS" ]	In this paper, we present a framework for studying the following fundamental question in network analysis: How should one assess the centralities of nodes in an information/influence propagation process over a social network? Our framework systematically extends a family of classical graph-theoretical centrality formulations, including degree centrality, harmonic centrality, and their "sphere-of-influence" generalizations, to influence-based network centralities. We further extend natural group centralities from graph models to influence models, since group cooperation is essential in social influences. This in turn enables us to assess individuals' centralities in group influence settings by applying the concept of Shapley value from cooperative game theory. Mathematically, using the property that these centrality formulations are Bayesian, we prove the following characterization theorem: Every influence-based centrality formulation in this family is the unique Bayesian centrality that conforms with its corresponding graph-theoretical centrality formulation. Moreover, the uniqueness is fully determined by the centrality formulation on the class of layered graphs, which is derived from a beautiful algebraic structure of influence instances modeled by cascading sequences. Our main mathematical result that layered graphs in fact form a basis for the space of influence-cascading-sequence profiles could also be useful in other studies of network influences. We further provide an algorithmic framework for efficient approximation of these influence-based centrality measures. Our study provides a systematic road map for comparative analyses of different influence-based centrality formulations, as well as for transferring graph-theoretical concepts to influence models.	cs.SI	cs	A Systematic Framework and Characterization of Influence-Based Network Centrality Wei Chen Shang-Hua Teng Microsoft Research University of Southern California Hanrui Zhang Duke University [email protected] [email protected] [email protected] October 24, 2018 Abstract In this paper, we present a framework for studying the following fundamental question in network analysis: How should one assess the centralities of nodes in an information/influence propa- gation process over a social network? Our framework systematically extends a family of classical graph-theoretical centrality for- mulations, including degree centrality, harmonic centrality, and their "sphere-of-influence" gen- eralizations, to influence-based network centralities. We further extend natural group centralities from graph models to influence models, since group cooperation is essential in social influences. This in turn enables us to assess individuals' centralities in group influence settings by applying the concept of Shapley value from cooperative game theory. Mathematically, using the property that these centrality formulations are Bayesian1, we prove the following characterization theorem: Every influence-based centrality formulation in this family is the unique Bayesian centrality that conforms with its corresponding graph-theoretical centrality formulation. Moreover, the uniqueness is fully determined by the centrality formu- lation on the class of layered graphs, which is derived from a beautiful algebraic structure of influence instances modeled by cascading sequences. Our main mathematical result that layered graphs in fact form a basis for the space of influence-cascading-sequence profiles could also be useful in other studies of network influences. We further provide an algorithmic framework for efficient approximation of these influence-based centrality measures. Our study provides a systematic road map for comparative analyses of different influence- based centrality formulations, as well as for transferring graph-theoretical concepts to influence models. 1 Introduction Network influence is a fundamental subject in network sciences [28, 15, 21, 7]. It arises from vast real-world backgrounds, ranging from epidemic spreading/control, to viral marketing, to innovation, and to political campaign. It also provides a family of concrete and illustrative examples for studying network phenomena -- particularly regarding the interplay between network dynamics and graph structures -- which require solution concepts beyond traditional graph theory [11]. As a result, 1Meaning that they are linear to the convex combination of influence instances. 1 network influence is a captivating subject for theoretical modeling, mathematical characterization, and algorithmic analysis [28, 15, 21, 10]. In contrast to some network processes such as random walks, network influence is defined not solely by the static graph structure. It is fundamentally defined by the interaction between the dynamic influence models and the static network structures. Even on the same static network, dif- ferent influence propagation models -- such as the popular independent cascade and linear-threshold models -- induce different underlying relationships among nodes in the network. The characteriza- tion of this interplay thus requires us to reformulate various fundamental graph-theoretical concepts such as centrality, closeness, distance, neighborhood (sphere-of-influence), and clusterability, as well as to identify new concepts fundamental to emerging network phenomena. In this paper, we will study the following basic question in network science with focusing on influence-based network centrality. Is there a systematic framework to expand graph-theoretical concepts in network sci- ences? 1.1 Motivations Network centrality -- a basic concept in network analysis -- measures the importance and the criticality of nodes or edges within a given network. Naturally, as network applications vary -- being Web search, internet routing, or social interactions -- centrality formulations should adapt as well. Thus, numerous centrality measures have been proposed, based on degree, closeness, betweenness, and random-walks (e.g., PageRank) (cf. [25]) to capture the significance of nodes on the Web, in the Internet, and within social networks. Most of these centrality measures depend only on the static graph structures of the networks. Thus, these traditional centrality formulations could be inadequate for many real-world applications -- including social influence, viral marketing, and epidemics control -- in which static structures are only part of the network data that define the dynamic processes. Our research will focus on the following basic questions: How should we summarize influence data to capture the significance of nodes in the dynamic propagation process defined by an influence model? How should we extend graph-theoretic centralities to the influence-based centralities? What does each centrality formulation capture? How should we comparatively evaluate different centrality formu- lations? At WWW'17, Chen and Teng [11] presented an axiomatic framework for characterizing influenced- based network centralities. Their work is motivated by studies in multiple disciplines, including social-choice theory [2], cooperative game theory [32], data mining [18], and particularly by [27] on measures of intellectual influence and [1] on PageRank. They present axiomatic characterizations for two basic centrality measures: (a) Single Node Influence (SNI) centrality, which measures each node's significance by its influence spread;2 (b) Shapley Centrality, which uses the Shapley value of the influence spread function -- formulated based on a fundamental cooperative-game-theoretical concept. Mathematically, the axioms are structured into two categories. 2The influence spread -- as defined in [21] -- of a group is the expected number of nodes this group can activate as the initial active set, called seed set. 2 • Principle Axioms: The set of axioms that all desirable influenced-based centrality formu- lations should satisfy. In [11], two principle axioms, Anonymity and Bayesian, are identified. Anonymity is an ubiquitous and exemplary principle axiom, which states that centrality mea- sures should be preserved under isomorphisms among influence instances. Bayesian states that influence-based centrality is a linear measure for mixtures of influence instances. • Choice axioms: A (minimal) set of axioms that together with the principle axioms uniquely determine a given centrality formulation. Such characterizations and the taxonomy of axioms precisely capture the essence of centrality formulations as well as their fundamental differences. In particular, the choice axioms succinctly distill the comparative differences between different centrality formulations. However, the axiomatic characterization in [11] has two limitations preventing it to be gen- eralized to study more influence-based centralities. First, it makes a significant simplification of the influence process: each influence instance I only captures the probability distributions of the final influenced nodes given any initial seed set, which we refer as the seeds-targets (probabilistic) profile. Essentially, it compressed out all intermediate steps in a cascading process and only takes the initial seed nodes and the final target nodes into account. This simplification is enough to study centrality measures concerning the final influence spread of the diffusion model, but is inadequate for characterizing influenced-based centrality measures that can capture the propagation details of network influences, such as neighborhood, closeness, sphere-of-influence centralities. Second, its choice axioms are based on a family of critical set instances, which do not have a graph-theoretical interpretation. This makes it less powerful in explaining the connection between graph-theoretical centralities and influence-based centralities. 1.2 Our Contributions In this paper, we address both of the above issues in [11] and significantly expand the charac- terization of influence-based network centrality. First, influence instance I is now defined as the probabilistic profile on the more detailed cascading sequences, as formally defined below. Definition 1.1 (Cascading Sequence). For a directed network G = (V, E) with n = V , a set sequence (S0, S1, . . . , Sn−1) is a cascading sequence if it is both (1) Monotonic: ∅ ⊂ S0 ⊂ S1 ⊂ ··· ⊂ St−1 ⊂ St = St+1 = ··· Sn−1 ⊆ V for some t = 0, 1, . . . , n − 1, and (2) G-continuous: for all t ∈ [T ], every node in ∆t = St − St−1 can be reached directly from some nodes in ∆t−1 (where ∆0 = S0). In the above definition, St represents the set of network nodes that become active by step t during the propagation, and ∆t = St − St−1 denotes the set of nodes newly activated at step t. Thus, the cascading sequence (S0, S1, . . . , Sn−1) provides a layered structure starting from seed set S0, similar to network broadcasting. However, unlike broadcasting, the layered cascading sequences in network influence are formed In each time step, already activated nodes stochastically activate more nodes in stochastically. the next step, and this stochastic propagation ends when no new nodes are activated in a step. Therefore, when describing an influence instance, we need to specify the probabilistic distribution of the possible cascading sequences. This is formally defined as influence profile below. 3 v1 v2 R0 v3 v4 v5 R1 v6 v7 R2 v8 v9 v10 R3 Figure 1: an example of a layered-graph instance with 4 layers (R0, R1, R2, R3) where R0 = {v1, v2}, R1 = {v3, v4, v5}, R2 = {v6, v7}, R3 = {v8, v9, v10}. Definition 1.2 (Influence Profile). An influence profile (also called an influence instance) is a tuple I = (V, E, PI), where G = (V, E) is a directed graph, and PI : (2V )n → R is a probability profile defining a probability distribution of cascading sequences for every seed set S0 ⊆ V . That is, PI(S0, S1, . . . , Sn−1) ∈ [0, 1] specifies the probability that the influence instance generates cascading sequence (S0, S1, . . . , Sn−1) when given seed set S0. Then an influence-based centrality measure is defined as a mapping from the above influence profiles to real-valued vectors assigning centrality values to each node. We further transfer the concept of graph distance to cascading distance for cascading sequences. By using the cascading- sequence based influence profile and cascading distance, we are able to extend a large family of graph theoretical centralities based on graph distances from individual nodes, such as degree, closeness, harmonic, reachability centralities to influence based centralities. We refer to them as the stochastic sphere-of-influence centralities, and use a generic distance function f to summarize all of them. Second, we provide a key technical contribution of the paper, which characterizes all influence instances by layered graphs. A layered graph is a directed graph with multiple layers and all nodes in one layer connect to all nodes in the next layer (Fig. 1). A layered graph instance is simply treating the graph as an influence instance following the breadth-first-search (BFS) propagation pattern. Surprisingly, we show that the set of all layered-graph instances form the basis in the vector space of all influence profiles, meaning that every influence profile is a linear combination of layered-graph instances. The result is instrumental to our centrality characterization, and is a powerful result by its own right. The above layered-graph characterization allows us to connect influence propagation with static graphs. Therefore, by combining the principle axioms (Axioms Anonymity and Bayesian), we are able to show that our extended stochastic sphere-of-influence centrality with a distance function f is the unique centrality satisfying Axioms Anonymity and Bayesian that conforms with the cor- responding graph-theoretical centrality with the same distance function f , and the centrality is uniquely determined by their values on layered-graph instances. This characterization illustrates that (a) our extension of graph-theoretical centralities to influence-based centralities is not only reasonable but the only feasible mathematical choice, and (b) layered graphs are the key family of graphs comparing different influence measures, since a graph-theoretical centrality measure on layered graphs fully determines the conforming influence-based centrality measure. The above centrality extension focuses on individual nodes. As group cooperation plays an im- portant role in influence propagation, we further extend individual centralities to group centralities, which provide a value for every subset of nodes in the network. Similar to individual centrality, we 4 Figure 2: Road map for the systematic extension of graph-theoretical distance-based centralities to influence-based centralities. provide a characterization theorem showing that influence-based group centrality is also uniquely characterized by their values on layered-graph instances, as long as they satisfy the group version of Axioms Anonymity and Bayesian. A group centrality measure has 2n dimensions, so we further use the Shapley value [32] in cooperative game theory to reduce it to n dimensions, and refer to it as the influence-based Shapley centrality. The Shapley centrality of a node measures its importance when the node collaborate with other nodes in groups. Due to the linearity of the Shapley value, we obtain the same characterization for the Shapley centrality: it is the unique one conforming with the graph-theoretical Shapley centrality and satisfying Axioms Anonymity and Bayesian, and the uniqueness is fully determined by its values on layered-graph instances. Figure 2 summarizes our systematic extension of graph-theoretical centralities (the lower three boxes) to influence-based centralities (the upper three boxes). Starting from the classical graph- theoretical distance-based individual centrality (e.g. harmonic centrality), by transferring the con- cept of graph distance to cascading distance, we could lift it to the stochastic sphere-of-influence individual centrality (e.g. influence-based harmonic centrality). From individual centralities (ei- ther graph-theoretical or influence-based), we could use group distance to extend them to group centralities. From group centralities, we could apply Shapley value to obtain Shapley centrali- ties. Therefore, Figure 2 provides a road map on how to extend many classical graph-theoretical centralities to influence-based centralities. In addition to studying the systematic framework and characterization of influence-based cen- tralities, we also address the algorithmic aspects of these centralities. We extend the approximation algorithm of [11] to cover all stochastic sphere-of-influence centralities introduced in this paper. The algorithm efficiently provides estimates for the centrality values of all nodes with theoretical guar- antees. To summarize, our contributions include: (a) a systematic extension of graph-theoretical distance- 5 based centralities to corresponding influence-based centralities, and further extending them to group centralities and Shapley centralities; (b) a key algebraic characterization of influence profiles by layered-graph instances; (c) unique characterization of stochastic sphere-of-influence centralities by their layered graph centralities and the principle axioms; and (d) an efficient algorithmic framework that approximates influence-based centralities for all centralities measures proposed in the paper. 1.3 Related Work Influence propagation is an important topic in network science and network mining. One well studied problem on influence propagation is the influence maximization problem [16, 29, 21], which is to find k seed nodes that generate the largest influence in the network. Influence maximization has been extensively studied for improving its efficiency or extending it to various other settings (e.g. [22, 23, 3, 12, 13]). However, influence maximization is different from the study of individual centrality as pointed out by [7]. Putting into our group centrality context, influence maximization can be viewed as the task of finding the group with the largest reachability group centrality. Our efficient centrality approximation algorithm is inspired from the scalable influence maximization based on the reverse-reachable set approach [8, 34, 33]. Network centrality has been extensively studied and numerous centrality measures are pro- posed (e.g. [25, 5, 20, 6, 9, 17]). As discussed in the introduction, most centrality measures (in- cluding group centrality [17] and Shapley centrality [24, 19]) are graph theoretical, and only focus on the static graph structure. A recent study by Chen and Teng [11] is the first systematic study combining network centrality with influence propagation. As already detailed in the introduction, our study is motivated by [11], and we aim at overcoming the limitations in [11] and extending the study of influence-based centrality to much wider settings. In particular, the SNI centrality and Shapley centrality studied in [11] are two instances related to reachability in the family of sphere-of-influence centrality measures we cover in this paper. Axiomatic approach has been used to characterize network centrality [30, 26, 4, 31, 1]. These characterizations are mainly for graph-theoretical centralities. Again, the study in [11] provides the first axiomatic characterization for a couple of influence-based centralities. The characterization we provide in this paper is novel in the sense that it connects general influence profiles with a family of classical layered graphs so that our characterization can be based on graph-theoretical centralities. 2 Algebraic Characterization of Network Influences In this section, we present our main technical result -- a surprising discovery during our research -- which provides a graph-theoretical characterization of the space of influence profiles. Specifically, we identify a simple set of classical graphs, and prove that when treated as BFS propagation instances, they form a linear basis in the space of all stochastic cascading-sequence profiles. This graph- theoretical characterization of influence models is instrumental to our systematic characterization of a family of influence-based network centralities. Moreover, we believe that this result is also important on its own right, and is potentially useful in other settings studying general influence propagation models. 6 2.1 Stochastic Diffusion Model Stochastic diffusion models describe how information or influence are propagated through a network. A number of models have been well-studied (cf. [21, 10]), and among them independent cascade (IC) and linear threshold (LT) models are most popular ones. Here, we illustrate the stochastic profile of cascading sequences with the triggering model of Kempe-Kleinberg-Tardos [21], which includes IC and LT models as special cases. Triggering model will also be the subject of our algorithmic study. In a triggering model, the static network structure is modeled as a directed graph G = (V, E) with n = V . Each node v ∈ V has a random triggering set T (v) drawn from distribution D(v) over subsets of v's in-neighbors N−(v). At time t = 0, triggering sets of all nodes are sampled from their distributions, and nodes in a given seed set S ⊆ V are activated. At any time t ≥ 1, a node v is activated if some nodes in its triggering set T (v) was activated at time t − 1. The propagation continues until no new nodes are activated in a step. In general, we can describe propagation in networks as a sequence of node activations, as described in the introduction. For convenience, we restate Definition 1.1 here. Definition 1.1 (Cascading Sequence). For a directed network G = (V, E) with n = V , a set sequence (S0, S1, . . . , Sn−1) is a cascading sequence if it is both (1) Monotonic: ∅ ⊂ S0 ⊂ S1 ⊂ ··· ⊂ St−1 ⊂ St = St+1 = ··· Sn−1 ⊆ V for some t = 0, 1, . . . , n − 1, and (2) G-continuous: for all t ∈ [T ], every node in ∆t = St − St−1 can be reached directly from some nodes in ∆t−1 (where ∆0 = S0). Note that St in a cascading sequence represents the nodes that are active by time t. The mono- tonicity requirement in the above definition means that (a) active nodes will not be deactivated. and (b) each step should have at least one new active node unless the cascade stops. This corre- sponds to the progress diffusion model in the literature. Since S0 cannot be empty, and the set must grow by at least one node in each cascading step we have at most n− 1 cascading steps in the sequence. The G-continuous condition means that the activation of any new node in ∆t at time t must be partially due to the activation of some nodes in ∆t−1 at the previous time step. At this level of abstraction, a stochastic diffusion model can be viewed as a probabilistic mechanism to generate cascading sequences. Similar to [10], in this paper we use the distribution of the cascad- ing sequences as the general specification of the diffusion model, as defined in Definition 1.2 and restated here. Definition 1.2 (Influence Profile). An influence profile (also called an influence instance) is a tuple I = (V, E, PI), where G = (V, E) is a directed graph, and PI : (2V )n → R is a probability profile defining a probability distribution of cascading sequences for every seed set S0 ⊆ V . That is, PI(S0, S1, . . . , Sn−1) ∈ [0, 1] specifies the probability that the influence instance generates cascading sequence (S0, S1, . . . , Sn−1) when given seed set S0. 0. (2) For every S0,(cid:80) (V, E, PI), we can define influence spread σI(S) =(cid:80) S1,...,Sn−1 Note that (1) if (S0, S1, . . . , Sn−1) is not a valid cascading sequence, then PI(S0, S1, . . . , Sn−1) = PI(S0, S1, . . . , Sn−1) = 1. We also use the notation PI(S0) to denote the distribution of sequence (S1, . . . , Sn−1) starting from S0. In influence propagation models, one key metric is the influence spread of a seed set S, which is the expected number of nodes activated when S is the seed set. For an influence profile I = S1,...,Sn−1 PI(S, S1, . . . , Sn−1) · Sn−1. 7 We remark that in [11], a more coarse-grained seed-target profile is used to study two influence- based centrality measures. The seed-target profile is only suitable for centrality measures addressing the final influence spread, but is not detailed enough to study other centrality extensions including extensions to degree, closeness, harmonic centralities, etc. Therefore, in this paper, we focus on the cascading sequence profile. Note that for any directed graph G = (V, E), we can equivalently interpret it as a diffusion model, where the diffusion is carried out by the breadth-first search (BFS). In particular, for any seed set S0, we have a deterministic cascading sequence (S0, S1, . . . , Sn−1), where St is all the nodes that can be reached within t steps of BFS. Thus, the probability profile PI is such that only this BFS sequence has probability 1, and all other sequences starting from S0 have probability 0. We call this instance the BFS influence instance corresponding to graph G, and denote it as I (BFS) . G 2.2 A Graph-Theoretical Basis of Influence Profiles Mathematically, each influence profile I of stochastic cascading sequences as defined in Defini- tion 1.2, can be represented as a vector of probabilities [PI(S0, S1, . . . , Sn−1)] over monotonic cascading sequences. In other words, the vector contains entries for each monotonic cascading sequences. Because for each set S0, all valid cascading sequences add up to 1, so one entry is redundant. We remove the entry PI(S0, S0, . . . , S0) from the vector, and express it implicitly. The resulting vector certainly has an exponential number of dimensions, as there are exponential number of monotonic set sequences. We use M to denote its dimension. The set of "basis" graphs are the layered graphs, as depicted in Figure 1. Formally, for a vertex set V , for an integer t ≥ 0, and t + 1 disjoint nonempty subsets R0, R1, . . . , Rt ⊆ V , a layered graph LV (R0, . . . , Rt) is a directed graph in which every node in Ri−1 has a directed edge pointing to every node in Ri, for i ∈ [t], and the rest nodes in V \ ∪t i=1Ri are isolated nodes with no connections to and from any other nodes. We say that the BFS influence instance of the layered graph LV (R0, . . . , Rt), namely I (BFS) LV (R0,...,Rt), is a layered-graph instance, and for convenience we also use IV (R0, . . . , Rt) to denote this instance. When the context is clear, we ignore V in the subscript. A trivial layered graph instance is when t = 0, in which case all nodes are isolated and there is no edge in the graph. We call this the null influence instance, and denote it as I N (or IV (V ) to make it consistent with the layered-graph notation). Technically, in I N , only PIN (S0, S0, . . . , S0) = 1, and all other probability values are 0, which means its corresponding vector form is the all-zero vector. Let L = {IV (R0, . . . , Rt) t = 1, . . . , n − 1,∅ (cid:54)= Ri ⊆ V, all Ri's are disjoint}, i.e. L is the set of all nontrivial layered-graph instances under node set V . As a fundamental characterization of the mathematical space of influence profiles, we prove the following theorem, which states that all nontrivial layered-graph instances form a linear basis in the space of all cascading-sequence based influence instances: Theorem 2.1 (Graph-Theoretical Basis). The set of vectors corresponding to the nontrivial layered- graph instances in L forms a basis in RM . Although the proof of this theorem is quite technical, its underlying principle is quite basic. Here we provide some intuitions. Note first that M = L. Thus, the central argument in the proof is to show that elements in L are independent, which we will establish using proof-by-contradiction: Suppose the profiles corresponding to layered graphs are not independent. That is, there are not- all-zero coefficients λI such that a linear combination (denoted by P =(cid:80)I∈L λIPI) of these profiles 8 is zero. We consider a carefully-designed inclusion-exclusion form linear combination of the entries of P and show that this combination is exactly some λI (cid:54)= 0, which means P (cid:54)= 0. 3 Influence-based Centrality: Single Node Perspective Recall that a graph-theoretical centrality, such as degree, distance, PageRank, and betweenness, summarizes network data to measure the importance of each node in a network structure. Likewise, the objective of influence-based centrality formulations is to summarize the network-influence data in order to measure the importance of every node in influence diffusion processes. Formally (in the cascading-sequence model): Definition 3.1 (Influence-based Centrality Measure). An influence-based centrality measure ψ is a mapping from an influence profile I = (V, E, PI), to a real-valued vector (ψv(I))v∈V ∈ Rn. The objective is to formulate network centrality measures that reflect dynamic influence propa- gation. Theorem 2.1 lays the foundation for a systematic framework to generalize graph-theoretical centrality formulation to network centrality measures. To this end, we first examine a unified cen- trality family that are natural for layered graphs. 3.1 A Unified Family of Sphere-of-Influence Centralities G (v) and N− In this subsection, we discuss a family of graph-theoretical centrality measures that contains various forms of "sphere-of-influence" and closeness centralities. These centrality measures have a common feature: the centrality of node v is fully determined by the distances from v to all nodes. Consider a directed graph G = (V, E). Let N + G (v) denote the set of out-neighbors and in-neighbors, respectively, of a node v. Let dG(u, v) be the graph distance from u to v in G. If v is not reachable from u in G then we set d(u, v) = ∞. Let dG(S, v) = minu∈S dG(u, v) be the distance from a subset S ⊆ V to node v. Let ΓG(S) be the set of nodes reachable from S ⊆ V in G. When the context is clear, we would remove G from the subscripts in the above notations. Recall that a graph-theoretical centrality measure is a mapping µ from a graph G to a real- valued vector (µv(G))v∈V ∈ Rn, where µv(G) denote the centrality of v in G. For every S ⊆ V and v ∈ V , let (cid:126)dG(S) be the vector in Rn consisting of the distance from S to every node u, i.e. (cid:126)dG(S) = (d(S, u))u∈V . Let (cid:126)dG(v) = (cid:126)dG({v}). We use R∞ to denote R ∪ {∞}. For each f : Rn∞ → R, we can define: Definition 3.2 (Distance-based Centrality). A distance-based centrality µind[f ] with function f : Rn∞ → R is defined as µind[f ]v(G) = f ((cid:126)dG(v)). Definition 3.2 is a general formulation. It includes several classical graph-theoretical centrality formulations as special cases: (a) The degree centrality (or immediate sphere of influence), µind-deg, G (v). It is defined by is defined as the out-degree of a node v in graph G, that is, µind-deg f deg( (cid:126)d) = {u ∈ V du = 1}. (b) The closeness centrality, µind-cls, is defined as the reciprocal of the average distance to other nodes, µind-cls . u∈V du If G is not strongly connected, then µind-cls (G) = 0 for any v that cannot reach all other nodes, and thus closeness centrality is not expressive enough for such graphs. (c) harmonic centrality, µind-har, is defined: µind-har . Note that harmonic centrality is closely related with closeness centrality, and is applicable to network with dG(v,u) . It is defined by f har( (cid:126)d) =(cid:80) 1(cid:80) u(cid:54)=v dG(v,u) . It is defined by f cls( (cid:126)d) = (G) =(cid:80) (G) = N + u∈V,du>0 1 du 1(cid:80) (G) = v v v v u(cid:54)=v 1 9 (G) = ΓG({v}), disjointed components. (d) The reachability centrality measure µind-rch: µind-rch which means the reachability centrality of v is the number of nodes v could reach in G. It is defined by f rch((cid:126)d) = {u ∈ V du < ∞}. (e) The sphere-of-influence centrality measure µind-SoI(δ): For a = {u : du ≤ δ}. It is defined by f SoI(δ)( (cid:126)d) = {u ∈ V du ≤ δ}. threshold parameter δ, µind-SoI(δ) Clearly, as δ varies from 1 to n− 1 (or ∞), the sphere-of-influence centrality interpolates the degree centrality and the reachability centrality. v v 3.2 Stochastic Sphere-of-Influence: Lifting from Graph to Influence Models Thus, Definition 3.2 represents a unified family of sphere-of-influence centralities for graphs. The function f -- which is usually a non-increasing function of distance profiles -- captures the scale of the impact, based on the distance of nodes from the source. By unifying these centralities under one general centrality class, we are able to systematically derive and study their generaliza- tion in the network-influence models. The key step is to transfer the graph distance in directed graph to cascading distance in cascading sequences. For any cascading sequence (S0, S1, . . . , Sn−1) starting from seed set S0, let du(S0, S1, . . . , Sn−1) = t if u ∈ ∆t = St \ St−1 (∆0 = S0), and du(S0, S1, . . . , Sn−1) = ∞ if u (cid:54)∈ Sn−1. We call du(S0, S1, . . . , Sn−1) the cascading distance from seed set S0 to node u, since it represents the number of steps needed for S0 to activate u in the cascading sequence. Then, we define the cascading distance vector (cid:126)d(S0, S1, . . . , Sn−1) as (du(S0, S1, . . . , Sn−1))u∈V ∈ Rn. In particular, when we consider a cascading sequence ({v}, S1, . . . , Sn−1) starting from a single node v, set ∆1 = S1 \ {v} can be viewed as the out-neighbor of v, set Sn−1 can be viewed as the all nodes reachable from v, and for every node u ∈ ∆t = St \ St−1, the distance from v to u is t. Definition 3.3 (Individual Stochastic Sphere-of-Influence Centrality). For each function f : Rn∞ → R, the influence-based individual stochastic sphere-of-influence centrality ψind[f ] is defined as: ψind[f ]v(I) = E(S1,...,Sn−1)∼PI ({v})[f ((cid:126)d({v}, S1, . . . , Sn−1))]. Definition 3.3 systematic extended the family of graph-theoretical centralities of Definition 3.2 to influence models. Natural influence-based centralities, e.g., the single-node influence (SNI) central- ity defined in [11] (using each node v's influence spread σI({v}) as the measure of its influence-based centrality), can be expressed by this extension: Proposition 3.1. ∀ influence profile I = (V, E, PI): SNI(I) = ψind[f rch](I). The influence-based centrality formulations of Definition 3.3 enjoy the following graph-theoretical conformity property. Definition 3.4 (Graph-Theoretical Conformity). An influence-based centrality measure ψ con- forms with a graph-theoretical centrality measure µ if for every directed graph G, ψ(I (BFS) ) = µ(G). G Proposition 3.2. For any function f : Rn∞ → R, ψind[f ] conforms with µind[f ]. 10 3.3 Characterization of Influence-Based Centrality Measures Given the multiplicity of the (potential) centrality formulations, "how should we characterize each formulation?" and "how should we compare different formulations?" are fundamental questions in network analysis. Inspired by the pioneering work of Arrow [2] on social choice, Shapley [32] on cooperation games and coalition, Palacios-Huerta & Volij [27] on measures of intellectual influence, and Altman & Tennenholtz [1] on PageRank, Chen and Teng [11] proposed an axiomatic framework for characterizing and analyzing influence-based network centrality. They identify two principle axioms that all desirable influenced-based centrality formulations should satisfy. Principle Axioms for Influenced-Based Centrality The first axiom -- ubiquitous axiom for centrality characterization, e.g. [30] -- states that labels on the nodes should have no effect on centrality measures. Axiom 3.1 (Anonymity). For any influence instance I = (V, E, PI), and permutation π of V , ψv(I) = ψπ(v)(π(I)), ∀v ∈ V. (1) In Axiom 3.1, π(I) = (π(V ), π(E), π(PI)) denotes the isomorphic instance: (1) ∀u, v ∈ V , (π(u), π(v)) ∈ π(E) if and only if (u, v) ∈ E, and (2) for any cascading sequence (S0, S1, . . . , Sn−1), Pπ(I)(π(S0), π(S1), . . . , π(Sn−1)) = PI(S0, S1, . . . , Sn−1). The second axiom concerns Bayesian social influence [11] through a given network: For any three influence profiles I, I1, I2 over the same vertex set V , we say I is a Bayesian of I1 and I2 if there exists α ∈ [0, 1] such that PI = αPI1 + (1 − α)PI2. In other words, I can be interpreted as a stochastic diffusion model where we first make a random selection -- with probability α of model I1 and with probability (1 − α) of model I2 -- and then carry out the diffusion process according to the selected model. We also say that I is a convex combination of I1 and I2. The axiom reflects the linearity-of-expectation principle. If an influence instance is a convex combination of two other influence instances, the centrality value of a vertex is the same convex combination of the corresponding centrality values in the two other instances. Axiom 3.2 (Bayesian). For any α ∈ [0, 1], for any influence profiles I, I1 and I2 over common vertex set V such that PI = αPI1 + (1 − α)PI2, ψv(I) = αψv(I1) + (1 − α)ψv(I2), Characterization of Influence-Based Centrality ∀v ∈ V. (2) We now use Theorem 2.1 and Axioms Anonymity and Bayesian to establish a complete charac- terization of the family of stochastic sphere-of-influence centralities formulated in Definition 3.3. Proposition 3.3. If a function f : Rn∞ → R is anonymous -- i.e., f ((cid:126)d) is permutation-invariant -- then ψind[f ] (as defined in Definition 3.3) satisfies Axiom Anonymity and µind[f ] (as defined in Definition 3.2) satisfies the graph-theoretical counterpart of Axiom Anonymity. Proposition 3.4. For any function f : Rn∞ → R, ψind[f ] satisfies Axiom Bayesian. Theorem 2.1 shows that all influence profiles can be represented as a linear combination of non- trivial layered-graph instances. The result enables us to study and compare centrality measures by looking at their instantiation in the simple layered graph instances. The Bayesian property together with the linear basis of nontrivial layered-graph instances leads to the following characterization theorem. 11 Theorem 3.1. A Bayesian influence-based centrality measure is uniquely determined by its values on layered-graph instances (including the null instance). Since layered-graph instances are all BFS instances derived from a special class of directed graphs, a direct implication of Theorem 3.1 is that any Bayesian centrality conforming with a classical graph-theoretical centrality is unique. Therefore, we have: Theorem 3.2 (Characterization of Individual Centrality). For any anonymous function f : Rn∞ → R, ψind[f ] (defined in Definition 3.3) is the unique influence-based centrality that conforms with µind[f ] (defined in Definition 3.2) that satisfies both Axiom Anonymity and Axiom Bayesian. The above uniqueness result show that our generalized definitions of influence-based centralities Definition 3.3 are not only reasonable but the only feasible choice (to the extent of Bayesian centralities). 4 Influence-based Centrality: Group Perspective and Shapley Cen- trality As highlighted in Domingos-Richardson [28, 15] and Kempe-Kleinberg-Tardos [21], social influ- ence propagation and viral marketing are largely group-based phenomena. Besides characterizing individuals' influential centralities, perhaps the more important task is to characterize the influen- tial centrality of groups, and individuals' roles in group cooperation. This is the group centrality and Shapley centrality introduced in this section. When distinction is necessary, we refer to the centrality defined in Section 3 as individual centrality. 4.1 Group Centrality Group centrality measures the importance of each group in a network. Formally, Definition 4.1 (Influence-based Group Centrality). An influence-based group centrality measure ψgrp is a mapping from an influence profile I = (V, E, PI) to a real-valued vector (ψgrp S (I))S∈2V ∈ R2n. For any function f : Rn∞ → R, both Definition 3.2 and Definition 3.3 have a natural extension: For S ⊆ V , µgrp[f ]S(G) = f ( (cid:126)dG(S)), ψgrp[f ]S(I) = E(S1,...,Sn−1)∼PI ({v})[f ((cid:126)d(S, S1, . . . , Sn−1))]. Axioms Anonymity and Bayesian extend naturally as well as the following characterization based on Theorem 2.1. Theorem 4.1 (Characterization of Group Centrality). For any anonymous function f : Rn∞ → R, ψgrp[f ] is the unique influence-based group centrality that conforms with µgrp[f ] and satisfies both Axiom Anonymity and Axiom Bayesian. Therefore, we can again reduce the analysis of an influence-based group centrality to the analysis of the measure on the particular layered-graph instances. 12 4.2 Cooperative Games and Shapley Value A cooperative game [32] is defined by tuple (V, τ ), where V is a set of n players, and τ : 2V → R is called characteristic function specifying the cooperative utility of any subset of players. In cooperative game theory, a ranking function φ is a mapping from a characteristic function τ to a vector (φv(τ ))v∈V ∈ Rn, indicating the importance of each individual in the cooperation. One famous ranking function is the Shapley value φShapley [32], as defined below. Let Π be the set of all permutations of V , and π ∼ Π denote a random permutation π drawn uniformly from set Π. For any v ∈ V and π ∈ Π, let Sπ,v denote the set of nodes in V preceding v in permutation π. Then, ∀v ∈ V : φShapley v (τ ) = 1 n! (τ (Sπ,v ∪ {v}) − τ (Sπ,v)) S!(n − S − 1)! (τ (S ∪ {v}) − τ (S)) (cid:88) (cid:88) π∈Π = S⊆V \{v} n! =Eπ∼Π[τ (Sπ,v ∪ {v}) − τ (Sπ,v)]. conditions: (1) Efficiency: (cid:80) The Shapley value of a player v measures the expected marginal contribution of v on the set of players ordered before v in a random order. Shapley [32] proved a remarkable representation theorem: The Shapley value is the unique ranking function that satisfies all the following four v∈V φv(τ ) = τ (V ). (2) Symmetry: For any u, v ∈ V , if τ (S ∪ {u}) = τ (S ∪{v}), ∀S ⊆ V \{u, v}, then φu(τ ) = φv(τ ). (3) Linearity: For any two characteristic functions (4) Null Player: For any v ∈ V , if τ and ω, for any α, β > 0, φ(ατ + βω) = αφ(τ ) + βφ(ω). τ (S ∪ {v}) − τ (S) = 0, ∀S ⊆ V \ {v}, then φv(τ ) = 0. Efficiency states that the total utility is fully distributed. Symmetry states that two players' ranking values should be the same if they have the identical marginal utility profile. Linearity states that the ranking values of the weighted sum of two coalition games is the same as the weighted sum of their ranking values. Null Player states that a player's ranking value should be zero if the player has zero marginal utility to every subset. 4.3 Shapley Centrality Shapley's celebrated concept -- as highlighted in [11] -- offers a formulation for assessing individu- als' performance in group influence settings. It can be used to systematically compress exponential- dimensional group centrality measures into n-dimensional individual centrality measures. Definition 4.2 (Influence-based Shapley Centrality). An influence-based Shapley centrality ψShapley is an individual centrality measure corresponding to a group centrality ψgrp: ψShapley v (I) = φShapley v = Eπ∼Π[ψgrp (ψgrp(I)) Sπ,v∪{v}(I) − ψgrp Sπ,v (I)]. We also denote it as ψShapley = φShapley ◦ ψgrp. In [11], Chen and Teng analyze the Shapley value of the influence-spread function, which is a special case of the following "Shapley extension" of Definition 3.3. Definition 4.3 (Shapley Centrality of Stochastic Sphere-of-Influence). For each f : Rn∞ → R, the Shapley centrality of Stochastic Sphere-of-Influence ψShapley[f ] is defined as: ψShapley[f ]v(I) = φShapley v (ψgrp[f ](I)). 13 Shapley centrality µShapley can also be defined similarly based on graph-theoretical group cen- trality (see, for example, [24]). We will refer to the extension of Definition 3.2 as µShapley[f ]. Using Theorem 2.1, we can establish the following characterization. Theorem 4.2 (Characterization of Shapley Centrality). For any anonymous function f : Rn∞ → R, ψShapley[f ] is the unique influence-based centrality that conforms with µShapley[f ] and satisfies both Axiom Anonymity and Axiom Bayesian. Theorem 4.2 systematically extends the work of [11] to all sphere-of-influence formulations. The SNI and Shapley centrality analyzed in [11] are ψind[f rch] and ψShapley[f rch], respectively. In our process of generalizing the work of [11], we also resolve an open question central to the axiomatic characterization of [11], which is based on a family of critical set instances that do not correspond to a graph-theoretical interpretation. In fact, the influence-spread functions of these "axiomatic" critical set instances used in [11] are not submodular. In contrast, influence-spread functions of the popular independent cascade (IC) and linear threshold (LT) models, as well as, the trigger models of Kempe-Kleinberg-Tardos, are submodular. The submodularity of these influence- spread functions plays an instrumental role in influence maximization algorithms [21, 10]. Thus, it is a fundamental and mathematical question whether influence profiles can be characterized by "simpler" influence instances. Our layered-graph characterization (Theorem 2.1) resolves this open question by connecting all influence profiles with simple BFS cascading sequence in the layered graphs, which is a special case of the IC model and possess the submodularity property. In summary, our layered-graph characterization is instrumental to the series of characterizations we could provide in this paper for influence-based individual, group, and Shapley centralities (Theorem 3.2, 4.1, 4.2). 5 Efficient Algorithm for Approximating Influence-based Central- ity Besides studying the characterization of influence-based centralities, we also want to compute these centrality measures efficiently. Accurate computation is in general infeasible (e.g. it is #P-hard to compute influence-based reachability centrality ψind[f rch] in the triggering model [35, 14]). Thus, we are looking into approximating centrality values. Instead of designing one algorithm for each centrality, we borrow the algorithmic framework from [11] and show how to adapt the framework to approximate different centralities. Same as in [11], the algorithmic framework applies to the triggering model of influence propagation. For efficient computation, we further assume that the u∈V g(du) for some scalar function g : R∞ → R satisfying g(∞) = 0. The degree, harmonic, and reachability centralities all satisfy this condition. u∈V gdeg(du), with gdeg(du) = 1 if du = 1 and gdeg(du) = 0 u∈V ghar(du), with ghar(du) = 1/du if du > 0 and ghar(du) = 0 otherwise; distance function f is additive, i.e. f ((cid:126)d) = (cid:80) In particular, we have f deg((cid:126)d) = (cid:80) otherwise; f har((cid:126)d) = (cid:80) and f rch((cid:126)d) =(cid:80) u∈V grch(du), where grch(du) = 1 if du < ∞ and grch(du) = 0 otherwise. The algorithmic framework for estimating individual and Shapley forms of sphere-of-influence centrality is given in Algorithm 1, and is denoted ICE-RR (for Influence-based Centrality Estimate via RR set). The algorithm uses the approach of reverse-reachable sets (RR sets) [8, 34, 33]. An RR set Rv is generated by randomly selecting a node v (called the root of Rv) with equal probability, and then reverse simulating the influence propagation starting from v. In the triggering model, it is simply sampling a random triggering set T (v) for v, putting all nodes in T (v) into Rv, and then recursively sampling triggering sets for all nodes in T (v), until no new nodes are generated. 14 Input: Network: G = (V, E); Parameters: random triggering set distribution {T (v)}v∈V , ε > 0, (cid:96) > 0, k ∈ [n], node-wise distance function g Output: ψv, ∀v ∈ V : estimated centrality value 1: {Phase 1. Estimate the number of RR sets needed } 2 · ε; θ0 = 0 2: LB = 1; ε(cid:48) = 3: estv = 0 for every v ∈ V 4: for i = 1 to (cid:98)log2 n(cid:99) − 1 do √ x = n/2i (cid:108) n·(((cid:96)+1) ln n+ln log2 n+ln 2)·(2+ 2 3 ε(cid:48)) (cid:109) θi = for j = 1 to θi − θi−1 do ε(cid:48)2·x generate a random RR set Rv rooted at v, and for each u ∈ Rv, record the distance dRv (u, v) from u to v in this reverse simulation. if estimating individual centrality then for every u ∈ Rv, estu = estu + g(dRv (u, v)) {estimating Shapley centrality} for every u ∈ Rv, estu = estu + φShapley (g(dRv (·, v))) else u 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 27: 28: 29: 30: end if end for est(k) = the k-th largest value in {estv}v∈V if n · est(k)/θi ≥ (1 + ε(cid:48)) · x then LB = n · est(k)/(θi · (1 + ε(cid:48))) break (cid:109) (cid:108) n(((cid:96)+1) ln n+ln 4)(2+ 2 end if 20: 21: end for 22: θ = 23: {Phase 2. Estimate the centrality value} 24: estv = 0 for every v ∈ V 25: for j = 1 to θ do 26: ε2·LB 3 ε) generate a random RR set Rv rooted at v, and for each u ∈ Rv, record the distance dRv (u, v) from u to v in this reverse simulation. if estimating individual centrality then for every u ∈ Rv, estu = estu + g(dRv (u, v)) for every u ∈ Rv, estu = estu + φShapley u else (g(dRv (·, v))) end if 31: 32: end for 33: for every v ∈ V , ψv = n · estv/θ 34: return ψv, v ∈ V Algorithm 1: ICE-RR: Efficient estimation of sphere-of-influence centralities via RR-sets, for the triggering model and additive distance function f ((cid:126)d) =(cid:80) u∈V g(du). The algorithm has two phases. In the first phase (lines 1 -- 22), the number θ of RR sets needed for the estimation is computed. The mechanism for obtaining θ follows the IMM algorithm in [33] 15 and is also the same as in [11]. In the second phase (lines 23 -- 34), θ RR sets are generated, and for each RR set R, the centrality estimate of u ∈ R, estu, is updated properly depending on the centrality type. Comparing to the algorithm in [11], our change is in lines 8 -- 14 and lines 26 -- 31. First, when generating an RR set Rv, we not only stores the nodes, but for each u ∈ Rv, we also store the distance from u to root v in the reverse simulation paths dRv (u, v). Technically, dRv (u, v) is the graph distance from u to v in the subgraph GRv , where GRv = (V, ERv ) with ERv = {(w, u) u ∈ Rv, w ∈ T (u)} is the subgraph generated by the triggering sets sampled during the reverse simulation. Note that with this definition, for u (cid:54)∈ Rv, we have dRv (u, v) = ∞. Next, if we are estimating individual centrality, we simply update the estimate estu by adding g(dRv (u, v)). If (g(dRv (·, v))), the we are estimating Shapley centrality, we need to update estu by adding φShapley Shapley value of u on the set function g(dRv (·, v)) : S ∈ 2V (cid:55)→ g(dRv (S, v)) ∈ R. We will show (g(dRv (·, v))) for all u ∈ Rv together is linear to Rv, so below that the computation of φShapley it is in the same order of generating Rv and does not incur significant extra cost. Note that the algorithm in [11] corresponds to our algorithm with g = grch. The correctness of the algorithm replies on the following crucial lemma. u u (cid:80) Lemma 5.1. Let Rv be a random RR set with root v generated in a triggering model instance I. Then, ∀u ∈ V , u's stochastic sphere-of-influence individual centrality with function f ((cid:126)d) = u∈V g(du) is ψ[f ]u(I) = n · E[g(dRv (u, v))], where the expectation is taking over the distribution of RR set Rv. Similarly, u's influence-based Shapley centrality with f is ψShapley[f ]u(I) = n · E[φShapley (g(dRv (·, v)))]. u u From Lemma 5.1, we can understand that lines 28 and 30 are simply accumulating empirical (g(dRv (·, v))) for individual centrality and Shapley centrality, values of g(dRv (u, v)) and φShapley respectively, and line 33 averages this cumulative value and then multiply it by n to obtain the final centrality estimate. With the above lemma, the correctness of the algorithm ICE-RR is shown by the following theorem. Theorem 5.1. Let (ψv)v∈V be the true centrality value for an influence-based individual or Shapley centrality with additive function f , and let ψ(k) be the k-th largest value in (ψv)v∈V . For any > 0, (cid:96) > 0, and k ∈ [n], Algorithm ICE-RR returns the estimated centrality ( ψv)v∈V that satisfies (a) unbiasedness: E[ ψv] = ψv,∀v ∈ V ; and (b) robustness: under the condition that ψ(k) ≥ 1, with probability at least 1 − 1 n(cid:96) :(cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). (3) In terms of time complexity, for individual centrality, lines 10 and 28 take constant time for each u ∈ Rv, so it has the same complexity as the algorithm in [11]. For Shapley centrality, the (g(dRv (·, v))) for all v is O(Rv), same following lemma shows that the computation of φShapley as the complexity of generating Rv, so it will not add complexity to the overall running time. Suppose Rv has ∆ levels in total (i.e., ∆ = max{dRv (u(cid:48), v) u(cid:48) ∈ Rv}), and let si = {u(cid:48) u(cid:48) ∈ Rv, dRv (u(cid:48), v) ≥ i}. Lemma 5.2. For any function g : R∞ → R with g(∞) = 0, u  (cid:88) 0≤j≤si − (cid:88) 0≤j≤si+1 si! j!  . si+1! j! φShapley u (g(dRv (·, v))) = 1 Rv g(k) + 1 Rv! (g(k) − g(i)) (cid:88) k<i≤∆ 16 In O(Rv) time we can compute this value for all nodes in Rv (assuming infinite precision). For (gdeg(dRv (·, v))) = 1/{w dRv (w, v) = 1} if dRv (u, v) = 1, and otherwise degree centrality, φShapley it is 0. For reachability centrality, φShapley (grch(dRv (·, v))) = 1/Rv. u u Therefore, the time complexity follows [11]: Theorem 5.2. Under the assumption that sampling a triggering set T (v) takes time at most O(N−(v)) time, and the condition (cid:96) ≥ (log2 k − log2 log2 n)/ log2 n, the expected running time of ICE-RR is O((cid:96)(m + n) log n · E[σ(v)]/(ψ(k)ε2)), where E[σ(v)] is the expected influence spread of a random node v drawn from V with probability proportional to the in-degree of v. Theorems 5.1 and 5.2 together show that our algorithm ICE-RR provides a framework to effi- ciently estimate all individual and Shapley centralities in the family of influence-based stochastic sphere-of-influence centralities. We further remark that, although algorithm ICE-RR is shown for computing individual centralities and Shapley centralities, it can be easily adapted to computing group centralities as well. Of course, a group centrality has 2n values, so it is not feasible to list all of them. But if we consider that the algorithm is to estimate n group centrality values for n given sets, then we only need to replace estu with estS for every S in the input, and change the lines corresponding to individual centrality (lines 10 and 28) to "for each S in the input, estS = estS + g(dRv (S, v))". This change is enough for estimating n group centrality values. 6 Future Work Many topics concerning the interaction between network centralities and influence dynamics can be further explored. One open question is how to extend other centralities that are not covered by sphere-of-influence to influence-based centralities. For example, betweenness centrality of a node v is determined not only by the distance from v to other nodes, but by all-pair distances, while PageRank and other eigenvalue centralities are determined by the entire graph structure. Therefore, one may need to capture further aspects of the influence propagation to provide natural extensions to these graph-theoretical centralities. Another open question is how to characterize centrality for a class of influence profiles, e.g. all submodular influence profiles, all triggering models, etc. Empirical comparisons of different influence-based centralities, as well as studying the applications that could utilize influence-based centralities, are all interesting and important topics worth further investigation. 17 References [1] Alon Altman and Moshe Tennenholtz. Ranking systems: The pagerank axioms. In ACM, EC '05, pages 1 -- 8, 2005. [2] K. J. Arrow. Social Choice and Individual Values. Wiley, New York, 2nd edition, 1963. [3] Shishir Bharathi, David Kempe, and Mahyar Salek. Competitive influence maximization in social networks. In WINE, volume 4858, pages 306 -- 311. 2007. [4] P. Boldi and S. Vigna. Axioms for centrality. Internet Mathematics, 10:222 -- 262, 2014. [5] Phillip Bonacich. Factoring and weighting approaches to status scores and clique identification,. Journal of Mathematical Sociology, 2:113 -- 120, 1972. [6] Phillip Bonacich. Power and centrality: A family of measures. American Journal of Sociology, 92(5):1170 -- 1182, 1987. [7] Stephen P. Borgatti. Identifying sets of key players in a social network. Computational and Mathematical Organizational Theory, 12:21 -- 34, 2006. [8] Christian Borgs, Michael Brautbar, Jennifer Chayes, and Brendan Lucier. Maximizing social influence in nearly optimal time. In ACM-SIAM, SODA '14, pages 946 -- 957, 2014. [9] Sergey Brin and Lawrence Page. The anatomy of a large-scale hypertextual web search engine. Computer Networks, 30(1-7):107 -- 117, 1998. [10] Wei Chen, Laks V. S. Lakshmanan, and Carlos Castillo. Information and Influence Propagation in Social Networks. Synthesis Lectures on Data Management. Morgan & Claypool Publishers, 2013. [11] Wei Chen and Shang-Hua Teng. Interplay between social influence and network centrality: A comparative study on shapley centrality and single-node-influence centrality. In WWW, April 2017. The full version appears in arXiv:1602.03780. [12] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for prevalent viral marketing in large-scale social networks. In KDD, pages 1029 -- 1038, 2010. [13] Wei Chen, Yifei Yuan, and Li Zhang. Scalable influence maximization in social networks under the linear threshold model. In ICDM, pages 88 -- 97, 2010. [14] Wei Chen, Yifei Yuan, and Li Zhang. Scalable influence maximization in social networks under the linear threshold model. In ICDM, pages 88 -- 97, 2010. [15] Pedro Domingos and Matt Richardson. Mining the network value of customers. In ACM, KDD '01, pages 57 -- 66, 2001. [16] Pedro Domingos and Matthew Richardson. Mining the network value of customers. In KDD, pages 57 -- 66, 2001. [17] M. G. Everett and S. P. Borgatti. The centrality of groups and classes. Journal of Mathematical Sociology, 23(3):181 -- 201, 1999. 18 [18] Rumi Ghosh, Shang-Hua Teng, Kristina Lerman, and Xiaoran Yan. The interplay between dynamics and networks: centrality, communities, and cheeger inequality. In ACM, KDD '14, pages 1406 -- 1415, 2014. [19] Daniel G´omez, Enrique Gonzalez-Aranguena, Conrado Manuel, Guillermo Owen, Monica del Pozo, and Juan Tejada. Centrality and power in social networks: a game theoretic approach. Mathematical Social Sciences, 46(1):27 -- 54, 2003. [20] Leo Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39 -- 43, March 1953. [21] David Kempe, Jon M. Kleinberg, and ´Eva Tardos. Maximizing the spread of influence through a social network. In KDD, pages 137 -- 146, 2003. [22] David Kempe, Jon M. Kleinberg, and ´Eva Tardos. Influential nodes in a diffusion model for social networks. In ICALP, pages 1127 -- 1138, 2005. [23] Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos, Jeanne M. VanBriesen, and Natalie S. Glance. Cost-effective outbreak detection in networks. In KDD, pages 420 -- 429, 2007. [24] Tomasz P. Michalak, Karthik V. Aadithya, Piotr L. Szczepanski, Balaraman Ravindran, and Nicholas R. Jennings. Efficient computation of the shapley value for game-theoretic network centrality. J. Artif. Int. Res., 46(1):607 -- 650, January 2013. [25] Mark Newman. Networks: An Introduction. Oxford University Press, Inc., New York, NY, USA, 2010. [26] U.J. Nieminen. On the centrality in a directed graph. Social Science Research, 2(4):371 -- 378, 1973. [27] Ignacio Palacios-Huerta and Oscar Volij. The measurement of intellectual influence. Econo- metrica, 72:963 -- 977, 2004. [28] Matthew Richardson and Pedro Domingos. Mining knowledge-sharing sites for viral marketing. In ACM, KDD '02, pages 61 -- 70, 2002. [29] Matthew Richardson and Pedro Domingos. Mining knowledge-sharing sites for viral marketing. In KDD, pages 61 -- 70, 2002. [30] G. Sabidussi. The centrality index of a graph. Psychometrika, 31(4):581 -- 603, 1966. [31] D. Schoch and U. Brandes. Re-conceptualizing centrality in social networks. European Journal of Applied Mathematics, 27:971 -- 985, 2016. [32] Lloyd. S. Shapley. A value for n-person games. In H. Kuhn and A. Tucker, editors, Con- tributions to the Theory of Games, Volume II, pages 307 -- 317. Princeton University Press, 1953. [33] Youze Tang, Yanchen Shi, and Xiaokui Xiao. Influence maximization in near-linear time: a martingale approach. In SIGMOD, pages 1539 -- 1554, 2015. 19 [34] Youze Tang, Xiaokui Xiao, and Yanchen Shi. Influence maximization: near-optimal time complexity meets practical efficiency. In SIGMOD, pages 75 -- 86, 2014. [35] Chi Wang, Wei Chen, and Yajun Wang. Scalable influence maximization for independent cascade model in large-scale social networks. DMKD, 25(3):545 -- 576, 2012. 20 Appendix A Omitted Proofs For convenience, we restate the theorems and lemmas in this section before the proofs. Proposition 3.1. ∀ influence profile I = (V, E, PI): SNI(I) = ψind[f rch](I). Proof. For any v ∈ V , SNIv(I) = σI(v) (cid:88) I[du ≤ ∞] = E(S1,...,Sn−1)∼PI = E(S1,...,Sn−1)∼PI f rch((cid:126)d({v}, S1, . . . , Sn−1)) = ψSNSSoI[f rch](I). u∈V Proposition 3.2. For any function f : Rn∞ → R, ψind[f ] conforms with µind[f ]. Proof. For any directed graph G = (V, E) and any set v ∈ V , let ({v}, Sv 1 , . . . , Sv sequence starting from v in G. Then we have n−1) be the BFS ψ[f ]v(IG) = E(S1,...,Sn−1)∼PIG ({v})[f ( (cid:126)d({v}, S1, . . . , Sn−1))] = f ( (cid:126)d({v}, Sv = f ( (cid:126)dG(v)) = µ[f ]v(G). 1 , . . . , Sv n−1)) Thus the lemma holds. Theorem 2.1 (Graph-Theoretical Basis). The set of vectors corresponding to the nontrivial layered- graph instances in L forms a basis in RM . Proof. All we have to show is independence since L = M . Suppose not, i.e., there is a nontrivial group of {λ(R0, . . . , Rt)} such that(cid:88) R0,...,Rt: I(R0,...,Rt)∈L λ(R0, . . . , Rt)PI(R0,...,Rt) = 0. Let I(R∗ 0, . . . , R∗ t∗) be a layered-graph instance t∗) (cid:54)= 0; 0, . . . , R∗ • Such that λ(R∗ • Among those satisfying the condition above, with the largest number of layers (i.e. t∗); • Among those satisfying the conditions above, with the largest number of vertices in the first layer (i.e. R∗ 0). 21 Note that fixing the seed set, the propagation on a layered-graph instance is deterministic. That is, there is exactly one cascading sequence with the fixed seed set, which happens with probability 1. Let SeqI(R0,...,Rt)(S0) be the unique BFS sequence which happens on I(R0, . . . , Rt) with seed set S0. We show that (−1)1+S0 ∅(cid:54)=S0⊆R∗ 0 R0,...,Rt: I(R0,...,Rt)∈L λ(R0, . . . , Rt)PI(R0,...,Rt)(SeqI(R∗ 0,...,R∗ t∗ )(S0)) (cid:54)= 0, which contradicts the assumption of non-independence and thereby concludes the proof. We now compute the left hand side of the above formula. (cid:88) (cid:88) (cid:88) (cid:88) ∅(cid:54)=S0⊆R∗ 0 R0,...,Rt: I(R0,...,Rt)∈L R0,...,Rt: I(R0,...,Rt)∈L λ(R0, . . . , Rt) (cid:88) ∅(cid:54)=S0⊆R∗ 0 t<t∗,R0,...,Rt: I(R0,...,Rt)∈L t>t∗,R0,...,Rt: I(R0,...,Rt)∈L λ(R0, . . . , Rt) λ(R0, . . . , Rt) R0,...,Rt∗ : I(R0,...,Rt∗ )∈L R0,...,Rt∗ : I(R0,...,Rt∗ )∈L λ(R0, . . . , Rt∗) λ(R0, . . . , Rt∗) λ(R0, . . . , Rt)PI(R0,...,Rt)(SeqI(R∗ 0,...,R∗ t∗ )(S0)) t∗ )(S0)) ∅(cid:54)=S0⊆R∗ 0 0,...,R∗ (−1)1+S0PI(R0,...,Rt)(SeqI(R∗ (−1)1+S0PI(R0,...,Rt)(SeqI(R∗ (cid:88) (cid:88) (cid:88) (cid:88) (−1)1+S0PI(R0,...,Rt∗ )(SeqI(R∗ (−1)1+S0PI(R0,...,Rt∗ )(SeqI(R∗ (−1)1+S0PI(R0,...,Rt)(SeqI(R∗ 0 0 ∅(cid:54)=S0⊆R∗ ∅(cid:54)=S0⊆R∗ 0,...,R∗ t∗ )(S0)) 0,...,R∗ t∗ )(S0)) 0,...,R∗ t∗ )(S0)) 0,...,R∗ t∗ )(S0)) = = + + = R0: R0∩R∗ = × (cid:88) ∅(cid:54)=S0⊆R∗ 0 ∅(cid:54)=S0⊆R∗ 1, . . . , Rt∗) 0 λ(R0, R∗ 1,...,Rt∗ )(SeqI(R∗ 0,...,R∗ t∗ )(S0)). Now consider the summand in the last line of the above equation. For R0 (cid:54)= R∗ 0, (−1)1+S0 (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) 0(cid:54)=∅, I(R0,R∗ 1,...,Rt∗ )∈L (−1)1+S0PI(R0,R∗ (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) 0∩R0 0∩R0 0\R0 0 ∅(cid:54)=S0⊆R∗ ∅(cid:54)=X⊆R∗ ∅(cid:54)=X⊆R∗ ∅(cid:54)=Y ⊆R∗ (−1)1+S0PI(R0,R∗ (cid:88) Y ⊆R∗ (−1)1+Y PI(R0,R∗ (−1)1+X (cid:88) 0\R0 Y ⊆R∗ (−1)1+Y × 0 0\R0 = + = = ∅(cid:54)=X⊆R∗ 0\R0 = 0. (−1)Y 22 1,...,R∗ t∗ )(SeqI(R∗ (−1)1+X+Y PI(R0,R∗ 0,...,R∗ t∗ )(S0)) 1,...,R∗ t∗ )(SeqI(R∗ 0,...,R∗ t∗ )(X ∪ Y )) 1,...,R∗ t∗ )(SeqI(R∗ 0,...,R∗ t∗ )(Y )) 0,R∗ 1,...,R∗ t∗ )(SeqI(R∗ 0,...,R∗ t∗ )(S0)) = (cid:88) ∅(cid:54)=S0⊆R∗ 0 (−1)1+S0 = 1. And for R0 = R∗ 0,(cid:88) ∅(cid:54)=S0⊆R∗ 0 Plugging these in, we get (−1)1+S0PI(R∗ (cid:88) (cid:88) ∅(cid:54)=S0⊆R∗ 0 = = (−1)1+S0 (cid:88) (cid:88) R0: R0∩R∗ 0(cid:54)=∅, I(R0,R∗ 1,...,R∗ t∗ )∈L R0: R0∩R∗ 0(cid:54)=∅, R0(cid:54)=R∗ 0, I(R0,R∗ 1,...,R∗ t∗ )∈L 0, . . . , R∗ t∗) = λ(R∗ (cid:54)= 0. λ(R0, . . . , Rt)PI(R0,...,Rt)(SeqI(R∗ 0,...,R∗ t∗ )(S0)) R0,...,Rt: I(R0,...,Rt)∈L λ(R0, R∗ 1, . . . , R∗ t∗) (cid:88) ∅(cid:54)=S0⊆R∗ 0 (−1)1+S0PI(R0,R∗ 1,...,R∗ 0,...,R∗ t∗ )(S0)) t∗ )(SeqI(R∗ t∗) × 1 λ(R0, R∗ 1, . . . , R∗ t∗) × 0 + λ(R∗ 0, R∗ 1, . . . , R∗ Proposition 3.3. If a function f : Rn∞ → R is anonymous -- i.e., f ((cid:126)d) is permutation-invariant -- then ψind[f ] (as defined in Definition 3.3) satisfies Axiom Anonymity and µind[f ] (as defined in Definition 3.2) satisfies the graph-theoretical counterpart of Axiom Anonymity. Proof. ψind[f ]π(v)(π(I)) = E(S1,...,Sn−1)∼Pπ(I)({π(v)})[f ((cid:126)d({π(v)}, π(S1), . . . , π(Sn−1)))] = E(S1,...,Sn−1)∼PI ({v})[f ((cid:126)d({v}, S1, . . . , Sn−1))] = ψind[f ]v(I). Proposition 3.4. For any function f : Rn∞ → R, ψind[f ] satisfies Axiom Bayesian. Proof. Suppose ψ[f ] is an influence-based distance-function centrality measure. Let instances I, I1, I2 be that PI = αPI1 + (1− α)PI2. Then for every sequence (S1, . . . , Sn−1) drawn from distribution PI({v}), it is equivalently drawn with probability α from PI1({v}), and with probability 1− α from PI2({v}). Therefore, ψ[f ]v(I) = E(S1,...,Sn−1)∼PI ({v})[f ((cid:126)d({v}, S1, . . . , Sn−1))] = α · E(S1,...,Sn−1)∼PI1 ({v})[f ( (cid:126)d({v}, S1, . . . , Sn−1))] + (1 − α) · E(S1,...,Sn−1)∼PI2 ({v})[f ((cid:126)d({v}, S1, . . . , Sn−1))] = α · ψ[f ]v(I1) + (1 − α) · ψ[f ]v(I2). Thus the lemma holds. 23 The proof of Theorem 3.1 relies on a general lemma about linear mapping as given in [11], as restated below. Lemma A.1 (Lemma 11 of [11]). Let ψ be a mapping from a convex set D ⊆ RM to Rn satisfying i=1 αi · i=1 αi · ψ((cid:126)vi). Suppose that D contains a set of linearly independent basis vectors of RM , i=1 λi·(cid:126)bi i=1 αi = 1, ψ((cid:80)s that for any vectors (cid:126)v1, (cid:126)v2, . . . , (cid:126)vs ∈ D, for any α1, α2, . . . , αs ≥ 0 and (cid:80)s (cid:126)vi) =(cid:80)s {(cid:126)b1,(cid:126)b2, . . . ,(cid:126)bM} and also vector (cid:126)0. Then for any (cid:126)v ∈ D, which can be represented as (cid:126)v =(cid:80)M (cid:33) for some λ1, λ2, . . . , λM ∈ R, we have (cid:33) λi · (cid:126)bi = λi · ψ((cid:126)bi) + · ψ((cid:126)0). λi (cid:32) 1 − M(cid:88) (cid:32) M(cid:88) ψ((cid:126)v) = ψ M(cid:88) i=1 i=1 i=1 Theorem 3.1. A Bayesian influence-based centrality measure is uniquely determined by its values on layered-graph instances (including the null instance). Proof of Theorem 3.1. By the definition of the null influence instance I N (same as the trivial layered-graph instance IV (V )), we can see that the vector representation of the null influence instance is the all 0 vector, because the entries corresponding to PIN (S0, S0, . . . , S0) are not included in the vector by definition. Then by Theorem 2.1 and the Lemma A.1, we know that for any Bayesian centrality measure ψ, its value on any influence instance I, ψ(I), can be represented as a linear combination of the I's values on layered-graph instances (including the null instance). Thus, the theorem holds. Theorem 3.2 (Characterization of Individual Centrality). For any anonymous function f : Rn∞ → R, ψind[f ] (defined in Definition 3.3) is the unique influence-based centrality that conforms with µind[f ] (defined in Definition 3.2) that satisfies both Axiom Anonymity and Axiom Bayesian. Proof. The theorem follows directly from Propositions 3.4, 3.3 and 3.2. Theorem 4.1 (Characterization of Group Centrality). For any anonymous function f : Rn∞ → R, ψgrp[f ] is the unique influence-based group centrality that conforms with µgrp[f ] and satisfies both Axiom Anonymity and Axiom Bayesian. Proof. The Bayesian part of the proof is essentially the same as the proof of Proposition 3.4, with subset S replacing node v. The Anonymity part of the proof is essentially the same as the proof of Proposition 3.3, with subset S replacing node v. The conformity part of the proof is essentially the same as the proof of Proposition 3.2, with subset S replacing node v. Theorem 4.2 (Characterization of Shapley Centrality). For any anonymous function f : Rn∞ → R, ψShapley[f ] is the unique influence-based centrality that conforms with µShapley[f ] and satisfies both Axiom Anonymity and Axiom Bayesian. Proof. We first show that ψShapley[f ] is Bayesian. Let I, I1, I2 be three influence instances with the same vertex set V , and α ∈ [0, 1], where PI = αPI1 + (1 − α)PI2. Then for every node v ∈ V , 24 we have ψShapley v v v (I) = φShapley = φShapley = φShapley = αφShapley = αψShapley (ψgrp(I)) (ψgrp(αPI1 + (1 − α)PI2)) (αψgrp(I1) + (1 − α)ψgrp(I2)) (ψgrp(I1)) + (1 − α)φShapley (I1) + (1 − α)ψShapley (I2), v v v v v (ψgrp(I2)) (4) where Eq. (4) is due to the linearity of the Shapley value, which is easy to verify by the following derivations: φShapley v (ατ1 + βτ2) = Eπ∼Π[(ατ1 + βτ2)(Sπ,v ∪ {v}) − (ατ1 + βτ2)(Sπ,v)] = αEπ∼Π[τ1(Sπ,v ∪ {v}) − τ1(Sπ,v)] + βEπ∼Π[τ2(Sπ,v ∪ {v}) − τ2(Sπ,v)] = αφShapley (τ1) + βφShapley (τ2). v v Now we show that ψShapley[f ] conforms with µShapley[f ]. For any directed graph G = (V, E) and any node v ∈ V , let ({v}, Sv n−1) be the BFS sequence starting from v in graph G. We have 1 , . . . , Sv ψShapley[f ]v(IG) = φShapley (ψgrp[f ](IG)) v = Eπ∼Π[ψgrp[f ]Sπ,v∪{v}(IG) − ψgrp[f ]Sπ,v (IG)] = Eπ∼Π[µgrp[f ]Sπ,v∪{v}(G) − µgrp[f ]Sπ,v (G)] = φShapley = µShapley[f ](G), (µgrp[f ](G)) v (5) where Eq. (5) is because influence-based group centrality ψgrp[f ] conforms with the structure-based group centrality µgrp[f ] (Theorem 4.1). Anonymity follows from anonymity of group centralities (Theorem 4.1). Uniqueness then follows from Theorem 3.1. Lemma A.2. For fixed nodes u, w ∈ V , suppose we generate a random RR set Rw rooted at w, according to a triggering model instance I. Then we have E[g(dRw (u, w))] = E(S1,...,Sn−1)∼PI ({u})[g(dw({u}, S1, . . . , Sn−1))]. Proof. We know that the triggering model is equivalent to the following live-edge graph model [21]: for every node v ∈ V , sample its triggering set T (v) and add edges (u, v) to a live-edge graph L for all u ∈ T (v) (these edges are called live edges). Then the diffusion from a seed set S is the same as the BFS propagation in L from S. Since reverse simulation for generating RR set Rw also do the same sampling of the triggering sets, we can couple the reverse simulation process with the forward propagation by fixing a live-edge graph L. For a fixed live-edge graph L, the subgraph GRw generated by reverse simulation from the fixed root w is simply the induced subgraph of L induced by all nodes that can reach w in L. Thus dRw (u, w) is the fixed distance from u to w in L, namely dL(u, w). On the other hand, with the fixed L, the cascading sequence starting from u is the fixed BFS sequence starting from u in L. Then in this BFS sequence dw({u}, S1, . . . , Sn−1) is the distance from u to w in this sequence, which is the same as the distance from u to w in 25 L. Therefore, dw({u}, S1, . . . , Sn−1) = dRw (u, w) for fixed live-edge graph L. We can then vary L according to its distribution, and obtain E[g(dRw (u, w))] = E(S1,...,Sn−1)∼PI ({u})[g(dw({u}, S1, . . . , Sn−1))]. (cid:80) Lemma 5.1. Let Rv be a random RR set with root v generated in a triggering model instance I. Then, ∀u ∈ V , u's stochastic sphere-of-influence individual centrality with function f ( (cid:126)d) = u∈V g(du) is ψ[f ]u(I) = n · E[g(dRv (u, v))], where the expectation is taking over the distribution of RR set Rv. Similarly, u's influence-based Shapley centrality with f is ψShapley[f ]u(I) = n · E[φShapley (g(dRv (·, v)))]. u Proof. Consider first the influence-based distance-function centrality. We have n · E[g(dRv (u, v))] = n ·(cid:88) = n ·(cid:88) (cid:88) w∈V = w∈V Pr{w = v}E[g(dRv (u, v)) w = v] E[g(dRw (u, w))] 1 n w∈V E(S1,...,Sn−1)∼PI ({u})[g(dw({u}, S1, . . . , Sn−1))] (cid:35) (cid:34)(cid:88) g(dw({u}, S1, . . . , Sn−1)) = E(S1,...,Sn−1)∼PI ({u}) = E(S1,...,Sn−1)∼PI ({u})[f ((cid:126)d({u}, S1, . . . , Sn−1))] = ψ[f ]u(I), w∈V (6) (7) (8) where Eq. (6) is by Lemma A.2. Next consider the influence-based distance-function Shapley centrality. n · E[φShapley u (g(dRv (·, v)))] = n ·(cid:88) = n ·(cid:88) w∈V w∈V Pr{w = v}E[φShapley u (g(dRv (·, v))) w = v] (g(dRw (·, w)))] (cid:33) = φShapley u E[g(dRw (·, w))] E[φShapley u 1 n (cid:32)(cid:88) w∈V (ψgrp[f ](I)) u = φShapley = ψShapley[f ]u(I), where Eq. (7) is by the linearity of the Shapley value, as already argued in the proof of Theorem 4.1, and Eq. (8) follows the similar derivation step as in the case of individual centrality above. Theorem 5.1. Let (ψv)v∈V be the true centrality value for an influence-based individual or Shapley centrality with additive function f , and let ψ(k) be the k-th largest value in (ψv)v∈V . For any > 0, (cid:96) > 0, and k ∈ [n], Algorithm ICE-RR returns the estimated centrality ( ψv)v∈V that satisfies (a) 26 unbiasedness: E[ ψv] = ψv,∀v ∈ V ; and (b) robustness: under the condition that ψ(k) ≥ 1, with probability at least 1 − 1 n(cid:96) :(cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). (3) Proof. The proof follows exactly the same proof structure as the proof in [11]. All we need to change is to use our crucial lemma connecting RR sets with centrality measures (Lemma 5.1) to replace the corresponding lemma (Lemma 23) in [11]. Lemma 5.2. For any function g : R∞ → R with g(∞) = 0, φShapley u (g(dRv (·, v))) = 1 Rv g(k) + 1 Rv! (g(k) − g(i)) (cid:88) k<i≤∆  (cid:88) 0≤j≤si − (cid:88) 0≤j≤si+1 si! j! si+1! j!  . In O(Rv) time we can compute this value for all nodes in Rv (assuming infinite precision). For (gdeg(dRv (·, v))) = 1/{w dRv (w, v) = 1} if dRv (u, v) = 1, and otherwise degree centrality, φShapley it is 0. For reachability centrality, φShapley (grch(dRv (·, v))) = 1/Rv. u u Proof. We prove only for general g. For u at the k-th level, π k<i≤∆ = = = k<i≤∆ π (cid:88) (g(k) − g(i)) 1 Rv! 1 Rv! φShapley u 1 Rv! 1 Rv g(k) + 1 Rv g(k) + (cid:88) (g(dRv (·, v))) (g(dRv (Sπ,u ∪ {u}, v)) − g(dRv (Sπ,u, v))) (cid:88) (cid:88) (cid:88) (cid:88) 1. Compute x! and(cid:80) 2. Now the term (cid:80) (cid:88)  (cid:88)  (cid:88) One possible way to compute the value above is: 1 Rv g(k) + 1 Rv g(k) + (g(k) − g(i)) (g(k) − g(i)) 1 Rv! 1 Rv! 1≤j≤si+1 1≤j≤Rv 0≤j≤si k<i≤∆ k<i≤∆ = = 0≤i≤x si! j! 0≤j≤si compute I[dRv (Sπ,u, v) = i](g(k) − g(i)) (cid:19) (cid:20)(cid:18) si (j − 1)! − j − 1 (cid:21) (j − 1)! si+1! (si+1 − j + 1)!  (cid:19) j − 1 (cid:18) si+1 − (cid:88)  . j! 1≤j≤si+1+1 si+1! si! (si − j + 1)! − (cid:88) 0≤j≤si+1 1 i! for all x ∈ [Rv] in time O(Rv). si! j! can be computed in additional constant time for any i. We can (cid:88) si+1! − (cid:88)  (cid:88) (g(k) − g(i))  si! j! 0≤j≤si 0≤j≤si+1 j! 1 Rv! k<i≤∆ 27 for all 0 ≤ k ≤ ∆ in additional total time O(∆) by first computing a suffix sum of si! j!  (cid:88)  (cid:88) 0≤j≤si g(i) − (cid:88) − (cid:88) 0≤j≤si+1 si! j! 0≤j≤si 0≤j≤si+1  . si+1! j!  . si+1! j! g(i) That is, for all k ∈ {0, . . . , ∆}. (cid:88) k<i≤∆ 3. Assign the values to vertices in each level in total time O(Rv). 28
1511.02586	1	1511	2015-11-09T07:28:33	S-PowerGraph: Streaming Graph Partitioning for Natural Graphs by Vertex-Cut	[ "cs.SI", "cs.DC" ]	One standard solution for analyzing large natural graphs is to adopt distributed computation on clusters. In distributed computation, graph partitioning (GP) methods assign the vertices or edges of a graph to different machines in a balanced way so that some distributed algorithms can be adapted for. Most of traditional GP methods are offline, which means that the whole graph has been observed before partitioning. However, the offline methods often incur high computation cost. Hence, streaming graph partitioning (SGP) methods, which can partition graphs in an online way, have recently attracted great attention in distributed computation. There exist two typical GP strategies: edge-cut and vertex-cut. Most SGP methods adopt edge-cut, but few vertex-cut methods have been proposed for SGP. However, the vertex-cut strategy would be a better choice than the edge-cut strategy because the degree of a natural graph in general follows a highly skewed power-law distribution. Thus, we propose a novel method, called S-PowerGraph, for SGP of natural graphs by vertex-cut. Our S-PowerGraph method is simple but effective. Experiments on several large natural graphs and synthetic graphs show that our S-PowerGraph can outperform the state-of-the-art baselines.	cs.SI	cs	S-PowerGraph: Streaming Graph Partitioning for Natural Graphs by Vertex-Cut Cong Xie Department of Computer Science and Engineering Shanghai Jiao Tong University 800 Dong Chuan Road Shanghai, China 200240 [email protected] Wu-Jun Li National Key Laboratory for Novel Software Technology Department of Computer Science and Technology Nanjing University [email protected] Zhihua Zhang Department of Computer Science and Engineering Shanghai Jiao Tong University 800 Dong Chuan Road Shanghai, China 200240 [email protected] 5 1 0 2 v o N 9 ] I S . s c [ 1 v 6 8 5 2 0 . 1 1 5 1 : v i X r a ABSTRACT One standard solution for analyzing large natural graphs is to adopt distributed computation on clusters. In dis- tributed computation, graph partitioning (GP) methods as- sign the vertices or edges of a graph to different machines in a balanced way so that some distributed algorithms can be adapted for. Most of traditional GP methods are of- fline, which means that the whole graph has been observed before partitioning. However, the offline methods often in- cur high computation cost. Hence, streaming graph parti- tioning (SGP) methods, which can partition graphs in an online way, have recently attracted great attention in dis- tributed computation. There exist two typical GP strate- gies: edge-cut and vertex-cut. Most SGP methods adopt edge-cut, but few vertex-cut methods have been proposed for SGP. However, the vertex-cut strategy would be a bet- ter choice than the edge-cut strategy because the degree of a natural graph in general follows a highly skewed power-law distribution. Thus, we propose a novel method, called S- PowerGraph, for SGP of natural graphs by vertex-cut. Our S-PowerGraph method is simple but effective. Experiments on several large natural graphs and synthetic graphs show that our S-PowerGraph can outperform the state-of-the-art baselines. Categories and Subject Descriptors A.3 [Distributed computing-cloud, map-reduce, MPI, others]: Big data; 7 [Data streams]; 13 [Graph mining] General Terms Algorithms, Experimentation Keywords Graph Partitioning, Streaming, Vertex-Cut 1. INTRODUCTION In this big data era, data mining and machine learning with massive datasets have captured more and more attentions from research community and industry. In particular, re- cent years have witnessed the emergence of large natural graphs in many applications like Twitter and Facebook. It is very difficult for a single commodity machine to handle such kinds of large-scale graphs. Hence, more and more works try to adopt parallel (distributed) computation on clusters for tasks with large graphs. To complete distributed computa- tion, we first need to partition the whole graph across the cluster. Hence, graph partitioning (GP) [14, 5, 6, 13] has become one of the hot topics in graph-related research. In distributed computation with clusters, the inter-machine communication cost is typically high. Hence, the goal of GP is to minimize the inter-machine (cross partition) communi- cation, and simultaneously keep the number of vertices (or edges) in every partition approximately balanced [14]. There exist two representative GP strategies: one is edge-cut and the other is vertex-cut. Edge-cut [14, 12, 15] divides the graphs by cutting the edges and assigns the vertices to dif- ferent partitions. To achieve the goal of GP, edge-cut tries to minimize the cross-partition edges and keep the number of vertices in every partition balanced. On the contrary, vertex-cut [5, 6] divides the graphs by cutting the vertices and assigns the edges to different partitions. To achieve the goal of GP, vertex-cut tries to minimize the number of replications of vertices and keep the number of edges in ev- ery partition balanced. Figure 1 illustrates the difference between these two strategies. In Figure 1 (a), the edges (A,C) and (A,E) are cut. The two partitions are node sets {A,B,D} and {C, E}, respectively. In Figure 1 (b), the node A is cut. The two partitions are edge sets {(A,B), (A,D)} and {(A,C), (A,E)}, respectively. It is easy to see that there exists essential difference between these two strategies. (a) Edge-Cut (b) Vertex-Cut Figure 1: Two strategies for graph partitioning. Shaded vertices are ghosts and mirrors respectively. Most of the traditional GP methods are offline [7], which means that the whole graph has been observed before parti- tioning. However, the offline methods often incur high com- putation cost and hence can not scale to big data applica- tions. Furthermore, the system often need to serially read the graph data from some disks onto a cluster or dynami- cally crawl the graph from some web sites. Hence, streaming graph partitioning (SGP) methods [14, 6, 12, 15], which can partition graphs in an online way, have recently attracted more and more attention from researchers. Representative methods include linear deterministic greedy (LDG) [14] and FENNEL [15]. However, most existing SGP methods adopt the edge-cut strategy for partitioning. Actually, many existing offline (non-streaming) works, such as PowerGraph [5], have found that the vertex-cut strategy can outperform edge-cut in applications with natural graphs which in general have highly skewed power-law degree dis- tributions. Even in the edge-cut based SGP work [14], the authors admitted that the edge-balanced partition, which can be achieved by vertex-cut, may be more preferable than edge-cut for power-law distributed graphs. However, few vertex-cut methods have been proposed to solve the SGP problem. In this paper, a novel method called S-PowerGraph 1 is pro- posed for SGP of natural graphs by vertex-cut. The main contributions of this paper are briefly outlined as follows: • This is the first work to systematically study and com- pare different heuristics for SGP of natural graphs by vertex-cut. • Although GraphBuilder [6] and PowerGraph [5] are not motivated by SGP and hence cannot be directly used for SGP, we propose methods to adapt them for SGP. • S-PowerGraph can be viewed as an improved version of the adapted PowerGraph, by making use of the prop- erty of power-law degree distribution for vertex-cut based SGP. • Experiments on several large natural graphs and syn- thetic graphs show that our S-PowerGraph can outper- form other baselines, including the adapted GraphBuilder and PowerGraph, to achieve the state-of-the-art per- formance in real applications. 2. PROBLEM FORMULATION Graph G is denoted by G = (V, E) where V = {v1, v2, . . . , vn} is a set of vertices and E ⊆ V × V denotes a set of edges. This implies that the number of vertices V = n, i.e., the cardinality of V is n. We say that vi and vj are neighbors if (vi, vj) ∈ E. Typically, many graph mining and learning al- gorithms can be expressed by computations on vertices of a graph, and the edges of such a graph represents dependency between the data samples stored on the vertices [10, 5]. The goal of GP is to balance the parallel work on each ma- chine and minimize the communication cost among differ- ent machines. In the following, we will briefly introduce 1The S in S-PowerGraph comes from the term streaming. two important GP strategies: edge-cut and vertex-cut. Fur- thermore, we will also discuss the SGP problem based on vertex-cut. 2.1 GP by Edge-Cut The traditional GP problem is to balance the number of vertices in each partition and minimize the edges across dif- ferent partitions, which is known as edge-cut based GP [7, 14, 12, 13, 15]. More specifically, the problem is defined as follows: (cid:12)(cid:12){e e = (vi, vj) ∈ E, vi ∈ Vx, vj ∈ Vy, x (cid:54)= y}(cid:12)(cid:12) (cid:80)p Vi i=1 Vi ≤ 1 + ξ, min A s.t. max i 1 p where · denotes the cardinality of a set, ξ is a constant with 0 ≤ ξ ≤ 1, c is the number of partitions, A = {V1, V2, . . . , Vp} is a partition of V , i.e., V1 ∪ ··· ∪ Vp = V and Vx ∩ Vy = ∅ for x (cid:54)= y. The objective function in the above formulation tries to min- imize the number of edges across different partitions, and the constraint makes the partitions as balanced as possible. The demonstration of edge-cut is shown in Figure 1 (a). Repre- sentative methods include the offline method Metis [7] and the streaming methods LDG [14] and FENNEL [15]. 2.2 GP by Vertex-Cut The edge-cut methods are efficient when partitioning sparse graphs, but they are not suitable when adapted for scale- free graphs. Typically, natural graphs are not only large but also skewed. The degree of the graph usually follows the power-law degree distribution: P (d) ∝ d −α, where P (d) is the probability that a vertex has degree d, α is a positive constant. For example, α ≈ 2.2 in the twitter- 2010 social network [8]. This skewed degree distribution might result in an imbalanced partitioning and challenge the traditional edge-cut methods. PowerGraph [5] defined a balanced p-way vertex-cut based graph partitioning model. Let each edge e ∈ E be assigned to a machine M (e) ∈ {1, . . . , p}. Then each vertex spans a set of different partitions A(v) ⊆ {1, . . . , p}. Thus, A(v) is the number of replications of v among different machines. The balanced vertex-cut based partitioning model is defined as follows: (cid:88) A(v) min A 1 V s.t. max m v∈V {e ∈ E M (e) = m} < λ E p , where λ ≥ 1 is an imbalance factor. In vertex-cut, the number of edges on each machine is used to estimate the computation cost of that machine, and the number of the replications of the vertices is used to estimate the communication cost. The objective is still to balance the workload among machines and minimize the communication cost. 2.3 SGP by Vertex-Cut The problem of SGP by vertex-cut tries to optimize the following conditional expectation: E(cid:104)(cid:88) v∈V A(v) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Mi, M (ei+1) = k (cid:105) argmin k s.t. max m {e ∈ E M (e) = m} < λ E p , , (1) where the k partitions. Thus, we define S as: (cid:40){1, ..., p} if A(u) = 0 or A(v) = 0 S = A( w) otherwise, w = argmin w D(w) s.t. w ∈ {u, v} for (u, v) = ei+1. Here D(u) is the degree of vertex u. In this paper, we simply consider the graph undirected. For particular applications, D(u) can be the degree for undirected graph or directed graph, and it can further be either in-degree or out-degree of u if the graph is directed. Additionally, since we do the partitioning job in a real on- line condition, the degree (in-degree and out-degree) of each vertex is not known in advance. The degree distribution is unveiled gradually when the edges of the graph arrive in stream. So we need to find a way to estimate the degree distribution of the whole graph. We use model in (2) and some kind of estimation of degree distribution to build the algorithm Random-Degree, Degree and DegreeIO in the next section. 3.2 Algorithms In this paper, we examine multiple partitioning algorithms. We define each of them as follows. Let P (k) denote the current set of edges belong to the kth partition, and let [p] = {1, . . . , p}. We define maxedges = max k∈[p] {P (k)} and minedges = min k∈[p] {P (k)}. We refer to a certain partition by its index idx. Then the idxth partition is denoted by Pidx. For each algorithm, and index idx will be decided by different heuristic and each edge will be assigned to the idxth partition. For the greedy algorithms, when assigning an edge, each algorithm will first compute a Score for each partition. It is the partition with the highest Score that the edge will finally be assigned to. The first three algorithms are adapted from some related works. Random The random method simply assigns each edge via a random hash function as implemented in PowerGraph. So it is the most naive algorithm. The edge e is assigned to Pidx where idx is decided by: idx = hash(e) hash(e) here is a randomized hash function. Grid GraphBuilder [6] provides a Grid-based Constrained Random Vertex-cut algorithm, which can also be adapted for streaming graph partitioning. In this Grid partitioning, a vertex is mapped into a certain shard which contains a constrained set of partitions. Then, the assignment will be chosen from the constrained set randomly. The constraint on the choice of partitions produces less replications than the naive randomized partitioning. The edge e is assigned to Pidx where idx is decided by: idx = Grid hash(e) where E denotes the expectation operation, k is the index number of a partition, Mi is the assignment of the previous i edges, ei+1 is the current edge in the stream. 3. METHODOLOGY We consider a streaming model in which edges arrive in a stream and each edge appears only once in the stream. The vertices also arrive in a stream order just like what a web crawler does. Moreover, each time a vertex arrives together with some of its edges. For SGP, the order that the vertex arrives is very important because it can influence the performance of the algorithms significantly [14, 15]. In our streaming setting, the vertices arrive in a particular order with all of their out-edges arriv- ing in a random order. As stated in existing works [14, 15], there are three typical orders in which the vertices arrive: • Random (Rnd): The vertices simply arrive in a ran- dom order or in an order given by a random permuta- tion of the vertices. • Breadth-first search (BFS): A starting vertex is selected from the graph and new vertices arrive if they are found by a breadth-first search. • Depth-first search (DFS): This order is just the same as the BFS ordering except that depth-first search will be used. 3.1 Model We observe that the replication factor, which is defined as the total number of replications divided by the total number of vertices, will be smaller if we cut more vertices with high degree. That is because when the degrees of vertices are under the power-law distribution, there exist many vertices with small degree and a relatively small amount of vertices are dense. So if choices are available, we always choose to cut a vertex with relatively higher degree. Then we can sim- ply use this strategy to improve the naive random (hashed) vertex-cut. Relative to the model in (1), we add a constraint to guar- antee if either vertex of the edge can be cut, the one with higher degree will be chosen. In particular, we defined the following model for streaming partitioning: A(v)(cid:12)(cid:12)(cid:12)Mi, M (ei+1) = k, k ∈ S (cid:105) . (2) E(cid:104)(cid:88) v∈V argmin k Here S is a set of partitions which the vertex with higher degree of ei+1 belongs to. If neither of the two vertices of ei+1 belongs to any partition, we let S simply contain all The hash function Grid hash(e) here is the same one im- plemented by GraphBuilder. Balance (i.e., PowerGraph) - An enhanced version of the greedy algorithm proposed in PowerGraph [5]. However some particular orders may result in relatively imbalance partitioning for the algorithm introduced by PowerGraph. So in practice the proposed greedy algorithm cannot be used as a baseline. We instead need an algorithm to enhances the constraint of imbalance which turns out to be the algo- rithm Balance(PowerGraph). A new constraint is added to avoid imbalance. This modified version plays a role as base- line which demonstrates the performance of the PowerGraph partitioning. Score(k) = 1{k ∈ A(u)} + 1{k ∈ A(v)} + balance(k), balance(k) = maxedges − P (k) maxedges − minedges + 1 , (u, v) = e, k ∈ [p]. where 1{#} = 1 if # is true and 1{#} = 0 otherwise. i max avgpi ≥ 1.1, then the edge e is assigned to Pidx where pi If idx is decided by: idx = argmax k∈[p] {balance(k)} Otherwise, the edge e is assigned to Pidx where idx is decided by: idx = argmax k∈[p] {Score(k)} Otherwise, the edge e is assigned to Pidx where idx is decided by: idx = argmax k∈[p] {Scorein} DegreeIO - The score of the kth partition is defined as follows: Scorein(k) = 1{k ∈ A(u)} + 1{k ∈ A(v)} + 1{k ∈ A(u), Din(u) ≤ Din(v)} + 1{k ∈ A(v), Din(v) ≤ Din(u)} + balance(k) Score(k) = 1{k ∈ A(u)} + 1{k ∈ A(v)} + 1{k ∈ A(u), D(u) ≤ D(v)} + 1{k ∈ A(v), D(v) ≤ D(u)} + balance(k), balance(k) = maxedges − P (k) maxedges − minedges + 1 , (u, v) = e, k ∈ [p]. For the current edge e, we let u be the source vertex and v be the target vertex. If Dout(u) ≥ p and Dout(v) is unknown which means vertex v has not arrived in the stream, then we do not assign the edge immediately and store it in a buffer until vertex v arrives. If Dout(v) is known, the edge e is assigned to Pidx where idx is decided by: {Score(k)} idx = argmax k∈[p] Otherwise, the edge e is assigned to Pidx where idx is decided by: The following algorithms are proposed by us which are based on the model in (2). idx = argmax k∈[p] {Scorein(k)} Random-Degree The edge e is assigned to Pidx where idx is decided by: 3.3 Theoretical Analysis Different algorithms have different motivations. idx = hash(vL) where vL = argmin w Din(w) s.t. w ∈ {u, v} Degree - Assign e to a partition of maximal score, breaking ties randomly. The score of the kth partition is defined as follows: Scorein(k) = 1{k ∈ A(u)} + 1{k ∈ A(v)} + 1{k ∈ A(u), Din(u) ≤ Din(v)} + 1{k ∈ A(v), Din(v) ≤ Din(u)} + balance(k), maxedges − P (k) maxedges − minedges + 1 , balance(k) = (u, v) = e, k ∈ [p]. i max avgpi ≥ 1.1, then the edge e is assigned to Pidx where pi If idx is decided by: idx = argmax k∈[p] {balance(k)} = 1 + The first three algorithms are from related works and rela- tively need less computation and memory. For all the greedy algorithms Balance(Powergraph), Degree and DegreeIO, when the graph is not very skewed, Balance is a good choice which requires less computation and memory. Random-Degree demonstrates the main idea of our model: decisions of which partition to be assigned to is always ac- cording to the vertex with lower degree. Although this al- gorithm cannot be used in the stream setting because the degree of each vertex is unknown in a stream, a theoret- ical analysis is available to show how our idea reduce the replication factor. Theorem 1. A random-degree vertex-cut on p machines has the expected replication factor: (cid:34) E 1 V A(v) (cid:88) v∈V (p − 1) V (cid:35) (cid:88) v∈V (cid:18) 1 − (1 − 1 p )(1−Ratio(v))D[v] (cid:19) Here D(v) is the degree of vertex v. And all edges are treated as undirected. Ratio(v) is a ratio of degree of each vertex. Proof. Our degree-random vertex-cut algorithm: for edge {D(w)w ∈ {u, v}} which is the vertex e(u, v), vL = arg min with lower degree. Then we assign the edge by some kind of hash function. So we have the assignment of e: w A(e) = hash(vL), Denote the indicator Pi as the event that vertex v has at least one of S(v) in the ith partition where S(v) is a subset of the adjacent edges of vertex v. Then follow the above procedure, the expectation E[Pi] is: E[Pi] = 1 − (1 − 1 p )Ds[v] where Ds(v) is the number of edges in S(v). We denote the fraction of the adjacent edges of v that satisfy D(v) ≤ D(u) as Ratio(v) where u is the another vertex of each adjacent edge e(u, v). Intuitively, if we assume P (d) = Cd−α I have: D(v)(cid:88) D(v)(cid:88) −α d 1 − Ratio(v) = P (d ≤ D(v)) = P (d) = C d=1 d=1 For some vertex v, Ratio(v) of its adjacent edges will be hashed to one certain partition, and the residual will be randomly hashed to all the p partitions. So for the Ratio(v) of the adjacent edges, the number of replication of v is simply 1. And for the residual 1 − Ratio(v) of the adjacent edges, we have E[Pi] = 1−(1− 1 p )(1−Ratio(v))D[v]. Here Pi will the other p− 1 partitions than the one that already have a replication. Put the two parts together, then we can have: E [A(v)] p−1(cid:88) = 1 + E[Pi] (cid:18) i=1 = 1 + (p − 1) 1 − (1 − 1 p (cid:19) )(1−Ratio(v))D[v] Thus the expected replication factor is: (cid:34) E = 1 V 1 V = 1 + (cid:35) (cid:88) (cid:88) v∈V A(v) (cid:18) 1 + (p − 1) (cid:18) (cid:88) v∈V v∈V (p − 1) V (cid:19)(cid:19) (cid:18) 1 − (1 − 1 p )(1−Ratio(v))D[v] (cid:19) 1 − (1 − 1 p )(1−Ratio(v))D[v] The two novel greedy algorithms Degree and DegreeIO can be viewed as a greedy version of Random-Degree in the stream setting. Degree is a greedy procedure which makes use of the in- degree distribution. The idea seems similar to that pro- posed in PowerLyra [3] which acts as a component of the distributed framework. However, the motivations of both are different. Unlike PowerLyra, we do not assume that the application adapted for the graph is known; i.e., we do not know which kind of edges (in-edges or out-edges) the com- putation involves in advance. Thus, we treat each edge as undirected. But the current problem is that the out-degree of a vertex is unknown until it arrives with all its out-edges. Fortunately, the degree distribution of the in-edges or the relationship of the degrees between vertices can be easily es- timated by the vertices arrived according to our streaming model. This is why we only consider the in-degree in this step. DegreeIO aims to meet the challenge that both in-degree and out-degree of natural graphs are skewed and to further improve the vertex-cut. To solve the problem that the out- degree of some vertex is unknown temporarily, the edge is buffered until the out-degree is known. For the vertex whose out-degree is small, we simply reuse the Degree step to re- duce the memory use. So if we are sure in advance that the in-degree of the graph is more skewed than the out-degree, Degree is a better choice with less computation and memory use. Otherwise DegreeIO is always a better choice with slightly more computation cost. The Degree and DegreeIO are two variants of our S-PowerGraph method. 4. EXPERIMENT In this section we demonstrate the setup and results of our experiments as well as the discussion of the results. We use a server with Intel Xeon cores, 64GB of RAM. 4.1 Datasets The collection of large graph datasets we use in the experi- ments consists of real-world graphs shown in Table 1(a) and synthetic power-law graphs shown in Table 2. Some of the real-world graphs were selected to be the same as the ex- periment of PowerLyra[3]. And some additional real-world graphs were from the UF Sparse Matrices Collection[4]. The synthetic power-law graphs were generated by a combination of two synthetic graphs. For one of them, the out-degree of each vertex is sampled randomly from a Zipf distribution[1] with parameter α and the in-degree of each vertex is under a uniform distribution. For the other one, the in-degree of each vertex is sampled randomly from a Zipf distribution with parameter β and the out-degree of each vertex is under a uniform distribution. The power-law constant (α and β) of the synthetic graphs ranges from 1.9 to 2.2. The smaller the α is, the skewer the graph will be. However the different power-law constant result in different partitioning perfor- mance. And the relationship between the distribution of in-degree and out-degree also acts as an important role. To demonstrate such fact, we separate the synthetic graph col- lection into two parts: the one with parameter α no smaller than β shown in Table 2 (a) and the one with parameter β greater than α shown in Table 2 (b). Table 1: Real-world graphs Twitter[8] It-2004[4] UK-2005[2] Alias Graph Tw It UK Arab Arabic-2005[4] Wiki Wiki[2] LJ HW Hollywood-2009[4] In As CP WG WebGoogle[9] In-2004[4] As-skitter[4] CoPapersCiteseer[4] LiveJournal[11] E V 42M 1.47B 41M 1.15B 40M 936M 22M 0.6B 5.7M 130M 5.4M 79M 1.1M 57M 1.4M 17M 1.7M 11M 0.4M 16M 0.9M 5.1M Table 2: Synthetic graphs with different power-law constants (a) Synthetic graphs with α ≥ β Alias S1 S2 S3 S4 S5 S6 S7 S8 S9 α 2.2 2.2 2.2 2.2 2.1 2.1 2.1 2.0 2.0 β 2.2 2.1 2.0 1.9 2.1 2.0 1.9 2.0 1.9 #Edges 71,334,974 88,305,754 134,881,233 273,569,812 103,838,645 164,602,848 280,516,909 208,555,632 310,763,862 (b) Synthetic graphs with α < β Alias S10 S11 S12 S13 S14 S15 α 2.1 2.0 2.0 1.9 1.9 1.9 β 2.2 2.2 2.1 2.2 2.1 2.0 #Edges 88,617,300 135,998,503 145,307,486 280,090,594 289,002,621 327,718,498 4.2 Baseline and Evaluation Metric 4.2.1 Baseline Table 3 lists several representative graph partitioning meth- ods and our S-PowerGraph method. Note that the Degree and DegreeIO are two variants of our S-PowerGraph. In our experiment, we take algorithm Random, Grid, and Bal- ance(PowerGraph) as the baseline. Note that these algo- rithms comes from PowerGraph and GraphBuilder listed in Table 3. The reason why the rest of them are not consid- ered is that the first two are of edge-cut which cannot be compared to the results of vertex-cut and the algorithm of PowerLyra cannot be used in the stream setting. 4.2.2 Evaluation Metric The result of the experiments can be described by three fac- tors: communication cost, imbalance and sensibility to num- ber of partitions. We used the replication factor to define the communication cost between different partitions which was introduced by PowerGraph[5]: (cid:88) v∈V λ = 1 V A(v) , where V is the set of all the vertices, V denotes the num- ber of vertices and A(v) denotes the number of mirrors (replications) of vertex v. We experiment the replication factor on all the real-world graphs and synthetic graphs. To make the result more clear, wee also illustrate the im- provement of our algorithms over the baseline. In all the three algorithms which act as the baseline, the algorithm Balance(PowerGraph) enjoys the lowest replication factor defined above. So we compare the replication factor of our algorithms to that of Balance(PowerGraph) and denote the improvement as: improvement = λP owerGraph − λAlgorithm λP owerGraph × 100% where λAlgorithm is denoted by the replication factor pro- duced by a certain algorithm. The second factor is the imbalance factor defined as: {e ∈ E A(e) = m} max m ρ = E/p , where A(e) denotes the partition that edge e is placed in , E is the number of edges of the graph and p is the number of partitions. However, by using the vertex-cut strategy, the imbalance is no longer a big deal. So we just show the results of the real-world graphs. Finally we tested the sensibility to the number of partitions of our algorithm. We used the twitter-2010 social network[8] and compared the replication factor by the increasing num- ber of partitions. And we also show that the algorithm is scalable with the power-law constant by taking experiments on synthetic graphs with different power-law constant. 4.3 Experimental Results In Figure 2 (a),(b) and (c), we compared the replication factor on different real-world graphs in different orders by different algorithms. In all the experiments the number of partitions is 48. Our algorithm DegreeIO produced lower replication factor in each experiment. Figure 2 (d),(e) and (f) shows the improvement of the algorithms Degree and DegreeIO compared to Balance(PowerGraph). In Figure 3 (a), (b), (c), (d), (e) and (f), we compared the replication factor on synthetic graphs with different power- law constants in different orders by different algorithms. In all the experiments the number of partitions is 48. Fig- ure 4 (a), (b), (c), (d), (e) and (f) shows the improvement of the algorithms Degree and DegreeIO compared to Bal- ance(PowerGraph). In Table 4, Table 5 and Table 6, we compared the imbalance factor on different real-world graphs in different orders by Method LDG FENNEL PowerGraph PowerLyra GraphBuilder S-PowerGraph Table 3: Some representative graph partitioning methods. Partition Strategy edge-cut edge-cut vertex-cut hybrid edge-cut /vertex-cut edge-balance vertex-cut vertex-cut edge-balance edge-balance Balance Type Streaming/Online vertex-balance vertex-balance edge-balance yes yes no no no yes (a) Rnd order (b) BFS order (c) DFS order (d) Rnd order (e) BFS order (f) DFS order Figure 2: Replication factor (first row) and improvement (second row) on real-world graphs in different stream orders. different algorithms. In all the experiments the number of partitions is 48. The imbalance factors of our algorithms are a little bit higher but still tiny. Most of them are smaller than 1.01. Table 4: Imbalance factor on real-world graphs in random order (Here PowerGraph is abbreviated to PowerG. for a better layout.) Graph Random Grid WG CP As In HW LJ Wiki Arab UK It Tw 1.00725 1.00341 1.00408 1.00394 1.00227 1.00173 1.00145 1.00057 1.0005 1.00043 1.00032 1.04038 1.01943 1.06652 1.02741 1.02349 1.00703 1.03101 1.01618 1.02545 1.01874 1.04214 Method PowerG. Degree 1.00024 1.00044 1.0005 1.00015 1.00024 1.00005 1.01349 1.00075 1.00008 1.00002 1.00005 1.00006 1.0001 1.00002 1.00055 1.00007 1.00144 1.00002 1.00031 1.00003 1 1 DegreeIO 1.00041 1.00122 1.00018 1.00298 1.00008 1.00005 1.00003 1.00006 1.00003 1.00005 1 In Figure 5 (a), (b) and (c), we evaluate our algorithm on Table 5: Imbalance factor on real-world graphs in BFS order (Here PowerGraph is abbreviated to PowerG. for a better layout.) Graph Random Grid WG CP As In HW LJ Wiki Arab UK It Tw 1.00725 1.00341 1.00408 1.00394 1.00227 1.00173 1.00145 1.00057 1.0005 1.00043 1.00032 1.04038 1.01943 1.06652 1.02741 1.02349 1.00703 1.03101 1.01618 1.02545 1.01874 1.04214 Method PowerG. Degree 1.00036 1.00015 1.00007 1.00032 1.00021 1.00003 1.0163 1.01396 1.0399 1.0245 1.00721 1.00001 1 1.00134 1.00462 1.0001 1.00069 1.00046 1.01175 1.00024 1 1 DegreeIO 1.10936 1.01744 1.00022 1.04447 1.00011 1.00006 1.00001 1.00848 1.00002 1.06839 1 WGCPASInHWLJWikiArabUKItTwitter0246810121416Replication Factor RandomGridPowerGraphDegreeDegreeIOWGCPASInHWLJWikiArabUKItTwitter0246810121416Replication Factor RandomGridPowerGraphDegreeDegreeIOWGCPASInHWLJWikiArabUKItTwitter0246810121416Replication Factor RandomGridPowerGraphDegreeDegreeIOWGCPASInHWLJWikiArabUKItTwitter05101520253035404550Improvement(%) 15.9%18.7%39.2%30.8%19%19.3%3.47%13.4%30%29.7%17.2%20%21.8%20.4%16.8%35.1%22%17.7%14.7%30.5%20.1%30.3%DegreeDegreeIOWGCPASInHWLJWikiArabUKItTwitter05101520253035404550Improvement(%) 10.1%22.1%33.7%39.5%21.6%22.4%36.7%16.5%16.4%4.4%15.7%5.86%9.88%28.2%0.552%12.3%18.4%12.5%27%DegreeDegreeIOWGCPASInHWLJWikiArabUKItTwitter05101520253035404550Improvement(%) 13.5%22.9%27.2%33.2%20.1%22.6%44%7.83%9.1%8.36%17.4%14.5%14.4%28.8%6.2%9.63%19.5%10.3%24.8%DegreeDegreeIO(a) Rnd order (b) BFS order (c) DFS order (d) Rnd order (e) BFS order (f) DFS order Figure 3: Replication factor on synthetic graphs in different stream orders. Table 6: Imbalance factor on real-world graphs in DFS order (Here PowerGraph is abbreviated to PowerG. for a better layout.) Graph Random Grid WG CP As In HW LJ Wiki Arab UK It Tw 1.00725 1.00341 1.00408 1.00394 1.00227 1.00173 1.00145 1.00057 1.0005 1.00043 1.00032 1.04038 1.01943 1.06652 1.02741 1.02349 1.00703 1.03101 1.01618 1.02545 1.01874 1.04214 Method PowerG. Degree 1.00043 1.00014 1.00027 1.00009 1.00015 1.00002 1.01655 1.0161 1.03225 1.04545 1.00534 1.00003 1.00002 1 1.00787 1.0003 1.00568 1.00022 1.047 1.00018 1 1 DegreeIO 1.00024 1.00035 1.00019 1.04168 1.00177 1.00007 1.00001 1.02141 1.00002 1.00005 1 Twitter-2010 in different order to compare the sensibility to the number of partitions. Our algorithms is always better and the improvement gets more obvious when the number of partitions increases. 4.4 Discussion Taking into account all the experiments, our algorithm De- greeIO is better than Balance(PowerGraph) in all datasets. And Degree improves the performance of Balance(PowerGraph) in most cases. For the real-world graphs, we can see that the improvement on WebGoogle is not very obvious. The reason may be that the WebGoogle graph is not so skewed and the power-law constant may be high. For the largest dataset Twitter-2010, Degree gets an improvement for about 15% to 20% while DegreeIO gets nearly 25% to 30%. Our algo- rithms produce more improvement than the algorithm intro- duced by PowerGraph[5] when the graphs are more skewed in the degree distribution which can be shown in the exper- iment on synthetic graphs. This kind of phenomenon is a consequence of the motivation of our algorithm. Recall that our algorithm makes use of the power-law degree distribu- tion to further reduce the vertex-cut. So our algorithm will not produce an obvious improvement if the distribution of degrees are nearly uniform, and it works better when the graph is more skewed. Finally the experiments on Twitter- 2010 shows good scalability for our algorithms. Note that the algorithm Degree does not always produce a better replication factor than Balance(PowerGraph). The reason is that Degree estimates the degree distribution only by the in-degree. Thus if the out-degree is more skewed than the in-degree, for example In-2004, Arabic-2005, It-2004 and the second part of the synthetic graph collection, the esti- mation will be wrong which will result in bad partitioning. And DegreeIO estimate the degree distribution by both in- degree and out-degree, so the performance is better in such datasets. For example, the improvement of Arabic-2005 is nearly 35%. 5. CONCLUSION In this paper we have studied streaming graph partition- ing by vertex-cut. In particular, we have proposed a novel streaming graph partitioning method for natural graphs by adopting vertex-cut strategy and called it S-PowerGraph. To the best of our knowledge, this is the first work to sys- S1S2S3S4S5S6S7S8S9051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIOS1S2S3S4S5S6S7S8S9051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIOS1S2S3S4S5S6S7S8S9051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIOS10S11S12S13S14S15051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIOS10S11S12S13S14S15051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIOS10S11S12S13S14S15051015202530Replication Factor RandomGridPowerGraphDegreeDegreeIO(a) Rnd order (b) BFS order (c) DFS order (d) Rnd order (e) BFS order (f) DFS order Figure 4: Improvement on synthetic graphs in different stream orders (compared with PowerGraph). (a) Rnd order (b) BFS order (c) DFS order Figure 5: Improvement on Twitter-2010 in different stream orders (compared with PowerGraph). tematically study the effectiveness of vertex-cut in streaming graph partitioning. Experiments on large graphs show that our method is more suitable for partitioning skewed natu- ral graphs than the previous approaches with an acceptable imbalance factor. Our method can be used to partition the raw graph data from a web crawler or the graph data loaded from a single storage device. In our future work, we would like to pursued more variants of the proposed strategy the- oretically and empirically. 6. REFERENCES [1] L. A. Adamic and B. A. Huberman. Zipf's law and the internet. Glottometrics, 3(1):143 -- 150, 2002. [2] P. Boldi and S. Vigna. The webgraph framework I: compression techniques. In Proceedings of the 13th international conference on World Wide Web, pages 595 -- 602, 2004. [3] R. Chen, J. Shi, Y. Chen, H. Guan, and H. Chen. Powerlyra: Differentiated graph computation and partitioning on skewed graphs. Technical Report IPADSTR-2013-001, Institute of Parallel and Distributed Systems, Shanghai Jiao Tong University, November 2013. [4] T. A. Davis and Y. Hu. The University of Florida sparse matrix collection. ACM Transactions on Mathematical Software, 38(1):1, 2011. [5] J. E. Gonzalez, Y. Low, H. Gu, D. Bickson, and C. Guestrin. Powergraph: Distributed graph-parallel S1S2S3S4S5S6S7S8S905101520253035404550Improvement(%) 5.56%9.64%9.23%12.2%14.7%16.5%21%20.4%7.97%12.8%14.2%16.7%19.6%20%9.74%17.4%16.3%20.1%DegreeDegreeIOS1S2S3S4S5S6S7S8S905101520253035404550Improvement(%) 6.53%12.2%9.88%14%13%17.8%21.5%21.1%8.86%14.9%13.1%18.3%18.2%20.6%9.79%19.2%15.7%21.3%DegreeDegreeIOS1S2S3S4S5S6S7S8S905101520253035404550Improvement(%) 5.72%12.2%9.33%14.7%14.8%19.7%20.8%21.9%7.88%15.7%13.8%19%19.6%23.1%8.32%20.7%15.1%22.7%DegreeDegreeIOS10S11S12S13S14S1505101520253035404550Improvement(%) 3.53%11.2%15.6%4.94%14.8%17.6%17%20.2%DegreeDegreeIOS10S11S12S13S14S1505101520253035404550Improvement(%) 4.16%14.8%0.111%16.4%4.19%17.4%18.6%21.1%19.6%DegreeDegreeIOS10S11S12S13S14S1505101520253035404550Improvement(%) 4.3%14.3%0.407%18.1%4.67%18%18.5%19.2%1.4%21.1%DegreeDegreeIO81632486405101520253035404550Improvement(%)Number of Partitions 9.1%18%15.3%24.7%18.8%29.2%20.1%30.3%20.3%30.5%DegreeDegreeIO81632486405101520253035404550Improvement(%)Number of Partitions 6.21%15.2%9.93%21.5%11.9%25.8%12.4%27%12.4%27.2%DegreeDegreeIO81632486405101520253035404550Improvement(%)Number of Partitions 7.08%11.3%10.1%17.9%11.2%22.9%11%24.8%10.6%25.3%DegreeDegreeIOcomputation on natural graphs. In Proceedings of the 10th USENIX Symposium on Operating Systems Design and Implementation (OSDI), pages 17 -- 30, 2012. [6] N. Jain, G. Liao, and T. L. Willke. Graphbuilder: scalable graph etl framework. In First International Workshop on Graph Data Management Experiences and Systems, page 4, 2013. [7] G. Karypis and V. Kumar. Multilevel graph partitioning schemes. In Proceedings of the International Conference on Parallel Processing, pages 113 -- 122, 1995. [8] H. Kwak, C. Lee, H. Park, and S. Moon. What is twitter, a social network or a news media. In Proceedings of the 19th international conference on World Wide Web, pages 591 -- 600, 2010. [9] J. Leskovec. Stanford large network dataset collection. URL http://snap. stanford. edu/data/index. html, 2011. [10] Y. Low, J. Gonzalez, A. Kyrola, D. Bickson, C. Guestrin, and J. M. Hellerstein. GraphLab: A new framework for parallel machine learning. In Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI), pages 340 -- 349, 2010. [11] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee. Measurement and analysis of online social networks. In Proceedings of the 7th ACM SIGCOMM conference on Internet measurement, pages 29 -- 42, 2007. [12] J. Nishimura and J. Ugander. Restreaming graph partitioning: simple versatile algorithms for advanced balancing. In Proceedings of the ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1106 -- 1114, 2013. [13] I. Stanton. Streaming balanced graph partitioning algorithms for random graphs. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1287 -- 1301, 2014. [14] I. Stanton and G. Kliot. Streaming graph partitioning for large distributed graphs. In Proceedings of the ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1222 -- 1230, 2012. [15] C. E. Tsourakakis, C. Gkantsidis, B. Radunovic, and M. Vojnovic. FENNEL: Streaming graph partitioning for massive scale graphs. In Proceedings of the ACM International Conference on Web Search and Data Mining, 2014.
1806.03232	2	1806	2019-04-24T12:05:04	Randomized Optimal Transport on a Graph: framework and new distance measures	[ "cs.SI", "cs.LG", "stat.ML" ]	The recently developed bag-of-paths (BoP) framework consists in setting a Gibbs-Boltzmann distribution on all feasible paths of a graph. This probability distribution favors short paths over long ones, with a free parameter (the temperature $T$) controlling the entropic level of the distribution. This formalism enables the computation of new distances or dissimilarities, interpolating between the shortest-path and the resistance distance, which have been shown to perform well in clustering and classification tasks. In this work, the bag-of-paths formalism is extended by adding two independent equality constraints fixing starting and ending nodes distributions of paths (margins). When the temperature is low, this formalism is shown to be equivalent to a relaxation of the optimal transport problem on a network where paths carry a flow between two discrete distributions on nodes. The randomization is achieved by considering free energy minimization instead of traditional cost minimization. Algorithms computing the optimal free energy solution are developed for two types of paths: hitting (or absorbing) paths and non-hitting, regular, paths, and require the inversion of an $n \times n$ matrix with $n$ being the number of nodes. Interestingly, for regular paths on an undirected graph, the resulting optimal policy interpolates between the deterministic optimal transport policy ($T \rightarrow 0^{+}$) and the solution to the corresponding electrical circuit ($T \rightarrow \infty$). Two distance measures between nodes and a dissimilarity between groups of nodes, both integrating weights on nodes, are derived from this framework.	cs.SI	cs	Randomized Optimal Transport on a Graph: framework and new distance measures (draft preprint paper accepted for publication in Network Science journal) Guillaume Guex ICTEAM, Universit´e catholique de Louvain, Belgium Ilkka Kivimaki Department of Computer Science, Aalto University, Finland & ICTEAM, Universit´e catholique de Louvain, Belgium Marco Saerens ICTEAM, Universit´e catholique de Louvain, Belgium 18/02/2019 Abstract The recently developed bag-of-paths (BoP) framework consists in set- ting a Gibbs-Boltzmann distribution on all feasible paths of a graph. This probability distribution favors short paths over long ones, with a free parameter (the temperature T ) controlling the entropic level of the dis- tribution. This formalism enables the computation of new distances or dissimilarities, interpolating between the shortest-path and the resistance distance, which have been shown to perform well in clustering and clas- sification tasks. In this work, the bag-of-paths formalism is extended by adding two independent equality constraints fixing starting and ending nodes distributions of paths (margins). When the temperature is low, this formalism is shown to be equivalent to a relaxation of the optimal transport problem on a network where paths carry a flow between two discrete distributions on nodes. The randomization is achieved by consid- ering free energy minimization instead of traditional cost minimization. Algorithms computing the optimal free energy solution are developed for two types of paths: hitting (or absorbing) paths and non-hitting, regu- lar, paths, and require the inversion of an n × n matrix with n being the number of nodes. Interestingly, for regular paths on an undirected graph, the resulting optimal policy interpolates between the deterministic optimal transport policy (T → 0+) and the solution to the corresponding electrical circuit (T → ∞). Two distance measures between nodes and a dissimilarity between groups of nodes, both integrating weights on nodes, are derived from this framework. Keywords: Network Science, Optimal Transportation, Bag of Paths, Randomized Shortest Path, Distances between Nodes, Link Analysis. 1 1 Introduction 1.1 General introduction and motivation Today, network data are studied in many different areas of science, including ap- plied mathematics, computer science, social science, physics, chemistry, pattern recognition, applied statistics, data mining and machine learning, to name a few (see, e.g., (Barab´asi, 2016; Chung & Lu, 2006; Estrada, 2012; Fouss, Saerens, & Shimbo, 2016; Kolaczyk, 2009; Lewis, 2009; Newman, 2010; Silva & Zhao, 2016; Thelwall, 2004; Wasserman & Faust, 1994)). In this context, one key problem is the definition of distances between nodes taking both direct and indirect connec- tions into account (Chebotarev, 2011, 2012, 2013; Fouss et al., 2016; Herbster & Lever, 2009; Fran¸coisse, Kivimaki, Mantrach, Rossi, & Saerens, 2017; Lu & Zhou, 2011; Alamgir & von Luxburg, 2011; Yen, Mantrach, Shimbo, & Saerens, 2008). This problem is faced in many applications such as link prediction, com- munity detection, node classification, and network visualization, among others. Now, it has been shown that the standard shortest path distance and the resistance distance (Klein & Randi´c, 1993) suffer from important drawbacks in some situations, which sometimes hinders their use as distance measures between nodes. More precisely, the shortest path distance does not integrate the concept of high connectivity between the two nodes (it only considers the shortest paths, see, e.g., (Fouss et al., 2016)), while the resistance distance provides useless results when dealing with large graphs (the so-called "lost-in- space effect" (von Luxburg, Radl, & Hein, 2010, 2014)). Another drawback of the shortest path distance is that it provides lots of ties when comparing distances, especially on unweighted undirected graphs. In order to avoid the drawbacks of the shortest path and resistance distances, new families of distance measures, interpolating between these two extremes, were recently suggested based on a bag-of-paths (BoP) framework (Fran¸coisse et al., 2017; Kivimaki, Shimbo, & Saerens, 2014; Lebichot, Kivimaki, Fran¸coisse, & Saerens, 2014; Mantrach et al., 2010). This framework defines a Gibbs- Boltzmann probability distribution over paths on a graph, which focuses on the shortest paths, but spreads also on longer paths and random walks. The spread of the distribution is controlled by a temperature parameter monitoring the bal- ance between choosing low-cost paths and a pure random behaviour. Different distance measures between nodes are then derived based on this distribution; other ones are described in the next, related work, subsection. Following this previous work, the effort is pursued in this paper with the introduction of weighted distance measures derived from new bag-of-paths (BoP) models. The weights of the distances are determined by introducing equality constraints on the path distribution margins, i.e., the a priori probabilities over starting nodes and ending nodes of paths. In other words, the new model assumes that the user knows only where paths on average start and where they on average end but not how they are distributed otherwise. In the original BoP model, which was developed for the unweighted BoP-based distances, the starting and ending node distributions are instead unconstrained, and can be inferred directly from the probabilities of the paths (Fran¸coisse et al., 2017). The model proposed in the current work will be called the margin-constrained bag-of-paths framework (abbreviated as cBoP). More precisely, the work defines two models by considering two different types of paths -- the first one is based 2 on regular, non-hitting paths, and the second on hitting paths, i.e. paths where the ending node cannot appear as an intermediate node. Weighting the nodes of a network for determining distances can be impor- tant in applications where each node represents a whole collection of items (like cities where nodes could be weighted by population). Moreover, in some situ- ations, it could be beneficial to weigh nodes by the reciprocal of their degree in order to avoid the hubness effect (Radovanovi´c, Nanopoulos, & Ivanovi´c, 2010a). This will be investigated experimentally in further work. In addition to defining weighted distances between nodes on a graph, a dissimilarity measure between groups of nodes on the graph is derived from the margin-constrained BoP framework. The margin-constrained BoP model can also be understood as defining a ran- domized policy for the optimal transport problem on a graph (Ahuja, Magnanti, & Orlin, 1993; Kantorovich, 1942; Villani, 2003, 2008), because the starting and ending node distributions can be considered as supply and demand distri- butions for goods meant to be transported over the graph. The randomization is achieved by finding the probability mass on the set of paths connecting starting and ending nodes that minimizes free energy (a balance between expected cost and entropy), subject to margin constraints corresponding to the predefined supply and demand distributions. As is common for such formulations, min- imizing this objective function results in a Gibbs-Boltzmann probability mass on paths. As discussed in more detail in (Saerens, Achbany, Fouss, & Yen, 2009), randomization from optimality can prove useful for several reasons, both in the context of transportation, as well as when measuring distance: • If the environment is changing over time (non-stationary), the system could benefit from randomization by performing continual exploration. • A deterministic policy makes behavior totally predictable; on the contrary, randomness introduces unpredictability and therefore renders interception more difficult. Randomization has proved useful for this reason in game theory (see, e.g., (Osborne, 2004)). • A randomized policy spreads the traffic on multiple paths, therefore re- ducing the danger of congestion. • A distance measure accounting for all paths -- and thus integrating the concept of high connectivity -- can be more useful, e.g. in social network analysis, than relying on the best paths only. 1.2 Related work The model proposed in this work builds on and extends previous work dedicated to the bag-of-paths (BoP) framework (Fran¸coisse et al., 2017; Mantrach et al., 2010), as well as the randomized-shortest-path (RSP) framework (Akamatsu, 1996; Kivimaki et al., 2014; Kivimaki, Lebichot, Saramaki, & Saerens, 2016; Saerens et al., 2009; Yen et al., 2008), and their variants (Bavaud & Guex, 2012; Guex & Bavaud, 2015; Guex, 2016); see also (Zhang, Zhao, & Wang, 2013) for a related proposition, called path integral. 3 The main motivation for using such models can be understood as follows (Lebichot & Saerens, 2018). Most of the traditional network measures are es- sentially based on two different paradigms about movement or communication occurring in the network: optimal communication based on shortest paths, and random communication based on a random walk on the graph. For instance, shortest path distance, as well as the standard betweenness centrality (Freeman, 1977) are defined from shortest paths, while resistance distance and random walk centrality (Brandes & Fleischer, 2005; Newman, 2005) are based on ran- dom walks (which have a strong analogy with electrical flow on the network (Doyle & Snell, 1984)). However, in reality, communication or movements over a network seldom occur either optimally or purely randomly. The BoP and RSP frameworks both relax these assumptions by interpolating between shortest paths and a pure random walk based on a temperature parameter. This enables the definition of measures with increased adaptability given by the temperature parameter. In addition to defining distances interpolating between the shortest path and resistance distances, (Fran¸coisse et al., 2017; Kivimaki et al., 2014; Yen et al., 2008), the models can also be used to define a centrality measure interpolating between a shortest path-based betweenness and the random walk betweenness (Kivimaki et al., 2016). Besides the works mentioned above, other new families of distances have re- cently been developed integrating information on both the proximity (shortest- path distance) and amount of connectivity between nodes (captured, e.g., by the resistance distance) (Chebotarev, 2011, 2012, 2013; Fouss et al., 2016; Hashimoto, Sun, & Jaakkola, 2015; Herbster & Lever, 2009; Li, Zhang, & Boley, 2013; Lu & Zhou, 2011; Nguyen & Mamitsuka, 2016; Alamgir & von Luxburg, 2011). Many of these measures indeed interpolate (up to a constant scaling fac- tor) between the shortest path distance (or length) and the resistance distance, therefore alleviating the previously mentioned lost-in-space effect. A short discussion of the standard, deterministic, optimal transport on a graph problem appears in Section 4. Methods based on the optimal trans- port problem using entropic regularization have recently been investigated in a number of pattern recognition and machine learning tasks (e.g., (Courty, Fla- mary, Tuia, & Rakotomamonjy, 2017; Solomon, Rustamov, Guibas, & Butscher, 2014)). For instance, (Cuturi, 2013; Ferradans, Papadakis, Peyr´e, & Aujol, 2014; Guex, Emmanouilidis, & Bavaud, 2017) propose to regularize the stan- dard objective function of the classical discrete optimal transport problem with an entropy term. They show on various problems, including image processing problems, that the resulting algorithm is much faster than the original one. Note that discrete entropy-regularized optimal transport problems were previ- ously studied in economics, transportation science and operations research (see, e.g., (Wilson, 1970; Erlander & Stewart, 1990; Kapur, 1989; Kapur & Kesavan, 1992; Fang, Rajasekera, & Tsao, 1997)). The main difference with these previ- ous contributions is that the present work defines the different quantities, such as entropy or cost, over full paths on the network by adopting a sum-over-paths formalism. Finally, the hubness effect, mentioned earlier, has been studied recently in various works (Radovanovi´c, Nanopoulos, & Ivanovi´c, 2010b; Radovanovi´c et al., 2010a; Suzuki, Hara, Shimbo, Matsumoto, & Saerens, 2012; Suzuki, Hara, Shimbo, Saerens, & Fukumizu, 2013; Tomasev, Radovanovic, Mladenic, 4 & Ivanovic, 2014; Hara et al., 2015). Hubness is a problem faced with high- dimensional data, e.g. when building nearest-neighbor graphs, as some nodes may become over-represented as hubs in such graphs due to concentration of distances in high-dimensional spaces. The weighting of distances provided by the margin-constrained BoP framework could help alleviate this effect in graph- based data analysis applications. 1.3 Main contributions This work defines weighted distance measures between graph nodes by develop- ing a margin-constrained bag-of-paths model. This model can be interpreted as a solution to the optimal transport problem on a graph involving a regularization term. The problem is tackled by using Kullback-Leibler divergence (also called relative entropy (Cover & Thomas, 2006)) as regularization term. Furthermore, two types of paths are considered: regular paths and hitting, absorbing, paths. The optimal randomized policy consists in the assignment of a probability distribution on the set of choices (deciding to follow an available edge) for each node of the network. It therefore defines (optimal) biased transition probabili- ties "attracting" the agents to the destination nodes. Furthermore, the model depends on a temperature parameter monitoring the balance between exploita- tion (expected cost) and exploration (entropy of paths) so that the solutions interpolate between the classical deterministic optimal transport solution (pure exploitation) and the random walk on the graph provided a priori by the user (pure exploration). Low temperatures correspond to (randomized) near-optimal solutions while high temperatures simply provide the predefined random walk behavior. Note that, when considering hitting paths, the model reduces to the standard randomized shortest path model when there is only one unique initial node and one unique destination node. The first contribution consists in deriving the probability distribution over paths minimizing expected cost under relative entropy regularization and mar- gin constraints, for both regular and hitting paths. Once the probability distri- bution over paths is derived, all the quantities of interest, such as • the policy (optimal routing transition probabilities), • the flow over the network based on the a priori starting and ending node distributions of paths, • a weighted distance measure between nodes, and • a dissimilarity between groups of nodes can be defined and computed by simple matrix expressions. Note that the present work is partly a re-interpretation of (Guex, 2016) in which the author already studied a similar optimal transport on a graph problem regularized by an entropic term. There, the entropic term at the node level was defined by considering, on each node, the relative entropy between the desired transition probabilities (the policy) and the reference transition probabilities corresponding to a natural random walk on the graph. Then, the global entropic regularization term was defined as a weighted sum of the relative entropies over all nodes. As in (Saerens et al., 2009), the weighting factor is set to the expected number of visits to the node, therefore putting more emphasis on frequently 5 visited nodes. In the current work, we adopt a paths-based formalism and the entropic term is instead defined according to the relative entropy over path distributions. Interestingly, it was found that the model derived in (Guex, 2016) is exactly equivalent to one of the two models introduced in this work, the one dealing with regular, non-hitting, paths, in the sense that they provide the same routing policy. Therefore, in comparison with (Guex, 2016), the present work reformu- lates the problem in terms of probabilities and relative entropy over paths in the network, instead of transition probabilities on nodes. It also introduces another algorithm for solving the problem and it derives a new algorithm for dealing with hitting paths. In short, the main contributions of this paper are • the development of a new margin-constrained bag-of-paths framework, con- sidering fixed probability distributions on starting and ending nodes, • the introduction of a randomized solution to the optimal transport on a graph problem for both regular and hitting paths, • the definition of a new distance measure between nodes and a dissimilarity between groups of nodes derived from this framework, and • some illustrative simulations to explore the potential of the framework. The remaining of the paper is as follows. Section 2 develops the formalism and derives the solution to the margin-constrained bag-of-paths problem on a graph for regular paths, while Section 3 extends the model to hitting paths. Then, Section 4 discusses the connections with the standard optimal transport on a graph problem. The derived distances are introduced in Section 5. Section 6 provides some illustrative simulations. Finally, Section 7 is the conclusion. 2 The margin-constrained bag-of-paths formal- ism 2.1 Background and notation This Subsection first sets the notation and terminology of the paper, after which the standard bag-of-paths (BoP) and randomized shortest-paths (RSP) frame- works are briefly reviewed (note that a discussion of the standard optimal trans- port problem is deferred to Section 4). Then the margin-constrained BoP (cBop) framework, and the relevant related results, are presented. Note that in this sec- tion regular, non-hitting, paths are considered whereas Section 3 restricts the set of paths to hitting, or absorbing, paths where the ending node may appear only once as final node. Notation. In this paper, we always assume a weighted, strongly connected, directed graph G = (V,E), with set of nodes V = {1, 2, . . . , n} and set of edges E = {(i, j)} containing m edges in total. The nonnegative weights on edges, noted aij, represent local affinities between nodes, and are contained in the weighted adjacency matrix A. Edge weights define a natural reference transition probabilities matrix Pref of a standard random walk on G, with Pref = D−1A, 6 where D is the diagonal matrix containing row sums of A (outdegrees). The Markov chain defined by these transition probabilities is assumed to be regular. Elementwise, we have pref ij = . (1) aij(cid:80) k∈V aik Along with weights, nonnegative edges costs, noted cij, are also provided. These costs are contained in the cost matrix C, and can be defined either independently from weights aij, or, e.g., thanks to cij = 1/aij. We define a t-length path on the graph G, denoted by ℘, as a sequence of nodes ℘ = (i0, . . . , it), where t > 0 and (iτ−1, iτ ) ∈ E for all τ = 1, . . . , t. Note that a node can appear several times on the path (including the ending node). We denote a path starting in node i and ending in node j by ℘ij. The likelihood πref (℘) of a t-length path ℘ = (i0, . . . , it) starting in i0 and ending in it is defined by πref (℘) (cid:44) Πt and its cost τ =1 ciτ−1,iτ . We further denote respectively by Pij and P, the set of paths starting in i and ending in j and the set of all paths in G, also named the bag-of-paths, with P = ∪i,j∈V Pij. By convention, zero-length paths starting and ending in the same node with a zero cost are also included in the set of paths (see (Fran¸coisse et al., 2017) for details). c(℘) by c(℘) (cid:44) (cid:80)t τ =1pref iτ−1,iτ All vectors will be column vectors and denoted in lowercase bold while ma- trices are in uppercase bold. The bag-of-paths and the randomized-shortest-path frameworks. The context defined above states the usual background of the bag-of-paths framework developed in (Fran¸coisse et al., 2017; Mantrach et al., 2010). In these works, a probability distribution over the set of all paths, P(℘) with ℘ ∈ P, was con- structed in order to favor paths of low cost subject to a constant relative entropy constraint. It provides the probability of drawing a particular path ℘ from a bag, with replacement. The distribution can equivalently be obtained by solving the following problem (see (Kivimaki et al., 2014))1: (cid:88) ℘∈P (cid:18) P(℘) (cid:19) Pref (℘) (2) (cid:88) ℘∈P minimize {P(℘)}℘∈P subject to (cid:80) ℘∈P P(℘) = 1, FE(P) = P(℘)c(℘) + T P(℘) log where T > 0, the temperature, is a free parameter defining the relative entropy level (Cover & Thomas, 2006), and Pref (℘) is the natural, reference, probability of a path ℘ depending on πref (℘) (see Equation (1)) and to be discussed later. FE(P) is called the (relative) free energy, due to its similarity with the statistical physics quantity. It corresponds to the expected cost, or energy, to which the relative entropy weighted by temperature is added. Strictly speaking, a non- negativity constraint should also be added to (2), but this is not necessary since the resulting probability distribution will automatically be non-negative. Indeed, following (Jaynes, 1957; Cover & Thomas, 2006; Kapur & Kesavan, 1992) and (Fran¸coisse et al., 2017; Kivimaki et al., 2014; Mantrach et al., 2010) for the paths formalism, the solution is a standard Gibbs-Boltzmann distribution 1Alternatively, it can also be obtained by following a maximum entropy argument (Jaynes, 1957; Cover & Thomas, 2006; Kapur & Kesavan, 1992). 7 of the form P(℘) = (cid:88) ℘(cid:48)∈P Pref (℘) exp [−θc(℘)] Pref (℘(cid:48)) exp [−θc(℘(cid:48))] , where c(℘) is the total cumulated cost along path ℘. It provides the probability of choosing any path ℘ ∈ P in the network. In addition, the randomized shortest path framework (Saerens et al., 2009; Kivimaki et al., 2014; Yen et al., 2008); inspired by (Akamatsu, 1996) restricts the set of paths to hitting paths (see next section for details) connecting only two predefined nodes i and j. This defines optimal randomized policies for reaching node j from i, ranging from shortest paths to a random walk. A method for computing the RSP on large sparse graphs by restricting the set to paths with a finite predefined length was developed in (Mantrach et al., 2011, Section 4). 2.2 Problem definition and σout, with(cid:80) In this Section, the BoP framework developed in (Fran¸coisse et al., 2017; Mantrach et al., 2010) is extended into a margin-constrained bag-of-paths framework by introducing two additional density vectors on nodes, provided by the user: σin i ≥ 0, ∀i ∈ V. These vec- tors define desired constraints on the distribution margins of our bag-of-paths probabilities, i.e. i = 1 and σin i∈V σout i , σout i∈V σin i =(cid:80) P(S = i) (cid:44)(cid:88) P(E = j) (cid:44)(cid:88) j∈V (cid:88) (cid:88) ℘ij∈Pij i∈V ℘ij∈Pij P(℘ij) = σin i P(℘ij) = σout j ∀i ∈ V, ∀j ∈ V, (3) (4) where S and E denote random variables containing respectively the starting and the ending nodes of the drawn path. In turn, it means that we want to constrain the probability of picking a path in the BoP starting from i to value σin i and the probability of picking a path ending in j to value σout . The intuition is as follows: the model assumes that we are carrying a unit of goods in the network from the set of nodes In = {i ∈ V : σin i > 0} (supply nodes) to the set j > 0} (demand nodes) in an optimal way by minimizing a Out = {j ∈ V : σout balance between expected cost and relative entropy of paths. A discussion of this model in the light of optimal transport on a graph appears later in Section 4. distribution, {P(℘)}, ℘ ∈ P, solving Altogether, we extend problem (2) and seek the optimal paths probability j FE(P) = minimize {P(℘)}℘∈P j∈V(cid:80) subject to (cid:80) (cid:80) i∈V(cid:80) Note that as we have(cid:80) ℘∈P ℘ij∈Pij ℘ij∈Pij i =(cid:80) P(℘)c(℘) + T P(℘) log ∀i ∈ V ∀j ∈ V. P(℘ij) = σin i P(℘ij) = σout j i = 1, the constraint(cid:80) ℘∈P P(℘) = 1 in (2) can be dropped. The goal of this problem is to find a probability distribution with fixed margins, such that it favors the paths of least cost when T → 0, and the paths with high likelihood πref (℘) when T → ∞. i∈V σout i∈V σin (cid:88) (cid:88) ℘∈P (cid:18) P(℘) (cid:19) Pref (℘) (5) 8 the definition is simply Pref (℘) = πref (℘)/(cid:80) In order for the high temperature bounds to be consistent, Pref (℘) must be defined according to πref (℘). In the usual, unconstrained, BoP formalism with a uniform a priori probability of choosing the starting and ending node, ℘(cid:48)∈P πref (℘(cid:48)) (Fran¸coisse et al., 2017)2. However, defining reference probabilities is not as trivial in the margin- constrained setting studied in this work because of the constraints. This defi- nition is the goal of the next section (note that the reader mainly interested in the randomized optimal transport problem can simply assume that the reference transition probabilities as given, skip the Section 2.3 and go directly to Section 2.4). The solution of problem (5) is then stated and proved in Section 2.4. 2.3 Reference probabilities with fixed margins The reference probability of a path, Pref (℘), should have appropriate margins in order to ensure the convergence P(cid:63)(℘) → Pref (℘) when T → ∞ (pure random walk). In other words, for consistency, the reference probabilities of paths should also satisfy the constraints, Pref (S = i) = Pref (℘ij) = σin i Pref (E = j) = Pref (℘ij) = σout j (cid:88) (cid:88) j∈V (cid:88) (cid:88) ℘ij∈Pij i∈V ℘ij∈Pij ∀i ∈ V, ∀j ∈ V, (6) (7) which further implies that the path probabilities Pref (℘ij) sum to one. In (Fran¸coisse et al., 2017), the reference distribution of a path is simply set proportional to path likelihood πref (℘), which can be interpreted as follows: the starting distribution Pref (S = i) is defined as uniform and the ending dis- tribution Pref (E = j) is equal to the proportion of time spent in each node for the Markov chain St, defined by transition probabilities matrix Pref , when t → ∞ (stationarity). Obviously, defining the reference probability in a similar way here would lead to a problem: the ending distribution Pref (E = j) depends entirely on the transition matrix of the Markov chain defined by Pref and will generally not yield the desired distribution σout. To address this problem we introduce a new killed Markov process which will lead to the desired ending distribution, while being as similar as possible, in a certain sense, to the original chain. More precisely, we design this killed Markov process in such a way that the random walker encounters exactly the same probabilities of jumping to any adjacent node as the original transition probabilities (Pref ) as long as he survives. 2.3.1 A particular killed Markov process Let us first define a killed random process which will be helpful later. Definition 2.1. From the reference, regular, Markov process St defined on the state space V, with initial distribution πin and transition matrix Pref , a killed Markov process, denoted by Kt, is defined as a new process with the same 2Note that non-uniform prior probabilities on the starting and ending node are briefly discussed in (Fran¸coisse et al., 2017). 9 initial distribution and following a substochastic transition matrix (cid:98)Pref , given by (cid:98)Pref (cid:44) (I − Diag(α))Pref , (8) where the probability to be killed after visiting node i is αi, with 0 ≤ αi ≤ 1 ∀i ∈ V. Diag(α) is the diagonal matrix containing vector α on its diagonal. In other words, the vector α = (αi) contains the killing rate of each node. This killed Markov process can be seen as adding a new (virtual) absorbing state, the cemetery ω, and following the rules P(Kt+1 = jKt = i) =(cid:98)pref P(Kt+1 = ωKt = i) = 1 −(cid:98)pref P(K0 = i) = πin i , ij = (1 − αi)pref ij , i• (cid:44) αi, P(Kt+1 = ωKt = ω) = 1, (9) (10) (11) (12) where the dot in (cid:98)pref We observe that (cid:98)pref nodes V). ij /(cid:98)pref i• means summation over the second index (over the set of ∞(cid:88) i• = pref ij for any α. This means that this killing process will behave similarly to the original process as long as it survives, thus arguing in favor of the similarity requirement between the two chains. How- ever, unlike the reference Markov chain, this killed Markov process possesses an "ending" distribution. Definition 2.2. A killed Markov process, as defined in Definition 2.1, possesses an ending distribution, denoted by πout = (πout ) and given by i πout i (cid:44) P(KM−1 = i) = P(Kt = i, Kt+1 = ω), (13) t=0 where M is the random variable containing the time where the process is killed (it reaches the cemetery state ω). This quantity denotes the probability of being killed in state i when starting it sums up the probability of being from initial nodes with probabilities πin i : killed after 0, 1, 2, . . . steps. Interestingly, it is possible to find the vector of killing rates α corresponding to a desired ending distribution πout = σout. This is important as it will allow us to design a proper reference probability distribution satisfying the predefined margins. But we first need the following preliminary lemma. ij =(cid:80)∞ Lemma 2.1. The expected number of visits to j before being killed, given that the t=0((cid:98)Pref )t = (I −(cid:98)Pref )−1 which is t=0 P(Kt = jK0 = i), process started from state i, that is, the quantity ¯nref is given by element i, j of matrix (cid:98)Zref (cid:44)(cid:80)∞ well-defined for a substochastic matrix and a strongly connected graph. In other words, ¯nref ij =(cid:98)zref ij . Proof. See (Fouss et al., 2016, Section 1.5.7). Then, the killing rates and the ending distribution are related by the fol- lowing proposition. This will allow us to determine the killing rates in order to satisfy a predefined ending distribution. 10 Proposition 2.1. For a killed Markov process on a strongly connected graph (see Definition 2.1), the following equality is satisfied (cid:80) α = πout ÷ ¯nref , (14) where ÷ is the elementwise division and column vector ¯nref = (¯nref j = ij , holds the expected number of visits to state i before being killed j ) with ¯nref i ¯nref i∈V πin which can be computed thanks to (cid:0)I − (Pref )(cid:62)(cid:1)¯nref = πin − (Pref )(cid:62)πout, (15) where Pref is the transition probability matrix defined in Equation (1). Proof. First, let us observe that the joint distribution for the starting node and the ending node of the killing process is P(K0 = i, KM−1 = j) = = ∞(cid:88) ∞(cid:88) t=0 t=0 (cid:88) P(K0 = i, Kt = j, Kt+1 = ω) P(Kt+1 = ωKt = j)P(Kt = jK0 = i)P(K0 = i) ((cid:98)Pref )t (cid:35) ij (cid:34) ∞(cid:88) i (cid:98)zref t=0 = πin i (cid:104)(cid:98)Zref(cid:105) αj = πin i αj ij (cid:88) i∈V i (cid:98)zref πin ij αj. where(cid:98)Zref is defined by Lemma 2.1. Thus the ending distribution (see Definition ij αj, = πin (16) 2.2) reads πout j = P(K0 = i, KM−1 = j) = (17) ij can be interpreted as the expected number of times the killed Markov process visits node j when starting from i. So if we define i∈V We know from Lemma 2.1 that(cid:98)zref (cid:44)(cid:88) ¯nref j i (cid:98)zref πin ij = (cid:88) i∈V i∈V i ¯nref πin ij , (18) the column vector ¯nref = (¯nref is in j before being killed. Then, (17) directly provides j ) holds the expected number of times the process α = πout ÷ ¯nref . (19) The second part of the proposition is obtained by starting from the definition of ¯nref and following (8), ¯nref = ((cid:98)Zref )(cid:62)πin =(cid:0)I − ((cid:98)Pref )(cid:62)(cid:1)−1 ⇒ ¯nref − ((cid:98)Pref )(cid:62)¯nref = πin, ⇒ ¯nref −(cid:0)(Pref )(cid:62) − (Pref )(cid:62)Diag(α)(cid:1)¯nref = πin. πin, Then, using Diag(α) ¯nref = πout (see Equation (14)), (cid:0)I − (Pref )(cid:62)(cid:1)¯nref = πin − (Pref )(cid:62)πout, which provides the same results as the expression derived in (Guex, 2016, propo- sition 1) from another perspective. 11 i = ¯nref The Equation (14) simply states that the probability of being killed in state i is equal to the expected number of visits to i times the probability of jumping to the cemetery state from i, πout i αi. This is similar to the computation of the absorption probabilities when starting from a transient state in an absorbing Markov chain (Doyle & Snell, 1984; Grinstead & Snell, 1997). In conclusion, it is possible to determine ¯nref and then α through Proposition 2.1 by considering an additional free parameter. Indeed, as (I−(Pref )(cid:62)) is rank- deficient (its rank is n− 1, as the initial reference chain is regular), we have (see, e.g., (Graybill, 1983)) ¯nref =(cid:0)I − (Pref )(cid:62)(cid:1)+(cid:0)πin − (Pref )(cid:62)πout (cid:1) + πref, (20) where(cid:0)I− (Pref )(cid:62)(cid:1)+ (cid:32) ≥ max (cid:33) . denotes the Moore-Penrose pseudoinverse, πref is the sta- tionary distribution of the regular, reference, Markov chain defined by Pref (i.e. a vector summing to 1 and generating the null-space of (I − (Pref )(cid:62))), and is an additional free parameter, named persistence, such that (cid:0)I − (Pref )(cid:62)(cid:1)+(cid:0)πin − (Pref )(cid:62)πout (cid:1) i − e(cid:62) πout i i i i πref i ≥ πout and thus αi ≤ 1, ∀i ∈ V. Intuitively, This last inequality ensures that ¯nref the persistence parameter reflects the difficulty for a process to be killed in nodes where αi > 0 (called killing nodes), and thus affects the expected length of the paths (see (Guex, 2016) for a discussion). In (Guex, 2016), the persistence is shown to have an electrical interpretation within the well-known analogy between random-walk models and electrical models (Doyle & Snell, 1984): it corresponds to the lowest electrical potential that can be defined on nodes. However, the effect of the persistence on the behavior of the model is beyond the scope of this work, and we will set it to its lower bound in our simulations. 2.3.2 Defining the reference probabilities over paths from the killed process ending distributions σin and σout, that is, (cid:80) (cid:80) i∈V(cid:80) We will now define the reference probabilities Pref (℘) over the set of paths, with the help of a killed Markov process, in order to obtain the desired starting and i and . We show that it suffices to set πin = σin and πout = σout in the previously defined killed Markov process and deduce the reference probabilities from it. Proposition 2.2. If the reference probabilities Pref (℘), ℘ ∈ P, are set to j∈V(cid:80) Pref (℘ij) = σout Pref (℘ij) = σin ℘ij∈Pij ℘ij∈Pij j Pref (℘ij) (cid:44) σin where (cid:98)πref (℘ij) is the modified likelihood of the path ℘ij, defined by i (cid:98)πref (℘ij) αj for each i, j ∈ V, (21) kτ−1kτ for any path ℘ = (k0, k1, . . . , kt), (22) (cid:98)πref (℘) (cid:44) t(cid:89) (cid:98)pref τ =1 12 Algorithm 1 Killing rates for non-hitting paths Input: -- The n × n non-negative adjacency matrix A of a strongly connected directed graph, containing edge affinities. -- An input distribution vector σin of size n. -- An output distribution vector σout of size n. -- A scalar persistence gap parameter g > 0. Output: -- The killing rates vector α of size n. -- The expected number of visits to each node for the reference random walk, ¯nref . 1. D ← Diag(Ae) 2. Pref ← D−1A (cid:46) the reference transition probabilities matrix 3. Q ← I − (Pref )(cid:62) 4. π ← null(Q); π ← π/sum(π) (cid:46) the auxiliary matrix Q (cid:46) the out-degree matrix (cid:46) computes the stationary distribution of the Markov chain whose transition matrix is Pref (π spans the null space of Q and its elements are non-negative and sum to one) 0 ← Q+(σin − (Pref )(cid:62)σout) node for the reference random walk 0 ) ÷ π) + g (cid:46) the Moore-Penrose pseudoinverse matrix of Q (cid:46) the initial expected number of visits to each (cid:46) the value of the persistence parameter 0 + π (cid:46) the expected number of visits to each node for the reference 5. Q+ ← pinv(Q) 6. ¯nref 7. ← max((σout − ¯nref 8. ¯nref ← ¯nref 9. α ← σout ÷ ¯nref 10. return α, ¯nref random walk (cid:46) the killing rates vector; ÷ is the elementwise division with (cid:98)Pref = (I − Diag(α))Pref , and where α is computed from Proposition 2.1 (with πin = σin and πout = σout), then we have Pref (S = i) = Pref (℘ij) = σin i Pref (E = j) = Pref (℘ij) = σout j ∀i ∈ V, ∀j ∈ V. (cid:88) (cid:88) j∈V (cid:88) (cid:88) ℘ij∈Pij i∈V ℘ij∈Pij Moreover, these reference probabilities are properly scaled as they sum to one. Proof. Observe that if we set the reference probabilities to Pref (℘ij) (cid:44) πin then, from (16), we have for the probability of picking a paths ℘ij ∈ Pij in the bag of all paths starting in i (S = i) and ending in j (E = j) (see (Fran¸coisse et al., 2017)) i (cid:98)πref (℘ij) αj, P(S = i, E = j) = Pref (℘ij) = πin i (cid:88) (cid:34) ∞(cid:88) ((cid:98)Pref )t ℘ij∈Pij (cid:35) = πin i (cid:88) ℘ij∈Pij (cid:98)πref (℘ij) αj αj = P(K0 = i, KM−1 = j), and therefore, t=0 ij P(S = i) = P(K0 = i) = πin i , P(E = j) = P(KM−1 = j) = πout j . 13 Moreover, it also follows that(cid:80) (cid:80)n ℘∈P Pref (℘) =(cid:80)n Pref (℘ij) = i,j=1 P(K0 = i, KM−1 = j) = 1 so that the reference probability distribution ℘ij∈Pij i,j=1 (cid:80) over paths is properly scaled. Note that the last quantity, P(S = i, E = j), was called the bag-of-paths probability matrix and played a key role in the bag-of-paths framework (see (Fran¸coisse et al., 2017) for details). It is also called the coupling matrix in optimal transportation (see later). The procedure for computing the vectors α and ¯nref allowing to obtain a desired ending distribution is sumarized in Algo- rithm 1. In addition to the predefined margins, it takes as input a persistence gap parameter g indicating to which extend persistence of flow is present in the network (see the discussion following Equation (20)). 2.4 Computation of the optimal probability distribution over paths In this section, now that we have found a proper reference distribution, we focus on the computation of the optimal probability distribution P(cid:63)(℘) solving problem (5). This solution is obtained through its Lagrange parameter vectors, which can be obtained from the constraints. 2.4.1 The optimal path probabilities The optimal probability distribution is obtained by the following proposition. Proposition 2.3. For reference path probabilities Pref (℘) as defined in Propo- sition 2.2, the minimization problem (5), as stated in Section 2.2, is solved by j αj(cid:98)πref (℘ij) exp(−βc(℘ij)) for each i, j ∈ V, i σin i µout P(cid:63)(℘ij) = µin (23) (cid:44) exp(−βλin i ), ) are two vectors derived from the Lagrange parameter vectors where β (cid:44) 1/T is the inverse temperature parameter, and µin µout λin, λout, associated with the constraints. (cid:44) exp(−βλout j j i Proof. We derive the solution for the optimal probability distribution P(cid:63)(℘) solving problem (5). By introducing Lagrange parameter vectors λin and λout, the Lagrange function associated to (5) is P(℘)c(℘) + T P(℘) log L(P, λin, λout) (cid:44) (cid:88) (cid:88) (cid:88) i∈V ℘∈P + (λin i − T 2 ) (λout j − T 2 ) + j∈V ℘∈P (cid:88) (cid:88) (cid:88) ℘ij∈Pij (cid:88) (cid:88) j∈V i∈V ℘ij∈Pij (cid:19) (cid:18) P(℘)  i Pref (℘) P(℘ij) − σin P(℘ij) − σout j  , (24) where the Lagrange parameters are shifted by T /2 to simplify the notation. Taking its partial derivative with respect to P(℘ij), setting the result to zero, 14 and defining the inverse temperature β (cid:44) 1/T , provides P(cid:63)(℘ij) = Pref (℘ij) exp(cid:2)−β(cid:0)c(℘ij) + λin i + λout j = µin j Pref (℘ij) exp(−βc(℘ij)), i µout (cid:1)(cid:3) where we defined µin i (cid:44) exp(−βλin i ) and µout i (cid:44) exp(−βλout i ), (25) which corresponds to a re-parametrization of the the Lagrange parameters that will be used instead of the original parameters. By inserting the reference prob- ability found in (21) in this last equation, we get the following form for the path probabilities j αj(cid:98)πref (℘ij) exp(−βc(℘ij)), where the(cid:98)πref (℘) are provided by Proposition 2.2. P(cid:63)(℘ij) = µin i µout i σin 2.4.2 Computing the Lagrange parameters i i ) and µout Proposition 2.4. The two vectors, defined as µin exp(−βλout The solution (23) requires the values of the Lagrange multipliers λin and λout, or alternatively the vectors µin and µout, which can be obtained from the equality constraints. Proposition 2.4 shows how to compute these vectors. i = exp(−βλin (cid:17) µin = e ÷(cid:16)(cid:98)Z (µout ◦ α) ) (see Proposition 2.2), verify the following equations (cid:17) µout = ¯nref ÷(cid:16)(cid:98)Z(cid:62) (µin ◦ σin) where (cid:98)Z (cid:44) (I −(cid:99)W)−1 is the fundamental matrix, obtained from (cid:99)W (cid:44) (cid:98)Pref ◦ exp[−βC], and ◦ is the elementwise product. Proof. The Lagrange parameters can be found by enforcing the constraints (3) and (4) on P(cid:63)(℘). By injecting (23) for P(cid:63)(℘) in (3) and (4) provides ∀i ∈ V, i = µout (26) (27) (28) , , σin i µin i = σin i j αj (cid:88) (cid:88) j∈V i∈V (cid:88) (cid:88) ℘ij∈Pij ℘ij∈Pij (cid:98)πref (℘ij) exp(−βc(℘ij)) (cid:98)πref (℘ij) exp(−βc(℘ij)) (cid:98)zij αjµout j µin i σin i = σout j ∀j ∈ V. (29) (cid:98)zij (cid:98)Z (cid:44) (I −(cid:99)W)−1 By further defining the fundamental matrix as with (cid:99)W (cid:44) (cid:98)Pref ◦ exp[−βC], where ◦ is the elementwise product, it can easily et al., 2010) that (cid:88) be shown by using a development similar to (Fran¸coisse et al., 2017; Mantrach (cid:98)πref (℘ij) exp(−βc(℘ij)) =(cid:98)zij = [(cid:98)Z]ij. (30) (31) ℘ij∈Pij 15 For computing the parameters, we use a variant (Kapur & Kesavan, 1992) of iterative proportional fitting procedure (discussed below) based on (28) and (29). Isolating σin in (28) and (29) after replacing αj by the result found with Proposition 2.1 (Equations (14) and (20)), i.e. αj = σout , in the second equation, we obtain i and σout j /¯nref j i µin = e ÷(cid:16)(cid:98)Z (µout ◦ α) (cid:17) µout = ¯nref ÷(cid:16)(cid:98)Z(cid:62) (µin ◦ σin) (cid:17) , . j j i and µout j > 0, respectively, and that µin Note that the above derivation is only valid for nodes i and j for which σin i > 0 and σout are not needed when these quantities are equal to zero because the path probabilities also vanish in this case (see Equations (23)). However, defining and computing the quantities µin i and µout for all nodes i and j according to (26) and (27) proves to be convenient for what follows, as these quantities appear in other meaningful expressions. For unconstrained nodes with σin j = 0, the Lagrange parameters are equal to zero, meaning that the corresponding µin Note that by further defining the matrix(cid:98)tij (cid:44) σin i (cid:98)zijαj, (28) and (29) can be rewritten in matrix form as Diag(µin)(Diag(σin)(cid:98)Z Diag(α))Diag(µout) e = Diag(µin)(cid:98)T Diag(µout) e = σin and Diag(µout)(cid:98)T(cid:62)Diag(µin) e = σout, where i = 1 and µout i = 0 or σout j = 1 Diag(x) is a diagonal matrix with x on its main diagonal and e is a column vector full of 1's. Therefore, we have in matrix form (cid:98)T = Diag(σin)(cid:98)Z Diag(α). (32) Thus, the computation of the Lagrange parameters reduces to the problem of finding two nonnegative row and column scaling vectors (µin and µout), reweighting the rows and the columns of (cid:98)T such that the row marginals and the column marginals of the new rescaled matrix are equal to σin and σout, just like in the context of standard optimal transport with entropy regularization (Cuturi, 2013). Indeed, this procedure is closely related to the solution of the standard, relaxed, optimal transport problem with entropy regularization when using the matrix (cid:98)T as the matrix containing the costs or rewards of transportation (Wilson, 1970; Kapur, 1989; Erlander & Stewart, 1990; Kapur & Kesavan, 1992; Cuturi, 2013). It is usually solved by iterative proportional fitting, ma- trix balancing or biproportional scaling ((Bacharach, 1965; Sinkhorn, 1967); see (Kurras, 2015; Pukelsheim, 2014) and the references therein for a more recent discussion). The iterative proportional fitting algorithm has guaranteed conver- gence to a unique solution under some mild conditions (see, e.g., (Kurras, 2015; Pukelsheim, 2014)). Our iterative procedure for solving Equations (26-27) consists of first fixing in from (26) at iteration τ . Thereafter, out, and then computing µ(τ ) an arbitrary µ(0) in is kept fixed and µ(τ +1) µ(τ ) out is computed from µ(τ ) in in (27). Finally, note that the optimal probability distribution P(cid:63)(℘) given by Propo- sition 2.3 is useless in practice, as there is an infinite number of paths. However, the different interesting and useful quantities can also be computed from the fundamental matrix and the Lagrange multipliers, as shown in the next section. 16 (cid:80) Moreover, it will be shown in Subsection 4.3 that the optimization problem can be reduced to the estimation of the joint probabilities P(cid:63)(S = i, E = j) = P(cid:63)(℘ij) (the coupling), therefore completely avoiding the introduction ℘ij∈Pij of the probability distribution. 2.5 Computation of other important quantities Similarly to (Fran¸coisse et al., 2017), other meaningful quantities can be com- puted in closed form after the convergence of µin and µout. This section provides expressions for computing them. The computation of the Lagrange multipliers together with the derived quantities, for the regular, non-hitting, BoP model, are summarized in Algorithm 2. 2.5.1 The coupling matrix First, the definition and computation of the coupling matrix is presented. Its name derives from the transportation science literature (Villani, 2003, 2008). Proposition 2.5. The coupling matrix, denoted by Γ = (γij) and defined by γij (cid:44) P(cid:63)(S = i, E = j) = P(cid:63)(℘ij), (33) (cid:88) ℘ij∈Pij (cid:88) where Pij is the set of paths starting in node i and ending in node j, can be computed by Γ = Diag(µin ◦ σin)(cid:98)Z Diag(µout ◦ α). (34) Proof. From the optimal probabilities (23) and the definition of the fundamental matrix (30), the probability of drawing a path starting in i and ending j is P(cid:63)(S = i, E = j) = P(cid:63)(℘ij) (cid:88) ℘ij∈Pij ℘ij∈Pij i σin = µin i µout j αj i (cid:98)zijµout = µin i σin j αj, (cid:98)πref (℘ij) exp(−βc(℘ij)) (35) (36) which in matrix form writes out as Equation (34). 2.5.2 The optimal free energy Proposition 2.6. The value of the free energy at the optimal paths probability distribution P(cid:63) is FE(P(cid:63)) = −(cid:0)λ(cid:62) inσin + λ(cid:62) outσout (cid:1) , (37) where the Lagrange parameter vectors λin, λin are obtained from the vectors µin, µin as stated in Proposition 2.3. Proof. By replacing P(cid:63)(℘ij) = µin i µout ergy expression (5), we get j Pref (℘ij) exp(−βc(℘ij)) into the free en- FE(P(cid:63)) = −(cid:88) λin i (cid:88) (cid:88) P(cid:63)(℘ij) −(cid:88) (cid:88) (cid:88) λout j P(cid:63)(℘ij) i∈V j∈V ℘ij∈Pij j∈V i∈V ℘ij∈Pij and, as the margins are fixed, we get the result. 17 An interesting interpretation of this proposition based on an optimal trans- portation analogy is discussed in Section 4. 2.5.3 The expected number of passages through an edge Let us define the matrix ¯N = (¯nij) as the matrix containing the expected num- ber of times an edge (i, j) appears on a path drawn from the optimal distribution P(cid:63)(℘). Formally, ¯nij (cid:44) (cid:88) (cid:88) k,l∈V ℘kl∈Pkl P(cid:63)(℘kl)η(i → j ∈ ℘kl), where η(i → j ∈ ℘kl) denotes the number of times the edge (i, j) appears on path ℘kl. In (Guex, 2016), ¯nij is interpreted as the flow on edges, creating a "stream of matter" going from supply nodes in In to destination nodes in Out. This interpretation will also be discussed further in Section 4. Proposition 2.7. The matrix ¯N, containing the expected number of times an edge (i, j) appears on a drawn path, is given by ¯N = Diag(¯nref ÷ µout) (cid:99)W (Diag(µin))−1, (38) Proof. Because costs are additive along paths, η(i → j ∈ ℘kl) = ∂c(℘kl)/∂cij. Then, by using (23), we obtain where (cid:99)W is defined after Equation (30). (cid:88) (cid:98)πref (℘kl) exp(−βc(℘kl)) ∂(cid:98)zkl ∂c(℘kl) ∂cij (39) (cid:88) (cid:88) ℘kl∈Pkl ¯nij = µin k σin k µout l αl µin k σin k µout l αl k,l∈V = − 1 β . ∂cij k,l∈V We know from direct calculus (Kivimaki et al., 2016, Equation (11)) that ∂(cid:98)zkl/∂cij = −β(cid:98)wij(cid:98)zki(cid:98)zjl. Thus, (cid:32)(cid:88) ¯nij = (cid:98)wij µin k σin k∈V (cid:33) l αl(cid:98)zjl µout (cid:33)(cid:32)(cid:88) l∈V k(cid:98)zki i (cid:98)wij ¯nref µout i µin j , (40) and by using (28)-(29) with αl = σout l /¯nref l (see Proposition 2.1), ¯nij = , (41) which provides the desired result. 18 Algorithm 2 Regular, non-hitting, margin-constrained bag-of-paths algorithm Input: -- The n × n non-negative adjacency matrix A of a strongly connected directed graph, containing edge affinities. -- The n × n non-negative cost matrix C, containing edge costs. -- An input distribution vector σin of size n. -- An output distribution vector σout of size n. -- A killing rate vector α of size n. -- The expected number of visits to each node for the reference random walk, ¯nref , of size n. -- A scalar inverse temperature parameter β > 0. -- The coupling matrix Γ of size n × n. -- The minimum free energy FEmin (a scalar). -- The matrix ¯N of size n × n, containing the expected number of visits to each edge (i, j). -- The vector ¯n of size n, containing the expected number of visits to each node. -- The resulting biased transition probabilities matrix P of size n× n (the policy). Output: (cid:46) the out-degree matrix (cid:46) the fundamental matrix 6. µout ← e 7. repeat 8. (cid:46) the initial Lagrange parameters vector for ouputs (cid:46) the killed reference transition probabilities matrix 1. D ← Diag(Ae) 2. Pref ← D−1A (cid:46) the reference transition probabilities matrix 3. (cid:98)Pref ← (I − Diag(α))Pref 4. (cid:99)W ← (cid:98)Pref ◦ exp[−βC] (cid:46) the auxiliary matrix (cid:99)W; ◦ is the elementwise product 5. (cid:98)Z ← (I −(cid:99)W)−1 µin ← e÷(cid:16)(cid:98)Z (µout ◦ α) (cid:17) (cid:17) µout ← ¯nref ÷(cid:16)(cid:98)Z(cid:62) (µin ◦ σin) 11. Γ ← Diag(µin ◦ σin)(cid:98)Z Diag(µout ◦ α) 13. ¯N ← Diag(¯nref ÷µout) (cid:99)W (Diag(µin))−1 β ((log[µin])(cid:62)σin + (log[µout])(cid:62)σout) 9. 10. until convergence 12. FEmin ← 1 (cid:46) the update of input vector; ÷ is the elementwise (cid:46) the update of output vector (cid:46) the coupling matrix (cid:46) the minimum free energy value (cid:46) the matrix containing the expected (cid:46) main iteration loop division number of visits to each edge 14. ¯n ← ¯nref ÷ (µin ◦ µout) 15. P ← pinv(Diag( ¯Ne)) ¯N (cid:46) the biased transition matrix (the policy) 16. return Γ, FEmin, ¯N, ¯n, P to each node (cid:46) the vector containing the expected number of visits 2.5.4 The expected number of visits to a node Let us further define the vector ¯n = (¯ni), containing the expected number of times node i is drawn under P(cid:63)(℘), by ¯ni (cid:44) (cid:88) ℘∈P P(cid:63)(℘) η(i ∈ ℘), where η(i ∈ ℘) denotes the number of times node i appears on path ℘. Proposition 2.8. The vector ¯n, containing the expected number of times node i is drawn, is ¯n = ¯nref ÷ (µin ◦ µout), (42) 19 where ÷ and ◦ are the elementwise division and product. Proof. In fact, we can decompose this quantity as η(i ∈ ℘) = η(i → j ∈ ℘) + δ(E(℘) = i), (43) (cid:88) j∈V (cid:88) ℘∈P where δ(E(℘) = i) is one if i is the ending node of the path ℘. This gives P(cid:63)(℘)η(i → j ∈ ℘) + P(cid:63)(℘)δ(E(℘) = i) = (cid:88) (cid:88) j∈V ℘∈P ¯ni = (cid:88) j∈V ¯nij + σout i . (44) ¯ni = µin k σin (cid:88) By using (40) we find (cid:32)(cid:88) (cid:33)(cid:88) k(cid:98)zki However, because (I −(cid:99)W)(cid:98)Z = I,(cid:80) j∈V (cid:98)wij(cid:98)zjl =(cid:98)zil − δil as shown in (Kivimaki (cid:32)(cid:88) k(cid:98)zki l αl(cid:98)zil − µout  + σout (cid:33) et al., 2016, Equation (13)); therefore (cid:98)wij(cid:98)zjl (cid:88) µout l αl + σout , µin k σin i αi ¯ni = µout . i k∈V j∈V l∈V i k∈V l∈V which, by using (28)-(29) and αi = σout result i /¯nref i (see Proposition 2.1), provides the ¯ni = i ¯nref i µout µin i . (45) 2.5.5 The optimal randomized policy The expected number of times edges are visited induces a biased random walk on the network with transition matrix P (cid:44) (pij) provided by ¯nij(cid:80) k∈V ¯nik pij (cid:44) . (46) Proposition 2.9. The biased random walk transition matrix, P, called the randomized routing policy, is provided by P = (cid:99)W ÷(cid:0)((e ÷ µin) − (µout ◦ α)) µ(cid:62) in (cid:1) . (47) Proof. The result is obtained in elementwise form by combining expression (46) with (41), (44) and (45) and using αi = σout (see Proposition 2.1): i /¯nref i i (cid:98)wij µin j (1 − µin µin . i µout i αi) pij = 20 We can observe that when β → 0, then µin → e, µout → e, (cid:99)W → (cid:98)Pref and we obtain pij −−−−→ T→∞ (cid:98)pref ij (1 − αi) = pref ij , as it should be. This biased random walk is the optimal policy that has to be followed for reaching nodes in Out from nodes in In, and can be interpreted as follows. When β → 0, the behavior becomes similar to the random walk defined by Pref , but as β increases, random walkers are more and more "attracted" by high σout nodes. These "pools of attraction", whose sizes are related to the components of σout, get less and less "fuzzy" as β increases, eventually forcing walkers to adopt quasi-deterministic, optimal, paths following the solution of the optimal transport on a graph problem (see Subsection 6.2). Therefore, this framework can be viewed as an extension of standard electrical networks, interpolating between an optimal behavior based on least cost paths and a random behavior based on the reference probabilities (Guex, 2016). i 3 The margin-constrained bag-of-hitting-paths formalism In (Fran¸coisse et al., 2017), the BoP formalism was defined for regular paths (as in previous section) as well as for hitting paths, i.e. paths where the final node j appears only once as last node of the path. In this section, we will now consider the margin-constrained problem for hitting paths and define, ac- cordingly, the margin-constrained bag-of-hitting-paths framework (abbreviated as cBoHP). This new model will yield interesting properties and analogies with other models, and will require less computation time in comparison to the non- hitting, regular, bag-of-paths model considered so far in this work. We will see that while both models are similar when T → 0, they are quite different when T → ∞, and that the hitting formalism has a somewhat more straightforward solution. Nevertheless, the hitting paths assumption can prove more practical and appropriate in practice. But, of course, the choice of whether to consider non-hitting or hitting paths depends on the application. 3.1 Problem definition Let P h ij be the set of all hitting paths starting in i and ending in j, i.e. all paths ℘ij = (i0, . . . , it) where i0 = i, it = j and iτ (cid:54)= j,∀τ (cid:54)= t where t is the length of the path. This means that, technically, the ending node is turned into a killing, absorbing, node from which we cannot escape (Fran¸coisse et al., 2017; Kivimaki et al., 2014; Fouss et al., 2016). We define the set of all hitting paths, also named bag-of-hitting-paths, by Ph = ∪i,j∈VP h ij. By analogy with (5), the problem here is to find the optimal hitting paths probability distribution, P(cid:63) h(℘), 21 solving (cid:88) FEh(Ph) = minimize {Ph(℘)}℘∈Ph subject to (cid:80) j∈V(cid:80) (cid:80) i∈V(cid:80) ℘∈Ph ℘ij∈P h ℘ij∈P h ij ij Ph(℘)c(℘) + T Ph(℘) log Ph(℘ij) = σin i Ph(℘ij) = σout j ∀i ∈ V, ∀j ∈ V. (cid:88) ℘∈Ph (cid:18) Ph(℘) (cid:19) Pref h (℘) (48) As probabilities for regular paths containing the final node more than once converge to zero in the non-hitting formalism when T → 0 (it is sub-optimal to visit several times the same node), we easily see that both problems are equivalent at this limit. However, this is not the case when T → ∞, due to the difference in reference probabilities between the two models and the structure of the paths, as shown in the next section. Yet another important difference between the hitting and the non-hitting formulations is that the former model is equivalent to the standard entropy regularized optimal transport problem (Wilson, 1970; Erlander & Stewart, 1990; Cuturi, 2013) based on the directed free energy distance (or potential) between nodes (Kivimaki et al., 2014; Fran¸coisse et al., 2017; Fouss et al., 2016), which can easily be pre-computed for the whole graph. This is detailed in Section 4 (see Equation (71)). A last difference is that the hitting paths formulation does not need the pre-processing step computing the reference probabilities with fixed margins described in Section 2.3 and Algorithm 1. 3.2 Reference probabilities with fixed margins Finding reference probabilities Pref h (℘) is at the heart of the difference between both formalisms, and is greatly simplified in the hitting case. In fact, it was shown in (Fran¸coisse et al., 2017) that the sum of likelihoods over all hitting paths between two nodes i and j is always equal to 1. In other words, (cid:88) ℘∈P h ij πref (℘) = 1, where πref (℘) (cid:44) Πt ij being the reference transition probabilities (1). From this, it is easy to observe that the reference probability defined by and with the pref τ =1pref iτ−1,iτ h (℘ij) (cid:44) σin Pref i σout j πref (℘ij), (49) yields the correct margins as expressed in the constraints of problem (48). 3.3 Computation of the optimal probability distribution over paths The reasoning for finding the solution follows the same rationale as before for the non-hitting case; therefore the details of the proofs are not repeated in this section. The main difference lies in the replacement of the reference probabilities with (49), and the following new expression for hitting paths (the equivalent of Equation (31)) πref (℘ij) exp(−βc(℘ij)) = zh ij, (50) (cid:88) ℘ij∈P h ij 22 with Zh = (zh ((Kivimaki et al., 2014), Equation (12)), obtained through ij) being the fundamental matrix for hitting paths as introduced in (51) where Z (cid:44) (I− W)−1, W (cid:44) Pref ◦ exp[−βC], and Dh (cid:44) Diag(Z), the diagonal matrix containing the main diagonal of Z. Elementwise, we have zh ij = zij/zjj. Zh (cid:44) ZD−1 h , 3.3.1 The optimal hitting-paths probabilities The optimal hitting-paths probabilities are obtained with the following propo- sition. Proposition 3.1. When the set of paths P is restricted to the set of hitting paths Ph and the reference path probabilities Pref (℘) are defined according to Equation (49), then the minimization problem (48) is solved by i µh,out j πref (℘ij) exp(−βc(℘ij)), σout h(℘ij) = µh,in P(cid:63) i σin where β (cid:44) 1/T is the inverse temperature parameter, and µh,in µh,out j vectors λh,in, λh,out, associated with the constraints. (52) (cid:44) exp(−βλh,in ), ) are two vectors derived from the Lagrange parameter (cid:44) exp(−βλh,out j j i i Proof. The proof is similar to the proof of Proposition 2.3. 3.3.2 Computing the Lagrange parameters Proposition 3.2. The two vectors µh,in and µh,out, defined by µh,in and µh,out ), verify the following expressions = exp(−βλh,out i i in = e ÷(cid:0)Zh (cid:0)µh out = e ÷(cid:0)Z(cid:62) (cid:0)µh µh µh h out ◦ σout in ◦ σin (cid:1)(cid:1) , (cid:1)(cid:1) . i = exp(−βλh,in i ) (53) (54) Proof. The proof is similar to the proof of Proposition 2.4, but by replacing the matrix Z with Zh. As for regular paths, these two expressions are recomputed iteratively until convergence. Note that in the hitting paths case, the matrix that needs to be rescaled in order to satisfy the margin constraints is Th = Diag(σin)ZhDiag(σout). (55) The computation of the Lagrange multipliers and the derived quantities for the hitting bag-of-paths model is summarized in Algorithm 3. 3.4 Computation of other important quantities The computation of the other interesting quantities is slightly different in the hitting formalism, and shows some interesting new properties. They are re- viewed in this section. 23 3.4.1 The coupling matrix Proposition 3.3. The coupling matrix for hitting paths Γh = (γh ij) is given by Γh = Diag(µh in ◦ σin) Zh Diag(µh out ◦ σout). (56) Proof. The proof is similar to the proof of Proposition 2.5. Notice that, for the hitting paths case, an alternative way of obtaining the coupling matrix by pre-computing the directed free energy distances is discussed in Subsection 4.3 (see Equation (71)). 3.4.2 The optimal free energy Proposition 3.4. The value of the free energy at the optimal distribution P(cid:63) h is out)(cid:62)σout Proof. The proof is similar to the proof of Proposition 2.6. in)(cid:62)σin + (λh FEh(P(cid:63) (57) h) = −(cid:0)(λh (cid:1) . ¯Nh =(cid:0)(e ÷ µh 3.4.3 The expected number of visits to an edge Proposition 3.5. The matrix ¯Nh, containing the expected number of times an edge (i, j) appears on a hitting path, is given by in)(cid:62) − (ZhDiag(σout)Zh)(cid:62)(cid:1) ◦ (Diag(Z)W), (58) where Diag(Z) is a diagonal matrix containing the main diagonal of Z and ◦ is the elementwise matrix product. out)(e ÷ µh Proof. Let ¯nh drawn according to Ph(℘), i.e., ij be the expected number of times edge (i, j) appears on a path (cid:44) (cid:88) (cid:88) k,l∈V ℘kl∈P h kl ¯nh ij Ph(℘kl) η(i → j ∈ ℘kl), where η(i → j ∈ ℘kl) denotes the number of times the edge (i, j) is visited along hitting path ℘kl. By a reasoning similar to (39) we get (cid:88) k,l∈V ij = − 1 ¯nh β µh,in k σin k µh,out l σout l ∂zh kl ∂cij . This time, we have (Kivimaki et al., 2016, Equation (11)), leading to ∂zh kl ∂cij = ∂ (zkl/zll) ∂cij = −βwij (cid:88) k,l∈V ¯nh ij = µh,in k σin k µh,out l zll σout l 24 (cid:19) , − zklzlizjl z2 ll (cid:18) zkizjl (cid:18) zll (cid:19) wijzjl. (59) zki − zklzli zll (cid:44) (cid:0) zki From (56), and recalling that zh γh kl/zkl, where γh thus (cid:19) k σin k µh,out (cid:18) zki kl = zkl/zll, we get (µh,in )/zll = kl is element k, l of the coupling matrix (see Equation (56)), (cid:88) (cid:1)wijzjl is the expected number of times (i, j) is visited (cid:88) − zli zll kl ¯n(kl) γh wijzjl = − zli k,l∈V k,l∈V σout (60) γh kl zkl ij l , l ¯nh ij = ij zll zkl and ¯n(kl) when the starting node is fixed to k and the ending node to l, as shown in (Kivimaki et al., 2016, Equation (12)). Within the constrained bag-of-hitting- paths formalism, this quantity is simply the average over all starting and ending nodes, weighted by the coupling probabilities, Γh. In fact, ¯nh ij can be seen as a weighted randomized shortest-paths (RSP) betweenness centrality for edges, compared to the unweighted RSP betweenness centrality defined in (Kivimaki et al., 2016). More precisely, this quantity provides a weighted group betweenness between the two sets of nodes, In and Out. While the decomposition (60) of ¯nh ij provides an interesting analogy, we will use another expression to actually compute this quantity. From (59) with the help of (53) and (54), we get (cid:32) ¯nh ij = 1 µh,out µh,in j i −(cid:88) l∈V (cid:33) σout l zh lizh jl ziiwij, (61) which provides the expression for computing ¯Nh (cid:44) (¯nh ij). 3.4.4 The expected number of visits to a node Proposition 3.6. The vector ¯nh, containing the expected number of times node i appears on a hitting path drawn from a bag of hitting paths, is provided by ¯nh = Diag(Z)(cid:2)(e ÷ (µh in ◦ µh (62) out)) − diag(ZhDiag(σout)Zh)(cid:3) + σout, (cid:44) (cid:88) Ph(℘) η(i ∈ ℘), where diag(X) is a column vector containing the main diagonal of matrix X. i , the expected number of times node i appears on a path Proof. Let us define ¯nh under Ph(℘), by where η(i ∈ ℘) denotes the number of times node i is visited along hitting path ℘. Using (43) and (44) again, we get which, after using (60) and (cid:80) Equation (13)), results in ¯nh i = ¯nh ij + σout i , (63) j∈V wijzjl = zil − δil (Kivimaki et al., 2016, (cid:88) ¯nh i = kl ¯n(kl) γh i + σout i , (cid:44)(cid:0) zki k,l∈V (cid:1)zil is the number of times i is visited when starting in where ¯n(kl) k and ending in l, as defined in (Kivimaki et al., 2016). Notice that ¯n(kl) zkl zll i − zli i = 0 ¯nh i ℘∈Ph (cid:88) j∈V 25 Algorithm 3 Margin-constrained bag-of-hitting-paths algorithm Input: -- The n × n non-negative adjacency matrix A of a strongly connected directed graph, containing edge affinities. -- The n × n non-negative cost matrix C, containing edge costs. -- An input distribution vector σin of size n. -- An output distribution vector σout of size n. -- A scalar inverse temperature parameter β > 0. -- The coupling matrix Γh of size n × n. -- The minimum free energy scalar FEmin -- The matrix ¯Nh of size n × n, containing the expected number of visits to each edge (i, j). -- The vector ¯nh of size n, containing the expected number of visits to each node. -- The resulting biased transition probabilities matrix Ph of size n× n (the policy). h . Output: (cid:46) the auxiliary matrix W; ◦ is the elementwise product (cid:46) the fundamental matrix (cid:46) the column-normalization matrix for hitting paths probabili- −1 h (cid:46) the out-degree matrix (cid:46) the column-normalized fundamental matrix 1. D ← Diag(Ae) 2. Pref ← D−1A (cid:46) the reference transition probabilities matrix 3. W ← Pref ◦ exp[−βC] 4. Z ← (I − W)−1 5. Dh ← Diag(Z) ties 6. Zh ← ZD out ← e 7. µh 8. repeat µh 9. µh (cid:0)µh in ← e ÷(cid:0)Zh out ← e ÷(cid:0)Z(cid:62) (cid:0)µh in)(cid:62) − (ZhDiag(σout)Zh)(cid:62)(cid:1) ◦ (Diag(Z)W) out)) − diag(ZhDiag(σout)Zh)(cid:1) + σout in ◦ σin) Zh Diag(µh in])(cid:62)σin + (log[µh out)(e ÷ µh (cid:46) main iteration loop out ◦ σout in ◦ σin (cid:46) the update of input vector (cid:46) the update of output vector 10. 11. until convergence 12. Γh ← Diag(µh 13. FEmin 14. ¯Nh ←(cid:0)(e ÷ µh 15. ¯nh ← Diag(Z)(cid:0)(e ÷ (µh h ← 1 β ((log[µh trix containing the expected number of visits to each edge (cid:46) the initial Lagrange parameters vector for ouputs out ◦ σout) out])(cid:62)σout) (cid:1)(cid:1) (cid:1)(cid:1) in ◦ µh h (cid:46) the coupling matrix (cid:46) the minimum free energy value (cid:46) the ma- (cid:46) the vec- tor containing the expected number of visits to each node 16. Ph ← pinv(Diag( ¯Nhe)) ¯Nh 17. return Γh, FEmin , ¯Nh, ¯nh, Ph h (cid:46) the biased transition matrix (the policy) when i = l and when k = l. Again, this quantity can be seen as a weighted RSP betweenness centrality for sets of nodes, by analogy with the unweighted RSP betweenness centrality defined in (Kivimaki et al., 2016, Equation (15)). However, for computation, it is more convenient to apply the equivalent of Equation (61). Following the same path as for the derivation of (61), we first j∈V wijzjl = zil − δil with (63) on (59), and then (53) and (54) in order use(cid:80) to get (cid:33) (cid:32) ¯nh i = 1 µh,out µh,in i i −(cid:88) l∈V σout l zh lizh il zii + σout i , (64) which provides the expression for computing the vector ¯nh (cid:44) (¯nh i ). 26 3.4.5 The optimal randomized policy Proposition 3.7. The biased random walk transition matrix, Ph = (ph is, the randomized routing policy, is given by ij), that (cid:18) (cid:16) ij(cid:80) ¯nh k∈V ¯nh ik = (cid:44) ph ij −(cid:80) −(cid:80) 1 µh,out i µh,in j 1 µh,out i µh,in i l∈V σout l l∈V σout l zh lizh jl zh lizh il (cid:19) (cid:17) wij. (65) Proof. We get the result by combining (61), (63) and (64). Again, we observed experimentally that this quantity converges to the ref- erence transition matrix of the reference random walk when β → 0, i.e., ij −−−→ ph β→0 pref ij . Conversely, when β → ∞, the problem becomes an optimal transport on a graph problem (see Subsection 6.2). Note that, for convenience, in Algorithm 3, matrix Ph (cid:44) (ph ij) is computed thanks to Ph = (Diag( ¯Nhe))−1 ¯Nh. Let us now turn to a discussion of the relations between the proposed models and the regularized optimal transport problem. 4 The regularized optimal transport problem anal- ogy In this section, we will show that both non-hitting (see Equation (5)) and hit- ting (see Equation (48)) problems correspond to two different kinds of regu- larization for the optimal transport problem (Ahuja et al., 1993; Guex et al., 2017; Kantorovich, 1942; Villani, 2003, 2008). It therefore generalizes dis- crete entropy-regularized optimal transport problems (Wilson, 1970; Erlander & Stewart, 1990; Kapur & Kesavan, 1992) to a graph structure. 4.1 The optimal transport problem General optimal transport is a well-known problem defined, for example, in (Kantorovich, 1942; Villani, 2003, 2008), and the special case where the space is a graph is easily derived from it (Guex et al., 2017). Assume we have a subset of nodes I n ∈ V, called sources, with a supply of a certain quantity of matter, while we observe a demand of the same matter in another subset of nodes, Out ∈ V, called targets. We suppose that the overall supply is equal to the overall demand, thus these quantities on nodes can be represented, without loss of generality, by their proportion of the total. In other words, supply and demand are represented respectively by two discrete distribution vector σin and i∈V σin i∈V σout i = 1 and σin i ≥ 0, ∀i ∈ V. i , σout The goal of the optimal transport problem is to find an optimal attribution plan or optimal coupling (Villani, 2003, 2008), i.e. a n × n matrix Γ, where γij represents the proportion of matter going from i to j, in order to fulfill supply σout, with(cid:80) i =(cid:80) 27 attribution plan, i.e. (cid:80) and demand. Optimality here means that the cost of transportation of this i,j∈V γij cij where cij is the cost of transportation from i to j, must be minimal. Altogether, we have (cid:80) subject to (cid:80) (cid:80) minimize Γ i,j∈V γij cij j∈V γij = σin i i∈V γij = σout j γij ≥ 0 ∀i ∈ V, ∀j ∈ V, ∀i, j ∈ V. (66) Another interesting interpretation can be found in the dual optimal transport problem (Guex et al., 2017; Villani, 2003, 2008), expressed by j σout ∀i, j ∈ V. i +(cid:80) subject to λin j∈V λout i∈V λin i + λout i σin j ≤ cij maximize (cid:80) λin,λout (67) j Here, the dual vectors λin = (λin ) can be interpreted respec- tively as the dual embarkment prices on sources and disembarkment prices on targets, as shown in (Villani, 2003, 2008). This is a common property of the dual problem in linear programming (Griva, Nash, & Sofer, 2009). i ) and λout = (λout i 4.2 The standard optimal transport flow on a graph prob- lem For completeness, let us recall the standard (exact) transport flow problem on a graph. The linear programming optimal transport flow problem is defined as (Ahuja et al., 1993) minimize subject to (cid:0)X − X(cid:62)(cid:1) e = σin − σout, e(cid:62)(X ◦ C(cid:48))e X xij ≥ 0 ∀i, j ∈ V, (68) where C(cid:48) is the cost matrix containing non-negative costs on the edges of the network with infinite components set to 0. As before, it is assumed that input and output flows are non-negative as well as (σin)(cid:62)e = (σout)(cid:62)e = 1. The idea is therefore to minimize the total cost of flows while satisfying the input and output constraints. The solution X = (xij) of this problem corresponds to the matrix containing directed flows on the edges. 4.3 The regularized optimal transport problem In order to show that the bag-of-paths formalism is closely related to the optimal transport problem, we need to compute again the minimum free energy value in terms of the elements of the coupling matrix γij, instead of the Lagrange parameters λin as in (37). i , λout j Let us start with the regular, non-hitting paths model. By inserting the form taken by the optimal path probability distribution P(cid:63)(℘), given by (23), in the free energy functional (5) (similarly to the proof following Proposition 2.6), we obtain FE(P(cid:63)) = T log(µin i µout j ) P(cid:63)(℘ij) , (cid:88) i,j∈V (cid:88) ℘ij∈Pij γij 28 γij =(cid:80) which directly provides from µin i µout j = γij/(σin ℘ij∈Pij P(cid:63)(℘ij) (Equation (33)) FE(P(cid:63)) = −T (cid:88) i,j∈V log((cid:98)zij) γij + T i αj(cid:98)zij) (see Equation (36)) and (cid:88) (cid:19) (69) γij log (cid:18) γij (cid:32) σin i αj . (cid:33) i,j∈V (cid:88) i,j∈V (cid:88) i,j∈V Following the same reasoning for the hitting paths formalism, we obtain FEh(P(cid:63) h) = −T log(zh ij) γh ij + T φT (i,j) γh ij log γh ij i σout σin j . (70) This shows that, instead of working with the whole probability distribution as required by Equation (5), it is sufficient to compute the elements of the coupling matrix γij if the quantities φT (i, j) are pre-computed. h) = −T log(zh Indeed, the resulting expression in the hitting formalism has a nice inter- pretation. Indeed, in (Kivimaki et al., 2014) it was shown that in the sim- ple randomized shortest-paths framework φT (i, j) (cid:44) FEh(P(cid:63) ij) is known to be the minimum free energy when problem (2) is restricted to hit- ting paths connecting a single source i to a single destination j. This quantity corresponds to the pairwise directed free energy distance between nodes of a graph introduced in (Kivimaki et al., 2014; Fran¸coisse et al., 2017), where it is proved that it is a distance metric. This distance provided competitive results in pattern recognition tasks (Fran¸coisse et al., 2017; Sommer, Fouss, & Saerens, 2016, 2017). Moreover, it is shown in (Fran¸coisse et al., 2017) that φT (i, j) converges to the directed shortest path distance between i and j when T → 0, and to the average first passage time (up to a scaling factor) between i and j when T → ∞. The free energy distances between all pairs of nodes can easily be computed in matrix form (Kivimaki et al., 2014; Fran¸coisse et al., 2017; Fouss et al., 2016). It has further been shown that, when computing the continuous time -- continuous state equivalent to the randomized shortest-paths model by densifying the graph, the minimum free energy becomes a potential attracting the agents to the goal state (Garc´ıa-D´ıez, Vandenbussche, & Saerens, 2011). From (70), once the directed free energy distances have been computed, we observe that problem (48) can be restated as ij + T(cid:80) minimize (cid:80) subject to (cid:80) (cid:80) j∈V γh i∈V γh ij ≥ 0 γh Γh i,j∈V φT (i, j)γh i,j∈V γh ij log ij = σin i ij = σout j ∀i ∈ V, ∀j ∈ V, ∀i, j ∈ V. (cid:33) (cid:32) γh ij i σout σin j (71) Knowing that the directed free energy distance φT (i, j) converges to the directed shortest path distance when T → 0, we conclude that problem (71) reduces to the optimal transport problem at this limit. Thus, problem (71) is actually a "soft" (entropy regularized) optimal transport problem, similar to the one studied in, e.g., (Wilson, 1970; Erlander & Stewart, 1990; Cuturi, 2013) and based on the directed free energy distance (Kivimaki et al., 2014; Fran¸coisse et al., 2017; Fouss et al., 2016), monitored by temperature. 29 matter of fact, when T → 0, (cid:98)zij and zh Therefore, in the case of hitting paths, an alternative way of solving the entropy-regularized optimal transport on a graph problem (48) is to pre-compute the free energy distances φT (i, j) and then solve problem (71) (see (Cuturi, 2013) for a recent discussion). The non-hitting problem is another regularization of the optimal transport problem as both (69) and (70) converge to the same limit when T → 0. As a ij have the same limit, as the effect of αi vanishes in the first case and zjj → 1 in the second case. In contrast, these two formalisms diverge when T → ∞. The hitting paths formalism converges to the problem described in (Guex et al., 2017), and thus results in the trivial, j when T → ∞, while the non-hitting paths independent, coupling γh formalism provides a more interesting solution, though harder to interpret (see Equation (69)). Actually, following the derivations appearing so far in this paper and results discussed in (Guex, 2016), it appears that the high temperature limit of the non-hitting formalism corresponds, in the case of an undirected graph, to the electrical circuit formalism, where sources and targets correspond to nodes with potentials fixed by the user (a high potential on sources and a low potential on targets) (see (Guex, 2016) for details). A study of this interesting question is left for further work. ij = σin i σout Interestingly, Equations (37) and (57) give two different, alternative, expres- sions for the minimal free energy, which are equivalent to the objective function of the dual optimal transport problem (with Lagrange multipliers correspond- ing to the dual variables multiplied by −1). It implies that, when T → 0, −λin and −λout (and equivalently, −λh out) converge respectively to the dual embarkment prices on sources and disembarkment prices on targets. Therefore, a mapping of these variables on nodes can highlight problematic sources and targets, in terms of optimal transport (see, e.g., (Guex et al., 2017)). in and −λh 4.4 The optimal transportation flow Within the context of the optimal transport problem, it is interesting to discuss the interpretation of the matrix containing the expected number of visits to edges, i.e., ¯N = (¯nij). Note that the discussion is developed within the non- hitting formalism, but remains valid in the hitting case. composed in η(i ∈ ℘) = (cid:80) leading to ¯ni =(cid:80) In Section 2.5.4, the number of times node i appears on a path ℘ was de- j∈V η(i → j ∈ ℘) + δ(E(℘) = i) (Equation (43)), (cid:88) (Equation (44)). However, it is also possible to η(j → i ∈ ℘) + δ(S(℘) = i), j∈V ¯nij + σout η(i ∈ ℘) = i write where δ(S(℘) = i) is 1 iff ℘ begins with i and 0 otherwise. This second version results in having and combining this result with (44), we obtain j∈V (cid:88) j∈V ¯ni = ¯nji + σin i , (cid:88) j∈V ¯nij −(cid:88) j∈V ¯nji = σin i − σout i , ∀i ∈ V. 30 This last equation shows that ¯nij can be interpreted as a directed flow on edges, as shown in (Guex, 2016). This flow is emitted by sources, absorbed by targets, and conserved everywhere else. Mapping this flow allows us to analyse the transportation of matter along the edges of the graph, and is illustrated in Section 6. In (Guex, 2016), it is shown that the net flow on an undirected graph, i.e. ¯nij − ¯nji, converges to the electrical flow for the non-hitting formalism when T → ∞. Furthermore, in the hitting formalism, summing the absolute values of the net flows over the edges results in a weighted randomized shortest path (RSP) net betweenness centrality (Kivimaki et al., 2016). For an unweighted graph, the standard, unweighted RSP net betweenness converges to the current flow betweenness in the limit T → ∞ (Newman, 2005; Brandes & Fleischer, 2005). 5 Derived distances and dissimilarities Two general families of distances are derived from our framework: distances between nodes and dissimilarities between groups of nodes (both for the non- hitting and the hitting case). 5.1 Distances between nodes Let us first discuss distances between nodes. 5.1.1 Definitions For both the hitting and non-hitting formalism, we can now define a distance named the surprisal distance, generalizing the one introduced in (Fran¸coisse et al., 2017; Kivimaki et al., 2014). The particularity here is that we can attach positive weights w = (wi) (with wi > 0 and summing to 1) to nodes, which affect the distances through σin and σout. More precisely, for a strongly connected graph, we define the margin-constrained bag-of-paths surprisal distance ∆sur = (∆ sur = (∆ sur ij ), and margin-constrained bag-of-hitting-paths surprisal distance, ∆h h,sur ij ) by, respectively, (cid:44) ∆ sur ij 2 (log(γij) + log(γji)) 0 ∆ h,sur ij (cid:44) 2 (log(γh 0 ij) + log(γh ji)) (cid:26) − 1 (cid:26) − 1 if i (cid:54)= j, if i = j, if i (cid:54)= j, if i = j, (72) (73) where Γ = (γij) and Γh = (γh ij) (the elements of the coupling matrix, see Sections 2.5.1 and 3.4.1) are obtained from, respectively, the non-hitting and hitting path formalisms with σin = σout = w. From this definition, each node acts as a source and a target and larger weights induce a stronger influence over the graph, as the flows along the paths starting and ending in a particular node will scale accordingly. Proposition 5.1. The margin-constrained bag-of-paths surprisal distance, ∆sur, defined in (72), and the margin-constrained bag-of-hitting-paths surprisal dis- tance, ∆h sur defined in (73), are both metrics. 31 sur ik + is trivially proven if i = k, i = k or j = k. So we will assume here that Proof. The triangle inequality for both surprisal distances, i.e., ∆ ∆ i (cid:54)= j (cid:54)= k (cid:54)= i. sur kj sur ij ≤ ∆ Non-hitting formalism. First, notice that the reasoning found in Appendix B of (Fran¸coisse et al., 2017) is still valid with the non-hitting reference prob- abilities derived in Section 2.3, namely, we have (cid:98)πref (℘ij) = (cid:98)πref (℘h ij)(cid:98)πref (℘jj) and c(℘ij) = c(℘h ij is the hitting path consisting of the first part of ℘ij, until it reaches j for the first time, and ℘jj is the remaining part of ℘ij. Thus, we also have ij) + c(℘jj), where ℘h (cid:44)(cid:80) (cid:98)zij =(cid:98)zh ij(cid:98)zjj, ij) exp(−βc(℘h ij(cid:98)πref (℘h where (cid:98)zh P(S = i, E = j) ≥ P(S = i, k ∈ ℘, E = j) (cid:44) (cid:88) the optimal path probabilities obtained in (23), ij∈P h ℘h ij ij)). Now, it is easy to see that, for (74) δ(k ∈ ℘ij)P(cid:63)(℘ij), (75) ℘ij∈Pij where δ(k ∈ ℘ij) is equal to 1 if node k lies on path ℘ij and 0 otherwise. By developing with (23), we obtain P(S = i, k ∈ ℘, E = j) = µin i σin δ(k ∈ ℘ij)(cid:98)πref (℘ij) exp(−βc(℘ij)) (cid:88) (cid:98)πref (℘kj) exp(−βc(℘kj)) ik)) (cid:88) ℘ij∈Pij ik) exp(−βc(℘h i µout j αj = µin i σin i µout j αj πref (℘h ℘kj∈Pkj (cid:88) ik(cid:98)zkj, ik∈P h ℘h ik j αj(cid:98)zh = µin i σin i µout where each path from i to j is again cut in two sequential sub-paths. Then, using Equations (74) and (36) provides P(S = i, k ∈ ℘, E = j) = µin i σin k αk(cid:98)zikµin i µout µin k σin k µout k σin k αk(cid:98)zkk k µout j αj(cid:98)zkj P(S = k, E = k) Using the inequality in (75) and P(S = k, E = k) ≤ 1, we get P(S = i, E = k)P(S = k, E = j) . = − log(P(S = i, E = j)) ≤ − log(P(S = i, E = k)) − log(P(S = k, E = j)), which proves the triangle inequality for the non-hitting surprisal distance for i (cid:54)= j (cid:54)= k (cid:54)= i. Hitting formalism. The reasoning is similar to the previous case. First let us consider Ph(S = i, E = j) = µh,in i σin i µh,out j σout j zh ij = µh,in i σin ≥ µh,in i σin i µh,out j σout j ℘ij∈Pij (cid:80) j σout j i µh,out δ(k ∈ ℘ij)πref (℘ij) exp(−βc(℘ij)) zij zjj . zjj (76) 32 And again, µh,in i σin i µh,out j σout j (cid:80) = µh,in i σin i µh,out j zjj σout j ℘ij∈Pij (cid:88) δ(k ∈ ℘ij)πref (℘ij) exp(−βc(℘ij)) zjj (cid:88) πref (℘h ik) exp(−βc(℘h ik)) πref (℘kj) exp(−βc(℘kj)) Ph(S = i, E = k)Ph(S = k, E = j) ℘kj∈Pkj j k ik = µh,in i σin i µh,out σout j zh kj k µh,out k σin σout ik∈P h ℘h ik µh,in σout k zh µh,in k µh,out k σin kk = 1. With (76) and Ph(S = k, E = k) ≤ 1, we finally obtain where we used zh − log(Ph(S = i, E = j)) ≤ − log(Ph(S = i, E = k)) − log(Ph(S = k, E = j)), which shows the triangle inequality for the hitting surprisal distance for i (cid:54)= j (cid:54)= k (cid:54)= i. Ph(S = k, E = k) = k k , 5.2 Distances between groups of nodes A different family of dissimilarities naturally arises from the optimal transport interpretation of the margin-constrained BoP formalism, namely dissimilarities between groups of nodes. These dissimilarities can be viewed as an extension of the Wasserstein distance between node distributions on a graph, also known under the name of the Monge-Kantorovich distance or the earth mover distance in the literature (see e.g. (Dobrushin, 1970; Villani, 2003, 2008; Zolotarev, 1983)). This dissimilarity is defined as the total cost of transportation in order to move from the distribution on source nodes, σin, to the distribution on target nodes, σout. Similarly to the usual free energy distance described in (Kivimaki et al., 2014; Fran¸coisse et al., 2017), interpolating between the shortest path distance and the commute cost distance (which is proportional to the resistance distance for undirected graphs), the margin-constrained BoP formalism uses the value of the free energy functional in order to derive a dissimilarity which interpolates between the Wasserstein distance and an electrical circuit-based dissimilarity between groups of node. with mig ≥ 0 and(cid:80)p 5.2.1 Definitions Let G be a directed, strongly connected, graph with n nodes, weighted by vector w. Suppose we have p groups of nodes, and the membership matrix M = (mig) g=1 mig = 1, represents the membership degree of node i to group g (fuzzy memberships are allowed). From that, we can compute the node distribution in group g, σg = (σg i ), as , e.g., (cid:80) σg i = wimig j∈V wjmjg . Then, as for the standard free energy distance between two nodes (Kivimaki et al., 2014, see this paper for details), the bag-of-paths free energy dissimilarity 33 between groups g and h is defined as the symmetrized minimum free energy between these two groups of nodes (cid:44) ∆ FE gh 2 (FEmin(σg, σh) + FEmin(σh, σg)) 0 if i (cid:54)= j, if i = j, (77) (cid:26) 1 (cid:26) 1 and the bag-of-hitting-paths free energy dissimilarity between groups g and h by ∆ h,FE gh (cid:44) 2 (FEmin 0 h (σg, σh) + FEmin h (σh, σg)) if i (cid:54)= j, if i = j, (78) where FEmin(σin, σout) is the non-hitting free energy (37), and FEmin (σin, σout) the hitting free energy (57), with starting and ending node flows respectively equal to σin and σout. By definition of the free energy, we are sure that this quantity is always positive. When T → 0, the dissimilarity between groups g and h will yield the optimal cost of transportation from group g to h and from group h to g, which is obviously a metric (Dobrushin, 1970; Villani, 2003, 2008; Zolotarev, 1983). It is, however, possible that this dissimilarity is not a metric anymore for other values of T . h From a computational point of view, there exists an important difference between the bag-of-hitting-paths and the bag-of-paths algorithms computing their respective free energy dissimilarities. As a matter of fact, in the hitting formalism, the matrix Zh is only computed once, and dissimilarities between each pair of groups can be obtained afterward by solely changing the values of σin and σout in the iterative procedure defined by (53) and (54). It is however impossible to proceed that way for the non-hitting formalism, as the computa- tion of(cid:98)Z requires the values of σin and σout. Therefore, the bag-of-hitting-paths distances are obtained in a more efficient way than the bag-of-paths distances. 6 Some illustrations Although the main contribution of this work lies in the theoretical development of the margins constrained bag-of-paths models, we provide here an illustration of the algorithms on a toy example. The bag-of-paths formalism (as well as the bag-of-hitting-paths, as they converge to the same solution when T → 0) defines an efficient way to find an approximate solution of the transportation problem on a graph. Moreover, by varying the temperature of the model, we can add uncertainty to the optimal paths and offer a flexible, stochastic, alternative to the optimal solution. With this feature, we can, e.g., pinpoint target nodes for which the optimal source coupling is most unclear. In a practical setting, when focusing on a deterministic transportation policy, this can help in evaluating the importance of each source-target coupling decision. Unlike other efficient optimal transport solvers, the algorithm here provides not only the coupling be- tween pairs of nodes, but also the flows on edges. Knowing the most frequented edges could be a major asset for real-life applications, for example in order to forecast network traffic. This section illustrates this idea on a toy graph (a 10 × 10 lattice). In addition, we evaluate the computational efficiency of Al- gorithms 2 and 3 by comparing the computation times with a baseline linear solver on lattices of different sizes. 34 (a) cBoP, β = 10−3 (b) cBoHP, β = 10−3 (c) cBoP, β = 10−1 (d) cBoP, β = 10 Figure 1: Resulting coupling Γ and edge flows ¯N for the margin-constrained bag-of-paths (cBoP) and the margin-constrained bag-of-hitting-paths (cBoHP) models, for different values of β. Source nodes are represented by a square and targets by a circle. Γ is shown on targets by a color interpolation representing their membership to corresponding sources (the membership of target j to source i is P(S = iE = j) = γij/σout ). ¯N is represented on edges. j 35 6.1 Illustrations on a 10 × 10 lattice j In this illustrative example, 5 random nodes were picked with σin i = 0.2 and i = 0.02. The resulting coupling Γ and flows on edges ¯N 50 others with σout are represented in Figure 1 for the constrained bag-of-paths (cBoP) and the constrained bag-of-hitting-paths (cBoHP), with different values of β. In this figure, target nodes j are colored to represent their distribution of membership over sources, i.e. P(S = iT = j) = γij/σout , and edge colors display the flows, ¯N. We observe that results obtained by the bag-of-paths model and the bag- of-hitting-paths model are quite different for β = 10−3 (Figure 1, top row): the bag-of-paths model displays a behavior similar to a diffusive process, with edges near sources drawn more frequently, which is known to be similar to the electrical current (Ahuja et al., 1993; Doyle & Snell, 1984; Guex, 2016). On the other hand, the bag-of-hitting-paths solution for β = 10−3 is quite trivial, with a uniform distribution of memberships of every target j to sources and the flow almost similar on every edge. In contrast, when the temperature is low, both models converge to the same solution and, to avoid redundancy, only the bag-of-paths model is shown here (Figure 1, bottom row). With β = 10, this model displays an optimal transport solution, with only shortest paths followed and almost deterministic distributions of targets-to-sources memberships. Therefore, in the present problem, the constrained bag-of-hitting-paths is perhaps less useful than the constrained bag-of-paths when the parameter β is close to zero. However, this depends on the application at hand and, essentially, on the desired behaviour of the system when β → 0, either the solution of an electrical circuit or the independence between sources and destinations. 6.2 Comparison of computation time To assess the computation time of the constrained bag-of-paths and bag-of- hitting-paths algorithms, we compare them to the open-source Computational Infrastructure for Operations Research (Coin-or) linear programming solver (clp) written in C++ (Lougee-Heimer, 2003), which is considered as an efficient base- line algorithm for finding the coupling and the flow of the exact optimal trans- port problem (see Equation (68)). We run the algorithms on lattices of various dimensions in order to increase the number of nodes n. The number of source nodes and target nodes are set to be both (cid:98)n/3(cid:99) (rounded down) with similar weights, and their locations are randomly selected. The results are presented in Figure 2. We can observe that on large graphs, both hitting and non-hitting bag-of-paths algorithms perform much faster than the linear programming baseline, with a slight advantage for the hitting al- gorithm. This was already observed in (Cuturi, 2013) for entropy regularized optimal transport problems. All results were obtained with Julia (version 0.5.0) running on an Intel Xeon with 2 × 8 3.6GHz processors and 128 GB of RAM. 7 Conclusion This work extends the bag-of-paths framework introduced in (Mantrach et al., 2010; Fran¸coisse et al., 2017) by allowing the user to set constraints on the start- ing and ending nodes of paths, σin and σout. Like its predecessor, this formalism 36 Figure 2: Computation times for, respectively, the open-source "Coin-or" linear programming solver (clp, blue), the margin-constrained bag-of-paths algorithm (cbop, red), and the margin-constrained bag-of-hitting-paths algorithm (cbopH, green) on lattices containing n nodes (x-axis), (cid:98)n/3(cid:99) sources and (cid:98)n/3(cid:99) targets with uniform weights. is derived for two types of paths, non-hitting paths and hitting paths. It also de- pends on a user-defined parameter, the temperature T > 0, according to which the model interpolates between a deterministic optimal policy and a completely random behavior. Both the non-hitting and hitting paths formalisms allow the computation of various quantities: the coupling, γij = P(S = i, E = j); the expected number of times a node appears on the paths (a betweenness value), ¯ni; and the optimal policy defining a biased random walk with transition prob- abilities pij. All these quantities are expressed in terms of three computational elements: a fundamental matrix Z, also found in (Fran¸coisse et al., 2017), and Lagrange multipliers λin and λout. The addition of the set of constraints over starting and ending node distri- butions adds flexibility to its unconstrained predecessor, and yields interesting connections with other models. When T → 0+, both the non-hitting and hit- ting formalisms are shown to be similar, and converge to a solution of the optimal transport on a graph problem. Unlike most algorithms solving the transportation problem, both bag-of-paths formalisms not only give sources-to- targets attributions, i.e. the coupling γij, but also corresponding embarkment and disembarkment prices (with λin and λout) and the flow on edges (¯nij), while running with a competitive computation time compared to a baseline lin- ear solver. In contrast, when T → 0, each formalism behaves differently, each having its own merits. The non-hitting formalism converges to the electrical solution, with starting and ending node distributions corresponding to differ- ent potentials defined on nodes, and the hitting formalism, which is faster to compute, converges to the trivial, independent coupling. These constraints also enlarge the range of applications of the bag-of-paths formalism, and it was shown here how to derive two families of dissimilarities from it. The first family of dissimilarities is defined as the surprisal distance between nodes, and constraints on starting and ending nodes provide a way to 37 llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll01020304002505007501000Number of NodesComputing Time [s]MethodlllcbopcbopHclpassociate weights on nodes. The second family of dissimilarities is the free energy dissimilarity between groups of nodes. For the moment, these dissimilarities are quite theoretical and their applications are not explored in this paper. However, future research will investigate the use of these new dissimilarities in semi- supervised classification, hierarchical clustering, as well as other applications. Generally speaking, the flexibility and the richness of this model could lead to different use cases, and future investigations will aim at finding various ap- plications of the different introduced quantities. An on-going study will also investigate the introduction of flow constraints in the bag-of-paths framework. Acknowledgements This work was partially supported by the Immediate and the Brufence projects funded by InnovIris (Brussels Region), as well as former projects funded by the Walloon region, Belgium. Ilkka Kivimaki was partially funded by Emil Aaltonen Foundation, Finland. We thank these institutions for giving us the opportunity to conduct both fundamental and applied research. We also thank the anonymous reviewers and the editor whose remarks al- lowed to improve significantly the manuscript. References Ahuja, R. K., Magnanti, T. L., & Orlin, J. B. (1993). Network flows: theory, algorithms, and applications. Prentice Hall. Akamatsu, T. (1996). Cyclic flows, Markov process and stochastic traffic as- signment. Transportation Research B , 30 (5), 369 -- 386. Alamgir, M., & von Luxburg, U. (2011). Phase transition in the family of p-resistances. In Advances in neural information processing systems 24: Proceedings of the NIPS '11 conference (p. 379-387). MIT Press. Bacharach, M. (1965). Estimating nonnegative matrices from marginal data. International Economic Review , 6 (3), 294 -- 310. Barab´asi, A.-L. (2016). Network science. Cambridge University Press. Bavaud, F., & Guex, G. Interpolating between random walks and shortest paths: a path functional approach. In International conference on social informatics (pp. 68 -- 81). (2012). Brandes, U., & Fleischer, D. (2005). Centrality measures based on current flow. In Proceedings of the 22nd annual symposium on theoretical aspects of computer science (STACS '05) (pp. 533 -- 544). Chebotarev, P. (2011). A class of graph-geodetic distances generalizing the shortest-path and the resistance distances. Discrete Applied Mathematics, 159 (5), 295 -- 302. Chebotarev, P. (2012). The walk distances in graphs. Discrete Applied Mathe- matics, 160 (10 -- 11), 1484 -- 1500. Chebotarev, P. (2013). Studying new classes of graph metrics. In F. Nielsen & F. Barbaresco (Eds.), Proceedings of the 1st international conference on geometric science of information (GSI '13) (Vol. 8085, pp. 207 -- 214). Springer. 38 Chung, F. R., & Lu, L. (2006). Complex graphs and networks. American Mathematical Society. Courty, N., Flamary, R., Tuia, D., & Rakotomamonjy, A. (2017). Optimal transport for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39 (9), 1853-1865. Cover, T. M., & Thomas, J. A. (2006). Elements of information theory, 2nd ed. John Wiley and Sons. Cuturi, M. (2013). Sinkhorn distances: lightspeed computation of optimal transport. In Advances in neural information processing systems 26: Pro- ceedings of the NIPS '13 conference (p. 2292-2300). MIT Press. Dobrushin, R. L. (1970). Prescribing a system of random variables by con- ditional distributions. Theory of Probability & Its Applications, 15 (3), 458 -- 486. Doyle, P. G., & Snell, J. L. (1984). Random walks and electric networks. The Mathematical Association of America. Erlander, S., & Stewart, N. (1990). The gravity model in transportation analysis. theory and extensions. VSP International Science Publishers. Estrada, E. Press. (2012). The structure of complex networks. Oxford University Fang, S., Rajasekera, J., & Tsao, H. (1997). Entropy optimization and mathe- matical programming. Springer. Ferradans, S., Papadakis, N., Peyr´e, G., & Aujol, J.-F. (2014). Regularized discrete optimal transport. SIAM Journal on Imaging Sciences, 7 (3), 1853 -- 1882. Fouss, F., Saerens, M., & Shimbo, M. (2016). Algorithms and models for network data and link analysis. Cambridge University Press. Fran¸coisse, K., Kivimaki, I., Mantrach, A., Rossi, F., & Saerens, M. (2017). A bag-of-paths framework for network data analysis. Neural Networks, 90 , 90 -- 111. Freeman, L. C. (1977). A set of measures of centrality based on betweenness. Sociometry, 40 (1), 35 -- 41. Garc´ıa-D´ıez, S., Vandenbussche, E., & Saerens, M. (2011). A continuous- state version of discrete randomized shortest-paths. Proceedings of the 50th IEEE International Conference on Decision and Control (IEEE CDC 2011), 6570-6577. Graybill, F. (1983). Matrices with applications in statistics. Wadsworth Inter- national Group. Grinstead, C., & Snell, J. L. (1997). Introduction to probability (2nd ed.). The Mathematical Association of America. Griva, I., Nash, S. G., & Sofer, A. (2009). Linear and nonlinear optimization: Second edition. Society for Industrial and Applied Mathematics (SIAM). Guex, G. (2016). Interpolating between random walks and optimal transporta- tion routes: Flow with multiple sources and targets. Physica A: Statistical Mechanics and its Applications, 450 , 264 -- 277. Guex, G., & Bavaud, F. (2015). Flow-based dissimilarities: shortest path, com- mute time, max-flow and free energy. In B. Lausen, S. Krolak-Schwerdt, & M. Bohmer (Eds.), Data science, learning by latent structures, and knowledge discovery (Vol. 1564, pp. 101 -- 111). Springer. Guex, G., Emmanouilidis, T., & Bavaud, F. (2017). Transportation clustering: a regularized version of the optimal transportation problem. (Submitted 39 for publication) Hara, K., Suzuki, I., Shimbo, M., Kobayashi, K., Fukumizu, K., & Radovanovic, M. (2015). Localized centering: Reducing hubness in large-sample data. In Proceedings of the 29th AAAI conference on artificial intelligence (AAAI '15) (pp. 2645 -- 2651). Hashimoto, T., Sun, Y., & Jaakkola, T. (2015). From random walks to distances on unweighted graphs. In Advances in neural information processing sys- tems 28: Proceedings of the NIPS '15 conference (p. 3429-3437). MIT Press. Herbster, M., & Lever, G. (2009). Predicting the labelling of a graph via minimum p-seminorm interpolation. In Proceedings of the 22nd conference on learning theory (COLT '09) (pp. 18 -- 21). Jaynes, E. T. (1957). Information theory and statistical mechanics. Physical Review , 106 , 620 -- 630. Kantorovich, L. V. (1942). On the translocation of masses. Doklady Akademii Nauk SSSR, 37 (7-8), 227 -- 229. Kapur, J. N. Wiley. (1989). Maximum-entropy models in science and engineering. Kapur, J. N., & Kesavan, H. K. (1992). Entropy optimization principles with applications. Academic Press. Kivimaki, I., Lebichot, B., Saramaki, J., & Saerens, M. (2016). Two between- ness centrality measures based on randomized shortest paths. Scientific Reports, 6 , srep19668. Kivimaki, I., Shimbo, M., & Saerens, M. (2014). Developments in the theory of randomized shortest paths with a comparison of graph node distances. Physica A: Statistical Mechanics and its Applications, 393 , 600 - 616. Klein, D. J., & Randi´c, M. (1993). Resistance distance. Journal of Mathematical Chemistry, 12 (1), 81 -- 95. Kolaczyk, E. D. (2009). Statistical analysis of network data: Methods and models. Springer. Kurras, S. (2015). Symmetric iterative proportional fitting. In Proceedings of the 18th international conference on artificial intelligence and statistics (AISTATS) (Vol. 38, pp. 526 -- 534). Lebichot, B., Kivimaki, I., Fran¸coisse, K., & Saerens, M. (2014). Semi- supervised classification through the bag-of-paths group betweenness. IEEE Transactions on Neural Networks and Learning Systems, 25 , 1173- 1186. Lebichot, B., & Saerens, M. (2018). A bag-of-paths node criticality measure. Neurocomputing, 275 , 224-236. Lewis, T. G. (2009). Network science. Wiley. Li, Y., Zhang, Z.-L., & Boley, D. (2013). From shortest-path to all-path: The IEEE Transactions on routing continuum theory and its applications. Parallel and Distributed Systems, 25 (7), 1745 -- 1755. Lougee-Heimer, R. (2003, Jan). The common optimization interface for opera- tions research: Promoting open-source software in the operations research community. IBM Journal of Research and Development, 47 (1), 57-66. doi: 10.1147/rd.471.0057 Lu, L., & Zhou, T. (2011). Link prediction in complex networks: A survey. Physica A: Statistical Mechanics and its Applications, 390 , 1150 -- 1170. 40 Mantrach, A., Yen, L., Callut, J., Fran¸coisse, K., Shimbo, M., & Saerens, M. (2010). The sum-over-paths covariance kernel: A novel covariance between nodes of a directed graph. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32 (6), 1112 -- 1126. Mantrach, A., Zeebroeck, N. V., Francq, P., Shimbo, M., Bersini, H., & Saerens, M. (2011). Semi-supervised classification and betweenness computation on large, sparse, directed graphs. Pattern recognition, 44 (6), 1212 -- 1224. Newman, M. E. (2005). A measure of betweenness centrality based on random walks. Social Networks, 27 (1), 39 -- 54. Newman, M. E. (2010). Networks: An introduction. Oxford University Press. Nguyen, C. H., & Mamitsuka, H. (2016). New resistance distances with global information on large graphs. In Proceedings of the 19th international con- ference on artificial intelligence and statistics (AISTATS) (pp. 639 -- 647). Osborne, M. J. (2004). An introduction to game theory. Oxford University Press. Pukelsheim, F. (2014). Biproportional scaling of matrices and the iterative proportional fitting procedure. Annals of Operations Research, 215 (1), 269 -- 283. Radovanovi´c, M., Nanopoulos, A., & Ivanovi´c, M. (2010a). Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11 (Sep), 2487 -- 2531. Radovanovi´c, M., Nanopoulos, A., & Ivanovi´c, M. (2010b). On the existence of obstinate results in vector space models. In Proceedings of the 33rd annual international ACM SIGIR conference on research and development in information retrieval (SIGIR '10) (pp. 186 -- 193). Saerens, M., Achbany, Y., Fouss, F., & Yen, L. (2009). Randomized shortest- path problems: Two related models. Neural Computation, 21 (8), 2363 -- 2404. Silva, T., & Zhao, L. (2016). Machine learning in complex networks. Springer. Sinkhorn, R. (1967). Diagonal equivalence to matrices with prescribed row and column sums. The American Mathematical Monthly, 74 (4), 402 -- 405. Solomon, J., Rustamov, R., Guibas, L., & Butscher, A. (2014). Wasserstein propagation for semi-supervised learning. In Proceedings of the 31 inter- national conference on machine learning (ICML '14) (p. 306-314). Sommer, F., Fouss, F., & Saerens, M. (2016). Comparison of graph node distances on clustering tasks. In Proceedings of the 25th international conference on artificial neural networks (ICANN 2016) (Vol. 9886, p. 192- 201). Sommer, F., Fouss, F., & Saerens, M. (2017). Modularity-driven kernel k- means for community detection. In Proceedings of the 26th international conference on artificial neural networks (ICANN 2017) (Vol. 10614, p. 423- 433). Suzuki, I., Hara, K., Shimbo, M., Matsumoto, Y., & Saerens, M. (2012). Inves- tigating the effectiveness of Laplacian-based kernels in hub reduction. In Proceedings of the 26th AAAI conference on artificial intelligence (AAAI '12) (pp. 1112 -- 1118). Suzuki, I., Hara, K., Shimbo, M., Saerens, M., & Fukumizu, K. (2013). Center- ing similarity measures to reduce hubs. In Proceedings of the international conference on empirical methods in natural language processing (EMNLP 2013) (p. 613-623). 41 Thelwall, M. (2004). Link analysis: An information science approach. Elsevier. Tomasev, N., Radovanovic, M., Mladenic, D., & Ivanovic, M. (2014). The role IEEE Transactions on of hubness in clustering high-dimensional data. Knowledge and Data Engineering, 26 (3), 739 -- 751. Villani, C. (2003). Topics in optimal transportation. American Mathematical Society. Villani, C. (2008). Optimal transport: old and new. Springer. von Luxburg, U., Radl, A., & Hein, M. (2010). Getting lost in space: large sample analysis of the commute distance. In Advances in neural infor- mation processing systems 23: Proceedings of the NIPS '10 conference (p. 2622-2630). MIT Press. von Luxburg, U., Radl, A., & Hein, M. (2014). Hitting and commute times in large random neighborhood graphs. Journal of Machine Learning Re- search, 15 (1), 1751 -- 1798. Wasserman, S., & Faust, K. (1994). Social network analysis: Methods and applications. Cambridge University Press. Wilson, A. (1970). Entropy in urban and regional modelling. Routledge. Yen, L., Mantrach, A., Shimbo, M., & Saerens, M. (2008). A family of dissim- ilarity measures between nodes generalizing both the shortest-path and the commute-time distances. In Proceedings of the 14th ACM SIGKDD international conference on knowledge discovery and data mining (KDD '08) (pp. 785 -- 793). Zhang, W., Zhao, D., & Wang, X. (2013). Agglomerative clustering via maxi- mum incremental path integral. Pattern Recognition, 46 (11), 3056 - 3065. Zolotarev, V. M. (1983). Probability metrics. Teoriya Veroyatnostei i ee Prime- neniya, 28 (2), 264 -- 287. 42
1905.12477	1	1905	2019-05-29T14:12:42	Understanding the Effectiveness of Data Reduction in Public Transportation Networks	[ "cs.SI", "cs.DS" ]	Given a public transportation network of stations and connections, we want to find a minimum subset of stations such that each connection runs through a selected station. Although this problem is NP-hard in general, real-world instances are regularly solved almost completely by a set of simple reduction rules. To explain this behavior, we view transportation networks as hitting set instances and identify two characteristic properties, locality and heterogeneity. We then devise a randomized model to generate hitting set instances with adjustable properties. While the heterogeneity does influence the effectiveness of the reduction rules, the generated instances show that locality is the significant factor. Beyond that, we prove that the effectiveness of the reduction rules is independent of the underlying graph structure. Finally, we show that high locality is also prevalent in instances from other domains, facilitating a fast computation of minimum hitting sets.	cs.SI	cs	Understanding the Effectiveness of Data Reduction in Public Transportation Networks Thomas Blasius, Philipp Fischbeck, Tobias Friedrich, and Martin Schirneck Hasso Plattner Institute, University of Potsdam, Potsdam, Germany [email protected] Abstract. Given a public transportation network of stations and con- nections, we want to find a minimum subset of stations such that each connection runs through a selected station. Although this problem is NP-hard in general, real-world instances are regularly solved almost com- pletely by a set of simple reduction rules. To explain this behavior, we view transportation networks as hitting set instances and identify two characteristic properties, locality and heterogeneity. We then devise a randomized model to generate hitting set instances with adjustable prop- erties. While the heterogeneity does influence the effectiveness of the re- duction rules, the generated instances show that locality is the significant factor. Beyond that, we prove that the effectiveness of the reduction rules is independent of the underlying graph structure. Finally, we show that high locality is also prevalent in instances from other domains, facilitat- ing a fast computation of minimum hitting sets. Keywords: Transportation networks · Hitting set · Graph algorithms · Random graph models 1 Introduction A public transportation network is a collection of stations along with a set of connections running through these stations. But beyond its literal definition, via bus stops and train lines, it also carries some of the geographical, social, and economical structure of the community it serves. Given such a network, we want to select as few stations as possible to cover all connections, i.e., each connection shall contain a selected station. This and similar covering problems arise from practical needs, e.g., when choosing stations for car maintenance, but their solu- tions also reveal some of the underlying structure of the network. Despite the fact that minimizing the number of selected stations is NP-hard, there is a surpris- ingly easy way to achieve just that on real-world instances: Weihe [19] showed for the German railroad network that two straightforward reduction rules simplify the network to a very small core which can then be solved by brute force. This is not a mere coincidence. Experiments have shown the same behavior on several other real-world transportation networks. Subsequently, the reduction rules be- came the standard preprocessing routine for many different covering problems. See the work of Niedermeier and Rossmanith [12], Abu-Khzam [1], or Davies and Bacchus [5], to name just a few. This raises the question as to why these rules are so effective. Answering this question would not only close the gap between theory and practice for the specific problem at hand, but also has the potential to lead to new insights into the networks' structure and ultimately pave the way for algorithmic advances in bordering areas. Our methodology for approaching this question is as follows. We first identify two characteristic properties of real-world transportation networks: heterogene- ity and locality; see Section 2.2 for more details. Then we propose a model that generates random instances resembling real-world instances with respect to het- erogeneity and locality. We validate our model by showing empirically that it provides a good predictor for the effectiveness of the reduction rules on real- world instances. Finally, we draw conclusions on why the reduction rules are so effective by running experiments on generated instances of varying heterogeneity and locality. Moreover, we show that our results extend beyond transportation networks to related problems in other domains. For our model, we regard transportation networks as instances of the hitting set problem. From this perspective, connections are mere subsets of the universe of stations and we need to select one station from each set. Note that this disre- gards some of the structure inherent to transportation networks: A connection is not just a set of stops but a sequence visiting the stops in a particular order. In fact, the sequences formed by the connections are paths in an underlying graph, which itself has rich structural properties inherited from the geography. Focusing on these structural properties, we also consider the graph-theoretic perspective. The working hypothesis for this perspective is that the underlying graphs of real-world transportation networks have beneficial properties that render the instances tractable. We disprove this hypothesis by showing that the underly- ing graph is almost irrelevant. This validates the hitting set perspective, which disregards the underlying graph. In Section 2, we formally state our findings on the graph-theoretic as well as the hitting set perspective. We study the hitting set instances of European transportation networks in Section 3, identifying heterogeneity and locality as characteristic features. In Section 4, we define and evaluate a model generating instances with these features. Section 5 extends our findings to other domains and Section 6 concludes this work. 2 Preliminary Considerations Before discussing the results regarding the two different perspectives, we fix some notations and state the reduction rules introduced by Weihe [19]. A public transportation network (or simply a network ) N = (S, C) consists of a set S of stations and a set C of connections which are sequences of stations. That is, each connection c ∈ C is a subset of S together with a linear ordering of its elements. Two stations s1, s2 ∈ S are connected in N if there exists a sequence of stations starting with s1 and ending in s2 such that each pair of consecutive stations shares a connection. The subnetworks induced by this equivalence relation are called the connected components of N . Given N = (S, C), the Station Cover problem is to find a subset S(cid:48) ⊆ S of minimum cardinality such that each connec- tion is covered, i.e., S(cid:48) ∩ c (cid:54)= ∅ for every c ∈ C. The reduction rules by Weihe [19] are based on notions of dominance, both between stations and connections. For two different stations s1, s2 ∈ S, s1 dominates s2 if every connection containing s2 also contains s1. If so, there is always an optimal station cover without s2, so it is never worse to select s1 instead. Thus, removing s2 from S and from every connection in C yields an equivalent instance. Similarly, for two different connections c1, c2 ∈ C, c1 dominates c2 if c1 ⊆ c2. Every subset of S cover- ing c1 then also covers c2. Removing c2 does not destroy any optimal solutions. Weihe's algorithm can thus be summarized as follows. Iteratively remove dom- inated stations and connections until this is no longer possible. The remaining instance, the core 1, is solved using brute force. Each connected component can be solved independently and the running time is exponential only in the number of stations. Thus, the complexity of an instance denotes the maximum number of stations in any of its connected components. The proofs of this section are in Appendix A. 2.1 Graph-Theoretic Perspective One way to represent a network N = (S, C) is via an undirected graph GN defined as follows. The stations S are the vertices of GN ; for each connection (s1, . . . , sk) ∈ C, GN contains the edges {si, si+1}1≤i<k. The basic hypothesis of the graph-theoretic perspective is that certain properties of GN make the real-world Station Cover instances easy. Consider a leaf u in GN , i.e., a degree-1 vertex. If there is a connection that contains only u, then this dominates all other connections containing u. Other- wise, all connections that contain u also contain its unique neighbor. Thus, u is dominated and removed by the reduction rules. We obtain the following propo- sition. The 2-core is the subgraph obtained by iteratively removing leaves [15]. Proposition 1. The reduction rules reduce any Station Cover instance N to an equivalent instance N(cid:48) such that GN(cid:48) is a subgraph of the 2-core of GN , with additional isolated vertices. Proposition 1 identifies the number of vertices in the 2-core of GN as an upper bound for the core complexity. The following theorem shows that this bound is arbitrarily bad. Supporting this assessment, we will see in Section 3 that the 2-cores of the graphs of real-world instances are rather large, while their core complexity is significantly smaller. Theorem 1. For every graph G, there exist two Station Cover instances N1 and N2 with G = GN1 = GN2 such that the core of N1 has complexity 1 while the core of N2 corresponds to the 2-core of G. 1 We note that the core is unique up to automorphisms. In particular, its size is independent of the removal order. Theorem 1 disproves the working hypothesis of the graph-theoretic perspective. For any connected graph that has no leaves, there is a Station Cover instance that is completely solved by the reduction rules, and another instance on the very same graph that is not reduced at all. Furthermore, unless the 2-core is small, the theorem shows that it is impossible to tell whether or not the reduction rules are effective on a given instance by only looking at the graph. So far, we have only focused on Weihe's algorithm. While our main goal is to explain the performance of this algorithm, one could argue that other methods exploiting different graph-theoretic properties are better suited to solve real-world instances. The next theorem, however, indicates that this is not the case. Even on "tree-like" graph classes Station Cover remains NP-hard. The reduction used to prove this theorem was originally given by Jansen [8]. Theorem 2 ([8], Theorem 5). Station Cover is NP-hard even if the cor- responding graph has treewidth 3 or feedback vertex number 2. 2.2 Hitting Set Perspective Another way to represent a network N is by an instance of the Hitting Set problem. Here, the connections C ⊆ 2S are regarded only as sets of stations (ignoring their order). An optimal cover is a minimum-cardinality subset of S that has a non-empty intersection with all members of C. This perspective turns out to be much more fruitful. In the next section, we analyze the Hitting Set instances stemming from 12 real-world networks. To summarize our results, we observe that the instances are heterogeneous, i.e., the number of connections con- taining a given station varies heavily. Moreover, the instances exhibit a certain locality, which probably has its origin in the stations' geographic positions. In more detail, for a station s ∈ S, let the number of connections in C that contain s be the degree of s. Conversely, for c ∈ C, c is its degree. The connection degrees of the real-world instances are rather homogeneous, i.e., every connection has roughly the same size. Although there are different types of connections they appear to have a similar number of stops. The station degrees, on the other hand, vary strongly. In fact, we observe that the station degree distributions roughly follow a power law. This is in line with observations that, e.g., the sizes of cities are power-law distributed [6]. To quantify the locality of an instance, we use a variant of the so-called bipartite clustering coefficient [14]. We conjecture that heterogeneity of stations and locality of the network are the crucial factors that make the reduction so effective. If the station degrees vary strongly, chances are that some high-degree station exists that dominates many low-degree ones. Moreover, if locality is high, there tend to be several connections differing only in few stations and stations appearing in similar sets of connections. This increases the likelihood of dominance among the elements of both S and C. To verify this hypothesis empirically, we propose a model for generating instances of varying heterogeneity and locality. Our findings suggest that higher heterogeneity decreases the core complexity, but the deciding factor is the locality. Finally, we observe that locality is also prevalent in other domains. As predicted by our model, preprocessing also greatly reduces these instances. Data Set sncf nl kvv vrs rnv athens petersburg warsaw luxembourg switzerland vbb db S 1742 4558 2115 5491 705 5729 4264 3944 2484 22 535 3031 514 S C 4.0 13.2 8.0 10.7 12.4 24.4 6.5 13.0 7.3 5.6 16.5 0.9 δS 2.2 1.5 2.1 1.9 1.4 1.8 2.5 1.8 2.7 2.0 β 3.3 3.8 3.5 3.5 4.2 3.9 4.0 5.9 2.9 4.5 1.4 15.7 12.4 2.0 KS 0.03 0.04 0.03 0.03 0.06 0.04 0.03 0.05 0.02 0.02 0.05 0.07 κ 2-core core 0.47 0.40 0.48 0.27 0.38 0.30 0.31 0.29 0.25 0.33 0.38 0.28 70% 0.3% 70% 2.8% 72% 0.8% 83% 0.1% 54% 0.1% 89% 4.7% 86% 8.3% 80% 5.9% 84% 0.2% 71% 1.7% 73% 1.8% 78% 0.2% Table 1: Transportation networks with atypical instances separated. Shown are the number S of stations, the station-connection ratio S/C, average station degree δS, estimated power-law exponent β, corresponding KS distance, bipartite clustering coefficient κ, relative 2-core size, and the relative core complexity. 3 Analysis of Real-World Networks We examined several public transportation networks from different cities (athens, petersburg, warsaw), rural areas (sncf, kvv, vrs, rnv, vbb), and countries (nl, luxembourg, switzerland, db). The networks are taken from the transitfeeds. com repository. The raw data has the General Transit Feed Specification (GTFS) format. It stores multiple connections for the same route, one for each time a train actually drives that route. For each route, only one connection was used. Table 1 gives an overview of the relevant features of the resulting networks. We reduced each instance to its largest component. For most of them, only a small fraction of stations and connections are disconnected from this component. A notable exception is the vbb-instance, representing the public transportation network of the city of Berlin, Germany. In total, it has 13 424 stations while its largest component has only 3031. The reason is that different modes of transport are separated in the raw data. As a result, vbb has rather uncommon features. Another unusual case is the db-instance of the German railway network. Table 1 shows that most instances have a station-connection ratio S/C of roughly 10. For db, however, this ratio is at 0.9 much smaller. Heterogeneity. The average station degree δS of the investigated networks is a small constant around 2, independent of the instances' complexity. The only exception is the db-network. This can be explained by the atypical value for S/C, and that each station is contained in much more connections. The average connection degree δC = δS · S/C (not explicitly given in the table) is roughly 20, due to the station-connection ratios all being of the same order. Fig. 1: (left) The blue line is the CCDF of station degrees for the data set kvv. The red line is the estimated power-law distribution. (right) The histogram of connection degrees for the data set kvv. Beyond small average degree, almost all instances exhibit strong hetero- geneity among the degrees of different stations. We take a closer look at the kvv-instance as a prototypical example, representing the public transportation network of Karlsruhe, Germany. Figure 1 (left) shows the complementary cumu- lative distribution function (CCDF) of the station degrees in a log-log plot. For a given value x, the CCDF describes the share of stations that have degree at least x. The CCDF closely resembles a straight line (in log-log scaling), indicating a power-law distribution. That means, there exists a real number β, the power-law exponent, such that the number of stations of degree x is roughly proportional to x−β. We estimated the power-law exponents using the python package pow- erlaw [2]. For kvv, the exponent β is approximately 3.5. The goodness of fit is measured by the Kolmogorov -- Smirnov distance (KS distance), which is the maximum absolute difference between the CCDFs of the measurement and of the assumed distribution. The KS distance for the kvv is 0.05. Table 1 reports both the power-law exponents and the corresponding KS distances. The estimated values of β, excluding the outlier vbb, indicate a high level of heterogeneity. As a side note, the power-law exponent for vbb is 4.1 with a KS distance of 0.03 when considering the whole network instead of the largest component. In contrast, the connection degrees are rather homogeneous, cf. e.g. kvv in Figure 1 (right). A possible explanation is that long-distance trains stop less frequently. Locality. To measure locality, we adapt the bipartite clustering coefficient [14]. Intuitively, it states how likely it is that two stations which share a connection are also contained together in a different connection, or that two connections containing the same station also have another station in common. For a formal definition, first note that we can interpret a Hitting Set instance (S, C) as a bipartite graph with the two partitions S and C and an edge joining s ∈ S and c ∈ C iff s ∈ c. Let #P3 denote the number of paths of length 3 and #C4 the number of cycles of length 4 in this graph. The bipartite clustering coefficient κ 10010110−2100stationdegreeratioofstations0204060020406080connectiondegreeconnectionsthen is defined as κ = 4· #C4/#P3. It is the probability that a uniformly chosen 3-path is contained in a 4-cycle. Before computing κ, we normalize the bipartite graph by reducing it to its 2-core, which removes any attached trees. In doing so, the measure becomes more robust for our purpose, as attached trees do not impact the difficulty of an instance (they get removed by the reduction rules) while they decrease the clustering coefficient. The clustering coefficients are reported in Table 1. All instances have a clus- tering coefficient of at least 0.25, which indicates a high level of locality. A pos- sible explanation are the underlying geographic positions of the stations, with nearby stations likely appearing in the same connection. Degree of Reduction. We measure the effectiveness of the reduction rules using the relative core complexity. It is the percentage of stations that remain after exhaustively applying the preprocessing. Table 1 shows that the resulting relative core complexity is very low for all 12 instances. This is in line with the original findings of Weihe [19], who applied the reduction rules on a few select European train networks. Moreover, it generalizes these results to networks of different scales, from urban to national. On the other hand, the 2-core is typically not much smaller than the original instance. This shows that Proposition 1 can- not explain the effectiveness of the reduction rules, which supports our previous assessment that the graph-theoretic perspective is not sufficient. Judging from Table 1, we believe that heterogeneity of the stations and high locality are the crucial properties rendering the preprocessing so effective. Notwithstanding, it is also worth noting that the reduction rules work well on all instances, including vbb which is not very heterogeneous. The clustering on the other hand is high for all instances, indicating that locality is more important. Also, the db and vbb outliers seem to show that the influence of the station- connection ratio and the average station degree is limited. Though looking at these 12 networks can provide clues to what features are most important, it is not sufficient to draw a clear picture. In the following, we thoroughly test the effect of different properties on the effectiveness of the reduction rules by generating instances with varying properties. 4 Analysis of Generated Instances This section discusses the generation and analysis of artificial Hitting Set in- stances. First, we present our model of generation which is based on the geometric inhomogeneous random graphs [4]. It allows creating networks with varying de- gree of heterogeneity and locality. We then analyze these instances with respect to the degree of reduction. 4.1 The Generative Model In the field of network science, it is generally accepted that vertex degrees in real- istic networks are heterogeneous [17]. A power-law distribution can be explained, inter alia, by the preferential attachment mechanism [3]. Beyond the generation of heterogeneous instances, different models have been proposed to also account for locality. The latter models typically use some kind of underlying geometry. One of the earliest works in that direction is by Watts and Strogatz [18]. More recently, and closer to our aim, Papadopoulos et al. [13] introduced the concept of popularity vs. similarity, making the creation of edges more likely, the more popular and similar the connected vertices are. They also observed that these two dimensions are naturally covered by the hyperbolic geometry, leading to hyper- bolic random graphs [10]. Bringmann, Keusch, and Lengler [4] generalized this concept to geometric inhomogeneous random graphs (GIRGs). There, each ver- tex has a geometric position and a weight. Vertices are then connected by edges depending on their weights and distances. Despite a plethora of models for gen- erating graphs, we are not aware of models generating heterogeneous Hitting Set instances. The closest is arguably the work by Gir´aldez-Cru and Levy [7], who generate SAT instances using the popularity vs. similarity paradigm. To generate Hitting Set instances with varying heterogeneity and locality, we formulate a randomized model based on GIRGs. Each station and connection has a weight representing its importance. Moreover, stations and connections are randomly placed in a geometric space. The distance between stations and connections then provides a measure of similarity. In the Hitting Set instance, some station s is a member of connection c with a probability proportional to the combined weights of s and c and inverse proportional to the distance between the vertices s and c. To make this more precise, let wS : S → R and wC : C → R be two weight functions; we omit the subscript when no ambiguity arises. For s ∈ S and c ∈ C, let dist(s, c) denote the geometric distance between the corresponding vertices. Finally, fix two positive constants a, T > 0. Then, station s is contained in connection c with probability (cid:40) (cid:18) (cid:19)1/T(cid:41) P (s, c) = min 1, . (1) a · w(s)w(c) dist(s, c) The parameter a governs the expected degree. The temperature T controls the influence of the geometry. For T → 0 the method converges to a step model, where s is contained in c if and only if dist(s, c)≤ aw(s)w(c). Larger temperatures soften this threshold, allowing s ∈ c for larger distances, and s /∈ c for smaller distances, with a low probability. Thus, T influences the locality of the instance. The remaining degrees of freedom are the choice of the underlying geometry and the weights. For the geometry, we use the unit circle. Positions for stations and connections are drawn uniformly at random from [0, 1] and the distance between x, y ∈ [0, 1] is min{x− y, 1−x− y}. This is arguably the simplest possible symmetric geometry. To choose the weights properly, it is important to note that the resulting degrees are expected to be proportional to the weights [4]. Thus, we mimic the real-world instances by choosing uniform weights for the connections and power- law weights, with varying exponent β, for the stations. It is not hard to see that for β → ∞, the latter converge to uniform weights as well. In summary, adjusting β controls the heterogeneity. 4.2 Evaluation We generate artificial networks and measure the dependence of their relative core complexity on the heterogeneity and locality. The size of an instance has three components: the (original) complexity S, the station-connection ratio S/C, and the average station degree δS. Note that these values also determine the number C of connections and average connection degree δC. From the model, we have the two parameters we are most interested in, the power-law exponent β and the temperature T . We also consider the limit case of uniform weights for all vertices; slightly abusing notation, we denote this by β = ∞. For the main part of the experiments, we used S = 2000, S/C = 10.0, and δS = 2.0, leaning on the respective properties for the real-world instances. We let T vary between 0 and 1 in increments of 0.05, and β between 2 and 5 in increments of 0.25. For each combination, we generated ten samples. In the following, we first validate the data. Then we examine the influence of hetero- geneity and locality. Afterwards we test whether our findings still hold true for different station-connection ratios and station degrees. Data Validation. There are two aspects to the data validation. First, the instances should approximately exhibit the properties we explicitly put in, i.e., the values of S, S/C, δS, and the power-law behavior. Second, the implicit properties should also be as expected. In our case, we have to verify that changing T actually has the desired effect on the bipartite clustering coefficient κ. Concerning the complexity S, note that a sampled instance per se does not need to be connected. If it is not, we again only use the largest component. Thus, when generating an instance with 2000 stations, the resulting complexity is actually a bit smaller. There are typically many isolated stations due to the small average station degree, this is particularly true for small power-law exponents. However, the complexity of the largest component never dropped below 1000 and usually was between 1400 and 1700 provided that β > 2.5. The transition to the largest component mainly meant ignoring isolated stations. Thus, also the station-connection ratio S/C decreases slightly. For β > 3 it was typically around 8 and always at least 7. For smaller β, it is never below 5. Recall that the average station degree δS is controlled by parameter a in Equation 1. It is a constant in the sense that it is independent of the considered station-connection pair. However, it does depend on other parameters of the model. As there is no closed formula to determine a from δS, we estimated it numerically. This estimation incurred some loss in accuracy but yielded values of δS between 1.9 and 2.1, very close to the desired δS = 2. Transitioning to the largest component typically slightly increases δS, as the largest component is more likely to contain stations of higher degree. Anyway, δS never went above 2.7. As with the real-world instances, we estimated the exponent β of the gen- erated instances. For small values, the estimates matched the specified values. Fig. 2: (left) The clustering coefficient κ depending on the temperature T for different β. (right) The relative core complexity depending on the power-law exponent β and temperature T . Fig. 3: The relative core complexity (left) depending on the temperature T (for different power-law exponents β), and (right) depending on β (for different T ). For larger exponents, the gap increases slightly, e.g., an estimated β = 5.7 for an instance with predefined parameter 5.0. Finally, we examined the dependency between the bipartite clustering coef- ficient κ and the temperature T ; see Figure 2 (left). As desired, the clustering coefficient increases with falling temperatures. More precisely, κ ranges between 0 and some maximum attained at T = 0. The value of this maximum depends on the power-law exponent β, with smaller β giving smaller maxima. Heterogeneity and Locality. For each instance described above, we computed the relative core complexity. To reduce noise, we look at the arithmetic mean of ten samples for each parameter configuration. The measured core complexi- ties and clustering coefficients in fact showed only small variance, with almost all values differing at most 5% from the respective mean. These measurements are presented in Figure 2 (right), showing the mean core complexity depending on the temperature T and the power-law exponent β. For all parameter con- 00.510.20.40.6Tκβ=2.5β=3.5β=5.0β=∞23450.30.60.90.20.40.60.8βTrel.corecomplexity00.5100.20.40.6Trelativecorecomplexityβ=2.5β=3.5β=5.0β=∞234500.20.40.6βrelativecorecomplexityT=0.1T=0.4T=0.7T=0.9Fig. 4: The relative core complexity depending on the power-law exponent β and temperature T under alternative model parameters. (left) The station- connection ratio is S/C = 4.0 (instead of 10.0). (right) The average station degree is δS = 5.0 (instead of 2.0). figurations, the relative core complexity was at most 50%. This is due to the low average station degree, which leads to dominant low-degree stations. More importantly, the core complexity varies strongly for different values of T and β. The complexity decreased both with lower temperatures and lower power-law exponents. This further supports our claim that heterogeneity and locality both have a positive impact on the effectiveness of the reduction rules. The locality, however, seems to be more vital. To make this precise, low temperatures lead to small cores, independent of β, as shown in Figure 3 (left). Even under uniform station weights (β = ∞), temperatures below 0.3 consistently produced instances with empty core. In contrast, the power-law exponent has only a minor impact. One can see in Figure 3 (right) that for low temperatures, the core complexity is (almost) independent of β. For higher temperatures, the core complexities remain high over wide ranges of β, except for very low exponents. In summary, high locality seems to be the most prominent feature that makes Station Cover instances tractable, independent of their heterogeneity. Hetero- geneity alone reduces the core complexity only slightly, except for extreme cases (very low power-law exponents). It is thus not the crucial factor. In the follow- ing, we verify this general behavior also for alternative model parameters such as station-connection ratio or average station degree. Station-Connection Ratio. Recall that we fixed a ratio of S/C = 10.0 for the main part of our experiments. To see whether our observations are still valid for different settings, we additionally generated data sets with mostly the same parameters as before, except for S/C = 4.0. The result are shown in Figure 4 (left). Comparing this to Figure 2 (right), one can see that general dependence on T and β is very similar. However, there are subtle differences. Under the smaller station-connection ratio the instances are tractable even for larger temperatures, up to T = 0.5 instead of the earlier 0.3, i.e., for lower locality. Also, the maximum core complexity over all combinations of T and β is 23450.30.60.90.20.40.60.8βTrel.corecomplexity23450.30.60.90.20.40.60.8βTrel.corecomplexityFig. 5: The relative core complexity depending on the clustering coefficient κ (left) for generated instances with different values of the power-law exponent β, and (right) for the real-world instances. larger than before, reaching almost 60%, compared to the 50% for S/C = 10.0. In summary, in the (more realistic) low-temperature regime, a lower station- connection ratio seems to further improve the effectiveness of the reduction rules. Average Degree. To examine the influence of the average station degree δS, we generated another instance set with the same parameters as the main one, except that we increased the δS from 2.0 to 5.0. The results are shown in Fig- ure 4 (right). Again the general behavior is similar, but now lower tempera- tures are necessary to render the instances tractable. Moreover, the maximum core complexity increases significantly, reaching up to almost 100%. Generally, a smaller average degree makes the reduction rules more effective. Intuitively speaking, the existence of stations with low degree increases the likelihood that the reduction rule of station dominance can be applied. Comparison with Real-World Instances. To compare generated and real- world instances more directly, we investigate the dependence between the rel- ative core complexity and the bipartite clustering coefficient κ, instead of the model-specific temperature T . Figure 5 shows the results. Several of the real- world networks are well covered by the model. For example, our findings from the generated instances can be directly transferred to sncf and kvv. They have power-law exponents of 3.3 and 3.5, respectively, as well as clustering coefficients of at least 0.47, which explains their small core complexity. Moreover, for lux- embourg, the model can explain the low complexity in spite of a small clustering coefficient of 0.25. The network exhibits a small exponent β = 2.9, which ben- efits the effectiveness of the preprocessing. On the other hand, petersburg has clustering κ = 0.31, but also a comparatively large core complexity of over 8%. Here, the main factor seems to be the high power-law exponent of 4.0. 00.20.40.600.20.40.6κrelativecorecomplexityβ=2.5β=3.5β=5.0β=∞00.20.40.600.20.40.6κrelativecorecomplexitysncfnlkvvvrsrnvathenspetersburgwarsawluxembourgvbbswitzerlanddbNotwithstanding, there are also some real-world instances that have an unex- pectedly low core complexity, which cannot be fully explained by the model. The vrs-instance has a low clustering coefficient of κ = 0.27 and a high power-law exponent of 3.5, but still a very low 0.1% core complexity. The reason seems to be its low average station degree of δS = 1.9. The switzerland-instance also has a low core complexity of 1.7%, despite its high power-law exponent of β = 4.5 and a clustering coefficient of κ = 0.33. Especially the high value of β would point to a much higher complexity, however, its station-connection ratio S/C = 5.6 is significantly lower than that of the generated instances. All in all, the net- works generated by the model are not perfectly realistic. However, the model does replicate properties that are crucial for the effectiveness of the reduction rules on real-world instances. Furthermore, in the interplay of heterogeneity and locality, it reveals locality as the more important property. 5 Impact on Other Domains Although the focus of this paper is to understand which structural properties of public transportation networks make Weihe's reduction rules so effective, our findings go beyond that. Our experiments on the random model predict that Hitting Set instances in general can be solved efficiently if they exhibit high locality. Moreover, if the instance is highly heterogeneous, a smaller clustering coefficient suffices; see Figure 5 (left). The element-set ratio and the average degree have, within reasonable bounds, only a minor impact on the effectiveness. Our experiments in Section 4.2 showed that instances are more difficult for a larger average degree and if the difference between the number of elements and the number of sets is high. Experiments not reported in this paper show that the latter is also true, if there are more sets than elements (i.e., the ratio is below 1). Data Sets. We consider Hitting Set instances from three different applica- tions; see Table 2. The first set of instances are metabolic reaction networks of Escherichia coli bacteria. The elements represent reactions and each set is a so-called elementary mode. Analyzing the hitting sets of these instances has applications in drug discovery. The corresponding data sets, labeled ec-, were generated with the Metatool [9]. In the second type of instance, the sets con- sist of so-called elementary pathways that need to be hit by interventions that suppress all signals, which is relevant, inter alia, for the treatment of cancer. The data sets, EGFR. and HER2.*, were obtained via the OCSANA tool [16]. The instances country-cover and language-cover are based on a country-language graph, taken from the network collection KONECT [11], with an edge between a country and a language if the language is spoken in that country. The corre- sponding Hitting Set instances ask for a minimum number of countries to visit to hear all languages, and for a minimum number of languages necessary to communicate with someone in every country, respectively. Data Set ec-acetate ec-succinate ec-glycerol ec-glucose ec-combined EGFR.short EGFR.sub HER2.short HER2.sub country-cover S 57 57 60 58 64 50 56 123 172 248 610 S/C 0.214 0.061 0.028 0.009 0.002 0.400 0.239 0.230 0.068 0.407 2.460 κ core 0.67 0.59 0.66 0.66 0.62 0.66 0.57 0.53 0.55 0.11 0.11 1.8% 1.8% 1.7% 36.2% 40.6% 2.0% 1.8% 9.8% 10.4% 0.4% 0.2% language-cover Table 2: Hitting Set instances from other domains. Listed are the number S of elements, the element-set ratio S/C, the bipartite clustering coefficient κ, and the relative core complexity. Evaluation. The basic properties of the instances and the effectiveness of the reduction rules are reported in Table 2. The results match the prediction of our model: most instances have a high clustering coefficient and the reduction rules are very effective. The only instances that stand out at first glance are ec-glucose, ec-combined, HER2.short, and HER2.sub, which are not solved completely by the reduction rules despite their high clustering coefficients, as well as country-cover and language-cover, which are solved completely despite the comparatively low clustering coefficient of κ = 0.11. However, a more detailed consideration reveals that these instances also match the predictions of the model. First, the two instances country-cover and language-cover are very heterogeneous with power-law exponent β = 2.2. As can be seen in Figure 2 (left), a clustering coefficient of κ = 0.11 is already rather high for this exponent, leading to a low core complexity; see Figure 5 (left). The instances ec-glucose and ec-combined have skewed element-set ratios (more than 100 times as many sets as elements) and a high average degree (30 for the sets; 3k and 13k for the elements, respectively). Thus, these instances at least qualitatively match the predictions of the model that the reduction rules are less effective if the element-set ratio is skewed or the average degree is high. One obtains a similar but less pronounced picture for HER2.short and HER2.sub. 6 Conclusion We explored the effectiveness of data reduction for Station Cover on trans- portation networks. Our main finding is that real-world instances have high locality and heterogeneity, and that these properties make the reduction rules effective, with locality being the crucial factor. This directly transfers to general Hitting Set instances. For future work, it would be interesting to rigorously prove that the reduction rules perform well on the model. References 1. Abu-Khzam, F.N.: A Kernelization Algorithm for d-Hitting Set. Journal of Com- puter and System Sciences 76, 524 -- 531 (2010) 2. Alstott, J., Bullmore, E., Plenz, D.: powerlaw: A Python Package for Analysis of Heavy-Tailed Distributions. PLOS One 9, e85777 (2014) 3. Barab´asi, A.L., Albert, R.: Emergence of Scaling in Random Networks. Science 286, 509 -- 512 (1999) 4. Bringmann, K., Keusch, R., Lengler, J.: Sampling Geometric Inhomogeneous Ran- dom Graphs in Linear Time. In: Proceedings of the 25th Annual European Sym- posium on Algorithms (ESA). pp. 20:1 -- 20:15 (2017) 5. Davies, J., Bacchus, F.: Solving MAXSAT by Solving a Sequence of Simpler SAT Instances. In: Proceedings of the 17th International Conference on Principles and Practice of Constraint Programming (CP). pp. 225 -- 239 (2011) 6. Gabaix, X.: Zipf's Law for Cities: an Explanation. The Quarterly Journal of Eco- nomics 114, 739 -- 767 (1999) 7. Gir´aldez-Cru, J., Levy, J.: Locality in Random SAT Instances. In: Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI). pp. 638 -- 644 (2017) 8. Jansen, B.M.P.: On Structural Parameterizations of Hitting Set: Hitting Paths in Graphs Using 2-SAT. Journal of Graph Algorithms and Applications 21, 219 -- 243 (2017) 9. von Kamp, A., Schuster, S.: Metatool 5.0: Fast and Flexible Elementary Modes Analysis. Bioinformatics 22, 1930 -- 1931 (2006) 10. Krioukov, D., Papadopoulos, F., Kitsak, M., Vahdat, A., Bogun´a, M.: Hyperbolic Geometry of Complex Networks. Physical Review E 82, 036106 (2010) 11. Kunegis, J.: KONECT - The Koblenz Network Collection. In: Proceedings of the 22nd International Conference on World Wide Web (WWW). pp. 1343 -- 1350 (2013) 12. Niedermeier, R., Rossmanith, P.: An Efficient Fixed-Parameter Algorithm for 3- Hitting Set. Journal of Discrete Algorithms 1, 89 -- 102 (2003) 13. Papadopoulos, F., Kitsak, M., ´Angeles Serrano, M., Bogun´a, M., Krioukov, D.: Popularity Versus Similarity in Growing Networks. Nature 489, 537 -- 540 (2012) 14. Robins, G., Alexander, M.: Small Worlds Among Interlocking Directors: Network Structure and Distance in Bipartite Graphs. Computational & Mathematical Or- ganization Theory 10, 69 -- 94 (2004) 15. Seidman, S.B.: Network Structure and Minimum Degree. Social Networks 5, 269 -- 287 (1983) 16. Vera-Licona, P., Bonnet, E., Barillot, E., Zinovyev, A.: OCSANA: Optimal Com- binations of Interventions from Network Analysis. Bioinformatics 29, 1571 -- 1573 (2013) 17. Voitalov, I., van der Hoorn, P., van der Hofstad, R., Krioukov, D.V.: Scale-free Networks Well Done. CoRR abs/1811.02071 (2018) 18. Watts, D.J., Strogatz, S.H.: Collective Dynamics of 'Small-World' Networks. Na- ture 393, 440 -- 442 (1998) 19. Weihe, K.: Covering Trains by Stations or the Power of Data Reduction. In: Pro- ceedings of the 1998 Algorithms and Experiments Conference (ALEX). pp. 1 -- 8 (1998) A Omitted Proofs In this appendix, we sketch the omitted proofs of the theorems in Section 2. Theorem 1. For every graph G, there exist Station Cover instances N1 and N2 with G = GN1 = GN2 such that the core of N1 has complexity 1 while the core of N2 corresponds to the 2-core of G. Proof (Proofsketch). We assume that G is connected, otherwise we apply the following to each component. For N1, we choose connections such that there is a single designated station s contained in each connection. Such connections can be constructed as follows. For an edge {u, v} of G, construct a simple path that contains s and the edge {u, v}. We use the stations on this path as a connection in N1 and repeat this for every edge. It can be easily verified that GN1 = G and that the core of N1 has complexity 1 (s dominates all other stations). For N2, first assume that G has no leaves, i.e., G is already the 2-core. It is then not hard to see that having one connection for each edge containing exactly its two endpoints prohibits any of the two reduction rules. In the general case, we have to be a bit more careful as applying reduction rules to leaves of G can lead to connections that contain only a single station. These then lead to further reductions that can potentially cascade through the 2-core. We prevent this as follows. For every edge in the 2-core of G, we add the connection containing its two endpoints (as before). For every station s not in the 2-core, we find a simple path that contains s and exactly one edge from the 2-core. We add the connection containing the stations of this path to N2. One can verify that indeed GN2 = G as every edge in G corresponds to a pair of consecutive stations in at least one connection. Moreover, exhaustively applying the reduction rules leads to the 2-core of G with one connection for each edge. Theorem 2 ([8], Theorem 5). Station Cover is NP-hard even if the cor- responding graph has treewidth 3 or feedback vertex number 2. Proof (Proofsketch). We prove this by reducing from 3-SAT. Let ϕ be a propo- sitional formula on n variables. We create two vertices x and x for every literal and connect them. Then, we add two vertices z0 and z1 which are connected to every x vertex. For every literal, we add a path (x, x). For every clause (a∨ b∨ c) we add a path (xa, z0, xb, z1, xc) where each xi is the vertex xi or xi based on whether the variable is inverted or not. Figure 6 illustrates this reduction. Now, if and only if there is a path cover by vertices of size n (where n is the number of variables), the SAT instance is solvable. The graph created has a feedback vertex number of 2: if we remove both z0 and z1 from the graph, the only edges left are those between the vertices x and x for each variable. The graph also has treewidth 3. Consider the following tree decomposition. For every variable x, we add a bag {x, x, z0, z1}. Our tree decomposition now consists of all of these bags in one single path in an arbitrary order. Each vertex is in some bag. For every edge, there is a bag that contains both of its vertices. Finally, the only vertices that occur in multiple bags are z0 and z1, and since they are in every bag, this forms a subtree. Each bag has a size of 4, thus the graph has a treewidth of 3. Fig. 6: Illustration of the proof of Theorem 2. The resulting graph has treewidth 3. The path shown in red represents the clause a ∨ b ∨ c. z1aabbccz0
1909.07316	1	1909	2019-09-16T16:20:10	Situational Awareness Enhanced through Social Media Analytics: A Survey of First Responders	[ "cs.SI", "cs.CY" ]	Social media data has been increasingly used to facilitate situational awareness during events and emergencies such as natural disasters. While researchers have investigated several methods to summarize, visualize or mine the data for analysis, first responders have not been able to fully leverage research advancements largely due to the gap between academic research and deployed, functional systems. In this paper, we explore the opportunities and barriers for the effective use of social media data from first responders' perspective. We present the summary of several detailed interviews with first responders on their use of social media for situational awareness. We further assess the impact of SMART-a social media visual analytics system-on first responder operations.	cs.SI	cs	Situational Awareness Enhanced through Social Media Analytics: A Survey of First Responders Luke S. Snyder Morteza Karimzadeh Christina Stober David S. Ebert Visual Analytics for Command, Deparment of Geography, Visual Analytics for Command, Visual Analytics for Command, Control, and Interoperability Environments (VACCINE) Center Purdue University West Lafayette, Indiana, USA [email protected] University of Colorado Boulder Boulder, Colorado, USA [email protected] Control, and Interoperability Environments (VACCINE) Control, and Interoperability Environments (VACCINE) Center Purdue University Center Purdue University West Lafayette, Indiana, USA West Lafayette, Indiana, USA [email protected] [email protected] Abstract -- Social media data has been increasingly used to facilitate situational awareness during events and emergencies such as natural disasters. While researchers have investigated several methods to summarize, visualize or mine the data for analysis, first responders have not been able to fully leverage research advancements largely due to the gap between academic research and deployed, functional systems. In this paper, we explore the opportunities and barriers for the effective use of social media data from first responders' perspective. We present the summary of several detailed interviews with first responders on their use of social media for situational awareness. We further assess the impact of SMART -- a social media visual analytics system -- on first responder operations. Keywords -- situational awareness, social media, survey I. INTRODUCTION There has been considerable research on the potential and use of social media data for situational awareness, such as disaster response, event monitoring, or public sentiment analysis [1 -- 5]. In particular, researchers have provided visual analytics approaches combining visualization and data analysis [2 -- 7], geoparsing for crisis-mapping [7, 8], cloud-based mobile services for on-the-ground information [14], and data mining and natural language processing techniques such as clustering and named entity recognition [15]. The majority of this research has focused on devising computational and/or visual methods for making sense of the vast amount of streaming social media data. However, the first responder community has not been able to fully leverage the state-of-the-art achievements made possible through such methods, mostly because (a) the advancements are not made available as user-friendly, operational and scalable systems, and (b) there is a gap between academic research (which strives to push the boundaries of science) and real-world requirements (which have different interests and may be constrained by various social, political, and ethical issues). In this paper, we build on our research on the development of the Social Media Analytics and Reporting Toolkit (SMART) [6, 9] -- which has been used by many the opportunities and obstacles in the use of social media data for real-time situational awareness. We leveraged our user community and conducted interviews with first responders responders -- to explore first XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE tasked with monitoring or using social media data for situational awareness (e.g. to report on an unfolding situation, coordinate action on the ground, and pass on information to relevant entities) to inform research and development in this field. The results of our the advantages and shortcomings of using social media data analytics for situational awareness in real-world operational settings. interviews highlight In what follows, we summarize the responses we received from interviewees in three major categories: (1) The use of social media platforms in first responder operations; (2) The impact of SMART social media analytics on first responder operations; and (3) Detailed feedback on the usefulness of SMART's features. The interviewees' responses shed light on future directions for successful research, development and adoption of social media analytics for situational awareness. II. METHODS We interviewed 13 first responders who had been introduced to and/or worked with SMART for situational awareness in emergency operation centers or command centers during the 2016-2019 period. All interviewees were active members of either law enforcement (8) or the United States Coast Guard (USCG) (5). Interviews were conducted over video calls or in- person, based on the location of interviewees. Per Institutional Review Board policies to protect human subjects, all survey responses have been anonymized to protect the interviewees' privacy by removing personally identifiable information (PII). As a result, some specifics from the responses have been removed from this paper. III. OVERVIEW OF SMART SMART [6, 9] was developed at Purdue University through an iterative, user-centered design by collecting feedback from the first responder community to ensure its usefulness. SMART provides a user-friendly, graphical web-based user interface, hiding the complexity of advanced algorithms filtering and summarizing social media data, which allows users to interactively explore, filter, and visualize trends and anomalies in real-time geo-tagged Twitter data. SMART also provides topic aggregation, natural language processing, and probabilistic event detection to analyze and refine the high volume of data. conventions, celebrations, political SMART has been successfully deployed by first responders during several major events in the United States, including holiday the 2017 Presidential Inauguration, and multiple Presidential State of the Union events. Using SMART, first responders have identified actionable information such as road closures for further investigations. Additionally, the USCG utilized SMART during the four hurricanes that made landfall in the U.S. in 2017, allowing them to respond more efficiently to developing situations. Fig. 1. SMART in a wildfire use-case. User-configured classifiers for keyword filtering, content lens for word cloud (topic) visualization in specific geographical locations, cluster lens for classifier keyword detection within tweet clusters, heat mapping to view the spatial distribution of tweets, temporal view (time bar graph) to visualize the time-evolution of tweet content, and the world cloud (LDA) topic modeling view to identify trending topics and associated keywords. IV. SURVEY RESULTS: USE OF SOCIAL MEDIA AND CROWDSOURCING BY FIRST RESPONDERS In this section, we review the results of our survey on the use of social media for situational awareness, using the platforms directly or tools other than SMART. In the next major section, we will focus on SMART's usefulness for this purpose. A. Purpose of/Reasons for Monitoring Social Media Most of the interviewees stated that they monitor social media (in general, or using SMART, specifically) during large events (such as street festivals, holiday events, or inclement- weather situations) to improve their situational awareness. Social media monitoring provides real-time information about developing situations, allows first responders to observe the overall atmosphere of the event, and may provide updates more quickly than official reports or calls from ordinary citizens; therefore, several interviewees found it useful for developing situational awareness more rapidly and efficiently than other methods such as calls from the public or patrols. According to our interviews, social media helps identify warnings or possible action-times, depending on the category. Safety/health issues can trigger immediate actions if reported on social media, while law enforcement concerns prompt further monitoring or corroboration in situations involving crimes. Large events and natural disasters require first responders to deal with rapidly changing situations and huge amounts of data. Social media, if filtered based on first responder needs, can provide information that might not otherwise reach officials in a timely manner. Some of our respondents reported that they monitor specific threats or concerns based on previous events; others monitor social media for a general overview of situations, good or bad, that might require a response from their agency or a referral of information to other related agencies (e.g. health and safety, law enforcement, fire department). B. Information Handling Process In this section of the survey, we discussed interviewees' typical information-gathering processes, including all social media monitoring as well as information from non-social media sources, and the ways those processes and information sources interacted. Some replies incorporated SMART monitoring as part of analyzing their overall processes. 1) Informational Channels: Social Media and Other Sources Only three of our interviewees used software platforms other than SMART specifically intended to monitor social media. Tweetdeck [16] and SocialBearing [17] were used for monitoring and Hootsuite [18] was used for posting to multiple departmental social media accounts simultaneously. Some interviewees searched social media platforms, e.g. Facebook or Twitter, directly; others relied on newsfeeds, posts or messages directly contacting responders, call stations set up to direct information to the relevant command centers, and official information sources such as the National Weather Service, utility company reports, and notices from other agencies. 2) Corroboration or Conflict between Social Media and Other Sources assumptions Interviewees reported that official news sources often corroborated information gleaned from social media (such as street closures or power outages reported on social media first). Similarly, event proceedings/mood based on official sources were often confirmed on social media that reflected people's real-time experience (for example, at an event with low attendance due to high heat and humidity, "weather" was one of the largest categories mentioned in social media posts). overall about the Five interviewees reported that information gleaned from social media was often more up-to-date than the official sources, which sometimes had a lag in reporting. For example, social media posts might indicate blocked roads before calls to emergency operation centers or before first responders explored the area; or traffic issues that might have gone unreported are often mentioned on social media. However, according to interviewees, social media posts tend to be much less specific than the official news sources and usually require follow-up or corroboration in order to be useful for direct action. Some interviewees had concerns about the veracity of social media implemented external processes to cross-check against officially-derived information, information; and therefore, such as sending personnel directly to the scene or contacting related agencies for status updates. C. Actions In this section, we review interviewees' actions taken based on the information-gathering processes detailed above. Some specific examples given in the surveys have been genericized to protect interviewees' privacy based on IRB guidelines. 1) Criteria for Dispatch/Action Almost all of the interviewees stated that they did not take direct action based solely on social media, either for policy reasons (law enforcement agencies cannot obtain warrants based on social media posts) or due to concerns about the accuracy of information gained from social media. However, several interviewees had implemented verification processes including sending personnel to monitor affected areas, where they could also respond to developing situations if necessary. Social media monitoring was generally used to either prompt further investigation or confirm/update information from other sources. In some cases, posts of concern were passed on to relevant law enforcement or safety departments. For example, during severe weather events, social media posts about power outages were passed on to the power utility company (after appropriate verification) so that the issue could be remedied quickly. 2) Successful Action based on Information Gleaned from Social Media While most interviewees did not take direct action based solely on social media, some posts did serve as prompts for further investigation. During severe weather events, social media posts might indicate downed trees or cars stranded in flooded areas; emergency operation centers could send out emergency services to verify and assist, or contact the relevant agency. Urgent medical issues (e.g. heatstrokes) have also been flagged through social media, allowing first responders to send medical personnel to the area to assist. 3) Erroneous Action based on Information Gleaned from Social Media None of our interviewees reported any erroneous action based on social media information. Many of them cited the policy limits on acting based on social media as the reason, meaning that only information whose veracity was confirmed was used for direct action. D. General Reflections 1) Major Advantages and Challenges of Monitoring Social Media When asked about advantages, many interviewees mentioned that social media provides a fuller picture of the situation, a sense of how the community is feeling and responding, as well as immediate concerns that might take time to be reported through official channels. Some interviewees also found it useful for investigation after the fact, and for extra information that could supplement official reports or records. When it came to challenges, several interviewees commented on the sheer amount of information coming through social media, and the need for specially-trained responders who knew how to use the platforms, social media mining tools, and specific flags to look for on social media. Lack of manpower was consistently cited as an issue, as monitoring social media requires having at least one person dedicated to the job and many agencies are already stretched thin. The possibility of missing things in the flood of data is a big concern, as is the fact that tools and social media platforms often limit the amount of information that can be accessed in the first place. Hoax/fake posts and certain types of slang can cause errors in filtering; for example, tools might flag the phrase "shoot on over for a free pizza" because of the word "shoot", while a post using an innocuous slang term for an illegal drug might be bypassed entirely due to not including any of the predetermined keywords. 2) Missing/Desired Information Several interviewees mentioned Facebook specifically as a site their agencies are interested in monitoring, since it is a widely used social media platform for some regions and population brackets. Due to Facebook's terms of service, however, monitoring programs cannot access Facebook posts, and legal or departmental restrictions may prevent agencies from directly monitoring the platform. Prohibitions in social media providers' terms of service often limit official/government access and monitoring, and the majority of social media platforms have privacy options that can prevent outside agencies from viewing specific posts or profiles. This allows privacy for citizens but may prevent first responders from quickly reacting to a situation due to not having enough timely information. Geographic information is not always reliable or particularly specific; for example, individuals might post on social media about a power outage caused by a storm, but the posts will not provide disaster response teams with information about how many houses are affected or what precisely caused the outage (a downed power line, an issue at the power plant, an issue at the house itself, etc.). Several interviewees mentioned a desire for better location information, which again, is a challenge due to privacy concerns. Many of our interviewees also mentioned the ability to monitor live street traffic as an interest and mentioned that the ability to link systems together so that tools monitoring Twitter, Facebook, live traffic, official weather alerts, etc. could all be viewed in one central location would be very useful. 3) Policy and Regulation Concerns the in which Many of our interviewees mentioned the lack of legal and policy clarity and inconsistency around first responder access to/use of social media posts as a concern. Interviewees described situations legal and policy compliance permissibility of accessing and acting on certain information is unclear, which may result in accessing information they legally should not, or missing information that they could have used. Some interviewees also mentioned a need for individual departmental policies to supplement the legal and political contexts. Several interviewees also raised ethical concerns surrounding privacy; especially in a law enforcement context, several interviewees had concerns about maintaining citizen privacy while still accessing necessary information. V. SURVEY RESULTS: IMPACT OF SMART ON USCG AND FIRST RESPONDER OPERATIONS In this section, we focus on the usefulness of SMART for situational awareness. At the time of this writing, SMART supports 376 users who may sporadically use the system. Throughout SMART's development and lifecycle, users have consistently attested to its value and effectiveness and continued to utilize it. In this section, we provide an overview of when and how users have applied SMART, and their overall opinion of the system. A. Events with SMART Use SMART has been used in many events, but most notably the following: Presidential Inauguration 2017, State of the Union 2017/2018/2019, Baltimore Fleet Week 2016, Coast Guard maritime events, Cincinnati's annual Riverfest, Thunder over Louisville's airshows and fireworks, 4th of July celebrations and other holiday events, Inclement weather (e.g. hurricane) events and multiple college football games. B. SMART Utilization based on a broader situation with information obtained or verified using other sources. SMART sometimes prompted responders to look further into a specific situation, but more often, users relied on it to expedite decision-making by improving awareness of the overall situation. Some first responders passed on tweets of concern to relevant agencies (for example, if SMART flagged a tweet about flash floods during a severe weather event, first responders might contact roadwork companies, or have a PR team send out safety messages to warn people in the area not to drive). 3) Successful Use of SMART Several users reported that SMART expedited their decision-making by up to 15-20 minutes during large events and severe weather, since they were able to use it to monitor developing situations with real-time data. In several cases, credible concerns identified through SMART prompted further research that might not have been performed in the absence of social media information. 4) Erroneous Use of SMART All of our interviewees stated that no (resulting) erroneous 1) SMART's Role in Event Monitoring action had been taken based on SMART. Interviewees who had used SMART incorporated it into their event command centers or emergency operations centers. Dedicated staff were tasked with situational awareness and coordination, sometimes monitoring SMART only, and sometimes monitoring other social media or newsfeeds. For large events, some recurring/long-term users used advanced monitoring leading up to the event, with 6-hour summaries and email alerts. Several newly introduced users also expressed interest in this feature, though they had not had the opportunity to use it as of the time we conducted our interviews. Most of the interviewees focused on smaller-scale monitoring (local events or games) and designated an individual to monitor and report anything of concern identified through the use of SMART. 2) SMART's Relationship to Other Tools None of the surveyed users directly linked SMART to other tools. Some of them monitored it alongside other tools (for example, the national weather service, and a departmental Facebook account) but did not connect them in any way aside from comparing instance of overlapping information. the occasional C. Process and Actions 1) SMART and Other Information Sources Most users reported no significant overlap between SMART and other information sources. In a few cases, it corroborated information (developing bad weather and medical issues) that had been received through phone calls or news feeds. Sometimes SMART provided complimentary, general information, and official reports and phone calls more specific information. No users reported instances where information gathered from SMART conflicted with other information sources. 2) Actions Initiated by SMART Most users did not take specific actions solely based on though SMART information gleaned consistently provided further investigation when an incident did arise. Actions were typically from SMART, information timely for D. General Reflections 1) Pros and Cons of Using SMART users the listed Several customized search keywords/indicators and the real-time analytics as major benefits of using SMART. First responders covering areas where Twitter (and Instagram linked to Twitter) are among major social media platforms mentioned that SMART's focus on these social media platforms made it more useful than other monitoring tools. Long-term users of SMART (who have used the system since as early as 2012) mentioned the consistent Purdue University research and development support and the ability to have one or two people monitoring it as benefits, as well as the ability to develop a consistent pattern for using the system and incorporating it into their event planning. Several interviewees mentioned that the system does require training and is not particularly intuitive. However, all but three repeat users of the system did not feel the need for re-training and all users were able to quickly re-familiarize themselves with the system again even if it had been months since the last time they had used SMART. The SMART filters primarily use user-specified keywords to narrow down the stream of tweets to the topical relevant ones. Keywords, however, may cause some contextual issues; slang, such as a tweet using "bomb" to mean "awesome", may cause false positives, or synonyms such as "slaughter" rather than "kill", not be caught by the classifiers. Based on user feedback, while these interviews were in progress, we extended SMART with interactive machine learning-based classifiers to leverage context beyond just the syntax of the words and include the meaning (semantic) of the keywords in the filtering process, which we believe will significantly improve this issue in future uses of the system [6]. Two interviewees mentioned location accuracy issues with tweet geotagging, as a result of approximate geotagging done by the user (e.g., when a user geotags their tweets with the State or City they are in, instead of geographic coordinates or postal address). The lack of Facebook monitoring (which is prohibited by Facebook terms of service and lack of public API, and not a SMART limitation) was also referenced by several users, who saw value in harvesting Facebook posts for situational awareness. Furthermore, users had concerns about the small percentage of tweets that are publicly available through the Twitter Streaming API (which SMART collects) compared to the entire Twitter firehose, as well as the minority of tweets being geotagged (usually between 1% and 2% of the entire tweets, up to 10% during crisis times when people intentionally turn on their geotags). 2) SMART Effectiveness All interviewees described SMART as effective for their goals, with features that were not actively used or required to execute the department's objectives being described as adequate (e.g., LDA topic modeling view and geofencing). Several mentioned that further training on some of the features such as the LDA topic modeling view and heatmap would likely enhance their use of the system, since most used it primarily for monitoring keywords and/or tweets in a specific geographic area rather than manually reviewing each individual tweet. VI. FEEDBACK ON SMART FEATURES We asked our interviewees about their perception of the usefulness of SMART's individual features, as well as their frequency of use of each tool. A. Features Ranked by Frequency of Use tool, but only one used lens. Two users mentioned the classifiers and content geofencing it regularly. Two interviewees mentioned the LDA topic modeling view (i.e., word cloud) as a tool they did not use frequently (since the map already provides that functionality). B. Features Ranked by Usefulness Fig. 3. SMART features ranked by the number of users who found them useful. Again, feature was the most useful the keyword search/classifier, though one interviewee described it as "fairly effective" and would have considered it more useful if the graphical user interface were more intuitive. Three interviewees also found the content lens extremely useful. While no interviewee mentioned the cluster lens in response to the frequency of use, two mentioned it as useful when they did access it. The temporal view and region selection had similar results to the frequency of use, with some interviewees liking them and others finding them irrelevant or inaccurate. Only one interviewee mentioned the LDA topic modeling view (word cloud) and ranked it very low on the scale of usefulness. Fig. 2. SMART features ranked by the number of interviewees who frequently used them. We used the feature names (e.g. content lens, heatmap) to ask our interviewees about their frequency of use and ranking. The interviewees who took the survey as part of a first-time demonstration were not able to provide a ranking due to having very little experience with the tool; and of those who did, many were unfamiliar with the official names for each tool. However, they were usually able to describe each feature even without knowing the official names, though in some cases we provided descriptions to resolve the ambiguity. Of the nine users who provided rankings, seven named the classifiers as a tool they used frequently, and five listed the content lens as well. The tweet tooltip was mentioned by only two users, but both described it as the tool they relied on most with other tools being secondary or less-frequently used. Only one user mentioned the temporal view (time bar graph), but used it as often as the VII. CONCLUSIONS In this paper, we provided an in-depth summary and discussion of key points from our interviews with several first responders tasked with situational awareness. Specifically, based on survey feedback, social media tools such as SMART play an important role in providing information that can be used to better assess and monitor events in a timely manner that might not be possible with major news sources. Interviewees found that social media often provided information up to 15-20 minutes before major news sources. Additionally, although social media was not used for direct action without corroboration, it did prompt further investigation for faster response. Feedback was positive for SMART, with interviewees remarking on its effective visualizations in locating important information. Further, SMART has been successfully used in several national events, such as the 2017 Presidential Inauguration and hurricane events to more effectively facilitate situational awareness. Interviewees, however, noted several limitations with SMART and social media in general. In particular, the majority of social media data is either not geotagged, making it difficult to determine the context of the message, or is unavailable due to privacy settings or lack of clarity on policies. Further, due to the vast amount of social media data and textual limitations such as misspellings or slang, gleaning relevant information is difficult. focus on (1) improving Based on collective survey feedback, we recommend social media analytics systems the identification of relevant social media messages, (2) increasing the amount of streaming data by utilizing well-known social media APIs such as Twitter and computationally enriching the information without explicit geo-information, and (3) integrating information such as traffic data, weather alerts, and news updates in order to enhance situational awareness and provide increased context for end users. For successful transition and independent utilization of social media analytics toolkits, the following additional steps would be beneficial: (1) Keeping the user manuals up to date through centralized hosting of online material (reflecting the latest version); (2) creating short training videos for introduction of each feature and for several use case scenarios; and (3) creating Standard Operating Procedures (SOP) material and documentation of the process to get approval for use of toolkits for different situations. ACKNOWLEDGMENTS This material is based upon work funded by the U.S. Department of Homeland Security (DHS) under Award No. 2009-ST-061-CI0003, DHS Cooperative Agreement No. 2014- ST-061-ML0001, and DHS Science and Technology Directorate Award No. 70RSAT18CB0000004. REFERENCES [1] M. Martínez-Rojas, M. del C. Pardo-Ferreira, and J. C. Rubio-Romero, "Twitter as a tool for the management and analysis of emergency situations: A systematic literature review," International Journal of Information Management, vol. 43, pp. 196 -- 208, 2018. [2] S. Vieweg, A. L. Hughes, K. Starbird, and L. Palen, "Microblogging during two natural hazards events: What twitter may contribute to situational awareness," In Proceedings of Conference on Human Factors in Computing Systems, vol. 2, pp. 1079 -- 1088, 2010. [3] M. A. Cameron, R. Power, B. Robinson, and J. Yin, "Emergency Situation Awareness from Twitter for crisis management," In Proceedings of WWW, pp. 695 -- 698, 2012. [4] N. Pogrebnyakov and E. Maldonado, "Didn't roger that: Social media message complexity and situational awareness of emergency responders," International Journal of Information Management, vol. 40, pp. 166 -- 174, 2018. [5] P. Panagiotopoulos, J. Barnett, A. Z. Bigdeli, and S. Sams, "Social Media in Emergency Management: Twitter as a Tool for communicating risks to the public," Technological Forecasting and Social Change, vol. 111, pp. 86-96, 2016. [6] L. S. Snyder, Y.-S. Lin, M. Karimzadeh, D. Goldwasser, and D. S. Ebert, "Interactive Learning for Identifying Relevant Tweets to Support Real-time Situational Awareness", IEEE Transactions on Visualization Computer Graphics, 2019. [7] H. Bosch, D. Thom, F. Heimerl, E. Püttmann, S. Koch, R. Krüger, M. Wörner, and T. Ertl, "ScatterBlogs2: Real-time monitoring of microblog messages through user-guided filtering," IEEE Transactions on Visualization and Computer Graphics, vol. 19, no. 12, pp. 2022 -- 2031, 2013. [8] S. Pezanowski, A. M. MacEachren, A. Savelyev, and A. C. Robinson, "SensePlace3: a geovisual framework to analyze place -- time -- attribute information in social media," Cartography and Geographic Information Science, vol. 45, no. 5, pp. 420 -- 437, 2018. [9] J. Zhang, J. Chae, C. Surakitbanharn, and D. S. Ebert, "SMART: Social Media Analytics and Reporting Toolkit", In The IEEE Workshop on Visualization in Practice 2017, pp.1-5, 2007. [10] J. Chae, D. Thom, H. Bosch, Y. Jang, R. Maciejewski, D. S. Ebert, and T. Ertl, "Spatiotemporal social media analytics for abnormal event detection and examination using seasonal-trend decomposition," In IEEE VAST, 143 -- 152, 2012. [11] Z. Zhou, X. Zhang, X. Zhou, and Y. Liu, "Visual Abstraction of Large- Scale Geo-Tagged Social Media Data," IEEE Access, vol. 7, pp. 114851 -- 114861, 2019. [12] M. Karimzadeh, S. Pezanowski, A. M. MacEachren, and J. O. Wallgrün, "GeoTxt: A scalable geoparsing system for unstructured text geolocation," Transactions on GIS, vol. 23, no. 1, pp. 118 -- 136, 2019. [13] S. E. Middleton, L. Middleton, and S. Modafferi, "Real-time crisis mapping of natural disasters using social media," IEEE Intelligent Systems, vol. 29, no. 2, pp. 9 -- 17, 2014. [14] H. N. Alshareef and D. Grigoras, "Using social media and the mobile cloud to enhance emergency and risk management," In Proceedings of the IEEE 16th International Symposium on Parallel and Distributed Computing, pp. 92 -- 99, 2016. [15] J. Yin, A. Lampert, M. Cameron, B. Robinson, and R. Power, "Using social media to enhance emergency situation awareness," IEEE Intelligent Systems, vol. 27, no. 6, pp. 52 -- 59, 2012. [16] "Tweetdeck," https://tweetdeck.twitter.com/. [17] "SocialBearing," https://socialbearing.com/. [18] "Hootsuite," https://hootsuite.com/. APPENDIX A. SMART Features In this appendix, we review SMART's features that first responders were surveyed on (Section VI of the main paper). 1) Keyword-based Filtering a) Keyword Search 2) Content Lens SMART users can enter multiple keywords of interest into the keyword filter box. Social media posts containing one or more user-input keywords are filtered and displayed in various views. b) Classifiers The content lens feature allows users to place lenses with word-cloud visualization over specific geographical locations. The area immediately surrounding the lens will be populated with the most frequently used keywords within the lens location, using integrated topic modeling for the geographic area under the lens. 3) Cluster Lens SMART Classifiers allow users to observe, search, and classify tweets by user-defined keywords (using inclusion and exclusion of such keywords). When a user creates a classifier, they are required to supply a name (e.g., "Security") and provide a set of keywords. Enabling a classifier will identify and display messages containing one or more of the classifier's keywords. The user can also union and intersect classifiers for further analysis. The cluster lens filters message clusters within the lens and visualizes the keywords related to the classifiers in a radial layout. The user can move the underlying map to investigate regions of interest while the position of the lens remains fixed on the screen. 4) Tweet Tooltip b) Region Selection The tweet tooltip provides users with the ability to view a message's content directly from the map by clicking on the message's symbol. 5) Heatmap SMART's region selection feature is provided on system start-up. Before loading the main system interface, users can select a specific region on the map to stream messages from during system use. 7) Temporal View (Time Bar Graph) The temporal view displays the number of tweets in different classifiers over time. It is aggregated at the minute level and reflects the evolution over the last two hours. 8) LDA Topic Modeling View SMART's heatmap provides an overview of the spatial distribution of the tweets. 6) Spatial Filtering a) Geo-fencing The geo-fencing tool allows users to define a geographical bound and remove messages outside of the bound from the visualizations. The LDA topic modeling view uses Latent Dirichlet Allocation to extract trending topics and associated keywords among the messages. Users can click on a colored keyword to locate messages containing it.
1805.10244	1	1805	2018-05-25T16:49:56	Detecting Influence Campaigns in Social Networks Using the Ising Model	[ "cs.SI", "physics.soc-ph", "stat.AP" ]	We consider the problem of identifying coordinated influence campaigns conducted by automated agents or bots in a social network. We study several different Twitter datasets which contain such campaigns and find that the bots exhibit heterophily - they interact more with humans than with each other. We use this observation to develop a probability model for the network structure and bot labels based on the Ising model from statistical physics. We present a method to find the maximum likelihood assignment of bot labels by solving a minimum cut problem. Our algorithm allows for the simultaneous detection of multiple bots that are potentially engaging in a coordinated influence campaign, in contrast to other methods that identify bots one at a time. We find that our algorithm is able to more accurately find bots than existing methods when compared to a human labeled ground truth. We also look at the content posted by the bots we identify and find that they seem to have a coordinated agenda.	cs.SI	cs	Detecting Influence Campaigns in Social Networks Using the Ising Model Operations Research Center, Massachusetts Institute of Technology, Cambridge, MA 02139, [email protected] Nicolas Guenon des Mesnards Sloan School of Management, Massachusetts Institute of Technology, Cambridge, MA 02139, [email protected] Tauhid Zaman We consider the problem of identifying coordinated influence campaigns conducted by automated agents or bots in a social network. We study several different Twitter datasets which contain such campaigns and find that the bots exhibit heterophily - they interact more with humans than with each other. We use this observation to develop a probability model for the network structure and bot labels based on the Ising model from statistical physics. We present a method to find the maximum likelihood assignment of bot labels by solving a minimum cut problem. Our algorithm allows for the simultaneous detection of multiple bots that are potentially engaging in a coordinated influence campaign, in contrast to other methods that identify bots one at a time. We find that our algorithm is able to more accurately find bots than existing methods when compared to a human labeled ground truth. We also look at the content posted by the bots we identify and find that they seem to have a coordinated agenda. Subject classifications : Social networks, Influence campaigns, Ising model. 1. Introduction Social networks face the challenge posed by automated bots which create spam in the network and result in a degraded user experience. However, recently these bots have become a serious threat to democracies. There have been multiple reports alleging that foreign actors attempted to penetrate U.S. social networks in order to manipulate elections (Parlapiano and Lee 2018, Shane 2017, Guilbeault and Woolley 2016, Byrnes 2016, Ferrara 2017). The perpetrators used bots to share politically polarizing content, much of it fake news, in order to amplify it and extend its reach, or directly interacted with humans to promote their agenda. While no one knows exactly how many 1 2 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) people were impacted by these influence campaigns, it has still become a concern for the U.S. and many other governments (Fandos and Shane 2017, Price 2018). Social network counter-measures are needed to combat these coordinated influence campaigns. Conventional methods of bot detection may not be sufficient because they can be fooled by modifying certain elements of bot behavior, such as the frequency of posting or sharing content. However, because many of these bots are coordinated, they may exhibit joint behaviors which are difficult to mask and which allow for more accurate and robust detection. These behaviors may not be observable by looking at accounts in isolation. Therefore, conventional algorithms which focus on individual detection may not find these bots. What is needed is an algorithm that can simultaneously detect multiple bots. 1.1. Our Contributions We provide evidence that modern bots in the social network Twitter coordinate their attacks. They do not create original content, but rather amplify certain human users by disproportionately retweeting them. We find that the bots exhibit heterophily, which means they interact with humans much more frequently than they interact with each other (Rogers 2010). This is the opposite of the more common homophily phenomenon, where similar individuals interact with each other. We use heterophily to design a new algorithm for bot detection. We utilize the Ising model from statistical physics to model the network structure and bot labels. We show that the maximum likelihood estimation of the bot labels in this Ising model reduces to solving a minimum cut problem, which allows for efficient estimation. Our algorithm scales to large networks and is much faster than modern bot detection algorithms (see Table 4). A natural application of our algorithm is understanding the agenda of the bots' influence cam- paigns. We do this by analyzing the content of the detected bots. We find that the bots seem to promote certain geo-political issues. We emphasize that our algorithm is language agnostic and relies only on the interaction net- work. Therefore, it can be used on social networks in many different countries without having prior Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 3 knowledge of the language, culture, or individuals involved. Also, our algorithm is not restricted to Twitter, and can be used on any online platform where individuals can interact with one another (Facebook, Reddit, Instagram, etc.). 1.2. Previous Work Originating from spam detection (Sahami et al. 1998), a significant subset of bot detection algorithms are based on recognizing patterns of automation (Heymann et al. 2007, Gyöngyi et al. 2004, Gyöngyi and Garcia-Molina 2005, Ntoulas et al. 2006, Castillo et al. 2007). It was observed by Malmgren et al. (2009) and Raghavan et al. (2014) that human beings present specific circadian patterns which can further be used for strategic timing or promotion campaigns (Li et al. 2014, Modaresi et al.), or, in our case, to discriminate humans from bots (Zhang and Paxson 2011, Chu et al. 2012, Ghosh et al. 2011). Posting at random times makes bots easily detectable, as shown in Gianvecchio et al. (2011), but bots that retweet tweets of actual humans will replicate human rhythms with a lag. Modern bots are no longer unintelligent spam accounts, and have learned how to mimic humans. A meticulous study of the rise of social bots is presented in Ferrara et al. (2016). Social bots are designed to interact with other users (Hwang et al. 2012, Messias et al. 2013, Boshmaf et al. 2013), and post human-like content (Freitas et al. 2015). Not only do bots look more human, but turn out to be half-human: Chu et al. (2012) mentions the concept of cyborgs, where a real person manages dozens of otherwise automated accounts. Such hybrid accounts1 make the detection task extremely challenging (Zangerle and Specht 2014, Syeed and Frier 2017). Bot detectors also became more sophisticated, from the earliest instances (Yardi et al. 2009) to the current state of the art (Davis et al. 2016) currently used in most applications today (Ferrara 2017, Mønsted et al. 2017, Vosoughi et al. 2018, Badawy et al. 2018). In Ferrara et al. (2016) the authors present a taxonomy of bot detectors, from crowd-sourcing (Wang et al. 2012, Stein et al. 2011, Elovici et al. 2014, Twitter b,a) and honeypot traps (Lee et al. 2011, Paradise et al. 2017), to 1 See the example @T eamT rumpRussia, https://twitter.com/teamtrumprussia. 4 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) user feature oriented classifiers (Davis et al. 2016, Chu et al. 2012, Benevenuto et al. 2010, Wang 2010, Egele et al. 2013, Viswanath et al. 2014, Thomas et al. 2011). Both approaches treat accounts individually, but do not detect coordinated attacks. Extant work exists for coordinated attacks, a few instances of which are CopyCatch for Facebook 'liked' pages (Beutel et al. 2013), Twitter memes (Ratkiewicz et al. 2011), and more generally Sybil detection in online communities (Benevenuto et al. 2009, Aggarwal 2014, Cao et al. 2014, Yang et al. 2014, Ghosh et al. 2012, Tran et al. 2009, Yu et al. 2008, Danezis and Mittal 2009, Yu et al. 2006, Wang et al. 2013, Alvisi et al. 2013, Cao et al. 2012). Our algorithm is strongly inspired from work on image segmentation from Zabih and Kolmogorov (2004), which was also applied to social networks in Marks and Zaman (2017) to detect user location. To the best of our knowledge, no model for bot-human heterophilic interactions on social networks exist, nor an efficient algorithm for joint bot detection based on such a model. 2. Data We collected Twitter data from six different controversial events that occurred in a variety of nations (US, France, Hungary), over various time periods (2015 to 2018), and for different durations. Some of the events were elections in the US and Hungary. Others were for politically motivated conspiracy theories or scandals, such as Pizzagate and Macron Leaks. Finally, there were activist groups such as Black Lives Matter (BLM). Details of the data are shown in Table 1. The data was collected using Twitter's REST and STREAM APIs. For the Pizzagate, BLM 2015 and Hungary Election datasets, we collected tweets containing relevant keywords. The US election dataset is from Littman et al. (2016). The Macron Leaks and BLM 2016 were collected by Summers (2017b,a). In addition to the basic content information, for each tweet we also collected the user profile. 2.1. Ground Truth Construction. To obtain a ground truth for bot identities, we manually labeled approximately 300 accounts per event. The accounts were randomly selected and only required to have a minimum activity level. For each account, the labeler was given three options: human, bot, or no idea. The decision is guided Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 5 Table 1 Twitter dataset descriptions. Descriptive statistics Dataset Description Time-period Number of tweets/users Pizzagate Pizzagate conspiracy Nov-Dec 2016 1,025,911 / 176,822 BLM 2015 Black Lives Matter Jan-Dec 2015 477,344 / 242,164 US election First presidential debate Sep-Oct 2016 2,435,886 / 995,918 Macron Leaks French election scandal May 2017 570,295 / 150,848 Hungary Parliamentary election Apr 2018 504,170 / 198,433 BLM 2016 Black Lives Matter Sep 2016 1,274,596 / 545,937 by the activity patterns (number of retweets versus original tweets, volume, etc.), content of tweets (level of creativity, replies, topic diversity), and other profile features (pictures, followers/friends ratio, etc.). Some events occurred two or three years ago, but the labeling is based on the most recent 2018 activity. Therefore, labelers were also given quick statistics about the account's activity during the event, but were explicitly told not to weigh it too much into their final decision. Statistics for the number of accounts labeled as bots on each event are found in Table 2. We found that approximately 10% of the accounts were labeled bots across the different datasets. Table 2 Statistics for hand labeled accounts for each dataset. Dataset Bot labels / Total number of accounts Pizzagate BLM2015 US Election Macron Leaks Hungary BLM2016 23/304 21/262 30/300 19/256 24/300 30/285 2.2. Heterophily. We study the collective behavior of the bots. One such behavior concerns who the bots interact with or retweet. Figure 1 shows different networks of retweeting interactions in the Pizzagate dataset by accounts labeled by humans (see Table 2). We place the bots on the peripheral ring, humans in 6 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) the central ring, and draw directed edges for the retweet interactions. Each network corresponds to different node types on the edges. We see qualitatively that bots prefer to retweet humans than other bots, and that humans prefer to retweet other humans instead of bots. This phenomenon where members of a group do not interact with each other, but do interact with members of different groups is known as heterophily. Figure 1 Retweet interaction networks for Pizzagate accounts labeled by humans. Each network consists of edges between a certain pair of account types (bots and humans). To obtain a more quantitative measure of heterophily, we broaden the set of ground truth labeled accounts by using a popular bot detection algorithm from Davis et al. (2016). This algorithm provides a probability of being a bot. We chose 0.5 as our threshold for being a bot. Then for each account, we calculated the total number of retweets it gives to humans/bots, normalized by the number of unique humans/bots it retweets. This measures the average number of retweets per human and retweets per bot for each account. We refer to this quantity as the retweet rate. We then look at the distribution of the retweet rates conditioned on the nodes on each end of the retweet edges: human to human, human to bot, bot to human, and bot to bot. In Table 3 we show the mean retweet rates for the different retweet types. As can be seen, the bot to human mean is much higher than the bot to bot mean. Also, the human to human mean is higher than the human Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 7 to bot mean. To further quantify the difference, we perform a Kolmogorov-Smirnov (KS) test on each pair of retweet rate distributions. The p-value for the tests on each dataset are also shown in Table 3. We find that there is a statistically significant difference between the distributions. This supports the hypothesis that there is heterophily for the bots. It also suggests that the humans exhibit homophily, meaning they are more likely to interact with each other than with bots. We will use these properties to design our bot detection algorithm. Table 3 Retweet rate and p-value of KS test for different rewteet networks. B stands for bot and H stands for human. The zeros for the p-value mean that the p-value was less than 10−6. Bots' Mean Humans' Mean KS-test p-value Retweet Rate Retweet Rate Dataset B → H B → B H → H H → B B →B vs. B →H H→H vs. H→B B→H vs. H→H H→B vs. B→B US Pizzagate 0.99 US BLM 2015 0.88 US election Macron Leaks Hungary 0.99 1.26 0.73 US BLM 2016 0.4 0.18 0.17 0.66 0.08 0.25 0.23 0.98 0.84 0.95 1.13 0.75 0.45 0.12 0.05 0.35 0.05 0.21 0.15 0 0 0 0 0 0 0 0 0 0 0 0 0 0.013 0 0 0 0 0 0 0 0 0 0 3. Joint Detection Using the Ising Model In this Section, we present our joint classification model. We assume we have an interaction graph G = (V, E) of accounts with vertex set V (the accounts) and edge set E (the interactions). For each node ui ∈ V we observe features xi and for each pair of users ui, uj ∈ V we observe interaction features zij. The interaction features could be the number of retweets, out degree, and in-degree. Each node i in the network has a latent variable ∆i which is one if i is a bot and zero other- wise. We use a factor graph to model the joint distribution of the observed features and latent Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 8 variables. We define functions φ(xi, ∆i) for each i ∈ V and ψ(zij, ∆i, ∆j) for i, j ∈ V . We refer to φ and ψ respectively as the node energy and link energy functions. Then given N accounts in an interaction graph/network with node features X = {xi}N the probability of the observed features and latent variables is i=1 and interaction features Z = {zij}N i,j=1, (cid:81)N i=1 e−φ(xi,∆i)(cid:81) (1) where Z is the partition function. A probability model of this form which involves products of , P(X, Z, ∆) = {i,j: i<j} e−ψ(zij ,∆i,∆j ) Z functions of single nodes and pairs of nodes has the structure of the Ising model from statistical physics (Ising 1925). Inference of the latent variables in general can be difficult because of the presence of the partition function. However, it has been shown that inference problem is much easier if one specifies certain characteristics of the model (Zabih and Kolmogorov 2004, Marks and Zaman 2017), which we do next. 3.1. Model Characteristics. Link Energy. We begin by defining the link energy functions. First there is the case where there is no edge between nodes ui and uj. In this case we assume that we can infer nothing about the latent variables ∆i and ∆j, so we set the link energy to be independent of the latent variables. For simplicity, we assume that ψ(zij, ∆i, ∆j) = 0 when there is no edge between i and j. For pairs of nodes with an edge, we choose a functional form for the link energy. Suppose node u1 has z1 retweets (its out-degree), node u2 receives z2 retweets (its in-degree), and u1 retweets u2 a total of w12 times. If either degree is small, then the retweet edge from u1 to u2 provides little information about node labels because occasional retweets are very common on Twitter. For example, let u1 be a human that retweeted u2 once. If u1 retweeted no one else, the edge (u1, u2) would increase the probability that it is a bot, though it is not. In simpler terms, the only edges that contain information are the ones where u2 happens to be a popular target and u1 a suspiciously active retweeter. Using these insights and following Marks and Zaman (2017) we define the link energy as ψ(z1,2, ∆1, ∆2u1 → u2) = 1 + exp((α1/z1 − 1) + (α2/z2 − 1)) w12γ , Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 9 where γ controls the weight of the link energy relative to the node energy, and α1, α2 represent thresholds for the in and out-degrees. As mentioned in Section 2.2, we require heterophily in our model between the bots, and homophily between the humans. This imposes the following constraint on the link energies when there is an edge from ui to uj (a high energy corresponds to small likelihood): ψ(zij, 1, 0ui → uj) ≤ ψ(zij, 0, 0ui → uj) ≤ ψ(zij, 1, 1ui → uj) ≤ ψ(zij, 0, 1ui → uj). (2) The constraints above simply say that if ui retweets uj, the most likely labeling is ui is a bot and uj is a human (ψ(zij, 1, 0ui → uj)), and the least likely is ui is a human and uj is a bot (ψ(zij, 0, 1ui → uj) ). To parameterize these energies, we introduce the constants < λ2 < λ1 ∈ [0, 1] and set the link energies to ψij = 1 + exp((α1/zi − 1) + (α2/zj − 1)) wijγ ,  ψ(zij, 0, 1ui → uj) = ψij ψ(zij, 1, 1ui → uj) = λ1ψij ψ(zij, 0, 0ui → uj) = λ2ψij ψ(zij, 1, 0ui → uj) = ψij. For pairs of nodes with reciprocate edges ui → uj and uj → ui, the energy of the pair is the aggregate of the two one-way energies. In order to efficiently infer the latent labels, the constants λ1, λ2 and must satisfy a set of constraints that we present in Section 4. Node Energy. We choose a relatively simple form for the node energy. If we have no prior knowl- edge on which accounts are bots, we set φ(xi, 1) = φ(xi, 0) = 0. If we have prior knowledge, perhaps from an expert or another bot detection algorithm, we can use it for the node energy. However, it turns out the simple uninformative energy performs quite well. 4. Inference via Minimum Cut To find the most likely values for the labels given observed values X and Z, we want to maximize equation (1), which is equivalent to solving (cid:88) i (cid:88) i<j ψ(zij, ∆i, ∆j). (3) minimize ∆ φ(xi, ∆i) + 10 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) It has been shown that performing the optimization in equation (3) is equivalent to finding the minimum capacity cut on an appropriately defined graph (Boykov et al. 2001). The expression in equation (3) can be viewed as the energy of the label configuration, so the graph on which the minimum cut is equal to the maximum likelihood label set is referred to as the energy graph. We now show how to construct this graph. Figure 2 illustrates how to map an interaction network/graph into an energy graph. There is a source node s and a sink node t in the energy graph. There are three types of edges in the energy graph. For each node ui in the interaction graph, there is an edge from the source and an edge to the sink: (s, ui) and (ui, t). There are also edges between every pair of nodes with an interaction edge. For each node ui in the interaction graph, every valid s− t cut in the energy graph must either cut the edge (s, ui) or (ui, t). If (s, ui) is cut, then ∆i = 0 and ui is a human. Otherwise, (ui, t) is cut, ∆i = 1 and ui is a bot. This is how a cut in the energy graph maps to a label configuration. By proper choice of the edge weights, a minimum weight cut in the energy graph will correspond to a maximum likelihood label configuration. We now define the edge weights following the results in Boykov et al. (2001). Figure 2 Energy graph representation of the label configuration energy. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 11 For each pair of users ui, uj with an edge from ui to uj in the interaction graph, we add edges (ui, uj) and (uj, ui) in the energy graph with weights: c(ui,uj ) = c(uj ,ui) = 1 2 (ψ(zij, 1, 0) + ψ(zij, 0, 1)− ψ(zij, 0, 0)− ψ(zij, 1, 1)) . (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) For each user ui, the weight of the edge (s, ui) is c(s,ui) = cout (s,ui) + cin (s,ui), where cout (s,ui) = 1 2 φ(xi, 0) + 1 2 and ψ(zij, 0, 0) + 1 4 {j:i→j} {j:i→j} (ψ(zij, 0, 1)− ψ(zij, 1, 0)), cin (s,ui) = φ(xi, 0) + 1 2 1 2 ψ(zij, 0, 0) + 1 4 {j:i←j} For each user ui, the weight of the edge (ui, t) is c(ui,t) = cout {j:i←j} (ui,t) + cin (ui,t), where (ψ(zij, 0, 1)− ψ(zij, 1, 0)). and cout (ui,t) = cin (ui,t) = 1 2 1 2 φ(xi, 1) + φ(xi, 1) + 1 2 1 2 ψ(zij, 1, 1) + ψ(zij, 1, 1) + 1 4 1 4 {j:i→j} {j:i→j} (ψ(zij, 1, 0)− ψ(zij, 0, 1)), (ψ(zij, 1, 0)− ψ(zij, 0, 1)). It can easily be shown that with these edge weights, the weight of an s− t cut in the energy graph {j:i←j} {j:i←j} equals the energy of the corresponding label configuration in the interaction graph. Energy Constraints. In order for the minimum weight cut to be the most likely labels, we need the energy graph edge weights to be positive. Also, from Boykov et al. (2001), it was shown that the link energies had to satisfy the condition ψ(z, 1, 1) + ψ(z, 0, 0) ≤ ψ(z, 1, 0) + ψ(z, 0, 1). This condition made the energy function from equation (3) submodular and also solvable via minimum cut. Finally, we have the heterophily constraint from equation (2). We express these constraints in terms of λ1, λ2 ,and . First, heterophily gives Submodularity gives 0 < < λ2 < λ1 < 1. = λ2 + λ1 − 1 + δ, for some δ ≥ 0. We choose δ = 0 and find that edge weight positivity gives 0 ≤ λ2 − 1 2 (1− ) ⇐⇒ 2 ≤ 3λ2 + λ1. (4) (5) (6) 12 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 5. Results In this section, we present the results of the performance of our bot detection algorithm based on the human labelers' ground truth. In addition to the 300 hand labeled accounts for each event, we also include all the Twitter verified users. Verified users have undergone robust identity checks from Twitter in order to receive such label. This increases the number of ground-truth labels by approximately one to two thousand. We compare our performance to the state of the art bot detection algorithm known as BotOrNot (Davis et al. 2016). In the results presented, (α1, α2) = (100, 100), except for BLM2016, where the volume of interactions is much bigger. There we set (α1, α2) = (100, 1000). For all events we set γ = 1, λ1 = 0.8, λ2 = 0.6 and = 0.4. (λ1, λ2) is picked as the barycenter of the feasible region defined by constraints (4), (6) while is determined by setting δ = 0 in (5) as we want as small as possible. As we consider the case of null node energies φ(xi, ∆i) = 0, the choice of γ does not impact the optimal cut, and only intervenes in computing the marginal probability of belonging to each class. We do not tune this parameter, and fix it to γ = 1, in order for the link energy differences to scale with the number of retweets wij. Figure 3 ROC curves for bot detection algorithms on different Twitter datasets. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 13 Table 4 AUC and runtime for our Ising model based bot detection algorithm and BotOrNot on different Twitter datasets. AUC Runtime Ising Bot Ising Bot Model or Not or Not 0.81 ≈5 min 3.5s×170,000 0.73 ≈10 min 3.5s×240,000 0.82 ≈20 min 3.5s×300,000 3.5s×150,000 0.72 ≈5 min 0.71 ≈5 min 3.5s×200,000 0.84 ≈4 hours 3.5s×545,000 Dataset Pizzagate BLM 2015 US election Model 0.91 0.67 0.83 Macron Leaks 0.84 Hungary BLM 2016 0.83 0.91 To evaluate the algorithms performance, we use the inferred labels to calculate the marginal probability of being a bot for each node. This allows us to calculate receiver operating characteristic (ROC) curves for our algorithm and compare it to BotOrNot. We show the ROC curves in Figure 3. In Table 4 we present the area under the curve (AUC) metric for the algorithms along with their run-time. We note that for BotOrNot, the run-time is the time it takes to query its API, which may include additional delays beyond computation time. Therefore, the runtimes for BotOrNot are an upper bound to the actual computational time. From Figure 3 and Table 4 it can be seen that our algorithm usually performs better than BotOrNot in terms of AUC. The one exception is BLM 2015. We suspect here that the coordinated bot campaigns we are trying to detect were not as prevalent. Our algorithm runs on several hundred thousand accounts in several minutes to a few hours. We find that we have much fewer false positives than BotOrNot. One of the reason is that BotOrNot labels news broadcasting accounts as bots. Examples of this in our datasets include the Twitter accounts for Newsweek, CNN, and the New York Times.2 Because our Ising model algorithm looks for who the accounts interact with and also who interacts with the accounts, we do not mistake 2 As per the documentation on their website, bot scores were updated on May 10, 2018. The later problem seems to have been fixed, since the aforementioned verified news broadcasting accounts are now given low bot scores. However, 14 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) these trusted news sources for bots. BotOrNot uses mainly the account behavior, so the high activity of these accounts make them appear like bots. In addition to the increased accuracy and fast runtime, our algorithm relies on much less infor- mation than BotOrNot (recall that we use uninformative node energies). This shows that there is a great deal of information in the interaction graph which maps out the retweet behavior of the accounts. Bot Agenda. A way to understand the underlying agenda of bot designers is by looking at the content of the tweets posted by the bots. After applying our bot detection algorithm to an event, we look at the hashtags used by the bots, but not by the humans. Figure 4 illustrates the results of such analysis on the Pizzagate and Macron Leaks datasets. In Pizzagate, the bots try to export the controversy outside of the US, by transforming the usual #WakeUpAmerica into # WakeUpBritain, and #WakeUpFrance. Similar intent can be seen in the Macron Leaks analysis, with the use of the #furherMerkel and #stopGlobalists hashtags. In both cases, it looks like bots intend to internationalize an otherwise local controversy in order to reach a broader audience, especially that of western Europe. It appears as if the bots are trying to shift the theme of the conversation to certain types of controversies (instability of Europe, Brexit, Frexit, etc.). 6. Conclusion We have presented an Ising model for the probability of bot labels in a social network. Our model was motivated by the observation of heterophily in bot interactions. We infer the bot labels efficiently by solving a minimum cut problem. We find our joint detection algorithm outperforms current existing algorithms. This may be due to the fact that the bots are coordinated. Individual detection does not allow one to take advantage of the coordinated behavior. We also find that our algorithm is able to discern bot accounts from just active trusted news source accounts, in contrast to other popular regular users that have no bad intentions, but use automation softwares to gain popularity would still be mistaken with bots, although they are in no way involved in any influence campaign. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 15 Figure 4 Word clouds of hashtags used during (left) Pizzagate and (right) Macron Leaks by bots and not by humans. algorithms. This is a result of our incorporation of the joint interactions of all accounts. Analysis of the content of the bots reveals that they seem to have an agenda focused on controversial topics. We feel that for these types of coordinated campaigns, joint detection is much more effective than individual detection. Acknowledgments This research was supported in part by the Office of Naval Research (ONR) and Charles Stark Draper Laboratory, Inc (Draper). The views presented here are those of the author and do not necessarily represent those of ONR, Draper, or MIT. 16 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) References Charu C Aggarwal. Data classification: algorithms and applications. CRC Press, 2014. Lorenzo Alvisi, Allen Clement, Alessandro Epasto, Silvio Lattanzi, and Alessandro Panconesi. Sok: The evolution of sybil defense via social networks. In Security and Privacy (SP), 2013 IEEE Symposium on, pages 382–396. IEEE, 2013. Adam Badawy, Emilio Ferrara, and Kristina Lerman. Analyzing the digital traces of political manipulation: The 2016 russian interference twitter campaign. arXiv preprint arXiv:1802.04291, 2018. Fabrício Benevenuto, Tiago Rodrigues, Virgílio Almeida, Jussara Almeida, and Marcos Gonçalves. Detecting spammers and content promoters in online video social networks. In Proceedings of the 32nd interna- tional ACM SIGIR conference on Research and development in information retrieval, pages 620–627. ACM, 2009. Fabricio Benevenuto, Gabriel Magno, Tiago Rodrigues, and Virgilio Almeida. Detecting spammers on twitter. In Collaboration, electronic messaging, anti-abuse and spam conference (CEAS), volume 6, page 12, 2010. Alex Beutel, Wanhong Xu, Venkatesan Guruswami, Christopher Palow, and Christos Faloutsos. Copycatch: stopping group attacks by spotting lockstep behavior in social networks. In Proceedings of the 22nd international conference on World Wide Web, pages 119–130. ACM, 2013. Yazan Boshmaf, Ildar Muslukhov, Konstantin Beznosov, and Matei Ripeanu. Design and analysis of a social botnet. Computer Networks, 57(2):556–578, 2013. Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222–1239, 2001. Nanette Byrnes. How the bot-y politic influenced this election. Technology Rev., 2016. Qiang Cao, Michael Sirivianos, Xiaowei Yang, and Tiago Pregueiro. Aiding the detection of fake accounts in large scale social online services. In Proceedings of the 9th USENIX conference on Networked Systems Design and Implementation, pages 15–15. USENIX Association, 2012. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 17 Qiang Cao, Xiaowei Yang, Jieqi Yu, and Christopher Palow. Uncovering large groups of active malicious accounts in online social networks. In Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security, pages 477–488. ACM, 2014. Carlos Castillo, Debora Donato, Aristides Gionis, Vanessa Murdock, and Fabrizio Silvestri. Know your neighbors: Web spam detection using the web topology. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval, pages 423–430. ACM, 2007. Zi Chu, Steven Gianvecchio, Haining Wang, and Sushil Jajodia. Detecting automation of twitter accounts: Are you a human, bot, or cyborg? IEEE Transactions on Dependable and Secure Computing, 9(6): 811–824, 2012. George Danezis and Prateek Mittal. Sybilinfer: Detecting sybil nodes using social networks. In NDSS, pages 1–15. San Diego, CA, 2009. Clayton Allen Davis, Onur Varol, Emilio Ferrara, Alessandro Flammini, and Filippo Menczer. Botornot: A system to evaluate social bots. In Proceedings of the 25th International Conference Companion on World Wide Web, pages 273–274. International World Wide Web Conferences Steering Committee, 2016. Manuel Egele, Gianluca Stringhini, Christopher Kruegel, and Giovanni Vigna. Compa: Detecting compro- mised accounts on social networks. In NDSS, 2013. Yuval Elovici, Michael Fire, Amir Herzberg, and Haya Shulman. Ethical considerations when employing fake identities in online social networks for research. Science and engineering ethics, 20(4):1027–1043, 2014. Nocholas Fandos and Scott Shane. Senator Berates Twitter Over 'Inadequate' Inquiry Into Russian Meddling . The New York Times, September 2017. URL https://www.nytimes.com/2017/09/28/us/politics/ twitter-russia-interference-2016-election-investigation.html?mtrref=www.google.com. Emilio Ferrara. Disinformation and social bot operations in the run up to the 2017 french presidential election. 2017. Emilio Ferrara, Onur Varol, Clayton Davis, Filippo Menczer, and Alessandro Flammini. The rise of social bots. Communications of the ACM, 59(7):96–104, 2016. 18 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) Carlos Freitas, Fabricio Benevenuto, Saptarshi Ghosh, and Adriano Veloso. Reverse engineering socialbot infiltration strategies in twitter. In Advances in Social Networks Analysis and Mining (ASONAM), 2015 IEEE/ACM International Conference on, pages 25–32. IEEE, 2015. Rumi Ghosh, Tawan Surachawala, and Kristina Lerman. Entropy-based classification of'retweeting'activity on twitter. arXiv preprint arXiv:1106.0346, 2011. Saptarshi Ghosh, Bimal Viswanath, Farshad Kooti, Naveen Kumar Sharma, Gautam Korlam, Fabricio Ben- evenuto, Niloy Ganguly, and Krishna Phani Gummadi. Understanding and combating link farming in the twitter social network. In Proceedings of the 21st international conference on World Wide Web, pages 61–70. ACM, 2012. Steven Gianvecchio, Mengjun Xie, Zhenyu Wu, and Haining Wang. Humans and bots in internet chat: measurement, analysis, and automated classification. IEEE/ACM Transactions on Networking (TON), 19(5):1557–1571, 2011. Douglas Guilbeault and Samuel Woolley. How twitter bots are shaping the election. The Atlantic, 1, 2016. Zoltan Gyöngyi and Hector Garcia-Molina. Web spam taxonomy. In First international workshop on adver- sarial information retrieval on the web (AIRWeb 2005), 2005. Zoltán Gyöngyi, Hector Garcia-Molina, and Jan Pedersen. Combating web spam with trustrank. In Pro- ceedings of the Thirtieth international conference on Very large data bases-Volume 30, pages 576–587. VLDB Endowment, 2004. Paul Heymann, Georgia Koutrika, and Hector Garcia-Molina. Fighting spam on social web sites: A survey of approaches and future challenges. IEEE Internet Computing, 11(6), 2007. Tim Hwang, Ian Pearce, and Max Nanis. Socialbots: Voices from the fronts. interactions, 19(2):38–45, 2012. Ernst Ising. Beitrag zur theorie des ferromagnetismus. Zeitschrift für Physik, 31(1):253–258, 1925. Kyumin Lee, Brian David Eoff, and James Caverlee. Seven months with the devils: A long-term study of content polluters on twitter. In ICWSM, 2011. Huayi Li, Arjun Mukherjee, Bing Liu, Rachel Kornfield, and Sherry Emery. Detecting campaign promoters on twitter using markov random fields. In Data Mining (ICDM), 2014 IEEE International Conference on, pages 290–299. IEEE, 2014. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 19 Justin Littman, Laura Wrubel, and Daniel Kerchner. 2016 united states presidential election tweet ids, 2016. URL https://doi.org/10.7910/DVN/PDI7IN. R Dean Malmgren, Jake M Hofman, Luis AN Amaral, and Duncan J Watts. Characterizing individual com- munication patterns. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 607–616. ACM, 2009. Christopher Marks and Tauhid Zaman. Building a location-based set of social media users. arXiv preprint arXiv:1711.01481, 2017. Johnnatan Messias, Lucas Schmidt, Ricardo Oliveira, and Fabrício Benevenuto. You followed my bot! transforming robots into influential users in twitter. First Monday, 18(7), 2013. Sina Modaresi, Juan Pablo Vielma, and Tauhid Zaman. Strategic timing of content in online social networks. Bjarke Mønsted, Piotr Sapieżyński, Emilio Ferrara, and Sune Lehmann. Evidence of complex contagion of information in social media: An experiment using twitter bots. PloS one, 12(9):e0184148, 2017. Alexandros Ntoulas, Marc Najork, Mark Manasse, and Dennis Fetterly. Detecting spam web pages through content analysis. In Proceedings of the 15th international conference on World Wide Web, pages 83–92. ACM, 2006. Abigail Paradise, Asaf Shabtai, Rami Puzis, Aviad Elyashar, Yuval Elovici, Mehran Roshandel, and Christoph Peylo. Creation and management of social network honeypots for detecting targeted cyber attacks. IEEE Transactions on Computational Social Systems, 4(3):65–79, 2017. Alicia Parlapiano and C. Lee, Jasmine. The Propaganda Tools Used by Russians to Influence the 2016 Election. The New York Times, February 2018. URL https://www.nytimes.com/interactive/2018/ 02/16/us/politics/russia-propaganda-election-2016.html. Molly Price. Democrats urge facebook and twitter to probe russian bots. CNET, January 2018. URL https: //www.cnet.com/news/facebook-and-twitter-asked-again-to-investigate-russian-bots/. Vasanthan Raghavan, Greg Ver Steeg, Aram Galstyan, and Alexander G Tartakovsky. Modeling temporal activity patterns in dynamic social networks. IEEE Transactions on Computational Social Systems, 1 (1):89–107, 2014. 20 Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) Jacob Ratkiewicz, Michael Conover, Mark R Meiss, Bruno Gonçalves, Alessandro Flammini, and Filippo Menczer. Detecting and tracking political abuse in social media. ICWSM, 11:297–304, 2011. Everett M Rogers. Diffusion of innovations. Simon and Schuster, 2010. Mehran Sahami, Susan Dumais, David Heckerman, and Eric Horvitz. A bayesian approach to filtering junk e-mail. In Learning for Text Categorization: Papers from the 1998 workshop, volume 62, pages 98–105, 1998. S Shane. The fake americans russia created to influence the election. The New York Times, 7, 2017. Tao Stein, Erdong Chen, and Karan Mangla. Facebook immune system. In Proceedings of the 4th Workshop on Social Network Systems, page 8. ACM, 2011. Ed Summers. Black lives matter tweets - dataset, 2017a. URL https://archive.org/details/ blacklivesmatter-tweets-2016.txt. Ed Summers. Macron leaks tweets - dataset, 2017b. URL https://archive.org/details/ MacronleaksTweets. Nafeesa Syeed and Sarah Frier. Pro-Russian Bots Sharpen Online Attacks for 2018 U.S. Vote. Bloomberg, September 2017. URL https://www.bloomberg.com/news/articles/2017-09-01/ russia-linked-bots-hone-online-attack-plans-for-2018-u-s-vote. Kurt Thomas, Chris Grier, Dawn Song, and Vern Paxson. Suspended accounts in retrospect: an analysis of twitter spam. In Proceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference, pages 243–258. ACM, 2011. Dinh Nguyen Tran, Bonan Min, Jinyang Li, and Lakshminarayanan Subramanian. Sybil-resilient online content voting. In NSDI, volume 9, pages 15–28, 2009. Twitter. Twitter unofficial botspot feed, a. URL https://twitter.com/hashtag/BotSpot. Twitter. Twitter official spam report tool, b. URL https://help.twitter.com/en/safety-and-security/ report-spam. Bimal Viswanath, Muhammad Ahmad Bashir, Mark Crovella, Saikat Guha, Krishna P Gummadi, Balachan- der Krishnamurthy, and Alan Mislove. Towards detecting anomalous user behavior in online social networks. In USENIX Security Symposium, pages 223–238, 2014. Guenon des Mesnards and Zaman: Detecting Influence Campaigns in Social Networks Using the Ising Model Article submitted to Operations Research; manuscript no. (Please, provide the mansucript number!) 21 Soroush Vosoughi, Deb Roy, and Sinan Aral. The spread of true and false news online. Science, 359(6380): 1146–1151, 2018. Alex Hai Wang. Detecting spam bots in online social networking sites: a machine learning approach. In IFIP Annual Conference on Data and Applications Security and Privacy, pages 335–342. Springer, 2010. Gang Wang, Manish Mohanlal, Christo Wilson, Xiao Wang, Miriam Metzger, Haitao Zheng, and Ben Y Zhao. Social turing tests: Crowdsourcing sybil detection. arXiv preprint arXiv:1205.3856, 2012. Gang Wang, Tristan Konolige, Christo Wilson, Xiao Wang, Haitao Zheng, and Ben Y Zhao. You are how you click: Clickstream analysis for sybil detection. In USENIX Security Symposium, volume 9, pages 1–008, 2013. Zhi Yang, Christo Wilson, Xiao Wang, Tingting Gao, Ben Y Zhao, and Yafei Dai. Uncovering social network sybils in the wild. ACM Transactions on Knowledge Discovery from Data (TKDD), 8(1):2, 2014. Sarita Yardi, Daniel Romero, Grant Schoenebeck, et al. Detecting spam in a twitter network. First Monday, 15(1), 2009. Haifeng Yu, Michael Kaminsky, Phillip B Gibbons, and Abraham Flaxman. Sybilguard: defending against sybil attacks via social networks. In ACM SIGCOMM Computer Communication Review, volume 36, pages 267–278. ACM, 2006. Haifeng Yu, Phillip B Gibbons, Michael Kaminsky, and Feng Xiao. Sybillimit: A near-optimal social network defense against sybil attacks. In Security and Privacy, 2008. SP 2008. IEEE Symposium on, pages 3–17. IEEE, 2008. Ramin Zabih and Vladimir Kolmogorov. Spatially coherent clustering using graph cuts. In Computer Vision and Pattern Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference on, volume 2, pages II–II. IEEE, 2004. Eva Zangerle and Günther Specht. Sorry, i was hacked: a classification of compromised twitter accounts. In Proceedings of the 29th Annual ACM Symposium on Applied Computing, pages 587–593. ACM, 2014. Chao Michael Zhang and Vern Paxson. Detecting and analyzing automated activity on twitter. In Interna- tional Conference on Passive and Active Network Measurement, pages 102–111. Springer, 2011.
1702.02462	1	1702	2017-02-08T15:15:24	Integrated Information as a Metric for Group Interaction: Analyzing Human and Computer Groups Using a Technique Developed to Measure Consciousness	[ "cs.SI", "q-bio.NC" ]	Researchers in many disciplines have previously used a variety of mathematical techniques for analyzing group interactions. Here we use a new metric for this purpose, called 'integrated information' or 'phi.' Phi was originally developed by neuroscientists as a measure of consciousness in brains, but it captures, in a single mathematical quantity, two properties that are important in many other kinds of groups as well: differentiated information and integration. Here we apply this metric to the activity of three types of groups that involve people and computers. First, we find that 4-person work groups with higher measured phi perform a wide range of tasks more effectively, as measured by their collective intelligence. Next, we find that groups of Wikipedia editors with higher measured phi create higher quality articles. Last, we find that the measured phi of the collection of people and computers communicating on the Internet increased over a recent six-year period. Together, these results suggest that integrated information can be a useful way of characterizing a certain kind of interactional complexity that, at least sometimes, predicts group performance. In this sense, phi can be viewed as a potential metric of effective group collaboration. Since the metric was originally developed as a measure of consciousness, the results also raise intriguing questions about the conditions under which it might be useful to regard groups as having a kind of consciousness.	cs.SI	cs	Integrated Information as a Metric for Group Interaction: Analyzing Human and Computer Groups Using a Technique Developed to Measure Consciousness David Engel1,2,3 and Thomas W. Malone1,2,* 1 Massachusetts Institute of Technology, Center for Collective Intelligence, Cambridge, MA 02142, USA 2 Massachusetts Institute of Technology, Sloan School of Management, Cambridge, MA 02142, USA 3 Now at Google, Inc., Zurich, Switzerland * Corresponding author. Email: [email protected] Abstract. Researchers in many disciplines have previously used a variety of mathematical techniques for analyzing group interactions. Here we use a new metric for this purpose, called "integrated information" or "phi." Phi was originally developed by neuroscientists as a measure of consciousness in brains, but it captures, in a single mathematical quantity, two properties that are important in many other kinds of groups as well: differentiated information and integration. Here we apply this metric to the activity of three types of groups that involve people and computers. First, we find that 4-person work groups with higher measured phi perform a wide range of tasks more effectively, as measured by their collective intelligence. Next, we find that groups of Wikipedia editors with higher measured phi create higher quality articles. Last, we find that the measured phi of the collection of people and computers communicating on the Internet increased over a recent six-year period. Together, these results suggest that integrated information can be a useful way of characterizing a certain kind of interactional complexity that, at least sometimes, predicts group performance. In this sense, phi can be viewed as a potential metric of effective group collaboration. Since the metric was originally developed as a measure of consciousness, the results also raise intriguing questions about the conditions under which it might be useful to regard groups as having a kind of consciousness. Introduction A vast number of phenomena in the world arise out of the interactions of individuals in groups, from the emotional tone of a family [1,2] to the productivity of an economy [3] to the spread of disease in a community [4], and researchers in a variety of disciplines have used many different mathematical tools to analyze these phenomena. For instance, psychologists have used Markov models to analyze the sequences of actions in small groups of people [5–7], economists have 1 used general equilibrium theory to analyze the interactions among buyers and sellers in a market [8], and sociologists have used graph theory to analyze various kinds of social networks [4,9]. In this paper, we examine another mathematical technique that has not previously been used for analyzing group interactions. This technique, based on information theory, is intriguing because it was developed as a physical measure that would correlate with the consciousness of a brain [10–14]. We will see, however, that the metric is general enough to apply to many other kinds of systems, and we focus here on using it to analyze groups of people and computers. What is integrated information? The metric we use is called "integrated information" or "phi" and was proposed by Tononi and colleagues [10–14]. There have been several successively refined versions of phi (summarized in [12]), but all the versions aim to quantify the integrated information in a system. Loosely speaking, this means the amount of information generated by the system as a whole that is more than just the sum of its parts. The phi metric does this by splitting the system into subsystems and then calculating how much information can be explained by looking at the system as a whole but not by looking at the subsystems separately. In other words, for a system to have a high value of phi, it must, first of all, generate a large amount of information. Information can be defined as the reduction of uncertainty produced when one event occurs out of many possible events that might have occurred [15]. Thus a system can produce more information when it can produce more possible events. This, in turn, is possible when it has more different parts that can be in more different combinations of states. In other words, a system needs a certain kind of differentiated complexity in its structure in order to generate a large amount of information. But phi requires more than just information; it also requires the information to be integrated at the level of the system as a whole. A system with many different parts could produce a great deal of information, but if the different parts were completely independent of each other, then the information would not be integrated at all, and the value of phi would be 0. For a system to be integrated, the events in some parts of the system need to depend on events in other parts of the system. And the stronger and more widespread these interdependencies are, the greater the degree of integration. For instance, a single photodiode that senses whether a scene is light or dark does not generate much information because it can only be in two possible states. But even a digital camera with a million photodiodes, which can discriminate among 21,000,000 possible states, would not produce any integrated information because each photodiode is independently responding to a different tiny segment of the scene. Since there are no interdependencies among the different photodiodes, there is no integrated information [13]. Tononi and colleagues argue that these two properties-differentiated information and integration-are both essential to the subjective experience of consciousness. For example, the conscious perception of a red triangle is an integrated subjective experience that is more than the sum of perceiving "a triangle but no red, plus a red patch but no triangle" [12]. The information is integrated in the sense that we cannot consciously perceive the triangle's shape independently 2 from its color, nor can we perceive the left visual hemisphere independently from the right. Said differently, integrated information in conscious experience results from functionally specialized subsystems that interact significantly with each other [16]. Even though there is not yet a general consensus among researchers that phi actually measures consciousness in humans (e.g., [17]), it does capture these important aspects of the experience of consciousness, and the mathematical behavior of phi is also consistent with many empirical observations of human consciousness [10,18–22]. Interestingly, the two properties used to define phi-differentiated information and integration- are similar to properties that are important in many other kinds of systems, too. For example, Adam Smith [23] observed that economic systems are often more productive when (a) division of labor leads different people to specialize in different kinds of work and (b) the "invisible hand" of the market integrates their diverse efforts. Lawrence and Lorsch [24] discussed the importance of differentiation and integration in large, hierarchical human organizations: (a) dividing the organization into specialized subunits and (b) integrating these units to achieve the goals of the overall organization. And in many fields of engineering and other kinds of design, effective problem solving often involves (a) dividing a problem into subparts and (b) integrating solutions for the subparts into a solution for the whole problem [25–27]. In other words, the mathematical concept of integrated information provides a quantitative way of measuring a combination of two properties that are important across a wide range of different types of systems. And whether phi is measuring consciousness or not, it is clearly measuring something that is of potential interest to many different disciplines. A mathematical formulation of integrated information The concept of integrated information, or phi, can be represented mathematically as follows [28]: ∅= 𝒓 𝒌!𝟏 𝑯𝑴𝟎𝒌 𝑴𝟏𝒌 −𝑯𝑿𝟎 𝑿𝟏 , (𝟏) where H(XY) is the entropy of variable X given knowledge of variable Y, X0 and X1 are the states of the whole system at time t0 and t1, respectively, and M0 completely partition the parts of X at these times. For example, H(M0 much of the uncertainty of subsystem k at time t0 cannot be explained by knowledge of the state of the subsystem at time t1. k are subsets of X that k) quantifies how k and M1 k M1 Summing over all subsystems (the first term in equation (1)) gives us the amount of entropy that cannot be explained by the subsystems themselves. The second term in equation (1) quantifies the conditional entropy of the whole system. Thus phi is high if there is a large amount of entropy that cannot be explained by looking at the subsystems separately but that is explained by looking at the system as a whole. The value calculated by equation (1) is the phi as defined by Tononi and colleagues if and only if the partitioning is chosen as the maximum information bipartition (MIB), that is, the 3 decomposition into two parts that are most independent. More thorough descriptions of phi can be found in [11–14,21,29]. Applying the phi metric As we saw above, phi provides a quantitative measure of properties that figure prominently in several theories of group performance, so we first test whether it is, in fact, correlated with performance in two different kinds of groups: (a) small groups of experimental subjects working together on shared laboratory tasks, and (b) groups of Wikipedia contributors improving Wikipedia articles over time. As a further test of the applicability of phi, we also examine whether it detects what we might assume would be the increasingly differentiated and integrated complexity of the Internet over time. We evaluate this by applying the phi metric to data about all the computers (and people) communicating over a specific Internet backbone during a six- year period. In order to apply the theoretical definition of phi, a complete model of the rules governing state transitions in the system is needed, but such models are rarely available for observational data. Therefore, we use two alternative versions of phi suggested by Barrett and Seth [28] that estimate conditional probabilities for state transitions from the actual observed data (see Methods). In each case, in order to use the phi metric, we needed to determine: (a) a characterization of the state of the system at different times, and (b) a time delay with respect to which phi will be calculated [10]. Results Study 1: Small work groups In Study 1, we applied the phi metric to interaction data we gathered from a previous study of groups performing a series of tasks designed to measure their collective intelligence [30]. Collective intelligence (CI) is a statistical factor for a group that predicts the group's performance on a wide range of tasks, just as individual intelligence does for individuals [31]. Following [31], we measured collective intelligence by asking 68 groups of 4 people each to perform a range of tasks, from brainstorming and memory tasks to solving visual and word puzzles. Then we did a factor analysis of the groups' performance on these tasks. Analogously to individual intelligence, we called the first factor that emerged from this factor analysis a group's collective intelligence. All groups did the tasks using a shared online tool. In the first condition, the groups also communicated face-to-face while collaborating on the tasks. In the second condition, groups only communicated via online text chat. In order to apply the phi metric, we characterized the state of the group in terms of which group member was communicating at which point in time. This yielded a binary state vector for each person (talking or not talking) in each time step (see Methods). In the face-to-face condition, we recorded separate audio files for each person and encoded which individuals were talking at each time. For the online groups we analyzed the chat log to determine which group members sent messages at each time. 4 We also needed to determine the time delay with respect to which phi would be calculated. For the online condition, we expect to see an influence of what is said in one comment on the next comment, so we set the time delay to one "timestep," that is, the time from one textual comment to the next one. For the face-to-face groups, we don't expect the actions of one group member to immediately influence the actions of another one. Instead, we would expect time delays on the order of a few seconds, the approximate time it takes for a person to hear and respond to what someone else says. To determine exactly how long this delay should be, we plotted average phi for different time delays (Fig 1). There is a clear peak at around 2 seconds, an intuitively plausible value, so we used this value as the time delay for analyzing phi. Fig 1. Average phi for face-to-face groups computed with different time delays. When we calculated phi, it was significantly correlated with the measured collective intelligence of the groups in both the face-to-face condition (r=0.401, p=0.047) and the online condition (r=0.352, p=0.035). The correlation was also very significant when we normalized both phi and CI scores and combined both conditions (r=0.372, p=0.003). Study 2: Groups of Wikipedia editors In Study 2, we analyzed the edit history of the articles from the "vital articles" list in Wikipedia [32] which at the time of retrieval contained 1000 articles. The quality of Wikipedia articles are classified by the community into the classes FA (Featured Article), which is the highest class, followed, in order of decreasing quality, by the categories A, GA (Good Article), B and C [33]. We examined all edits made in the 30-, 60-, and 90-day periods before an article was promoted to its current quality level, discarding the periods in which only one or two editors were active 5 (see descriptive statistics in S3 Table). All editors who edited an article during a given time period were considered members of the "group" for that article. Since the number of editors in these groups can be large (up to several hundred), computing phi for the minimum information bipartition is no longer computationally feasible since it would entail testing all possible bipartitions. Consequently we used an alternative measure of phi that treats each node as a separate partition (see Methods). Using this measure, it is clear that, in general, groups of editors who produce higher quality articles also have significantly higher phi (Fig 2 and S1 Fig; 30 days: Kendall Tau 0.0697, p < 0.00001; 60 days: Kendall Tau 0.113; p < 0.0001; 90 days Kendall Tau 0.1267, p< 0.00001, respectively). Fig. 2: Average phi for groups editing Wikipedia articles of different quality levels in the 60-day period before the articles were promoted to their current quality level. Quality levels are arranged in order of increasing quality. Error bars show standard error. To test whether this effect on phi is a result of the number of editors or of the number of edits they made, we created a regression model that predicted phi for each article from the number of editors, the number of edits per editor and four variables encoding the quality of the article. Note, that we used the categorical variables with standard treatment contrasts. That means the first level (quality C) was used as a reference and did not get parameter estimation. All other categorical quality parameters were computed relative to quality C. We did not expect the quality levels that are not significantly different from each other (see above) to be separate significant factors by themselves, and we indeed found that only quality levels B and A are significant predictors of phi (Table 1). 6 (Intercept) Number of editors Edits per editor Quality B Quality GA Quality A Quality FA Adjusted R-squared = 0.127, F = 70.11, p < 2.2e-16 Std. Error 0.207 0.001 0.009 0.240 0.311 0.359 0.378 1.594 0.012 -0.123 0.604 0.345 1.291 0.370 Estimate t Value Pr(>t) 7.708 1.75e-14 * 18.076 < 2e-16 * -14.46 < 2e-16 *** 2.515 0.012 * 1.11 0.267 3.592 0.0003 * 0.978 0.328 Table 1: Regression results when predicting phi for each article from number of edits, average number of edits per editor, and newly acquired quality level of the article. The effect of the newly attained quality level in the presence of covariates was also assessed by a likelihood ratio test between the model without quality as a variable and the full model. This showed that the article quality was still significantly correlated with phi, even when controlling for the other factors (F = 3.6847, p = 0.0053). More specifically, pairwise Wilcoxon ranksum tests show that the groups editing FA and A articles have significantly higher phi values than GA and B articles which, in turn, are significantly higher than C (Wilcoxon z-Statistics = 5.6024 p < 0.00001 between C and B and z- statistic = 3.5132, p = 0.0004 between GA and A). Study 3: Groups of computers and people on the Internet In Study 3, we applied the phi metric to a sample of the Internet traffic that passed through one Internet backbone over a six-year period [34]. We sampled one-minute segments of traffic separated by approximately six-month intervals over this period (see Methods). We encoded the state of the system in terms of whether a given machine was active (i.e., sent a data packet) at a given time. We picked a time delay of 1 time step and chose the time step size that maximized phi averaged over all years in the dataset (see Fig 3, Methods). In this case, the maximum is at 100 ms which is reassuring since it coincides well with average response times observed over the Internet (see Methods). In order to keep the computations of phi tractable, we also reduced the number of machines by randomly sampling subsets of machines that communicated with each other (see Methods). 7 Fig. 3. Average phi for Internet traffic computed with different time step sizes. When computing phi, there appears to be a steady upward trend over time. For example, Fig 4 shows one example of a highly significant relationship between the date and phi (β = 1.779, p<0.0001). Similar results were obtained for numerous other sampling methods and parameters (see Methods, S2 – S4 Figs). These results all take into account adjustments for a discontinuity between 2011 and 2012 due to a change in the hardware configuration at the recording site [35]. It is important to note that the results do not arise simply from an increasing number of machines in the Internet over time, since the number of machines in the samples analyzed is constant in each case (see Methods). 8 Fig. 4: Average phi computed on Internet traffic data over a span of 6 years (node sampling = random walk, node sample size = 100, time step size = 100 ms. See details in Methods). A change in hardware at the recording site between 2011 and 2012 caused a drop in subsequent recorded traffic [35]. The actual traffic in subsequent years is indicated by a horizontal black line and light red bars. The red bars and the red line show values adjusted to compensate for this change (see Methods). Discussion Together, these results suggest that the concept of integrated information, as formalized by the phi metric, can be usefully applied to group interactions. To begin with, the time delays at which this measure is maximized are intuitively plausible for a measure of interaction: 2 seconds for face-to-face human groups and 100 ms for machines on the Internet. Predicting group performance More importantly, phi is correlated with various measures of group performance. In 4-person work groups, it is correlated with the groups' collective intelligence. Previous work has shown that collective intelligence, in turn, predicts a group's performance on a wide range of other tasks [30,31,36]. Furthermore, in groups of Wikipedia editors, phi is correlated with the quality of the articles the groups edited. Since phi can be calculated from a relatively small sample of group interactions, this suggests that it might be possible to predict many kinds of group performance, long before a group's output is complete, merely by measuring phi. This possible use of phi seems plausible because we could interpret phi as a measure of group collaboration, and it seems likely that the degree of collaboration in a group could be a good predictor of the group's performance in many situations. 9 This use of phi would be analogous to the use of intelligence tests for individuals [37] or groups [31] to predict performance on future tasks. But these intelligence tests are interventional measures; they require people to do specific testing activities they would not otherwise have done in order to predict their performance on another task. The versions of phi used here, on the other hand, are observational measures; as we have seen, they can be calculated merely by observing what people are doing anyway. In this sense, then, phi could provide a relatively easy way of measuring how well a group is working together and using that to predict how well the group will perform on other tasks in the future. Of course, it is certainly possible that other metrics would have predictive power similar to that of phi. Therefore, we believe an important task for future research is to investigate the predictive power of various other metrics. For instance, it is possible that some of the information theoretic or correlational quantities used to compute phi would, themselves, predict performance as well as phi does. Or, perhaps, other measures of complexity (e.g., [38–41]) would be better predictors. And it will certainly be important to compare the predictive power of phi (or its components) with other potential explanatory variables such as (a) the relative participation of different group members [31], (b) the amount of effort and ability members devote to the group's tasks [42] and (c) different measures of the network topology of the group's interactions [4]. Measuring the complexity of group interaction In addition to predicting group performance in two cases, phi also provided quantitative confirmation of the plausible hypothesis that the Internet is increasing in a particular kind of complexity over time. This kind of complexity involves both increasing differentiation among parts and also increasing integration among the parts. And, as we saw, it is not simply a result of the increasing size of the Internet, since the comparisons all included the same number of nodes. Measuring the consciousness of groups We conclude this discussion with some speculations suggested by the interpretation of phi in the way it was originally intended-as a measure of consciousness. Of course, our results don't prove that groups can be conscious. That is primarily a philosophical question, not an empirical one. But our results do suggest that regarding groups as conscious, in the sense measured by integrated information, would lead to conclusions that are at least consistent with some of our intuitions about how conscious entities behave. Humans, for instance, generally perform better on a variety of tasks when they are more conscious, and we found a similar result for groups. One might also speculate that, if there is a sense in which the Internet is conscious, then its level of consciousness should be increasing over time. Our results are consistent with this speculation as well. From a philosophical perspective, the question of whether groups should be considered conscious involves whether doing so would lead to consequences that are unintuitive or otherwise undesirable. For instance, two recent philosophical papers take opposite positions on this question. List [43] argues that, even though groups may well be conscious in the functional sense of awareness, they do not have the subjective experience of consciousness. He supports this view, in part, by speculating that the integrated information of most groups would probably 10 be much lower than that of individuals. He also argues that if we consider groups as conscious, then we might be morally obligated to give them certain legal rights formerly reserved for individuals (such as the free-speech rights in the US Supreme Court's Citizens United decision). Schwitzgebel [44], on the other hand, argues that not allowing the possibility that groups could be conscious leads to conclusions that are incompatible with other materialist assumptions about the world. For instance, he argues that, unless you believe consciousness involves some kind of immaterial spirit, then it should be possible, in principle at least, for it to exist in animals, in hypothetical alien creatures, and in groups of humans. Believing otherwise, he asserts, leads either to inconsistencies or to arbitrarily excluding groups from being conscious simply by definition. The developers of integrated information theory have taken a philosophical position on this question that is referenced by both List and Schwitzgebel. Tononi and colleagues [12,16,45]) include, in their most recent version of the theory, an "exclusion postulate," which says that we should only regard a system as "conscious" if it is at the level of aggregation that maximizes phi. For example, if the phi value for a human brain is greater than the phi value for a group in which that human participates, then we should regard the human as conscious but not the group. Of course, it's possible to define consciousness in this way, and doing so avoids a possibility that many people find unintuitive: that groups might be conscious. But intuition is not an infallible guide here. As Huebner [46] writes: "it is hard to imagine that collectivities can be conscious; but it is just as hard to imagine that a mass of neurons, skin, blood, bones, and chemicals can be ... conscious." And the exclusion postulate has the undesirable consequence that a system can become unconscious, even without any change in its value of phi, if another system that contains it (or that it contains) attains a slightly higher value of phi. For instance, imagine that humans will someday be part of a very complex participatory democracy in which people are massively connected online. According to the exclusion postulate, if the participatory democracy becomes more and more richly integrated and someday reaches a level of phi greater than that of the individual humans, then all the individual humans would suddenly become "unconscious." [16,44]. From a scientific point of view, one interesting way to proceed might be to formulate an alternative version of integrated information theory that substantially weakens or eliminates the exclusion postulate. Such a theory might, for instance, allow us to say that conscious humans could also be part of higher-level groups that had a certain level of consciousness themselves. And it seems possible, at least, that such a theory could lead to interesting insights about groups as well as individuals. For example, as List [43] points out, corporations often act as more efficient profit maximizers than individual humans, and perhaps viewing the corporations as doing this consciously would provide useful insights. Conclusion In this paper, we have seen how the mathematical concept of integrated information formalizes observations about the importance of differentiation and integration that have arisen, more or less 11 independently, in a number of different disciplines. We have also seen how applying this metric to empirically analyze group interactions can lead to potentially useful predictions of group performance, measurements of group complexity, and measurements of something we might even want to call group consciousness. Much work remains to be done, but, perhaps, applying the concept of integrated information to large groups will be especially useful in understanding the complex kinds of hybrid human- computer systems that are becoming increasingly important in our modern world. Methods The research was approved by Massachusetts Institute of Technology's Institutional Review Board. Study 1: Small work groups Measuring the collective intelligence of small work groups The data about collective intelligence were collected during a previous study that tested the impact of mode of communication on general group performance [30]. In this study, groups of four people worked together on a set of diverse online tasks. The tasks included both verbal and nonverbal activities of the following types: generating, choosing, remembering, sensing, and taking physical actions. For instance, tasks included brainstorming uses for a brick, solving Raven's Matrices problems from a standardized intelligence test, remembering features of complex videos and images, and copying complex text passages into a shared online editor. Detailed task descriptions and descriptive statistics are included in [30], and summary task descriptions are in S1 and S2 Tables. All group members used individual laptop computers to work on the shared online tasks. In one condition, the group members were seated near each other and were able to communicate face- to-face while solving the tasks. In the other condition, the group members were seated far apart and were only able to communicate via the text chat functionality built into the online system. To determine the collective intelligence scores for the different groups, we performed a factor analysis of the groups' scores on the different tasks. As with previous work [31], the first factor in these analyses explained around 40% of the variance in the groups' performance on all the tasks. We treated each group's score on this first factor as the group's collective intelligence. This collective intelligence score, therefore, is a weighted average of the group's scores on all the tasks with the weights chosen to maximize the predictive power for performance on all the tasks. In this sense, the collective intelligence score for a group is exactly analogous to individual intelligence test scores for individuals [31,37]. 12 Calculating phi for small work groups For analyzing the data from Study 1, we used the phi metric that Barrett and Seth [28] call ΦE ("empirical phi"). This metric is based upon the theoretical definition of phi by Balduzzi and Tononi [13] and assumes that the system being analyzed is stationary. It can be written as: where X is a stochastic system, 𝜏 is the time delay with respect to which phi is measured, Xt is Φ![X; τ] = I (𝑋!!!;𝑿!)− the state of the system at time t, and M1 and M2 are subsets of X chosen such that they constitute a minimum information bipartition (MIB) of X. 𝐼 (𝑀!!!! ;𝑀!!) !!!! I (X,Y) is the mutual information between X and Y which is defined as the reduction in uncertainty (entropy), about X, knowing the outcome of Y: I (X; Y) = H (X) – H (XY). Thus ΦE is another way of calculating the information generated by the system as a whole that is more than just the sum of its parts. To use this metric, we recorded communication in different ways for the two conditions. For the face-to-face condition, each group member had an individual microphone. This resulted in four time-aligned audio tracks. We first used software to split the audio tracks into time steps of 200 ms each. The software then determined for each time step, who, if anyone, was speaking. To do this, the software analyzed which group members' audio volumes were above a threshold level. This level was optimized based on "ground truth" data obtained from human observer ratings of who was speaking for a limited subset of the data. The next step suppressed the audio tracks that picked up muted versions of someone else's speech. The final step merged speaking turns of a single speaker that were 400ms or less apart (e.g. someone making a brief pause during a speaking turn). This procedure thus yielded, for each team, a state vector that encoded everyone who was speaking at a specific time step with a 1 and everyone else with a 0. We then applied the phi metric to this state vector. For the online condition, we used software to analyze the chat transcripts. We encoded each line of chat as one time step. During this time step, the group member who chatted is encoded as 1 (for active) and all other group members are encoded as 0 (for inactive). This encoding leads to a situation (unlike with the face-to-face groups) where only one person can be active at any given time. We then computed the phi metric on this state dataset. Note that, in this study and the other two, we assume, but do not test for, stationarity of the time series of state vectors that are used to calculate phi. We believe that the results reported above about correlations between phi and other variables are of interest, in any case, even if they are caused, in part, by factors that led to non-stationarity in the systems. However, as noted in the Discussion above, we also believe that an important focus for future work would be to examine many alternative factors that might explain our results, including any that might have involved non-stationarity in the systems. 13 Study 2: Groups of Wikipedia editors Measuring quality of Wikipedia articles In this study, we analyzed the edit histories of the articles in Wikipedia's Vital Articles list [32]. At the time of downloading, this included 1000 articles spanning a wide range of rated quality levels, topics, and popularity. We discarded the Wikipedia front-page article since it had an order of magnitude more edits than any other article in the list and thus was as a clear outlier. This left 999 articles that we analyzed. From the edit histories of these articles, we parsed the quality level of the articles for each edit step giving us the points in time when changes in quality occurred. We then analyzed the 30-, 60-, and 90-day periods before each quality change, discarding the periods in which only one or two editors were active (see descriptive statistics in S3 Table). Calculating phi for groups of Wikipedia editors We computed phi for each article in the 30-, 60-, and 90-day periods before each quality change. As with the chat transcripts in Study 1, we encoded each edit as a single time step. An editor was considered to be active if he or she edited the Wikipedia article in question at that time step and inactive otherwise. However, we could not compute phi for Study 2 using the ΦE metric used in Study 1 for two reasons. First, as the number of nodes in the network grows, it becomes increasingly difficult to obtain enough data to accurately estimate all the relevant entropies using ΦE [28]. To deal with this problem, we used the phi metric that Barrett and Seth call ΦAR ("auto-regressive phi"). This metric provides reasonable estimates for both Gaussian and non-Gaussian systems with smaller amounts of data [28] and can be written as: 𝜑!"𝑋;𝜏,{𝑀!,𝑀!} = 12log det det (𝑋) (𝐸!) − 12! !!! log det det (𝑀!) (𝐸!!) where M1 and M2 are a bipartition of the data, detΣ(X) is the determinant of the covariance matrix of X, and EMk and EX are residuals in regression equations that estimate states of the system at one time based on knowledge of the system state at another time. To compute this version of phi, we used a MATLAB toolbox provided by Adam Barrett [28]. The second problem that arises with large systems is that determining the minimum information bipartition (MIB) requires enumerating all possible bipartitions of the dataset. Since the number of these bipartitions grows exponentially with the size of the network, this method quickly becomes computationally infeasible. To avoid these problems, we used "atomic" partitions in determining phi as recommended by [29,47]. With this approach, each node is considered as its own partition Mk and the summation in the second term is done over all these. 14 We verified the validity of this atomic measure on our dataset by computing the normal phi and the atomic phi for all edit histories from the Wikipedia dataset with 14 editors or less, 14 being the largest number where enumerating all bipartitions is still computationally feasible for all articles. The values of the atomic phi are higher but the correlation between the original phi and the atomic phi was highly significant (r = 0.83, p<0.001). In some cases in our data, the ΦAR algorithm became numerically unstable and was unable to return a value at all or returned a value that was theoretically impossible (that is, less than 0 or greater than the number of nodes). These problems usually occurred in cases where many nodes had (little or) no variance in their activities (e.g., the nodes were almost always on or almost always off). In these cases, the state vector matrices often became rank deficient, and the algorithm was unable to compute a valid value for phi. Since nodes with (little or) no variance have (little or) no entropy, they also have (little or) no effect on phi. Therefore, in cases where the algorithm did not return a valid value for phi, we simply dropped the 5% of nodes with the least variance and reran the computation, repeating this procedure until the algorithm did return a valid value. In Study 2, these problems occurred in 19.5% of the cases, and when they occurred, we had to repeat the procedure 2.27 times on average. In other words, we corrected for numerical instabilities in the calculation of phi by removing a small number of low-variance nodes that would have had little effect on phi in any case. This insured that all the data analyzed was for groups of nodes for which valid values of phi could be computed. Study 3: Groups of computers and people on the Internet To analyze Internet traffic, we used a database compiled by the Cooperative Association for Internet Data Analysis (CAIDA) [34]. This database includes records of the data logged by two high-speed monitors on a commercial backbone link on the Internet. The monitors are in Chicago and San Jose, and we chose the one in San Jose since it provides a longer undisrupted history (from 2008 to the present). We analyzed datasets separated by approximately 6-month intervals during the period (usually every March and September). Since the volume of Internet backbone traffic is huge, the database includes only one hour of data for each month, and we limited our analysis to one minute of this data for the months we analyzed. We picked the fourth minute of each hour to avoid any unusual activities in the first minute of the hour (such as special programs that operate automatically at the beginning of each hour). The database contains a trace of each packet of information sent, including an anonymized version of the Internet Protocol (IP) address for the origin and destination node of each packet. Each node (or "host") is a different computer, such as an end user's laptop, a mail server, or a web server for Google, Amazon, and other web service providers. The IP addresses for these nodes are anonymized in such a way that each real IP address always matches to the same anonymized counterpart. Descriptive statistics for the dataset, after unpacking and parsing (including, for example, removing IPv6 and unreadable packets) are shown in S4 Table. 15 Calculating phi We calculated phi for Study 3 using the same phi metric used for Study 2. To do this, we characterized the state of the system in terms of which nodes were active (in the sense of sending an information packet) at a given time. We also determined a time delay with respect to which to calculate phi. In order to do these things, several other steps were also needed. Sampling nodes As shown in S4 Table, the number of nodes sending packets in the months we analyzed ranged from about 200,000 to 1.6 million. We know of no method for calculating phi that is computationally feasible and numerically stable for systems with anything remotely approaching this number of nodes, so, before calculating phi, we needed to subsample the nodes to be analyzed. Ideally, these sampling methods should select subsets of nodes whose activity relationships are representative of those in the whole sample. Therefore, the first two methods we used were the two methods for sampling from large graphs that were found by Leskovec and Faloutsos [48] to best retain the network properties of the graphs. packets, and by A ⊂ S the subsample of nodes to be analyzed. Bold lower case letters indicate a) Random walk. We first pick a random node x ∈ S, add it to A, and make it the active node. We then randomly pick a new active node y ∈ D(x) and add it to A. At each step, single nodes. D(a) refers to the set of all destination nodes to which node a sent a packet in the sampled period and D(A) is the set of all destinations for any node in A. We do not allow duplicate nodes in A. In describing these methods, we denote by S the set of all nodes in a sample of information The two methods we used were: we continue by doing one of two things. With probability 0.85, we pick a new active node from the destinations of the current active node. And with probability 0.15, we return to x and start a new path from there. If we run out of new nodes to visit (e.g. in the case of a small isolated subset) we pick a new starting node x. This method is repeated until we have reached the sampling goal. As stated by [48] the return probability of 0.15 is the standard value picked in literature. node. Next we pick a random number n from a geometric distribution with mean 2.3 (the value suggested by [48]), randomly pick n nodes from D(x), and add these nodes to A. The procedure continues by selecting new active nodes from A and repeating the process until the required number of nodes is reached. If at any point, there are no nodes left in b) Forest fire. We first pick a random starting node x ∈ S, add it to A, and make it the active D(A) that are not already in A, then a new random starting node x ∈ S is selected and c) Breadth first. We randomly pick x ∈ S as our starting node and add it to A. We then new starting node x ∈ S and continue from there. iteratively add to A all nodes to which nodes in A sent packets (i.e. D(A)) until we reach our sampling goal. If there are no more nodes in D(A) that are not already in A, we pick a added to A. For comparison, we also used two other simple sampling methods: 16 d) Random nodes. We randomly pick x ∈ S and add it to A until we reach our sampling goal. Note that this method selects a small number of nodes (e.g., 100) completely randomly from a much larger set (e.g., several hundred thousand nodes). Therefore, even if there are substantial interactions among nodes of the type phi measures, this node sampling method may not detect them very well. However, we still include it for comparison purposes. For each date and each node sampling method, we created 100 different random subsamples of nodes. We then computed phi on the resulting state vectors and averaged the results across all 100 different random subsamples. Determining time step size and time delay To characterize the state of the system, we needed to determine the size δ of the time steps into which activity data will be grouped (i.e. for which we assume all the data packets are sent at the same time). We also needed to pick a time delay τ with respect to which phi will be calculated. These two factors depend on each other logically. For instance, if there are true interactions at a timescale of 100 ms, we could detect them with phi by, for example, setting δ = 100 ms and τ = 1 time step or by setting δ = 50 ms and τ = 2 time steps. To make the search space of possibilities more manageable, we fixed the time delay τ = 1 and selected the time step size δ that maximized phi when averaged across all the dates in our analysis. In calculating phi for this purpose, we made the following assumptions: (a) node sampling was done using the random walk method, and (b) the other corrections described below were made. This resulted in a time step size δ of 100 ms (see Fig 3). We also obtained similar results for other combinations of parameters. This corresponds very well with typical response times observed on the Internet. As noted by [49], the typical "round trip" time for data on the Internet to travel from point A to point B and back is about 200 ms. If we make the reasonable assumption that the processing time on the remote machine is minimal, then the delay is almost entirely due to time spent travelling back and forth on the network, so each one-way trip would be about 100 ms. The time delay relevant for calculating phi is the delay for one-way travel plus the time for the remote machine to respond, so these numbers correspond very well. Determining node sample size Based on preliminary experiments with our data, we found that the computations for phi often became numerically unstable and very computationally expensive at around 200 nodes. To avoid these problems we picked a standard node sample size of 100 nodes. As noted below, however, the results were also similar with samples of 150 and 200 nodes. 17 Correcting for numerical instabilities We used the same method to correct for numerical instabilities as used in Study 2. In Study 3, invalid values occurred initially in 67.14% of the cases, but they disappeared after repeating an average of 2.12 times the procedure of dropping low variance nodes. Correcting for hardware change at the recording site As mentioned in the main text, the hardware at the recording site was upgraded in the time period between September 2011 and March 2012, which led to a noticeable drop in the phi values. To correct for this, we added an indicator variable to our linear model that indicates if the date is before or after March 2012. This allowed us to extrapolate the corrected phi according to the model. For readability reasons, the graphs in Fig 4 and S2-S4 Figs show the extrapolated values in red, the uncorrected values in light red, the regression line for the corrected values as a red line and the regression line for the uncorrected values as a black line. Results Using the procedures just described, we calculated the value of phi over time for four node sampling methods (random walk, forest fire, breadth first, and random nodes). The resulting graphs are shown in Fig 4 and S2 Fig. In all cases except random node sampling, the relation between phi and year is positive and very significant (see Table 2). As noted above, we did not expect the random node sampling method to be very effective at detecting interactions of the sort phi measures, so it is not surprising that the results were not significant in this case. Regression coefficient * = p < 10-8 Random Walk 1.675* Node sampling method Forest Fire Breadth First Random Nodes 1.676* 1.715*** -0.26 Table 2. Regression coefficients for predicting phi from date with four different node sampling methods Robustness check for time step size As noted above, the main results were calculated with a time step size δ = 100 ms which maximized the value of phi. However, S3 Fig shows that using time step sizes of 50 ms or 150 ms also yields similar results. 18 Robustness check for node sample size As noted above, the main results were calculated with a node sample size of 100 nodes. However, S4 Fig shows that using sample sizes of 150 or 200 also yield similar results. Robustness check for number of packets sampled As shown in S4 Table, the number of packets sent in the minutes we studied is not constant over the dates we studied. To investigate whether the variable number of packets could have affected the results, we also investigated a different method for sampling packets. With this alternate method, we analyzed only the first 10,000,000 packets in each minute, since this is the maximum (round) number of packets present for all dates. S5 Table shows the regression coefficients for this sampling method. We see again that date is a significant predictor of phi in this case for all four sampling methods. Acknowledgments This work was made possible by financial support from the National Science Foundation (grant numbers IIS-0963285, ACI-1322254, and IIS-0963451), the U. S. Army Research Office (grant numbers 56692-MA, 64079-NS, and W911NF-15-1-0577) and Cisco Systems, Inc., through their sponsorship of the MIT Center for Collective Intelligence. We wish to especially thank Adam Barrett from the University of Sussex for providing the MATLAB toolbox to compute phi. We also thank Stephan Gade for help with the statistical analysis and Larissa Albantakis, Adam Barrett, Tomaso Poggio, and two anonymous referees for providing feedback on previous versions of the paper. 19 References 1. Bakeman R, Gottman JM. Observing interaction: an introduction to sequential analysis. 2nd ed. New York: Cambridge University Press; 1997. 2. Gottman JM. What predicts divorce?: the relationship between marital processes and marital outcomes. Hillsdale, N.J: Lawrence Erlbaum Associates; 1994. 3. Solow RM. Technical Change and the Aggregate Production Function. Rev Econ Stat. 1957;39: 312. doi:10.2307/1926047 4. Easley D, Kleinberg J. Networks, crowds, and markets: reasoning about a highly connected world. New York: Cambridge University Press; 2010. 5. Estes WK. Markov learning models for multiperson interactions, by Patrick Suppes and Richard C. Atkinson, Stanford University Press, Stanford, California, 1960, 296 pp., $8.25. Nav Res Logist Q. 1961;8: 314–315. doi:10.1002/nav.3800080313 6. Thomas EA, Malone TW. On the dynamics of two-person interactions. Psychol Rev. 1979;86: 331–360. doi:10.1037/0033-295X.86.4.331 7. Parker KC. Speaking turns in small group interaction: A context-sensitive event sequence model. J Pers Soc Psychol. 1988;54: 965–971. doi:10.1037/0022-3514.54.6.965 8. Arrow KJ, Debreu G. Existence of an Equilibrium for a Competitive Economy. Econometrica. 1954;22: 265. doi:10.2307/1907353 9. Wuchty S, Jones BF, Uzzi B. The Increasing Dominance of Teams in Production of Knowledge. Science. 2007;316: 1036–1039. doi:10.1126/science.1136099 10. Tononi G. An information integration theory of consciousness. BMC Neurosci. 2004;5: 42. doi:10.1186/1471-2202-5-42 11. Tononi G. Consciousness as Integrated Information: A Provisional Manifesto. Biol Bull. 2008;215: 216. doi:10.2307/25470707 12. Oizumi M, Albantakis L, Tononi G. From the Phenomenology to the Mechanisms of Consciousness: Integrated Information Theory 3.0. Sporns O, editor. PLoS Comput Biol. 2014;10: e1003588. doi:10.1371/journal.pcbi.1003588 13. Balduzzi D, Tononi G. Integrated Information in Discrete Dynamical Systems: Motivation and Theoretical Framework. Sporns O, editor. PLoS Comput Biol. 2008;4: e1000091. doi:10.1371/journal.pcbi.1000091 14. Balduzzi D, Tononi G. Qualia: The Geometry of Integrated Information. Friston KJ, editor. PLoS Comput Biol. 2009;5: e1000462. doi:10.1371/journal.pcbi.1000462 15. Shannon CE. A Mathematical Theory of Communication. Bell Syst Tech J. 1948;27: 379– 423. doi:10.1002/j.1538-7305.1948.tb01338.x 20 16. Tononi G, Koch C. Consciousness: here, there and everywhere? Philos Trans R Soc B Biol Sci. 2015;370: 20140167–20140167. doi:10.1098/rstb.2014.0167 17. Koch C, Tononi G. Can a Photodiode Be Conscious? The New York Review of Books. Mar 2013. Available: http://www.nybooks.com/articles/archives/2013/mar/07/can-photodiode- be-conscious/. Accessed 11 Jan 2015. 18. Tononi G, Edelman GM. Consciousness and Complexity. Science. 1998;282: 1846–1851. doi:10.1126/science.282.5395.1846 19. Tononi G, McIntosh AR, Russell DP, Edelman GM. Functional clustering: identifying strongly interactive brain regions in neuroimaging data. NeuroImage. 1998;7: 133–149. doi:10.1006/nimg.1997.0313 20. Alkire MT, Hudetz AG, Tononi G. Consciousness and anesthesia. Science. 2008;322: 876– 880. doi:10.1126/science.1149213 21. Casali AG, Gosseries O, Rosanova M, Boly M, Sarasso S, Casali KR, et al. A Theoretically Based Index of Consciousness Independent of Sensory Processing and Behavior. Sci Transl Med. 2013;5: 198ra105-198ra105. doi:10.1126/scitranslmed.3006294 22. King J-R, Sitt JD, Faugeras F, Rohaut B, El Karoui I, Cohen L, et al. Information Sharing in the Brain Indexes Consciousness in Noncommunicative Patients. Curr Biol. 2013;23: 1914–1919. doi:10.1016/j.cub.2013.07.075 23. Smith A. An inquiry into the nature and causes of the wealth of nations. William Strahan, Thomas Cadell; 1776. 24. Lawrence PR, Lorsch JW. Differentiation and Integration in Complex Organizations. Adm Sci Q. 1967;12: 1. doi:10.2307/2391211 25. Simon HA. The sciences of the artificial. 3rd ed. Cambridge, Mass: MIT Press; 1996. 26. Browning TR. Applying the design structure matrix to system decomposition and integration problems: a review and new directions. IEEE Trans Eng Manag. 2001;48: 292– 306. doi:10.1109/17.946528 27. Eppinger SD, Browning TR. Design structure matrix methods and applications. Cambridge, Mass: MIT Press; 2012. 28. Barrett AB, Seth AK. Practical Measures of Integrated Information for Time-Series Data. Sporns O, editor. PLoS Comput Biol. 2011;7: e1001052. doi:10.1371/journal.pcbi.1001052 29. Edlund JA, Chaumont N, Hintze A, Koch C, Tononi G, Adami C. Integrated Information Increases with Fitness in the Evolution of Animats. Graham LJ, editor. PLoS Comput Biol. 2011;7: e1002236. doi:10.1371/journal.pcbi.1002236 21 30. Engel D, Woolley AW, Jing LX, Chabris CF, Malone TW. Reading the mind in the eyes or reading between the lines? Theory of Mind predicts effective collaboration equally well online and face-to-face. PLOS ONE. 2014; 31. Woolley AW, Chabris CF, Pentland A, Hashmi N, Malone TW. Evidence for a collective intelligence factor in the performance of human groups. Science. 2010;330: 686–688. 32. Wikipedia Vital Articles [Internet]. Available: http://en.wikipedia.org/wiki/Wikipedia:Vital_articles 33. Wikipedia:Version 1.0 Editorial Team/Assessment. retrieved 09.10.2014 [Internet]. Wikipedia, the free encyclopedia. 2014. Available: http://en.wikipedia.org/w/index.php?title=Wikipedia:Version_1.0_Editorial_Team/Assessm ent 34. CAIDA. The CAIDA UCSD Anonymized Internet Traces 2008-2014: 2008-08-21, 2009- 03-31, 2009-09-17, 2010-03-25, 2010-09-16, 2011-03-24, 2011-09-15, 2012-03-15, 2012- 09-20, 2013-03-21, 2013-09-21, 2014-03-20 Available from web addresses in the following format, for each year from 2008 to 2014 [Internet]. Available: http://www.caida.org/data/passive/passive_2014_dataset.xml 35. CAIDA. CAIDA Data Usage Frequently Asked Questions (Retrieved 2014-11-03) Section 6.1 [Internet]. Available: http://www.caida.org/data/data-usage-faq.xml#faq- WhyisthepacketratefortheSanJos 36. Kim YJ, Engel D, Williams Woolley A, Lin J, McArthur N, Malone TW. Work together, play smart: Collective intelligence in League of Legends teams. Collective Intelligence Conference. 2015. 37. Deary IJ. Looking Down on Human Intelligence: From Psychometrics to the Brain. New York: Oxford University Press; 2000. 38. Seth AK. Causal connectivity of evolved neural networks during behavior. Netw Bristol Engl. 2005;16: 35–54. 39. Tononi G, Sporns O, Edelman GM. A measure for brain complexity: relating functional segregation and integration in the nervous system. Proc Natl Acad Sci. 1994;91: 5033– 5037. 40. Lloyd S, Pagels H. Complexity as thermodynamic depth. Ann Phys. 1988;188: 186–213. doi:10.1016/0003-4916(88)90094-2 41. Klinkner K, Shalizi C, Camperi M. Measuring Shared Information and Coordinated Activity in Neuronal Networks. Advances in Neural Information Processing Systems. 2005. pp. 667–674. 42. Hackman JR, Morris CG. Group Tasks, Group Interaction Process, and Group Performance Effectiveness: A Review and Proposed Integration. Advances in Experimental Social 22 Psychology. Elsevier; 1975. pp. 45–99. Available: http://linkinghub.elsevier.com/retrieve/pii/S0065260108602488 43. List C. What is it like to be a group agent? 2015; 44. Schwitzgebel E. If Materialism Is True, the United States Is Probably Conscious. Philos Stud. forthcoming; Available: http://faculty.ucr.edu/~eschwitz/SchwitzAbs/USAconscious.htm 45. Tononi G. Integrated information theory of consciousness: an updated account. Arch Ital Biol. 2012;150: 56–90. 46. Huebner B. Macrocognition: a theory of distributed minds and collective intentionality. Oxford: Oxford University Press; 2014. 47. Joshi NJ, Tononi G, Koch C. The Minimal Complexity of Adapting Agents Increases with Fitness. Ay N, editor. PLoS Comput Biol. 2013;9: e1003111. doi:10.1371/journal.pcbi.1003111 48. Leskovec J, Faloutsos C. Sampling from large graphs. Knowl Discov Data Min. 2006; 631. doi:10.1145/1150402.1150479 49. Internet Traffic Report [Internet]. [cited 1 Nov 2014]. Available: http://www.internettrafficreport.com/faq.htm#response 23 SUPPORTING INFORMATION S1 Table. Task categories and verbal vs. non-verbal dimensions in the Collective Intelligence task battery (reproduced from (23)) Task Category Verbal Non-Verbal 1. Generating Brainstorming Words Brainstorming Uses for a Brick Brainstorming Equations 2. Choosing Unscramble Words Judgment Slogans 3. Executing Typing Text 4. Remembering Memory Words* Matrix Reasoning Sudoku Judgment Picture Judgment Pages Typing Numbers Memory Video Memory Images 5. Sensing Detection Words Detection Images * Due to technical problems with administration of the "Memory Words" task, it was excluded from the analysis. 24 S2 Table. Description of tasks used to measure collective intelligence of groups. Task Type Brainstorming Groups had to collectively come up with as many as possible of the following: Description (a) uses for a brick, (b) words that started with S and ended with N, and (c) equations that equal 10 with certain constraints on operators and values used. Points were given for number of answers and uniqueness of answers. Unscrambling The subjects were awarded points for correctly identifying words whose letters were randomly scrambled Solving a set of Raven's Advanced Progressive Matrices, a standardized test of general fluid intelligence. Solving a Sudoku puzzle. Points were awarded for number of correct answers Groups had to predict how a larger population would rate the quality of images and slogans. They also had to estimate the number of pages in a book based on a picture of the book. The groups had to copy a complex text passage and a complex series of numbers into a shared workspace similar to Google Docs. Scoring was based on number of items copied correctly with significant penalties for incorrect and skipped items. Therefore, it was important for groups to carefully coordinate their work to avoid duplications and long sequences of skipped items. Groups were shown complex videos, images, and sequences of words and then asked to answer a set of questions about the items they had seen. Groups answered questions about a grid of small images such as "What is the most frequent object in the grid?" Matrix Reasoning Sudoku Judgment Typing Memory Detection 25 B C FA A 108 185 535 136 GA S3 Table: Descriptive statistics of the number of articles, editors, edits, and edits per editor in various periods before quality changes in the Wikipedia dataset (time windows of 30-, 60-, and 90-days shown in panels A, B, and C, respectively). Number of articles Number of Mean editors per Min Max article Std Number of Mean edits Min per article Max Std Number of Mean edits per Min editor per Max article Std (A) 30-day period before quality change 16.206 2 72 12.075 201.544 4 1049 210.452 12.153 1.75 72.8 11.186 10.984 3 63 8.589 110.676 8 825 137.086 10.909 1.8 71.8 11.857 12.806 3 56 9.216 94.241 8 861 136.702 8.039 1.833 172.2 17.157 9.164 2 51 7.111 50.206 4 857 78.652 5.356 1.667 159.2 9.422 169 6.728 2 18 3.382 34.817 6 318 44.514 5.611 1.75 58.333 8.765 A FA Number of articles Number of Mean editors Min per article Max Std Number of Mean edits per Min article Max Std Number of Mean edits per Min Editor per Max article Std (B) 60-day period before quality change 138 25.92 3 115 19.411 326.768 9 2341 345.283 12.655 2 70.939 10.908 GA B C 215 16.893 2 97 13.851 170.995 8 969 175.6 12.052 2 104 13.133 699 13.701 2 77 11.39 79.495 4 1635 124.957 5.66 1.75 272.5 11.846 273 9.183 2 31 5.611 45.388 4 350 52.16 4.931 1.75 47.5 6.19 129 20.163 3 89 16.355 152.24 7 1637 213.7 9.036 1.75 272.833 24.181 26 FA Number of articles Number of Mean editors Min per article Max Std Number of Mean edits per Min Max article Std Number of Mean edits per Min editor per Max article Std (C) 90-day period before quality change 139 34.576 6 152 26.284 413.281 21 2550 407.634 12.466 2.231 67.105 10.067 A GA B C 135 26.837 4 110 20.8 203.541 11 1653 243.445 9.051 2.2 236.143 20.515 226 22.845 3 121 18.114 229.615 9 1721 231.389 12.47 1.8 245.857 18.921 773 18.107 2 123 15.609 106.585 3 1661 151.653 5.647 1.5 237.286 9.684 345 11.533 2 41 8.222 54.939 3 420 61.581 4.591 1.5 46.667 5.049 27 S4 Table: Number of information packets, origin nodes and destination nodes for each month analyzed. Year Month 8 3 9 3 9 3 9 3 9 3 9 3 2008 2009 2009 2010 2010 2011 2011 2012 2012 2013 2013 2014 = p < 10-6 Number of information packets 10,419,879 15,640,702 17,402,179 15,322,907 14,570,116 18,094,613 20,132,417 11,242,632 10,503,070 11,195,621 20,739,677 18,803,582 Number of origin nodes 284,987 373,440 495,864 343,577 294,594 559,263 411,129 294,890 220,584 1,675,419 274,875 413,410 Number of destination nodes 248,892 551,702 622,101 399,145 173,678 352,122 412,778 164,138 96,248 150,134 204,683 164,051 S5 Table: Regression coefficients for predicting phi from date with four different node sampling methods while only looking at a fixed number of packets (10,000,000 packets). Node sampling method Random Walk 3.487* Forest Fire 3.025* Breadth First 3.850* Random Nodes 7.334* Regression coefficient * = p < 10-6 28 A B S1 Fig. Average phi for groups editing Wikipedia articles of different quality levels in the 30-day period (A) and the 90-day period (B) before the articles were promoted to their current quality level. 29 S2 Fig: Average phi over time for various node sampling methods (top left to bottom right: Random Walk, Forest Fire, Breadth First and Random Nodes). In all cases node sample size = 100, and time step size δ = 100 ms. 30 S3 Fig: Average phi plotted over time (node sampling = random walk, node sample size = 100, time step size δ = 50 ms (left) and 150ms (right). For 50ms, β = 0.8227, p = 0.000007; for 150ms, β = 1.333, p=0.002. S4 Fig: Average phi plotted over time (node sampling = random walk, node sample size = 150 (left) and 200 (right), time step size δ = 100 ms). For 150 nodes, β = 0.4477, p = 0.018; for 200 nodes, β = 0.6535, p=0.00026. 31
1802.03997	3	1802	2019-07-25T13:19:32	GEMSEC: Graph Embedding with Self Clustering	[ "cs.SI" ]	Modern graph embedding procedures can efficiently process graphs with millions of nodes. In this paper, we propose GEMSEC -- a graph embedding algorithm which learns a clustering of the nodes simultaneously with computing their embedding. GEMSEC is a general extension of earlier work in the domain of sequence-based graph embedding. GEMSEC places nodes in an abstract feature space where the vertex features minimize the negative log-likelihood of preserving sampled vertex neighborhoods, and it incorporates known social network properties through a machine learning regularization. We present two new social network datasets and show that by simultaneously considering the embedding and clustering problems with respect to social properties, GEMSEC extracts high-quality clusters competitive with or superior to other community detection algorithms. In experiments, the method is found to be computationally efficient and robust to the choice of hyperparameters.	cs.SI	cs	GEMSEC: Graph Embedding with Self Clustering Benedek Rozemberczki, Ryan Davies, Rik Sarkar and Charles Sutton School of Informatics, The University of Edinburgh {benedek.rozemberczki,ryan.davies}@ed.ac.uk,{rsarkar,csutton}@inf.ed.ac.uk 9 1 0 2 l u J 5 2 ] I S . s c [ 3 v 7 9 9 3 0 . 2 0 8 1 : v i X r a Abstract -- Modern graph embedding procedures can efficiently process graphs with millions of nodes. In this paper, we propose GEMSEC -- a graph embedding algorithm which learns a clustering of the nodes simultaneously with computing their embedding. GEMSEC is a general extension of earlier work in the domain of sequence-based graph embedding. GEMSEC places nodes in an abstract feature space where the vertex features minimize the negative log-likelihood of preserving sam- pled vertex neighborhoods, and it incorporates known social network properties through a machine learning regularization. We present two new social network datasets and show that by simultaneously considering the embedding and clustering problems with respect to social properties, GEMSEC extracts high-quality clusters competitive with or superior to other community detection algorithms. In experiments, the method is found to be computationally efficient and robust to the choice of hyperparameters. Index Terms -- community detection, clustering, node embed- ding, network embedding, feature extraction. I. INTRODUCTION Community detection is one of the most important problems in network analysis due to its wide applications ranging from the analysis of collaboration networks to image segmentation, the study of protein-protein interaction networks in biology, and many others [1], [2], [3]. Communities are usually defined as groups of nodes that are connected to each other more densely than to the rest of the network. Classical approaches to community detection depend on properties such as graph metrics, spectral properties and density of shortest paths [4]. Random walks and randomized label propagation [5], [6] have also been investigated. Embedding the nodes in a low dimensional Euclidean space enables us to apply standard machine learning techniques. This space is sometimes called the feature space -- implying that it represents abstract structural features of the network. Embed- dings have been used for machine learning tasks such as label- ing nodes, regression, link prediction, and graph visualization, see [7] for a survey. Graph embedding processes usually aim to preserve certain predefined differences between nodes encoded in their embedding distances. For social network embedding, a natural priority is to preserve community membership and enable community detection. Recently, sequence-based methods have been developed as a way to convert complex, non-linear network structures into formats more compatible with vector spaces. These methods sample sequences of nodes from the graph using a randomized mechanism (e.g. random walks), with the idea that nodes that are "close" in the graph connectivity will also frequently appear close in a sampling of random walks. The methods then proceed to use this random-walk-proximity information as a basis to embed nodes such that socially close nodes are placed nearby. In this category, Deepwalk [8] and Node2Vec [9] are two popular methods. While these methods preserve the proximity of nodes in the graph sense, they do not have an explicit preference for preserving social communities. Thus, in this paper, we develop a machine learning approach that considers clustering when embedding the network and includes a parameter to control the closeness of nodes in the same community. Figure 1(a) shows the embedding obtained by the standard Deepwalk method, where communities are coherent, but not clearly separated in the embedding. The method described in this paper, called GEMSEC, is able to produce clusters that are tightly embedded and separated from each other (Fig. 1(b)). (a) DeepWalk (b) GEMSEC Fig. 1. Zachary's Karate club graph [10]. White nodes: instructor's group; blue nodes: president's group. GEMSEC produces embedding with more tightly clustered communities. A. Our Contributions GEMSEC is an algorithm that considers the two problems of embedding and community detection simultaneously, and as a result, the two solutions of embedding and clustering can inform and improve each other. Through iterations, the embedding converges toward one where nodes are placed close to their neighbors in the network, while at the same time clusters in the embedding space are well separated. The algorithm is based on the paradigm of sequence- based node embedding procedures that create d dimensional feature representations of nodes in an abstract feature space. Sequence-based node embeddings embed pairs of nodes close to each other if they occur frequently within a small window of each other in a random walk. This problem can be formu- lated as minimizing the negative log-likelihood of observed neighborhood samples (Sec. III) and is called the skip-gram optimization [11]. We extend this objective function to include a clustering cost. The formal description is presented in Subsection III-A. The resulting optimization problem is solved with a variant of mini-batch gradient descent [12]. The detailed algorithm is presented in Subsection III-B. By enforcing clustering on the embedding, GEMSEC reveals the natural community structure (e.g. Figure 1).Our approach improves over existing methods of simultaneous embedding and clustering [13], [14], [15] and shows that community sensitivity can be directly incorporated into the skip-gram style optimization to obtain greater accuracy and efficiency. In social networks, nodes in the same community tend to have similar groups of friends, which is expressed as high neighborhood overlap. This fact can be leveraged to produce clusters that are better aligned with the underlying communi- ties. We achieve this effect using a regularization procedure -- a smoothness regularization added to the basic optimization achieves more coherent community detection. The effect can be seen in Figure 3, where a somewhat uncertain community affiliation suggested by the randomized sampling is sharpened by the smoothness regularization. This technique is described in Subsection III-C. In experimental evaluation we demonstrate that GEMSEC outperforms -- in clustering quality -- the state of the art neigh- borhood based [8], [9], multi-scale [16], [17] and community aware embedding methods [13], [14], [15]. We present new social datasets from the streaming service Deezer and show that the clustering can improve music recommendations. The clustering performance of GEMSEC is found to be robust to hyperparameter changes, and the runtime complexity of our method is linear in the size of the graphs. To summarize, the main contributions of our work are: 1) GEMSEC: a sequence sampling-based learning model which learns an embedding of the nodes at the same time as it learns a clustering of the nodes. 2) Clustering in GEMSEC can be aligned to network neigh- borhoods by a smoothness regularization added to the optimization. This enhances the algorithm's sensitivity to natural communities. 3) Two new large social network datasets are introduced -- from Facebook and Deezer data. 4) Experimental results show that the embedding process runs linearly in the input size. It generally performs well in quality of embedding and in particular outperforms existing methods on cluster quality measured by modu- larity and subsequent recommendation tasks. We start with reviewing related work in the area and relation to our approach in the next section. A high-performance Tensorflow reference implementation of GEMSEC and the datasets that we collected can be accessed online1. II. RELATED WORK There is a long line of research in metric embedding -- for example, embedding discrete metrics into trees [18] and 1https://github.com/benedekrozemberczki/GEMSEC into vector spaces [19]. Optimization-based representation of networks has been used for routing and navigation in domains such as sensor networks and robotics [20], [21]. Represen- tations in hyperbolic spaces have emerged as a technique to preserve richer network structures [22], [23], [24]. features for Recent advances in node embedding procedures have made large real-world it possible to learn vector graphs [8], [16], [9]. Features extracted with these sequence- based node embedding procedures can be used for predicting social network users' missing age [7], the category of scientific papers in citation networks [17] and the function of proteins in protein-protein interaction networks [9]. Besides supervised learning tasks on nodes the extracted features can be used for graph visualization [7], link prediction [9] and community detection [13]. Sequence based embedding commonly considers variations in the sampling strategy that is used to obtain vertex sequences -- truncated random walks being the simplest strategy [8]. More involved methods include second-order random walks [9], skips in random walks [17] and diffusion graphs [25]. It is worth noting that these models implicitly approximate matrix factorizations for different matrices that are expensive to factorize explicitly [26]. Our work extends the literature of node embedding algo- rithms which are community aware. Earlier works in this category did not directly extend the skip-gram embedding framework. M-NMF [14] applies computationally expensive non-negative matrix factorization with a modularity constraint term. The procedure DANMF [15] uses hierarchical non- negative matrix factorization to create community-aware node embeddings. ComE [13] is a more scalable approach, but it assumes that in the embedding space the communities fit a gaussian structure, and aims to model them by a mixture of Gaussians. In comparison to these methods, GEMSEC provides greater control over community sensitivity of the em- bedding process, it is independent of the specific neighborhood sampling methods and is computationally efficient. III. GRAPH EMBEDDING WITH SELF CLUSTERING For a graph G = (V, E), a node embedding is a mapping f : V → Rd where d is the dimensionality of the embedding space. For each node v ∈ V we create a d dimensional representation. Alternatively, the embedding f is a V × d real-valued matrix. In sequence-based embedding, sequences of neighboring nodes are sampled from the graph. Within a sequence, a node v occurs in the context of a window ω within the sequence. Given a sample S of sequences, we refer to the collection of windows containing v as NS(v). Earlier works have proposed random walks, second-order random walks or branching processes to obtain NS(v). In our experiments, we used unweighted first and second-order random walks for node sampling [8], [9]. Our goal is to minimize the negative log-likelihood of ob- serving neighborhoods of source nodes conditional on feature vectors that describe the position of nodes in the embedding space. Formally, the optimization objective is: f X min v∈V − log P (NS(v)f (v)) (1) for a suitable probability function P (··). To define this P , we consider two standard properties (see [9]) expected of the embedding f in relation to NS. First, it should be possible to factorize P (NS(v)f (v)) in line with conditional indepen- dence with respect to f (v). Formally: P (NS(v)f (v)) = Y ni∈NS(v) P (ni ∈ NS(v) f (v), f (ni)). (2) Second, it should satisfy symmetry in the feature space, mean- ing that source and neighboring nodes have a symmetric effect on each other in the embedding space. A softmax function on the pairwise dot products of node representations with f (v) to get P (ni ∈ NS(v) f (v), f (ni)) express such a property: O(V 2) runtime complexity. Because of this, we approximate the partition function term with negative sampling which is a form of noise contrastive estimation [11], [27]. (a) Node capture. (b) Empty initialization. Fig. 2. Potential issues with cluster cost weighting and cluster initialization. Different node colors denote different ground truth community memberships and the computed cluster boundary is denoted by the dashed line. In Subfigure 2a a single white node is captured in a cluster with the blue nodes due to clustering weight γ being high. In Subfigure 2b an empty cluster is initialized with no nodes in it. It is plausible that the cluster center remains empty throughout the optimization process. P (ni ∈ NS(v) f (v), f (ni)) = exp(f (ni) · f (v)) exp(f (u) · (f (v)) Pu∈V ∂L ∂f (v∗) = . (3) Substituting (2) and (3) into the optimization function, we get: min f Xv∈V  ln Xu∈V exp(f (v) · f (u))! −Xni∈NS (v) f (ni) · f (v)  . The partition function in Equation (4) enforces nodes to be embedded in a low volume space around the origin, while the second term forces nodes with similar sampled neighborhoods to be embedded close to each other. (4) A. Learning to Cluster Next, we extend the optimization to pay attention to the clusters it forms. We include a clustering cost similar to k- means, measuring the distance from nodes to their cluster centers. This augmented optimization problem is described by minimizing a loss function over the embedding f and position of cluster centers µ, that is, min f,µ L, where: exp(f (v) · f (u))! − Xni∈NS (v) Embedding cost kf (v) − µck2 f (ni) · f (v)  } (5)  ln Xu∈V L = Xv∈V + γ ·Xv∈V min c∈C Clustering cost {z {z . } In Equation (5) we have C the set of cluster centers -- the cth cluster mean is denoted by µc. Each of these cluster centers is a d-dimensional vector in the embedding space. The idea is to minimize the distance from each node to its nearest cluster center. The weight coefficient of the clustering cost is given by the hyperparameter γ. Evaluating the partition function in the proposed objective function for all of the source nodes has a exp(f (v∗) · f (u)) · f (u) Pu∈V Pu∈V exp(f (v∗) · f (u)) {z Partition function gradient f (v∗) − µc + γ · kf (v∗) − µck2 } {z Closest cluster direction f (ni) − X {z ni∈NS(v∗) } Neighbor direction } (6) The gradients of the loss function in Equation 5 are important in solving the minimization problem. As a result we can obtain the gradients for node representations and cluster centers. Ex- amining in more detail, the gradient of the objective function L with respect to the representation of node v∗ ∈ V is described by Equation (6) if µc is the closest cluster center to f (v∗). The gradient of the partition function pulls the represen- tation of v∗ towards the origin. The second term moves the representation of v∗ closer to the representations of its neighbors in the embedding space while the third term moves the node closer to the closest cluster center. If we set a high γ value the third term dominates the gradient. This will cause the node to gravitate towards the closest cluster center which might not contain the neighbors of v∗. An example is shown in Figure 2a. If the set of nodes that belong to cluster center c is Vc, then the gradient of the objective function with respect to µc is described by ∂L ∂µc = −γ · X v∈Vc f (v) − µc kf (v) − µck2 . (7) In Equation 7 we see that the gradient moves the cluster center by the sum of coordinates of nodes in the embedding space that belong to cluster c. Second, if a cluster ends up empty it will not be updated as elements of the gradient would be zero. Because of this, cluster centers and embedding weights are initialized with the same uniform distribution. A wrong initialization just like the one with an empty cluster in Subfigure 2b can affect clustering performance considerably. Data: G = (V, E) -- Graph to be embedded. N -- Number of sequence samples per node. l -- Length of sequences. ω -- Context size. d -- Number of embedding dimensions. C -- Number of clusters. k -- Number of noise samples. γ0 -- Initial clustering weight coefficient. α0, αF -- Initial and final learning rate. Result: f (v), where v ∈ V where c ∈ C µc, 1 Model ← Initialize Model(V , d, C) 2 t ← 0 3 for n in 1:N do 4 5 6 7 8 9 10 11 bV ← Shuffle(V ) for v in bV do t ← t + 1 γ ← Update γ (γ0, t, w, l, N, V ) α ← Update α (α0, αF , t, w, l, N, V ) Sequence ← Sample Nodes(G, v, l) Features ← Extract Features( Sequence, ω) Update Weights(Model, Features, γ, α, k) end 12 13 end Algorithm 1: GEMSEC training procedure B. GEMSEC algorithm We propose an efficient learning method to create GEM- SEC embeddings which is described with pseudo-code by Algorithm 1. The main idea behind our procedure is the following. To avoid the clustering cost overpowering the graph information (as in Fig. 2a), we initialize the system with a low weight γ0 ∈ [0, 1] for clustering, and through iterations anneal it to 1. The embedding computation proceeds as follows. The weights in the model are initialized based on the number of vertices, embedding dimensions and clusters. After this, the algorithm makes N sampling repetitions in order to generate vertex sequences from every source node. Before starting a sampling epoch, it shuffles the set of vertices. We set the clustering cost coefficient γ (line 7) according to an exponential annealing rule described by Equation (8). The learning rate is set to α (line 8) with a linear annealing rule (Equation (9)). γ = γ0 ·(cid:16)10 −t·log10 γ0 w·l·V ·N (cid:17) α = α0 − (α0 − αF ) · t w · l · V · N (8) (9) 10). The extracted features, gradient, current learning rate and clustering cost coefficient determine the update to model weights by the optimizer (line 11). In the implementation we utilized a variant of stochastic gradient descent -- the Adam optimizer [12]. We approximate the first cost term with noise contrastive estimation to make the gradient descent tractable, drawing k noise samples for each positive sample. If the node sampling is done by first-order random walks the runtime complexity of this procedure will be O((ω · k + C) · l · d · V · N ) while DeepWalk with noise contrastive estimation has a O(ω · k · l · d · V · N ) runtime complexity. C. Smoothness Regularization for coherent community detec- tion We have seen in Subsection III-A that there is a tension between what the clustering objective considers to be clusters and what the real communities are in the underlying social network. We can incorporate additional knowledge of social network communities using a machine learning technique called regularization. We observe that social networks have natural local prop- erties such as homophily, strong ties between members of a community, etc. Thus, we can incorporate such social network- specific properties in the form of regularization to find more natural embeddings and clusters. This regularization effect can be achieved by adding a term Λ to the loss function: Λ = λ ·X(v,u)∈ES w(v,u) · kf (v) − f (u)k2 , (10) where the weight function w determines the social network cost of the embedding with respect to properties of the edges traversed in the sampling. We use the neighborhood overlap of an edge -- defined as the fraction of neighbors common to two nodes of the edge relative to the union of the two neighbor sets2. In experiments on real data, neighborhood overlap is known to be a strong indicator of the strength of relation between members of a social network [28]. Thus, by treating neighborhood overlap as the weight wv,u of edge (v, u), we can get effective social network clustering, which is confirmed by experiments in the next section. The coeffeicient λ lets us tune the contribution of the social network cost in the embedding process. In experiments, the regularized version of the algorithms is found to be more robust to changes in hyperparameters. The effect of the regularization can be understood intuitively through an example. For this exposition, let us consider matrix representations of the social network describing closeness of nodes. In fact, other skip-gram style learning processes like [8], [9] are known to approximate the factorization of The sampling process reads sequences of length l (line 9) and extracts features using the context window size ω (line 2Neighbor sets N (a) and N (b) of nodes a and b, the neighborhood overlap of (a, b) is defined as the Jaccard similarity N(a)∩N(b) N(a)∪N(b) . a similarity matrix M such as [26]: Mu,v = log   vol(G) ω ωX r=1 PP ∈P r v,u Qa∈P \{v} 1 deg(a) deg(v)   − log(k) where P r v,u is the set of paths going from v to u with length r. Elements of the target matrix M grow with number of paths of length at most ω between the corresponding nodes. Thus M is intended to represent level of connectivity between nodes in terms of a raw graph feature like number of paths. The barbell graph in Figure 3a is a typical example with an obvious community structure we can use to analyze the matter. The optimization procedure used by Deepwalk [8] aims to converge to a target matrix Mu,v shown in Figure 3b. Observe that this matrix has fuzzy edges around the communities of the graph, showing a degree of uncertainty. An actual approxi- mation by running the Deepwalk is shown in Figure 3c, which naturally incorporates further uncertainty due to sampling. A much more clear output with sharp communities can be obtained by applying a regularized optimization. This can be seen in Figure 3d. STATISTICS OF SOCIAL NETWORKS USED IN THE PAPER. TABLE I Source Facebook Deezer Dataset Politicians Companies Athletes Media Celebrities Artists Government TV Shows Croatia Hungary Romania V 5,908 14,113 13,866 27,917 11,565 50,515 7,057 3,892 54,573 47,538 41,773 Density 0.0024 0.0005 0.0009 0.0005 0.0010 0.0006 0.0036 0.0023 0.0004 0.0002 0.0001 Transitivity 0.3011 0.1532 0.1292 0.1140 0.1666 0.1140 0.2238 0.5906 0.1146 0.0929 0.0752 IV. EXPERIMENTAL EVALUATION In this section we evaluate the cluster quality obtained by the GEMSEC variants, their scalability, robustness and predictive performance on a downstream supervised task. Results show that GEMSEC outperforms or is at par with existing methods in all measures. A. Datasets For the evaluation of GEMSEC real-world social network datasets are used which we collected from public APIs specif- ically for this work. Table I shows these social networks have a variety of size, density, and level of clustering. We used graphs from two sources: • Facebook page networks: These graphs represent mutual like networks among verified Facebook pages -- the types of sites included TV shows, politicians, athletes, and artists among others. • Deezer user-user friendship networks: We collected friendship networks from the music streaming site Deezer and included 3 European countries (Croatia, Hungary, and Romania). For each user, we curated the list of genres loved based on the songs liked by the user. B. Standard parameter settings A fixed standard parameter setting is used our experi- ments, and we indicate any deviations. Models using first order random walk sampling strategy are referenced as GEMSEC and Smooth GEMSEC, second order random walk variants are named as GEMSEC2 and Smooth GEMSEC2. Random walks with length 80 are used and 5 truncated random walks per source node were used. Second-order random walk control hyperparameters [9] return and in- out were chosen from (cid:8)2−2, 2−1, 1, 2, 4(cid:9). A window size of 5 is used for features. Each embedding has 16 dimen- sions and we extract 20 cluster centers. A parameter sweep over hyperparameters was used to obtain the highest aver- age modularity. Initial learning rate values are chosen from learning rate is cho- (cid:8)10−2, 5 · 10−3, 10−3(cid:9) and the final sen from(cid:8)10−3, 5 · 10−4, 10−4(cid:9). Noise contrastive estimation ficient is chosen from (cid:8)10−1, 10−2, 10−3(cid:9). The smoothness uses 10 negative examples. The initial clustering cost coef- regularization term's hyperparameter is 0.0625 and Jaccard's coefficient is the penalty weight. C. Cluster Quality Using Facebook page networks we evaluate the clustering performance. Cluster quality is evaluated by modularity -- we assume that a node belongs to a single community. Our results are summarized in Table II based on 10 experimental repetitions and errors in parentheses correspond to two stan- dard deviations. The baselines use the hyperparameters from the respective papers. We used 16-dimensional embeddings throughout. The embeddings obtained with non-community- aware methods were clustered after the embedding by k- means clustering to extract 20 cluster centers. Specifically, comparisons are made with: 1) Overlap Factorization [29]: Factorizes the neighborhood overlap matrix to create features. 2) DeepWalk [8]: Approximates the sum of the adjacency matrix powers with first order random walks and implic- itly factorizes it. 3) LINE [16]: Implicitly factorizes the sum of the first two powers for the normalized adjacency matrix and the resulting node representation vectors are concatenated together to form a multi-scale representation. 4) Node2vec [9]: Factorizes a neighbourhood matrix ob- tained with second order random walks. The in-out and return parameters of the second-order random walks were chosen from the (cid:8)2−2, 2−1, 1, 2, 4(cid:9) set to maxi- mize modularity. 5) Walklets [17]: Approximates with first order random walks each adjacency matrix power individually and implicitly factorizes the target matrix. These embeddings (a) The graph (b) Target Matrix of Deepwalk. (c) Approximation by Deep- walk. (d) Regularized Approxima- tion Fig. 3. An example Barbell graph with the corresponding target matrix factorized (window size of 3) by DeepWalk [26] and the reconstructed target matrices obtained with standard DeepWalk and Smooth DeepWalk. Regularized optimization produces more well defined communities. While the standard DeepWalk model has less well defined clusters. MEAN MODULARITY OF CLUSTERINGS ON THE FACEBOOK DATASETS. EACH EMBEDDING EXPERIMENT WAS REPEATED TEN TIMES. ERRORS IN THE PARENTHESES CORRESPOND TO TWO STANDARD DEVIATIONS. IN TERMS OF MODULARITY Smooth GEMSEC2 OUTPERFORMS THE BASELINES. TABLE II Politicians Companies Overlap Factorization DeepWalk LINE Node2Vec Walklets ComE M-NMF DANMF Smooth DeepWalk GEMSEC Smooth GEMSEC GEMSEC2 Smooth GEMSEC2 0.810 (±0.008) 0.840 (±0.015) 0.841 (±0.014) 0.846 (±0.012) 0.843 (±0.014) 0.830 (±0.008) 0.816 (±0.014) 0.810 (±0.020) 0.849 (±0.017) 0.851 (±0.009) 0.855 (±0.006) 0.852 (±0.010) 0.859 (±0.006) 0.553 (±0.010) 0.637 (±0.012) 0.651 (±0.009) 0.664 (±0.008) 0.655 (±0.012) 0.654 (±0.005) 0.646 (±0.007) 0.648 (±0.005) 0.667 (±0.007) 0.662 (±0.013) 0.683 (±0.009) 0.667 (±0.008) 0.684 (±0.009) Athletes 0.601 (±0.020) 0.649 (±0.012) 0.665 (±0.007) 0.669 (±0.007) 0.664 (±0.007) 0.665 (±0.007) 0.655 (±0.008) 0.650 (±0.009) 0.669 (±0.007) 0.674 (±0.009) 0.692 (±0.009) 0.683 (±0.008) 0.692 (±0.007) Media 0.471 (±0.016) 0.481 (±0.022) 0.558 (±0.012) 0.565 (±0.011) 0.562 (±0.009) 0.573 (±0.005) 0.561 (±0.004) 0.560 (±0.006) 0.541 (±0.006) 0.536 (±0.011) 0.567 (±0.009) 0.551 (±0.008) 0.571 (±0.010) Celebrities 0.551 (±0.01) 0.631 (±0.011) 0.642 (±0.010) 0.643 (±0.013) 0.621 (±0.043) 0.635 (±0.010) 0.628 (±0.006) 0.628 (±0.011) 0.643 (±0.008) 0.636 (±0.014) 0.649 (±0.008) 0.638 (±0.009) 0.649 (±0.011) Artists 0.474 (±0.018) 0.508 (±0.029) 0.557 (±0.014) 0.560 (±0.010) 0.548 (±0.016) 0.560 (±0.011) 0.535 (±0.021) 0.532 (±0.019) 0.523 (±0.020) 0.528 (±0.020) 0.559 (±0.011) 0.562 (±0.020) 0.562 (±0.017) Government TV Shows 0.608 (±0.024) 0.686 (±0.024) 0.690 (±0.017) 0.692 (±0.017) 0.689 (±0.019) 0.696 (±0.010) 0.668 (±0.011) 0.673 (±0.015) 0.707 (±0.008) 0.705 (±0.020) 0.710 (±0.008) 0.712 (±0.010) 0.712 (±0.010) 0.786 (±0.008) 0.811 (±0.005) 0.813 (±0.010) 0.827 (±0.016) 0.819 (±0.015) 0.806 (±0.011) 0.813 (±0.008) 0.812 (±0.014) 0.835 (±0.008) 0.833 (±0.010) 0.841 (±0.004) 0.838 (±0.010) 0.847 (±0.006) are concatenated to form a multi-scale representation of nodes. positive effect on the clustering performance of Deepwalk, GEMSEC and GEMSEC2. 6) ComE [13]: Uses a Gaussian mixture model to learn an embedding and clustering jointly using random walk features. 7) M-NMF [14]: Factorizes a matrix which is a weighted sum of the first two proximity matrices with a modular- ity based regularization constraint. 8) DANMF [15]: Decomposes a weighted sum of the first two proximity matrices hierarchically to obtain cluster memberships with an autoencoder-like non-negative ma- trix factorization model. Smooth GEMSEC, GEMSEC2 and Smooth GEMSEC2 consis- tently outperform the neighborhood conserving node embed- ding methods and the competing community aware methods. The relative advantage of Smooth GEMSEC2 over the bench- marks is highest on the Athletes dataset as the clustering's modularity is 3.44% higher than the best performing baseline. It is the worst on the Media dataset with a disadvantage of 0.35% compared to the strongest baseline. Use of smoothness regularization has sometimes non-significant, but definitely D. Sensitivity Analysis for hyperparameters We tested the effect of hyperparameter changes to clustering performance. The Politicians Facebook graph is embedded with the standard parameter settings while the initial and final learning rates are set to be 10−2 and 5 · 10−3 respectively, the clustering cost coefficient is 0.1 and we perturb certain hyper- parameters. The second-order random walks used in-out and return parameters of 4. In Figure 4 each data point represents the mean modularity calculated from 10 experiments. Based on the experimental results we make two observations. First, GEMSEC model variants give high-quality clusters for a wide range of parameter settings. Second, introducing smoothness regularization makes GEMSEC models more robust to hyper- parameter changes. This is particularly apparent across varying the number of clusters. The length of truncated random walks and the number of random walks per source node above a certain threshold has only a marginal effect on the community detection performance. 0.86 0.85 0.84 0.83 0.82 y t i r a l u d o M 4 12 20 28 36 44 52 60 Clusters 2 4 6 8 10 Context size y t i r a l u d o M 0.856 0.852 0.848 0.844 16 32 48 64 80 96 112128 Dimensions 40 80 120 Walk length 160 0.855 0.845 0.835 0.825 y t i r a l u d o M y t i r a l u d o M 0.8 0.7 0.6 0.5 0.86 y t i r a l u d o M 0.84 0.82 0.8 0.86 0.85 0.84 0.83 y t i r a l u d o M 0.4 0.6 0.2 Cluster cost coefficient GEMSEC 0.8 1 Smooth GEMSEC 2 4 6 8 10 Number of radom walks GEMSEC2 Smooth GEMSEC2 Fig. 4. Sensitivity of cluster quality to parameter changes measured by modularity. E. Music Genre Recommendation Node embeddings are often used for extracting features of nodes for downstream predictive tasks. In order to investigate this, we use social networks of Deezer users collected from European countries. We predict the genres (out of 84) of music liked by people. Following the embedding, we used logistic regression with ℓ2 regularization to predict each of the labels and 90% of the nodes were randomly selected for training. We evaluated the performance of the remaining users. Numbers reported in Table III are F1 scores calculated from 10 experimental repetitions. GEMSEC2 significantly outperforms the other methods on all three countries' datasets. The perfor- mance advantage varies between 3.03% and 4.95%. We also see that Smooth GEMSEC2 has lower accuracy, but it is able to outperform DeepWalk, LINE, Node2Vec, Walklets, ComE, M-NMF and DANMF on all datasets. F. Scalability and computational efficiency To create graphs of various sizes, we used the Erdos-Renyi model and with an average degree of 20. Figure 5 shows the log of mean runtime against the log of the number of nodes. Most importantly, we can conclude that doubling the size of the graph doubles the time needed for optimizing GEMSEC, thus the growth is linear. We also observe that embedding algorithms that incorporate clustering have a higher cost, and regularization also produces a higher cost, but similar growth. ) s ( e m i t n u r n o i t a z i m i t p O 2 g o l 13 11 9 7 5 3 1 −1 −3 7 9 11 13 15 log2 Vertex number DeepWalk Smooth GEMSEC Smooth DeepWalk M-NMF GEMSEC ComE Fig. 5. Sensitivity of optimization runtime to graph size measured by seconds. The dashed lines are linear references. V. CONCLUSIONS We described GEMSEC -- a novel algorithm that learns a node embedding and a clustering of nodes jointly. It extends existing embedding modes. We showed that smoothness regu- larization is used to incorporate social network properties and produce natural embedding and clustering. We presented new social datasets, and experimentally, our methods outperform a number of strong community aware node embedding base- lines. VI. ACKNOWLEDGEMENTS Benedek Rozemberczki and Ryan Davies were supported by the Centre for Doctoral Training in Data Science, funded by EPSRC (grant EP/L016427/1). REFERENCES [1] T. Van Laarhoven and E. Marchiori, "Robust community detection methods with resolution parameter for complex detection in protein protein interaction networks," Pattern Recognition in Bioinformatics, pp. 1 -- 13, 2012. [2] L. Backstrom, D. Huttenlocher, J. Kleinberg, and X. Lan, "Group formation in large social networks: Membership, growth, and evolution," in Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2006, pp. 44 -- 54. [3] S. Papadopoulos, Y. Kompatsiaris, A. Vakali, and P. Spyridonos, "Com- munity detection in social media," Data Mining and Knowledge Discov- ery, vol. 24, no. 3, pp. 515 -- 554, 2012. [4] J. Leskovec, A. Rajaraman, and J. D. Ullman, Mining of Massive Datasets. Cambridge University Press, 2014. [5] P. Pascal and M. Latapy, In International Symposium on Computer and Information Sciences. Springer Berlin Heidelberg, 2005, ch. Computing Communities in Large Networks Using Random Walks., pp. 284 -- 293. [6] S. Gregory, "Finding overlapping communities in networks by label propagation," New Journal of Physics, vol. 12, no. 10, p. 103018, 2010. [7] P. Goyal and E. Ferrara, "Graph embedding techniques, applications, and performance: A survey," arXiv preprint arXiv:1705.02801, 2017. [8] B. Perozzi, R. Al-Rfou, and S. Skiena, "Deepwalk: Online learning of social representations." in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining., 2014. [9] A. Grover and J. Leskovec, "Node2vec: Scalable feature learning for networks," in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 855 -- 864. MULTI-LABEL NODE CLASSIFICATION PERFORMANCE OF THE EMBEDDING EXTRACTED FEATURES ON THE DEEZER GENRE LIKES DATASETS. PERFORMANCE IS MEASURED BY AVERAGE F1 SCORE VALUES. MODELS WERE TRAINED ON 90% OF THE DATA AND EVALUATED ON THE REMAINING 10%. ERRORS IN THE PARENTHESES CORRESPOND TO TWO STANDARD DEVIATIONS. GEMSEC MODELS CONSISTENTLY HAVE GOOD PERFORMANCE. TABLE III CROATIA HUNGARY ROMANIA Overlap Factorization DeepWalk LINE Node2Vec Walklets ComE M-NMF DANMF Smooth DeepWalk GEMSEC Smooth GEMSEC GEMSEC2 Smooth GEMSEC2 Micro 0.319 (±0.017) 0.321 (±0.006) 0.331 (±0.013) 0.348 (±0.012) 0.363 (±0.013) 0.326 (±0.012) 0.336 (±0.005) 0.340 (±0.007) 0.329 (±0.006) 0.328 (±0.006) 0.333 (±0.006) 0.381 (±0.007) 0.373 (±0.005) Macro Weighted 0.026 0.208 (±0.002) (±0.010) 0.026 (±0.002) 0.028 (±0.002) 0.032 (±0.003) 0.043 (±0.003) 0.028 (±0.002) 0.028 (±0.001) 0.027 (±0.002) 0.028 (±0.002) 0.027 (±0.002) 0.028 (±0.002) 0.046 (±0.003) 0.044 (±0.002) 0.207 (±0.004) 0.212 (±0.010) 0.235 (±0.010) 0.270 (±0.012) 0.217 (±0.009) 0.217 (±0.003) 0.210 (±0.002) 0.215 (±0.006) 0.212 (±0.004) 0.215 (±0.004) 0.287 (±0.005) 0.276 (±0.006) Micro 0.361 (±0.007) 0.361 (±0.004) 0.374 (±0.007) 0.393 (±0.008) 0.397 (±0.007) 0.363 (±0.010) 0.369 (±0.015) 0.365 (±0.011) 0.375 (±0.006) 0.377 (±0.004) 0.379 (±0.006) 0.407 (±0.005) 0.409 (±0.004) Macro Weighted 0.029 0.227 (±0.001) (±0.006) 0.029 (±0.002) 0.033 (±0.002) 0.037 (±0.002) 0.051 (±0.001) 0.033 (±0.001) 0.032 (±0.002) 0.031 (±0.002) 0.032 (±0.002) 0.032 (±0.002) 0.034 (±0.002) 0.050 (±0.003) 0.053 (±0.002) 0.228 (±0.002) 0.250 (±0.005) 0.267 (±0.011) 0.307 (±0.006) 0.246 (±0.007) 0.239 (±0.011) 0.242 (±0.008) 0.244 (±0.004) 0.244 (±0.004) 0.250 (±0.004) 0.310 (±0.007) 0.314 (±0.006) Micro 0.275 (±0.025) 0.307 (±0.008) 0.332 (±0.007) 0.346 (±0.008) 0.361 (±0.011) 0.323 (±0.008) 0.330 (±0.016) 0.335 (±0.009) 0.321 (±0.008) 0.332 (±0.008) 0.334 (±0.008) 0.378 (±0.009) 0.376 (±0.008) Macro Weighted 0.020 0.167 (±0.003) (±0.017) 0.023 (±0.002) 0.028 (±0.002) 0.031 (±0.002) 0.050 (±0.005) 0.028 (±0.001) 0.028 (±0.002) 0.029 (±0.002) 0.026 (±0.002) 0.028 (±0.002) 0.029 (±0.002) 0.049 (±0.003) 0.049 (±0.003) 0.186 (±0.006) 0.212 (±0.006) 0.229 (±0.008) 0.281 (±0.012) 0.212 (±0.006) 0.209 (±0.013) 0.210 (±0.012) 0.204 (±0.006) 0.213 (±0.006) 0.215 (±0.006) 0.289 (±0.007) 0.287 (±0.007) [10] W. W. Zachary, "An information flow model for conflict and fission in small groups," Journal of anthropological research, vol. 33, no. 4, pp. 452 -- 473, 1977. [23] C. De Sa, A. Gu, C. R´e, and F. Sala, "Representation tradeoffs for hyperbolic embeddings," Proceedings of machine learning research, vol. 80, p. 4460, 2018. [24] W. Zeng, R. Sarkar, F. Luo, X. Gu, and J. Gao, "Resilient routing for sensor networks using hyperbolic embedding of universal covering space," in 2010 Proceedings IEEE INFOCOM. IEEE, 2010, pp. 1 -- 9. [25] B. Rozemberczki and R. Sarkar, "Fast sequence-based embedding with diffusion graphs," in International Workshop on Complex Networks. Springer, 2018, pp. 99 -- 107. [26] J. Qiu, Y. Dong, H. Ma, J. Li, K. Wang, and J. Tang, "Network embedding as matrix factorization: Unifying deepwalk, line, pte, and node2vec," in Proceedings of the Eleventh ACM International Confer- ence on Web Search and Data Mining. ACM, 2018, pp. 459 -- 467. [27] M. Gutmann and A. Hyvarinen, "Noise-contrastive estimation: A new estimation principle for unnormalized statistical models," in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, 2010, pp. 297 -- 304. [28] J.-P. Onnela, J. Saramaki, J. Hyvonen, G. Szab´o, D. Lazer, K. Kaski, J. Kert´esz, and A.-L. Barab´asi, "Structure and tie strengths in mobile communication networks," Proceedings of the national academy of sciences, vol. 104, no. 18, pp. 7332 -- 7336, 2007. [29] A. Ahmed, N. Shervashidze, S. Narayanamurthy, V. Josifovski, and A. J. Smola, "Distributed large-scale natural graph factorization," in Proceedings of the 22nd international conference on World Wide Web, 2013, pp. 37 -- 48. [11] T. Mikolov, K. Chen, G. Corrado, and J. Dean, "Efficient estimation of word representations in vector space," 2013. [12] J. B. Diederik P. Kingma, "Adam: A method for stochastic optimiza- tion," in Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. [13] S. Cavallari, V. W. Zheng, H. Cai, K. C.-C. Chang, and E. Cambria, "Learning community embedding with community detection and node embedding on graphs," in Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, 2017, pp. 377 -- 386. [14] X. Wang, P. Cui, J. Wang, J. Pei, W. Zhu, and S. Yang, "Community preserving network embedding." in AAAI, 2017, pp. 203 -- 209. [15] F. Ye, C. Chen, and Z. Zheng, "Deep autoencoder-like nonnegative matrix factorization for community detection," in Proceedings of the 27th ACM International Conference on Information and Knowledge Management. ACM, 2018, pp. 1393 -- 1402. [16] J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei, "Line: Large-scale information network embedding," in Proceedings of the 24th International Conference on World Wide Web, 2015, pp. 1067 -- 1077. [17] B. Perozzi, V. Kulkarni, H. Chen, and S. Skiena, "Don't walk, skip!: Online learning of multi-scale network embeddings," in Proceedings of the 2017 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining 2017, 2017, pp. 258 -- 265. [18] J. Fakcharoenphol, S. Rao, and K. Talwar, "A tight bound on approx- imating arbitrary metrics by tree metrics," Journal of Computer and System Sciences, vol. 69, no. 3, pp. 485 -- 497, 2004. [19] J. Matousek, Lectures on discrete geometry. Springer, 2002, vol. 108. [20] X. Yu, X. Ban, W. Zeng, R. Sarkar, X. Gu, and J. Gao, "Spherical representation and polyhedron routing for load balancing in wireless sensor networks," in 2011 Proceedings IEEE INFOCOM. IEEE, 2011, pp. 621 -- 625. [21] K. Huang, C.-C. Ni, R. Sarkar, J. Gao, and J. S. Mitchell, "Bounded stretch geographic homotopic routing in sensor networks," in IEEE IN- FOCOM 2014-IEEE Conference on Computer Communications. IEEE, 2014, pp. 979 -- 987. [22] R. Sarkar, "Low distortion delaunay embedding of trees in hyperbolic plane," in International Symposium on Graph Drawing. Springer, 2011, pp. 355 -- 366.
1911.05494	1	1911	2019-09-17T05:15:23	Concept Drift Adaptive Physical Event Detection for Social Media Streams	[ "cs.SI", "cs.CY", "cs.LG", "stat.ML" ]	Event detection has long been the domain of physical sensors operating in a static dataset assumption. The prevalence of social media and web access has led to the emergence of social, or human sensors who report on events globally. This warrants development of event detectors that can take advantage of the truly dense and high spatial and temporal resolution data provided by more than 3 billion social users. The phenomenon of concept drift, which causes terms and signals associated with a topic to change over time, renders static machine learning ineffective. Towards this end, we present an application for physical event detection on social sensors that improves traditional physical event detection with concept drift adaptation. Our approach continuously updates its machine learning classifiers automatically, without the need for human intervention. It integrates data from heterogeneous sources and is designed to handle weak-signal events (landslides, wildfires) with around ten posts per event in addition to large-signal events (hurricanes, earthquakes) with hundreds of thousands of posts per event. We demonstrate a landslide detector on our application that detects almost 350% more land-slides compared to static approaches. Our application has high performance: using classifiers trained in 2014, achieving event detection accuracy of 0.988, compared to 0.762 for static approaches.	cs.SI	cs	Concept Drift Adaptive Physical Event Detection for Social Media Streams Abhijit Suprem1, Aibek Musaev2, and Calton Pu1 1 Georgia Institute of Technology, GA 30332, USA 2 University of Alabama, AL, 35487, USA [email protected] Abstract. Event detection has long been the domain of physical sensors operat- ing in a static dataset assumption. The prevalence of social media and web ac- cess has led to the emergence of social, or human sensors who report on events globally. This warrants development of event detectors that can take advantage of the truly dense and high spatial and temporal resolution data provided by more than 3 billion social users. The phenomenon of concept drift, which caus- es terms and signals associated with a topic to change over time, renders static machine learning ineffective. Towards this end, we present an application for physical event detection on social sensors that improves traditional physical event detection with concept drift adaptation. Our approach continuously up- dates its machine learning classifiers automatically, without the need for human intervention. It integrates data from heterogeneous sources and is designed to handle weak-signal events (landslides, wildfires) with around ten posts per event in addition to large-signal events (hurricanes, earthquakes) with hundreds of thousands of posts per event. We demonstrate a landslide detector on our ap- plication that detects almost 350% more landslides compared to static ap- proaches. Our application has high performance: using classifiers trained in 2014, achieving event detection accuracy of 0.988, compared to 0.762 for static approaches. Keywords: Concept drift, Machine Learning Event Detection, Disaster Detec- tion. 1 Introduction The ubiquitous presence of web data and increase in users sharing information in social media has created a global network of human reporters who report on live events. Such human reporters can be considered as social sensors that provide infor- mation about live physical events around the globe [4-7]. Development of applica- tions that can take advantage of social sensors to perform physical event detection are a clear next step. The primary challenge lies in the actual event detection: since social sensor data is a noisy text stream, machine learning models are required. Further, there is the phenomena of concept drift, where the distribution of real-world data 2 changes with respect to time. This is especially apparent with text data. We provide an example with landslide detection on social media: the word landslide can refer to the disaster event or elections, among others. Since October and November are often election seasons in the United States, classification models that are tuned to ignore social media data with election related landslide keywords are more appropriate in this data window. In other months, the presence of election-related landslide tweets is scarce, which can cause increase false negatives in models overfit for election-related landslide tweets. This instance of changing data distributions is a form of concept drift. So, social sensors constitute a challenge for the traditional approaches to text clas- sification, which involve static machine learning classifiers that are never updated. Event detection systems for social sensors without concept drift adaptation face per- formance deterioration. We can consider Google Flu Trends (GFT) as an example; GFT was originally created to identify seasonal trends in the flu season [9]. However, the models did not incorporate changes in Google's own search data, causing increas- ing errors after release [9-11]. Our application addresses these challenges by incorpo- rating concept drift adaptation. Additionally, we develop automated concept drift adaptation techniques to automatically generate training data for machine learning model updates. This is necessary due to the sheer volume of social media data -- it is impractical to manually label the millions of social media posts per day. Our approach allows us to perform drift adaptation without human labelers, which significantly reduces training bottlenecks. Specifically, we have the following contributions: 1. We present a drift-adaptive event detection application that performs physical event detection on social sensors using machine learning. We also show automated classifier updates for concept drift adaptation. 2. We develop a procedure to combine news articles and physical sensor data (e.g. rainfall data from NOAA and earthquake data from USGS) to perform automated training data generation for concept drift adaptation. Our application uses the low- latency, abundant social sensor data to perform physical event detection with ma- chine learning classifiers, and the high-latency, scarce physical sensor data to tune and update classifiers. We demonstrate our event detection application with landslide detection. We select landslides because they do not have dedicated physical sensors (in contrast to tsuna- mis or earthquakes); however, they cause large monetary and human losses each year. Landslides are a also what we call a weak-signal disaster: landslide-related social media data has significant noise in social media streams. Also, usage of landslide keywords to reference disasters is small compared to usage to reference irrelevant topics such as election landslides. We compare our landslide detection application (LITMUS-adaptive) to the static approach in [2], which we call LITMUS-static. Our approach, LITMUS-adaptive, detects 350% more landslide events than LITMUS-static. We also evaluate our adap- tive classifiers' accuracy compared to the static classifiers in LITMUS-static. We train classifiers with data in 2014, and compare performance of static and adaptive 3 approaches in 2018. LITMUS-adaptive has f-score of 0.988 in 2018, compared to f- score of 0.762 in LITMUS-static, showcasing improvements our drift adaptive ap- proach makes. The rest of the paper is organized as follows: Section 2 covers related work. Sec- tion 3 covers data sources used in our application. Section 4 and 5 provide implemen- tation details for our application. Section 6 and 7 evaluate our application quantita- tively and qualitatively, respectively. Section 8 presents our conclusions. 2 Related Work 2.1 Physical Event Detection on Social Sensors Earthquake detection using social sensors was initially proposed in [1]. There have also been attempts to develop physical event detectors for other types of disasters, including flooding [2], flu [3, 4], infectious diseases [5], and landslides [6, 7]. In most cases, the works focus on large-scale disasters or health crises, such as earthquakes, hurricanes [8], and influenza that can be easily verified and have abundant reputable data Our application is general purpose, as it can handle small-scale disaster such as landslides and large-scale disasters. The existing approaches also assume data without concept drift. However such assumptions, made in Google Flu Trends [9, 10] degrade in the long term. 2.2 Concept Drift Adaptation Recent drift adaptation approaches evaluate their methods with synthetic data [11- 14]. Such data is perturbed to include specific, known forms of drift. Several mecha- nisms have been developed for handling concept drift with numeric, sensor data. Windowing, or sliding windows, is a common technique for adaptation. This ap- proach uses multiple data memories, or windows of different lengths sliding over incoming data. Each window has an associated model. The SAM-KNN approach uses k-NN classifier to select window closest to a new data sample for classification [15]. Nested windows are considered in [16] to obtain training sets. Adaptive Random Forests augment a random forest with a drift detector. Drift detection leads to forest pruning to remove tress that have poor performance. Pruned trees are replaced with new weak classifiers [17]. Knowledge Maximized Ensemble (KME) uses a combination of off-the-shelf and created drift detectors to recognize various forms of drift simultaneously. Models are updated when enough training data is collected and removed if they perform poor- ly [18]. Most methods approach concept drift with an eye towards detection and subse- quent normalization. Updating or rebuilding a machine learning model facing drift involves two bottlenecks in the classification pipeline: data labeling and model train- ing; of these, data labeling is the greater challenge due to its oracle requirements. Such wait-and-see models that perform corrections once errors have been detected entail periodic performance degradation before they are corrected with model up- 4 dates; this may be infeasible in mission-critical applications. Active learning strate- gies counteract this bottleneck in part; the tradeoff is between highly accurate models and clustered, knowledge-agnostic representations that consider data on distance without subject matter expertise. 3 Data Sources Our application, LITMUS-adaptive, combines physical sensor and news data, which have high-latency and are scarce, with social sensor data, which have low- latency and are abundant. The social sensor data is used for event detection through machine learning classifiers, while the physical sensors and news data are used to update machine learning classifiers. 3.1 Physical Sensors and News 1. News: News articles are downloaded from various online RSS feeds. Each source is described by the article link, the publish date, the article headline, and the publi- cation name. Locations are extracted from the text using Named Entity Recogni- tion. Publisher sources include international feeds from agencies (e.g. BBC, CNN, ABC, Reuters), as well as local news sources (some sample snippets are provided in Figure 1). 2. Rainfall Reporting: We download rainfall data from NOAA and earthquake data from USGS to perform to validate landslide detections. 3. Landslide Predictions: The National Oceanic and Atmospheric Administration (NOAA) and USGS provide landslide predictions in select locations where there is enough terrain and rainfall data. LITMUS uses this to perform localized landslide tracking and labeling. Fig. 1. Snippets of news articles about landslides. Each retrieved article is geolocated using NER to identify locations and indexed spatiotemporally. 5 3.2 Social Sensors 1. Twitter: a keyword streamer is used to download tweets continuously for Twitter. Keywords include the words 'landslide', 'mudslide', and 'rockslide' as well as their lemmas (some examples are provided below). 2. Facebook: a general keyword streamer is used to download public Facebook posts. Existing web crawlers are leveraged to improve retrieval efficiency Fig. 2. Sample of raw tweets. The left side are relevant tweets. The right side are irrelevant tweets for landslide detection. 4 Approach Fig. 3. Our application detects real-time events from social sensors. It remains drift-adaptive by integrating physical sensors with social sensors to continuously update machine learning classi- fiers. The dataflow for our application is shown in Figure 3. We perform physical event detection by performing classification on social media data. We use binary classifiers that detect whether a given social sensor post is relevant to a given event or not (e.g. landslides). The latency between a physical event's occurrence and social sensor post about the event is significantly lower than latency with news reports and physical sensors, which often require expert confirmation. In contrast, social sensor data has low latency. However, it lacks the reputability of physical sensors and news reports. We rely on the social sensors for event detection, and physical sensors to continuous- ly tune machine learning models. 6 Traditional approaches perform event detection under a static data assumption. Our contribution is in generating updated machine learning classifiers without any manual intervention so that our application can adapt to concept drift. 4.1 Social Sensor Download Social sensor download operates in real-time. Our application has streaming end- points for several short-text social media systems, such as Twitter and Facebook, with an extensible framework for integrating custom streaming sources. Social sensor downloads operate in a high-volume streaming setting. Each downloaded post object contains at least five fields: (i) the post content as a Unicode string, (ii) an array of named locations within the post (this is usually null and is filled during data pro- cessing step), (iii) timestamp of post, (iv) array of hyperlink content within the post, and (v) user-id or screen name of human reporter who created the post. 4.2 Physical Sensor and News Download Physical sensors and news sources are dedicated physical, social, and web sensors providing event information with human annotations. In contrast to social sensors, these are trustworthy sources. We further distinguish physical sensors and news from social sensors: physical sensor and news data is highly structured and contains de- tailed event information. In our domain (landslides), most physical sensors and news sources provide geographical coordinates and time of landslide disaster. In most mis- sion-critical applications, such physical sensor and news data appears long after an event takes place, once reputable sources have confirmed the event. Additionally, these sensors have lower volume. For physical sensors, it is trivial to insert the physical event provided by the sen- sors into our ground truth database by extracting timestamp and location information. News articles provide topic tags that can be mined for an application's event; event reports (e.g. earthquake or large landslide report by USGS) provide detailed infor- mation about events, including locations, timestamps, event range, and event impact. We use Named Entity Recognition as well as location tags of news articles to extract location information to identify event location. These events are also stored in the ground truth event database as confirmed landslide events. 4.3 Social Sensor Data Processing Social sensor data has low-context, which hinders location extraction and classifi- cation. Additionally, Named Entity Recognition (NER) often fails on short-text be- cause there are too few words for the NER algorithms. We augment natural language extractors by sharing information between ground truth events and social sensor data processing. We provide an example with location extraction. Since social posts have few words, location extraction is not accurate on the short-text and often misses loca- tions provided in a post's text content. So any location identified by NER is saved in memory for several days as a string. This string is used to augment location extraction 7 with substring match; intuitively, if there is one social post about an event in a loca- tion, there could be others. Similarly, any ground-truth event locations are also added to the short-term memory to augment social sensor location extraction. 4.4 Automated Training Data Generation We integrate social sensor data with physical sensors and news sources to automat- ically generate labeled data. During live operation, social sensor data is passed to the Machine Learning classifiers for event detection. Concept drift causes performance deterioration in these classifiers. So, our application performs model updates at regu- larly scheduled intervals. A model update requires labeled data, and a classifier's own labels cannot be used for updating itself. As noted before, it is also impractical to manually label the large volume of social sensor data (on Twitter alone, there are >500M tweets per day). Our application matches historical social sensor data to ground truth events detect- ed from physical sensors and news reports, which are highly trustworthy. Intuitively, social posts with landslide keywords that have similar space-time coordinates as ground truth events are very likely relevant to a real landslide (as opposed to irrele- vant posts such as election landslides). At the end of each data window (one-month windows in our landslide application), ground truth events of the window are stored as cells (coordinates of events are mapped to 2.5-min cell grids on the planet [7]). Social media data from the window is localized by time into 6-day bins. Note that in data processing step, NER augmenta- tion only occurs forward in time, i.e. when a new location is added to the memory for substring matching, only subsequent social sensor posts are processed with the new location. So, during the automated training data generation stage, we have access to archived social sensor data. Table 1. Automatically labeled data in each data window Data Window Data Samples Labeled 2014-Training Data 2014-Test Data July 2018 August 2018 September 2018 October 2018 November 2018 December 2018 26,953 6464 378 212 386 498 1770 446 13028 3266 189 106 193 249 885 223 We take advantage of this by re-processing data from the prior window. Locations extracted from each ground-truth event within days are used as a substring filter to extract locations from social posts. We then perform automated labeling by matching a social media post's space-time coordinates to true-event location and time. Location 8 matching is achieved with the 2.5-min cell grid superposition using strong supervi- sion. This is in contrast to weak-supervision [21] as our supervisory labeling is do- main-specific, instead of domain-agnostic. Table 1 below shows statistics about the generated training data (labeled data) in each window. 5 Event Detection with Machine Learning Classifiers Our application uses machine learning classifiers to perform event detection. We employ a variety of statistical and deep learners, including SVMs, Logistic Regres- sion, Decision Trees, and Neural Networks. Our drift adaptation consists of two com- plementary parts: Classifier generation (to create new ML models) and Classifier updates (to update older models with new data). We cover them below. We will first describe the classifier generation/update procedure. Then we will cover update sched- ules which govern when new classifiers are generated or updated. 5.1 Classifier generation We define a window as the collection of social sensor data between two updates. At the end of a data window, training data is generated for the previous window using procedures in Automated Training Data Generation. The labeled samples are used to train new ML classifiers. Existing classifiers are copied, and the copies are updated with the new data. Both new and updated classifiers are saved to a database using key-value scheme. The classifiers function as values and the training data as the key. Currently, instead of storing the entire training data, we store the training data cen- troid as the key for a classifier. 5.2 Update Schedule We support three types of classifier update schedules: user-specified, detector- specified, and hybrid, described below. These schedules allow for continuous classifi- er generation and updates to combat concept drift. User-Specified. Users can set up an update schedule (daily, weekly, monthly, etc). The application tracks the internal time, and when an update is triggered, procedures in Classifier Generation are followed to create new classifiers and update existing ones. Detector-Specified. Some classifiers types provide confidence values with predic- tions. Neural networks with softmax output layer provide class probabilities. Drift can be detected by tracking frequency of low confidence labels. For linear classifiers (in- cluding SVMs), higher density of data points close to the separating hyperplane over time can indicate signal drift. If drift time exceeds a threshold, procedures in Classifi- er Generation are followed to create new classifiers and update existing ones. 9 5.3 ML-Based Event Detection Classifiers are retrieved using two approaches: (i) recency or (ii) relevancy. Recen- cy performs lookup on most recently created classifiers. Relevancy performs k-NN (k nearest neighbors) search to find training data (using the stored data centroids) that is closest to prediction data; the k-closest data are then used to look up respective classi- fiers in the classifier database. These classifiers are used as an ensemble). Our machine learning event detectors use multiple classifiers with votes to perform predictions, since ensembles perform better than lone classifiers. Our ensemble classi- fier supports several weight assignment options: Unweighted average. The class labels predicted by each classifier in the ensemble (0 for irrelevant and 1 for relevant physical event) are summed and averaged. indicates majority of classifiers consider the input post as relevant to the physical event Weighted average. Classifiers can be weighted by domain experts based on which algorithm they implement. Weak classifiers (random forests) can be given lower weights than better classifiers (SVMs, neural networks). Model-weighted. We can determine classifier weights using their prior performance: where is classifier , is the weight of classifier in data window , and is the validation accuracy of on the testing data in the data window ( is the window where the model was last trained). 6 Evaluation of Drift Adaptive Approach We first demonstrate the need for concept drift adaptive event detection with evidence of concept drift in our data. As shown in Table 1, we have social sensor data from 2014 through 2018. Each labeled post's text is converted to a high-dimensional, nu- meric representation using word2vec [22]. The post vector is dimensionally reduced with tSNE. The tSNE-based reducer measures pairwise similarities between data points and shows the event characteristics separation between 2014 and 2018 data. Our real-world live data will continue to evolve over time. Such drift in text data is difficult to predict due to the phenomena of lexical diffu- sion [23]. Current drift adaptation methods use synthetic data with bounded and pre- dictable drift to evaluate methods; they do not focus on adapting to unbounded drift in live data. 10 Fig. 4. TSNE of datasets from 2014 through 2018 We evaluate our drift adaptive event detectors with two approaches summarized in Table 2. In each window, the drift adaptive learner is provided with classifiers trained under each approach and tuned with optimal hyperparameters obtained using grid search. Table 2. Summary of approaches Approach N_RES (Static) Description Non-drift resilient approach with static classifiers RES (Adaptive) Drift resilient approach using generated training data for updates Training Data 2014 Data 2014 Data -- 2018 Data (Separated by windows) We compare performance of each approach in Table 2 over subsequent windows in our data. As we show in Figure 5, ensembles with resilience (adaptive approach) out- perform non-resilient counterparts (static approach) throughout. RES under both sta- tistical and deep learners maintains high f-score across multiple years. N_RES has significantly higher variance in performance, and is generally poor at adapting to the live data's drift. 11 Fig. 5. The Adaptive methods significantly outperform static counterparts in the 2018 data. N_RES classifiers face deterioration across all metrics without access to generated training data to update their parameters (Figure 6). Fig. 6. Decay is apparent in non-resilient classifiers. The f-score is high during the offline win- dow, but degrades without access to generated training data. 7 Landslide Detection: Results We built our application with physical event detection in mind; evaluation is per- formed with landslide detection. In the previous section, we validated our drift adap- tive approach. Here, we compare our application -- LITMUS-adaptive, to LITMUS- static, the traditional approach. Figure 7 shows the raw event comparisons between LITMUS-static and LITMUS-adaptive. LITMUS-adaptive outperforms LITMUS- static, and over time, the share of events detected only by LITMUS-adaptive increases (Figure 8). We see that by December 2018, LITMUS-adaptive detects 71% of events, 12 compared to LITMUS-static's 28% of events, an increase by 350%. We show global coverage in Figure 9. Fig. 7. Raw additional events comparison between LITMUS-adaptive and LITMUS-static. Both applications are events detected by both LITMUS-static and LITMUS-adaptive. Every event detected by LITMUS-static was also detected by LITMUS-adaptive. In addition, LITMUS-adaptive also detects several additional events (LITMUS-adaptive only). The sum of the two are shown in LITMUS-adaptive Total Fig. 8. Events normalized as fractions of total. Over data windows, fraction of data missed by LITMUS-static increases as its classifiers deteriorate. However, with continuous updates, LITMUS-adaptive can adapt to drift and maintain higher accuracy and thus, better event detec- tion. 13 Fig. 9. Our application has global coverage. We are able to detect events across the globe, and as mentioned, every event in LITMUS-static is also detected in LITMUS-adaptive, along with hundreds of additional events LITMUS-static misses (Figure 7 and Figure 8). 8 Conclusions We proposed a physical event detector for social sensor data that remains resilient to concept drift. Our approach combines social sensors with physical sensors and news data to perform continuous learning to maintain model currency with the data distri- bution. Our application's drift adaptation takes advantage of human annotations in existing reputable sources (physical sensors and news data) to augment and generate training data. This removes the need for humans to perform manual labeling, signifi- cantly reducing cost and labeling bottlenecks. Our application uses an ML-based event processing classifiers that continuously adapt to changes in live data. We believe the application and the methods presented in this paper can be useful for a variety of social-sensor based physical event detection. We demonstrated a disaster detection application that is designed for landslide detec- tion. Our application, LITMUS-adaptive, improves upon static approaches such as LITMUS-static. LITMUS-adaptive adapts to changing event characteristics in social sources and detects almost 350% more landslide events than LITMUS-static. Moreo- ver, LITMUS-adaptive achieves f-score of 0.988 by December 2018, compared to f- score of 0.762 for static approaches. 14 9 Acknowledgement This research has been partially funded by National Science Foundation by CISE's SAVI/RCN (1402266, 1550379), CNS (1421561), CRISP (1541074), SaTC (1564097) programs, an REU supplement (1545173), and gifts, grants, or contracts from Fujitsu, HP, Intel, and Georgia Tech Foundation through the John P. Imlay, Jr. Chair endowment. Any opinions, findings, and conclusions or recommendations ex- pressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation or other funding agencies and companies mentioned above. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] T. Sakaki, M. Okazaki, and Y. Matsuo, "Earthquake shakes Twitter users: real-time event detection by social sensors," in Proceedings of the 19th international conference on World wide web, 2010, pp. 851-860: ACM. B. Jongman, J. Wagemaker, B. R. Romero, and E. C. de Perez, "Early flood detection for rapid humanitarian response: harnessing near real-time satellite and Twitter signals," ISPRS International Journal of Geo-Information, vol. 4, no. 4, pp. 2246- 2266, 2015. S. Wakamiya, Y. Kawai, and E. Aramaki, "Twitter-Based Influenza Detection After Flu Peak via Tweets With Indirect Information: Text Mining Study," JMIR public health and surveillance, vol. 4, no. 3, 2018. S. Yang, M. Santillana, and S. C. Kou, "Accurate estimation of influenza epidemics using Google search data via ARGO," Proceedings of the National Academy of Sciences, vol. 112, no. 47, pp. 14473-14478, 2015. H. Hirose and L. Wang, "Prediction of infectious disease spread using Twitter: A case of influenza," in Parallel Architectures, Algorithms and Programming (PAAP), 2012 Fifth International Symposium on, 2012, pp. 100-105: IEEE. A. Musaev, D. Wang, S. Shridhar, and C. Pu, "Fast text classification using randomized explicit semantic analysis," in Information Reuse and Integration (IRI), 2015 IEEE International Conference on, 2015, pp. 364-371: IEEE. A. Musaev, D. Wang, and C. Pu, "LITMUS: Landslide detection by integrating multiple sources," in ISCRAM, 2014. D. Thom, H. Bosch, S. Koch, M. Wörner, and T. Ertl, "Spatiotemporal anomaly detection through visual analysis of geolocated twitter messages," in Visualization Symposium (PacificVis), 2012 IEEE Pacific, 2012, pp. 41-48: IEEE. D. Lazer, R. Kennedy, G. King, and A. Vespignani, "The parable of Google Flu: traps in big data analysis," Science, vol. 343, no. 6176, pp. 1203-1205, 2014. D. Lazer and R. Kennedy, "What we can learn from the epic failure of Google Flu Trends," Wired. Conde Nast, vol. 10, 2015. P. R. Almeida, L. S. Oliveira, A. S. Britto Jr, and R. Sabourin, "Adapting dynamic classifier selection for concept drift," Expert Systems with Applications, vol. 104, pp. 67-85, 2018. 15 S. H. Bach and M. A. Maloof, "Paired learners for concept drift," in Data Mining, 2008. ICDM'08. Eighth IEEE International Conference on, 2008, pp. 23-32: IEEE. J. Gama, P. Medas, G. Castillo, and P. Rodrigues, "Learning with drift detection," in Brazilian symposium on artificial intelligence, 2004, pp. 286-295: Springer. J. P. Göpfert, B. Hammer, and H. Wersing, "Mitigating concept drift via rejection," in International Conference on Artificial Neural Networks, 2018, pp. 456-467: Springer. M. Markou and S. Singh, "Novelty detection: a review -- part 1: statistical approaches," Signal processing, vol. 83, no. 12, pp. 2481-2497, 2003. M. M. Lazarescu, S. Venkatesh, and H. H. Bui, "Using multiple windows to track concept drift," Intelligent data analysis, vol. 8, no. 1, pp. 29-59, 2004. H. M. Gomes et al., "Adaptive random forests for evolving data stream classification," Machine Learning, vol. 106, no. 9-10, pp. 1469-1495, 2017. S. Ren, B. Liao, W. Zhu, and K. Li, "Knowledge-maximized ensemble algorithm for different types of concept drift," Information Sciences, vol. 430, pp. 261-281, 2018. J. Shan, H. Zhang, W. Liu, and Q. Liu, "Online Active Learning Ensemble Framework for Drifted Data Streams," IEEE transactions on neural networks and learning systems, no. 99, pp. 1-13, 2018. I. Žliobaitė, A. Bifet, B. Pfahringer, and G. Holmes, "Active learning with evolving streaming data," in Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2011, pp. 597-612: Springer. M. Dehghani, H. Zamani, A. Severyn, J. Kamps, and W. B. Croft, "Neural ranking models with weak supervision," in Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2017, pp. 65-74: ACM. T. Mikolov, K. Chen, G. Corrado, J. Dean, L. Sutskever, and G. Zweig, "word2vec," URL https://code. google. com/p/word2vec, 2013. J. Eisenstein, B. O'Connor, N. A. Smith, and E. P. Xing, "Diffusion of lexical change in social media," PloS one, vol. 9, no. 11, p. e113114, 2014. [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23]
1609.07526	4	1609	2017-01-12T21:13:36	Balancing Speed and Coverage by Sequential Seeding in Complex Networks	[ "cs.SI", "physics.soc-ph" ]	Information spreading in complex networks is often modeled as diffusing information with certain probability from nodes that possess it to their neighbors that do not. Information cascades are triggered when the activation of a set of initial nodes (seeds) results in diffusion to large number of nodes. Here, several novel approaches for seed initiation that replace the commonly used activation of all seeds at once with a sequence of initiation stages are introduced. Sequential strategies at later stages avoid seeding highly ranked nodes that are already activated by diffusion active between stages. The gain arises when a saved seed is allocated to a node difficult to reach via diffusion. Sequential seeding and a single stage approach are compared using various seed ranking methods and diffusion parameters on real complex networks. The experimental results indicate that, regardless of the seed ranking method used, sequential seeding strategies deliver better coverage than single stage seeding in about 90% of cases. Longer seeding sequences tend to activate more nodes but they also extend the duration of diffusion. Various variants of sequential seeding resolve the trade-off between the coverage and speed of diffusion differently.	cs.SI	cs	Balancing Speed and Coverage by Sequential Seeding in Complex Networks Jarosław Jankowski1,2,, Piotr Br ´odka1, Przemysław Kazienko1, Boleslaw K. Szymanski1,3, Radosław Michalski1, and Tomasz Kajdanowicz1 1Wrocław University of Science and Technology, Department of Computational Intelligence, Wrocław, 50-370, Poland 2West Pomeranian University of Technology, Department of Computer Science and Information Technology, Szczecin, 71-210, Poland 3Rensselaer Polytechnic Institute, Department of Computer Science, Troy, 12180, United States [email protected] ABSTRACT Information spreading in complex networks is often modeled as diffusing information with certain probability from nodes that possess it to their neighbors that do not. Information cascades are triggered when the activation of a set of initial nodes -- seeds -- results in diffusion to large number of nodes. Here, several novel approaches for seed initiation that replace the commonly used activation of all seeds at once with a sequence of initiation stages are introduced. Sequential strategies at later stages avoid seeding highly ranked nodes that are already activated by diffusion active between stages. The gain arises when a saved seed is allocated to a node difficult to reach via diffusion. Sequential seeding and a single stage approach are compared using various seed ranking methods and diffusion parameters on real complex networks. The experimental results indicate that, regardless of the seed ranking method used, sequential seeding strategies deliver better coverage than single stage seeding in about 90% of cases. Longer seeding sequences tend to activate more nodes but they also extend the duration of diffusion. Various variants of sequential seeding resolve the trade-off between the coverage and speed of diffusion differently. Introduction The process of making the complex decisions is difficult, so it is often worth making partial decisions and to track their consequences before proceeding further. Such strategy was proven useful in areas such as: general theory of decision making,1, 2 financial markets,3, 4 epidemiology5 and marketing.6 Here, we show that sequential, consecutive approach is also highly efficient in choosing the individuals, called seeds, that when activated will widely spread information or opinion in a social network. The current research on influence maximization and information spread in complex networks focuses mainly on single stage seed initiation. An exception is new product adaptation with early diffusion of product samples7,8 to benefit from consumer responses and product spread. The main challenge is finding a method for selection of seeds to maximize the final spread of information within the network. If the total number of seeds to be used is limited, e.g. due to restricted budget, a typical approach is to rank all nodes in the network according to some criteria, select top n nodes as seeds and activate them at once to initiate the diffusion. Influence maximization problem in complex networks was defined by Kempe.9 Analyses of various factors affecting the diffusion and social influence in complex networks include the efficiency of using different centrality measures for ranking influencers for selection,10 impact of homophily for successful seeding,11 and heterogeneous thresholds on congestion,12 finding the critical initiator fraction beyond which the cascade becomes global13 or importance of different network features in predicting spread.14 Selection of initial seeds was also analyzed, including incentives for innovators to start diffusion15 and the multi-market entry perspective.16 While most of research is related to marketing, the problem has been generalized to ranking nodes for target set selection in the domain of combinatorial optimization of theoretical computer science.17,18,19 The influence maximization problem is also explored in physics from the perspective of diffusion in complex networks.20 Other studies discuss the role of communities21 and propose the influence maximization through optimal percolation.22 Some initial research has been carried out to identify efficient seeds to initiate spread in temporal23,24 and multilayer social networks.25 Several comparative studies discuss seeding strategies26 and marketing word of mouth programs.28 Typically, seeding strategies applied in viral marketing, social campaigns, political campaigns, diffusion of information, etc. are based on the assumption that selected seeds are initiated at once, and then the diffusion continues without any additional support.29 Resent research takes into account adaptive approaches with two- stage stochastic model exploring the potential of neighboring nodes,30 further extended towards more scalable approach.31 Other preliminary studies were performed to demonstrate the potential of using seeds after the first stage.32, 33 The potential of multi-period spraying algorithm for routing in delay-tolerant networks was also discussed.34 Some authors noticed the benefit of not initiating a node that is likely to be activated by other seeds. Hence, they proposed structural heuristics for static seed ranking that avoid choosing as a new seed (1) a node that is a neighbor of or close to the already chosen seed,27,36 or (2) a node that is in a local cluster already containing another seed.37,21 Here, we introduce sequential seeding for influence maximization in complex networks and compare it to classical single stage approach. In our approach seeds are divided into several packages, and seeds from each package start either in fixed intervals or when diffusion stops for the previous package. We verify the performance of the proposed approach for different parameters related to network structure and characteristics of diffusion. Our results show that deploying the same number of seeds in stages as in at once approach yields larger spread, than achieved in a single stage seeding thanks to use of diffusion between stages. This improvement is however offset by the longer time needed to finish diffusion and to reach the maximal coverage. The best sequential seeding strategy depends on the relative importance of the spread coverage and the time available for spread execution. A typical single seeding approach starts with ranking all nodes according to some structural features ranging from a simple degree ranking29 to more advanced solutions.27,36,37,21 Actual seed selection commences with initiating all n top ranking nodes at once at the beginning of campaign. In contrast, sequential seeding utilizes the same ranking as the corresponding single stage method but initiates nodes that are top ranked and not yet initiated in stages. Between those stages, the diffusion proceeds from initiated nodes. Thus, any top ranking nodes initiated by such diffusion is not used as seeds and in its place other high ranked but uninitiated node is chosen to be activated. The diffusion model that is used here is stochastic. Even for single stage seeding, when the same set of seeds is initiated with repeated executions, the coverage will vary as each activated node will attempt to activate its neighbor with certain probability, but no certainty. Still, it can be proven that there broad sufficient conditions under which there exists a sequential seeding using the same node ranking method with average coverage exceed- ing that of the single stage seeding. One such sufficient but not necessary condition is that there is a seed in the single stage seeding that is reachable from other seeds by with probability smaller than 1. Indeed, by initiating all seeds except such reachable seed in the first stage, we can observe the following. For all execution cases in which uninitiated seed is not activated by diffusion after the first stage, the average coverage over such cases will be the same for both methods. However, in all other execution cases (guaranteed to exist since the node is reachable with probability less than 1), the single stage seeding will have average coverage equal to the coverage of the first stage of the sequential seeding. However, in the second stage, the latter will initiate a new node in place of already active reachable seed (and perhaps more nodes in the subsequent diffusion), making the overall average coverage of the sequential seeding larger than the single stage one. Results Sequential seeding strategies are considered here for an independent cascade model representing stochastic dif- fusion of information over the network initiated by seeds.9 A basic, commonly used, single stage (SN ) seeding consists of only one stage with n seeds. A seeding stage starts with activation of all n seeds at the beginning to initiate diffusion and lasts until diffusion stops. According to the definition of independent cascade model each diffusion step consists of a single attempt performed by each node activated in the previous step to activate this node direct non-active neighbors with a given propagation probability. Due to the fact that we measure diffusion time in the number of steps, we assume that each such step takes a unit time to execute regardless of the network size. It means that time corresponds directly to the number of diffusion steps. Let TS N denote the number of diffusion steps of single stage seeding. All seeding methods considered here use the same fixed number of n seeds. Five sequential seeding methods are introduced here as presented in Table 1; see the Methods section for more details. In the single stage method, the n seeds are utilized at once at the beginning, whereas new sequential approaches deploy them in a sequence of multiple consecutive stages. In general, there are two types of proposed sequential seeding. The first type uses constant seeding stage time and includes an approach denoted as SQ kPS with k seeds initiated in each seeding stage. The other approach of this type denoted SQ TSN , uses the number of stages defined by the seeding time of single stage approach. The second type uses revival mode, in which the stage ends when the diffusion finishes and the next seeding stage starts immediately thereafter. This type includes 2/13 approaches denoted as SQ kPS R, SQ TSN R and the revival mode with buffering denoted as SQ kPS B. Strategy Description SQ kPS k Seeds Per Stage -- the sequential seeding consisting of n stages; each activating k additional seeds at its beginning and lasting one diffusion step SQ kPS B SQ kPS R k Seeds Per Stage with Revival -- with k seeds invoking the stage and reviv- ing the previously stopped diffusion; the stage ends when the diffusion stops k Seeds Per Stage with Buffering -- the seed set is distributed equally: kk seeds are assigned to each diffusion step which starts with k seeds; the next seeding stage is applied after the diffusion started after the previ- ous stage stops; the buffer collects for later use the unused seeds initi- ated by diffusion invoked by the previous seeding stages; the last seed- ing stage is in the nth diffusion step Time from Single Stage -- the number of steps TS N in the corresponding single stage seeding (SN) defines the number of seeding stages; seeds are distributed equally -- the number of seeds in each stage is the total number of seeds n divided by the sequence length: n/TS N SQ TSN SQ TSN R Time from Single Stage with Revival -- the seeds are allocated among stages the same way as in SQ TSN, but a new stage starts and another set of (n/TS N ) seeds is initiated after the previously started diffusion stops ∗∗ ∗∗ No. of seeding stages n n Steps per stage∗ 1 ≥ 1 ≤ n ≤ n min(n, TS N )∗∗ 1 min(n, TS N )∗∗ ≥ 1 Table 1. New seeding strategies, all based on sequential approach either without or with revival mode; n - denotes the fixed number of seeds, while TS N - the number of diffusion steps in the corresponding single stage approach (SN); ∗ the number of steps in the last seeding stage is not defined and can be greater than 1, since the diffusion proceeds as long as possible. ∗∗ if n < TS N , k Seeds Per Stage (SQ kPS) approach with n stages is used; otherwise seeding consists of TS N stages. Figure 1. (A) k per stage sequential seeding strategy (SQ kPS) with k = 1 compared with the single stage approach (SN). (B) Sequential strategy based on the reference time (SQ TSN) compared with the single stage approach (SN). (C) Sequential strategy with revival mode (SQ TSN R) compared with the single stage approach (SN). The example of one seed per stage sequential seeding (SQ 1PS) in the relation to single stage seeding (SN) with the same diffusion parameters is shown in Fig. 1(A). The use of a single seed in each step of simulation instead of using all seeds at the beginning results in the increased coverage CS Q kPS , when compared to single stage coverage CS N . Time -- more precisely the number of diffusion steps -- when the maximal coverage is reached, denoted as TS Q 1PS is also typically longer than the reference time for the single stage TS N . Time TS Q 1PS ,S N denotes the number of steps when SQ 1PS process reaches the maximal coverage CS N from single stage process. 3/13 The second approach (SQ TSN) is illustrated in Fig. 1(B), uses shorter sequence with the total number of stages equal to the number of steps needed to reach maximal coverage CS N in the single stage seeding (SN). The number of seeds s used in each step is equal to the total number of available seeds n divided by the number of steps TS N in the corresponding single stage process. An extension, based on the revival mode presented in Fig. 1(C), uses additional seeds once the diffusion dies out and needs to be reseeded. The revival mode may also be extended with buffering: the number of seeds designated for initiation but not used because of their initiation in earlier diffusion steps is cumulatively stored in the buffer and the equivalent number of new not activated by diffusion nodes are initiated as seeds when the current stage diffusion terminates. Sequential seeding uses dynamic selection and any node ranking method including simple heuristics or their extensions can be applied. To show the advantages of sequential seeding over single stage approach, we use for both the same nodes ranking, and compare their performance for the following five rankings: random selection (R), degree (D), second level degree (D2) -- it takes into account node degree and the degrees of its neighbors, Page Rank (PR) and eigenvector (EV). In each seeding stage, only nodes that have been not activated yet are selected in the order of their initial ranking as seeds, making the sequential seeding set often different from the set used by the single stage approach. Rankings of nodes may be computed at the beginning of the process based on the structural measures within the initial network. The results achieved with the sequential seeding strategies are compared, Table 2, with the single stage seeding (SN) for 1,875 versatile configurations. For each of the 15 networks, we use the same diffusion parameters (five propagation probabilities), the total number of seeds (five values of SP) and five node rankings. The simulation outcomes from all sequential strategies compared with single stage approach manifests the improvement in 89.8% configurations, with the largest benefit, in terms of better coverage, of 10.1% on average but exceeding 50% in some cases. Simultaneously, sequential methods increase the diffusion duration by 5.2 times on average. It is worth noticing that the average coverage achieved with sequential seeding SQ 1PS R applied to any seed ranking method is greater than the average coverage provided by the single seeding based on the same ranking. We observed two main factors affecting the performance of sequential seeding. The first is the extent to which sequential seeding avoids selecting nodes that would be activated through diffusion. The gain arises when a top ranked but already activated node is removed from selection and in its place another not yet activated node becomes a seed, Fig. 3(B), ), stages IV -- VI. It is most likely to happen in happen in the revival mode, which maximally delays the next seeding stage until diffusion stops in the current stage. The performance improvement depends also on selecting new seeds that have high potential of activating their neighbors in place of those seeds that were already activated by diffusion. The gain from sequential approach is commonly greater for configurations with low overall coverage, i.e. low propagation probability, see Fig. 2(D2), networks N1, N4, N5, N7 described in Supplementary Information, Fig. 2(D4) as well as for large seed sets, see Fig. 2(D3). Distributing the seeds over time makes it possible to take advantage of the progress the diffusion made before new seeds are selected. It is important especially when the budget, i.e. the number of initial seeds, is fixed. In such a case under single seeding approach the entire budget of seed is consumed at the outset and recovery is not possible if diffusion stops without reaching the desired coverage. Gradual spending allows to fund a campaign for a longer time period and to reach more customers. The resolution of trade-off between the final coverage and speed of diffusion depends on individual preferences for the importance of both coverage and the process duration in specific applications. Better coverage can be achieved by sequential approach, but it takes more time to finish or even to reach the same coverage as for single stage seeding. The more stages the sequential seeding process has, the more nodes might be activated, but the longer diffusion lasts. If the process is optimized for time, the length of sequence should be short. For example during the elections, time is fixed and all possible efforts should be taken to acquire as many voters as possible by the voting date. With high speed epidemics, vaccination should be applied quickly to protect as many people as possible. In critical situations, when initial actions are not sufficient, additional resources can be engaged and additional seeds activated. An experimental setup runs agent-based simulations on 15 static real networks N1-N15, containing from 899 to 16,726 nodes and from 2,742 to 147,547 edges, see Supplementary Information for details. The independent cascades model (IC)9 was used with propagation probability PP(a, b) that node a activates (influences or infects) node b in the step t+1 under condition that node a was activated at time t.35 The main reason for selecting this model was a relatively small number of seeds needed in each stage for the optimal decomposition of initial set of seeds. With this model even a single seed can induce diffusion, while in linear threshold model9 (LT) a small seeds packages would often have no effect. Parameters used in simulations characterizing diffusion, 4/13 networks and strategies are presented in Table 2. Symbol Parameter No. of distinct values Variants N PP SP S Network Propagation probability Seeds percentage Node ranking 15 5 5 5 for seed selection Real networks from various areas 0.05, 0.1, 0.15, 0.20, 0.25 1%, 2%, 3%, 4%, 5% random (R), degree (D), second level degree (D2), Page Rank (PR), eigenvector (EV) Table 2. Networks and parameters of diffusion used in simulations Simulation parameters create an experimental space N PP SP S, resulting in 1,875 configurations. Each configuration was independently applied for each seeding strategy from Table 1 and the results were av- eraged over 100 runs for each case. The agent based model was used, with agents connected according to the networks specifications. Each step of simulation includes the selection of additional seeds (if additional seed- ing is allowed), visiting all nodes activated in the previous step and newly selected seeds, and finally activating their neighbors according to the propagation probability (PP). Within the experimental space sequential seeding algorithms were tested including constant seeding stage time and revival mode, see Table 1. The simulation results achieved in sequential seeding were compared with the single stage seeding (SN) on the same network (N) and with the same parameters including propagation probability (PP), seeding percentage (SP) and node ranking strategy (S). Reference values for comparison were based on the coverage achieved in the single stage seeding CS N with the duration representing the number of steps in the single stage process TS N when the coverage CS N is achieved, see Fig. 1 and the Methods section. During experiments several variants of sequential seeding approach were verified including k seeds per stage with k equal to 1, 2, 4 and 8 with both revival (SQ kPS R) and non-revival mode (SQ kPS). For k = 1 buffered approach was used (SQ 1PS B). Reference based strategy was verified with revival (SQ TSN R) and non-revival mode (SQ TSN). Finally 11 variants of sequential strategies were used. The best results in terms of coverage were delivered by the sequential strategy with k = 1 seed used per stage with revival mode (SQ 1PS R) but they required the largest number of steps to finish. The opposite results with the shortest duration and the lowest coverage were achieved by the sequential approach with reference time from single stage (SQ TSN), see Fig. 5(D) for the results from both strategies. The SQ 1PS R method achieved the highest decomposition of the seeding and the highest usage of single seeds which resulted in the highest coverage, but it also have diffusion taking the longest time to finish. For SQ TSN, the number of seeding stages and the coverage ware usually low. Both strategies were compared with the coverage of single stage strategy (SN) achieved for the same simulation parameters, Fig. 2(A). The results from 1,875 configurations for both strategies are ordered by coverage obtained in the single stage method represented by red line. For the parameters with relatively low coverage, below 5%, results for the first 250 cases for all strategies are quite similar since the differences are small, while with growing coverage the differences are much more visible. Sequential seeding is almost always better than its single stage equivalent with the same parameters. The global results for all networks, strategies and parameters show better results than SN in 95.3% of simulation cases delivered by SQ 1PS R. The gain was also observed in 85.49% cases for SQ TSN. Fig. 2(B) illustrates coverage performance of sequential seeding strategies, SQ 1PS R and SQ TSN represented as a ratio of coverage obtained by these two methods to the one achieved the single stage (SN). The length of sequence positively affects the total results. In 1,697 cases from 1,875 (90.5%), one seed per stage method (SQ 1PS R) outperformed seeding with shorter sequences (SQ TSN). The example comparison of simulation results between SQ TSN, SQ 1PS R and single stage seeding is presented in Fig. 2(C). Sequential strategies outperform the single stage approach even though they need more time (steps) to achieve that. The results from simulations demonstrate that the improvement can even exceed 50% with the use of the same number of seeds as in the single stage seeding. The coverage of diffusion in SQ 1PS R is on average 13.1% better than when using SN approach with the same parameters. The SQ TSN increased the coverage, on average by 7.15%, comparing to the single stage seeding. The performance of proposed methods is dependent on parameters of diffusion and network characteristics. The results are also dependent on seed ranking strategy used. For example, the coverage gain and its distance from the results of the single stage seeding for both SQ 1PS R and SQ TSN was the highest for degree based ranking and the lowest for random node ranking as seen Fig. 2(D1). The overall coverage level depends only little on strategy or seed percentage and it is much more impacted by propagation probability and network profile. Experiments were 5/13 Figure 2. (A) Average coverage from all 1,875 configurations for k = 1 per stage seeding (SQ 1PS R) and the sequential approach based on single stage time reference (SQ TSN), compared with the single stage method (SN) for all simulation cases ordered by single stage coverage; (B) Performance of k = 1 per stage seeding (SQ 1PS R) and reference time based seeding (SQ TSN) in the relation to single stage (SN) for all configurations; (C) Cumulative coverage results from 100 steps of simulation with degree based ranking (D) for network N2 with propagation probability PP=0.1 and seeding percentage SP=5%; (D1) Coverage increase for SQ 1PS R and SQ TSN based on various seed ranking methods: random R, degree D, second level degree D2, Page Rank PR, Eigenvector EV; red part shows single stage seeding (SN) coverage, blue part and font show sequential SQ TSN improvement while green part and font show sequential SQ 1PS R gain; (D2) Coverage increase for SQ 1PS R and SQ TSN for all used propagation probabilities with values 0.05, 0.1, 0.15, 0.20 and 0.25; (D3) Results for seed selection percentages from the range 1% to 5% for SQ 1PS R and SQ TSN; (D4) Average coverage from single stage seeding together with coverage increase from sequential seeding SQ 1PS R and SQ TSN for networks N1 - N15; (E) Distribution of simulation cases in terms of coverage and time ratios for k = 1 per stage seeding (SQ 1PS R) and reference time based seeding (SQ TSN) for all networks, where TS Q and CS Q denote time and coverage in appropriate sequential seeding: SQ 1PS R or SQ TSN, while TS N and CS N stand for time and coverage in single stage seeding carried out on a wide range of networks and parameters, including cases with very low performance/coverage, e.g. the propagation probability PP=0.05 or seeding percentage SP=1%. For such parameters, it is very difficult to trigger a cascade regardless of seeding strategy used. The highest improvement was observed for the lowest propagation probability PP=0.05 at 13% for SQ 1PS R compared to single stage seeding with SQ TSN delivering 8% improvement. With the highest propagation probability PP=0.25 the improvement decreases to 7% and 4% for both strategies respectively, see Fig. 2(D2). In configurations with high propagation probability, diffusion reaches high coverage in a very short time, regardless of what seeding strategy is used - there is almost no space for any improvement. In terms of seeding percentage (SP) performance of sequential seeding grows together with the number of seeds selected. The lowest seeding percentage SP=1% delivered 4% and 2% improvement while for the SP=5% the improvement increases to 12% and 8% respectively, Fig. 2(D3). In general, the obtained results were strongly dependent on the network profile with the overall 6/13 coverage and performance of sequential seeding illustrated in Fig. 2(D3) and in Fig. 5(A). One of the reasons for a generally good coverage performance of sequential methods, is related to identifying new seeds with the highest potential to activate other nodes. Sequential seeding efficiently utilizes the potential of diffusion, but spreading seeds over time increases the diffusion duration, see Fig. 2(C). The greatest coverage can be achieved by means of one per stage strategy with revival SQ 1PS R, but it results in the longest duration since every single seed waits until diffusion started by the previous seed terminates. On the other hand, the coverage performance of the SQ TSN strategy is lower, but it takes less time to reach the coverage greater than in the single stage approach, Fig. 2(C). The final decision which strategy to use depends on how the trade-off between time and coverage is resolved. If the process is restricted in time, the seeding sequence length should be shorter. For SQ TSN, the diffusion is on average only 1.4 times longer than for the single stage approach SN whereas the duration of the longest sequences based on SQ 1PS R is on average 11.9 times longer. Gain in coverage (positive coverage ratio) and loss in time (positive time ratio) for both sequential strategies seen in Fig. 2(E) presents a trade-off between acceptable extended time and desired increased coverage for two sequential methods SQ 1PS R and SQ TSN shown in Fig. 2(E). It can be seen that SQ 1PS R provides greater variety of possible solutions and much greater gain in coverage, at the cost of the substantial expense of speed. While the results from two opposite approaches SQ 1PS R and SQ TSN are presented in this section, their extensions based on revival mode and buffering are described in the Discussion section. The SQ kPS strategy and usage of k parameter other than one (2, 4, 8) makes it possible to accordingly adjust both duration and the coverage. Methods The common approach for seeding in complex networks uses a single stage seeding (SN), Fig. 3(A). In the first stage, all n initial seeds are selected using a given node ranking. In the next and last stage, diffusion starts, and then continues without any additional support until it stops at certain time to which we refer as saturation time TS N , achieving coverage CS N . An sample network with 30 nodes was generated using Baraba´si-Albert model. At the beginning, n = 6 nodes with the highest degrees, representing 20% of all nodes, are selected as seeds. Then diffusion is simulated using the Independent Cascades model with propagation probability P=0.5, see Fig. 3(AI) and (AII). In the sequential seeding approach, introduced here, the seeds are split among several seeding stages inter- spersed with diffusion stages so each seeding stage is followed by diffusion stage. The selection of seeds in each stage is based on inactive node ranks. The first sequential method -- one seed per stage (SQ 1PS) utilizes the highest distribution of seeds assigning a single seed to each seeding stage. The number of stages equals the num- ber of seeds - n. To illustrate this process the same network with the same number of seeds as for SN is used in Fig. 3(B). The seeding consists of six stages with one seed per stage. In each stage, a new seed with the highest degree is selected from not activated yet nodes. It means that the already activated nodes are not considered for selection, even if they possess the high rank. This allows avoiding to seed nodes that would be activated anyway by diffusion. The diffusion with sequential seeding (SQ 1PS) reaches 24 nodes, while the single stage seeding (SN) activates just 18. All those nodes, except node 11 that have been activated by the SN method, were also activated by sequential strategy which in addition activated seven more nodes. However, since diffusion is simulated using the probabilistic Independent Cascades model, the outcome of sequential seeding strategy might be different, or even worse, than the final result of single stage seeding strategy. Nevertheless, as proved by experimental results, the sequential seeding strategy outperforms the single stage method in over 90% of cases. The higher coverage and the longer duration compared to single stage SN shown in Fig. 4(A) is illustrated for the k per stage sequential seeding process SQ 1PS in Fig. 4(B). The questions arise what is the increased coverage COPS how much it is greater than CS N and how much longer is the time TOP S to achieve this coverage greater than TS N . Additionally, we can assess the lower speed of the SQ 1PS process by measuring the time TOP S ,S N when the maximal coverage CS N from the SN process is reached. Commonly, TOPS ,S N >TS N . Another sign of lower speed is that the coverage CS N ,1PS of the SQ kPS process is achieved in time TS N and typically, CS N ,kPS <CS N . While the sequences with the SQ 1PS are relatively long, the second proposed approach is based on the refer- ence time TS N from the single stage seeding applied to sequential approach as SQ TSN, Fig. 4(C). The number of seeds used in each step is equal to the number of seeds s used in single stage seeding divided by the number of steps TS N needed to reach the saturation point in SN. The sequence of seeding starts at the first stage and finishes at T th S N stage. The coverage CT S N ,S N of the SQ TSN process at the time TS N can be compared with CS N . The slower 7/13 Figure 3. (A) Single stage seeding process (SN) with top degree ranking. (I) An initial set of 6 seeds (red color) consists of the nodes with the highest degree ranking. The seeds activate with probability PP=0.5 their neighbors causing 10 of them to become active. In the next step (II), all newly activated nodes are contacting their non-active neighbors, activating two of them. This process ends with the total of 18 activated nodes. (B) One per stage sequential seeding strategy (SQ 1PS) with top degree node ranking. (I) A node with the highest degree is selected as the first seed. Since four nodes have the same degree, node 7 is selected randomly. It tries to activate its neighbors with PP=0.5 and nodes 2 and 9 become active. Node 2, seeded in SN (AI) is activated by diffusion. (II) Node 3 is the second seed. At this stage, three nodes being seeds in the SN process, i.e. 1,2,6 are activated by other nodes; three other nodes can take their place as seeds. (III) Node 4 is taken as the third seed. (IV) The fourth seed is node 24 -- it did not belong to seeds in the SN process. It could become a seed because all nodes preceding it in the ranking have already been activated. So far, 20 nodes are activated in four steps comparing to 18 in the SN strategy, and there are two more seeds left to activate. (V) The fifth seed is node 11. Since the other nodes do not activate any new ones, the process would terminate now with totally 22 activated nodes. However, one more seed is still available. (VI) The last seed is node 22 which activates node 27. It cannot activate anyone else, so diffusion stops with 24 activated nodes. New parts of the network reached with the sequential seeding are marked. speed of the SQ TSN means that CT S N ,S N <CS N . However, CT S N ,S N >COPS ,S N , i.e. SQ TSN is faster than SQ kPS. It can be observed that TOP S ,S N >TT S N ,S N >TS N . On the other hand, regarding the total coverages, we have: COPS >CT S N >CS N . Overall, the greater coverage is, the lower speed is. Sequential seeding with revival and buffering modes. SQ kPS and SQ TSN are able to improve coverage thanks to the allocation of additional seeds, even if diffusion is still ongoing. To utilize the acquired knowledge about pro- 8/13 Figure 4. (A) Diffusion based on the single stage seeding. (B) k per stage sequential seeding strategy SQ kPS. (C) Sequential seeding based on reference time TS N -- SQ TSN. (D) k per stage sequential seeding strategy with the revival mode SQ kPS R. (E) The TS N -based sequential seeding with the revival mode SQ TSN R. (F) k per stage sequential seeding with buffering SQ kPS B. cess dynamics, the sequential strategy with revival mode is proposed and applied to modify sequential strategies with constant seeding stage time: SQ kPS R and SQ TSN R. In both cases, an additional seeding is suspended un- til the most recent diffusion stops, i.e., the increase of coverage in two following steps ∆C is zero, see Fig. 4(D) and (E). It means that the proposed approaches extend the execution time of individual seeding stages and the total process duration. Only one seed per stage is used in SQ kPS R and similarly to SQ TSN, the allocated package of s = n/TS N seeds is used in SQ TSN R. According to the nature of the independent cascades model, each recently activated node has only one chance to infect its neighbors. Hence, if the process stops, such a node would never be re-activated in the future stages. The only way to restart diffusion is by initiating additional seed (SQ kPS R) or seeds (SQ TSN R). This can improve the final diffusion coverage, but at the price of spread increased duration. Yet another proposed approach SQ kPS B is based on one seed per stage method and accumulates in a buffer the number of seeds unused because they were initiated by diffusion. The equal number of inactive nodes is initiated after diffusion terminates and requires recovery. The main goal of SQ kPS B is to limit the process duration and make it equal to SQ kPS, duration, simultaneously using the potential of diffusion, Fig. 4(F). If we compare SQ kPS B with SQ kPS, the diffusion time of both is similar but the coverage for the buffering mode -- SQ kPS B is commonly greater because of the better exploitation of diffusion. Discussion Sequential seeding with one seed per stage SQ 1PS outperforms coverage of the single stage seeding SN in over 85% of configurations, providing on average nearly 9% better coverage but its duration was over 7 times longer, Table 3. The revival mode applied to single seed per stage (SQ 1PS R) achieves the best coverage among all sequential methods - better by over 13% on average with gained observed for more than 95% of configurations but lasts the 9/13 SQ 1PS R compared to other sequential Cases Duration increase with SQ Method SQ 1PS SQ 2PS SQ 4PS SQ 8PS SQ 1PS R SQ 2PS R SQ 4PS R SQ 8PS R SQ TSN SQ TSN R SQ 1PS B Average Sequential compared to single stage SN Cases Duration increase with TS Q/TS N Coverage increase CS Q > CS N CS Q vs. CS N 85.8% 86.9% 84.3% 82.2% 95.3% 95.1% 94.2% 93.1% 85.5% 94.2% 91.3% 89.9% 8.8% (∆=67.6) 9.0% (∆=66.1) 8.0% (∆=61.3) 7.0% (∆=53.9) 13.1% (∆=85.5) 12.9% (∆=85.9) 12.1% (∆=82.2) 11.1% (∆=79.2) 7.2% (∆=54.1) 11.4% (∆=77.7) 10.7% (∆=76.1) 10.1% 7.3 (∆=102) 3.9 (∆=46.5) 2.3 (∆=20) 1.6 (∆=8.0) 11.9 (∆=176) 7.7 (∆=107) 5.2 (∆=65.5) 3.6 (∆=40.5) 1.4 (∆=7.5) 3.2 (∆=42) 9.3 (∆=133) 5.2 CS Q 1PS R > CS Q 82.5% 85.3% 87.9% 89.0% -- 51.4% 59.0% 66.8% 90.5% 65.5% 63.2% 74.1% Coverage increase CS Q 1PS R vs. CS Q 6.2% (∆=14.2) 3.8% (∆=17.2) 4.8% (∆=22.8) 5.9% (∆=30.5) -- 0.2% (∆=0.5) 0.9% (∆=3.2) 2.1% (∆=7.4) 5.5% (∆=31.5) 1.8% (∆=8.2) 2.2% (∆=7.0) 3.3% TS Q 1PS R/TS Q 1.7 (∆=66) 2.9 (∆=126) 4.7 (∆=155) 6.7 (∆=168) -- 1.5 (∆=68) 2.1 (∆=108) 2.9 (∆=134) 8.8 (∆=169) 3.5 (∆=134) 1.3 (∆=35) 3.6 Table 3. Coverage and duration of various sequential methods compared to the worse single stage seeding SN and the best sequential approach SQ 1PS R; ∆ - Hodges-Lehmann estimator used in Wilcoxon signed rank test as a measure of the difference between two result groups; values ∆ > 0 demonstrate significantly higher values for coverage and duration of all sequential strategies compared to single stage approach SN and SQ 1PS R compared to other sequential approaches; all analysis achieved statistical significance p-value <2e-16. longest (over 10 times longer than SN). In general, the revival mode applied to any k per stage methods (SQ kPS) outperforms single stage approach in 10% more configurations with 43%-59% increase of coverage gain compared to their non-revival version, see also Fig. 5(B). The similar improvement is observed for application of revival mode to SQ TSN, Fig. 5(B2). The revival mode provides greater coverage than its reference version since it makes better use of interstage diffusion. On the other hand, the revival mode significantly extends the process duration: by 60% to 130%. The buffered mode SQ 1PS B gains little compared to the regular one per stage method SQ 1PS at the small expense of time so it provides limited SQ 1PS improvement, much weaker than provided by the revival mode, Fig. 5(B1). The experiments include some configurations with very low final coverage, which result from low propagation probability PP=0.05 combined with low seeding percentage SP=1%. It means that the total number of activations will always be low, no matter what strategy is used. For example, for a small network N10 with SP=1% only 9 seeds are activated, so it is very difficult to induce any diffusion with PP=0.05. An opposite case arises for networks with very high node degrees and relatively high propagation probability, i.e. PP=0.25 and SP=5%, where all strategies perform very well resulting in a high number of activated nodes in a very short time. Taking into account the above conditions, an increase by 13% on average for SQ 1PS R can be considered substantial. Among all sequential strategies SQ TSN is the fastest -- it increases duration in reference to SN by only 1.4 times, but its coverage gain is relatively little: 7%. Hence, two strategies: SQ 1PS R and SQ TSN can be treated as upper and lower sequential boundaries, respectively, both for coverage and time, Fig. 5(D). The space between them may be explored by using different values of seeds per stage k either with or without revival: SQ kPS R or SQ kPS for k = 2, 4, 8, etc. Overall, the experimental results reveal that the proper seeding strategy is a trade-off between time and cover- age, which can be resolved according to the user needs by choosing appropriate sequential seeding strategy. Fig. 5 and Table 3 provide details on performance of the introduced sequential seeding strategies in terms of trade-off values. The greatest coverage can be achieved by means of longer sequences resulting in longer execution. If the process is limited in time that is very short, the sequential strategies may be less justified. Here we showed both experimentally and theoretically that the average coverage of any strategy of ranking nodes for selection of all seeds at the beginning of the process can be improved with sequential seeding using the same ranking. This is because the proposed method is dynamic as it adapts the set of activated seeds to the progress of the execution of diffusion active between stages of seeding. 10/13 Figure 5. (A) Coverage for SQ 1PS R and SQ TSN compared with single stage (SN) individually for each network; (B) Speed and coverage ratios for one per stage strategy with(SQ 1PS R) and without revival mode (SQ 1PS) in reference to single stage approach SN; TS Q, CS Q -- time and coverage in suitable sequential seeding;(B1) Time and coverage ratios for sequential strategy with revival SQ 1PS R vs. sequential strategy with buffering SQ 1PS B; (B2) Reference based sequential strategy SQ TSN vs. SQ TSN B ; (C) Coverage ratios for SQ kPS strategies with 2, 4 and 8 seeds per stage in relation to one per stage approach SQ 1PS; (C1) Duration of diffusion for SQ kPS strategies with 2, 4 and 8 seeds per stage in relation to one per stage approach SQ 1PS; (C2) Duration of SQ kPS R strategies with revival in the relation to SQ kPS without revival; (D) Normalized average time and coverage ratios for all used sequential strategies; (D1) Normalized ratios for random (R) node ranking; (D2) Normalized ratios for degree based (D) node ranking; (D3) Normalized ratios for second level degree (D2) based node ranking; (D4) Normalized ratios for PageRank (PR) based node ranking; (D5) Normalized ratios for eigenvector based node ranking. References 1. Wald, A. Sequential Analysis (John Wiley and Sons, 1947), 1st edn. 2. Siegmund, D. Sequential analysis : tests and confidence intervals. Springer series in statistics (Springer-Verlag, New York, 1985). 3. Cyert, R. M., DeGroot, M. H. & Holt, C. A. Sequential investment decisions with bayesian learning. Manage- ment Science 24, 712 -- 718 (1978). 4. Prastacos, G. P. Optimal sequential investment decisions under conditions of uncertainty. Management Science 29, 118 -- 134 (1983). 5. Price, W. H. Sequential immunization as a vaccination procedure against dengue viruses. American Journal of Epidemiology 88, 392 -- 397 (1968). 6. Elberse, A. & Eliashberg, J. Demand and supply dynamics for sequentially released products in international markets: The case of motion pictures. Marketing Science 22, 329 -- 354 (2003). 7. Bass, F. M. A new product growth for model consumer durables. Management Science 15, 215 -- 227 (1969). 8. Rogers, E. M. Diffusion of innovations (Free Press, 2003), 5th edn. 11/13 9. Kempe, D., Kleinberg, J. & Tardos, ´E. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, 137 -- 146 (ACM, 2003). 10. Kiss, C. & Bichler, M. Identification of influencers - measuring influence in customer networks. Decis. Support Syst. 46, 233 -- 253 (2008). 11. Nejad, M. G., Amini, M. & Babakus, E. Success factors in product seeding: The role of homophily. Journal of Retailing 91, 68 -- 88 (2015). 12. Karampourniotis, P. D., Sreenivasan, S., Szymanski, B. K. & Korniss, G. The impact of heterogeneous thresh- olds on social contagion with multiple initiators. PloS one 10, e0143020 (2015). 13. Singh, P., Sreenivasan, S., Szymanski, B. K. & Korniss, G. Threshold-limited spreading in social networks with multiple initiators. Scientific reports 3 (2013). 14. Liu-Thompkins, Y. Seeding viral content. Journal of Advertising Research 52, 465 -- 478 (2012). 15. Lehmann, D. & Esteban-Bravo, M. When giving some away makes sense to jump-start the diffusion process. Marketing Letters 17, 243 -- 254 (2006). 16. Libai, B., Muller, E. & Peres, R. The role of seeding in multi-market entry. International Journal of Research in Marketing 22, 375 -- 393 (2005). 17. Ackerman, E., Ben-Zwi, O. & Wolfovitz, G. Combinatorial model and bounds for target set selection. Theoret- ical Computer Science 411, 4017 -- 4022 (2010). 18. Ben-Zwi, O., Hermelin, D., Lokshtanov, D. & Newman, I. Treewidth governs the complexity of target set selection. Discret. Optim. 8, 87 -- 96 (2011). 19. Chiang, C.-Y., Huang, L.-H., Li, B.-J., Wu, J. & Yeh, H.-G. Some results on the target set selection problem. Journal of Combinatorial Optimization 25, 702 -- 715 (2013). 20. Galstyan, A., Musoyan, V. & Cohen, P. Maximizing influence propagation in networks with community struc- ture. Phys. Rev. E 79, 056102 (2009). 21. He, J.-L., Fu, Y. & Chen, D.-B. A novel top-k strategy for influence maximization in complex networks with community structure. PloS one 10, e0145283 (2015). 22. Morone, F. & Makse, H. A. Influence maximization in complex networks through optimal percolation. Nature 524, 65 -- 68 (2015). 23. Michalski, R., Kajdanowicz, T., Br ´odka, P. & Kazienko, P. Seed selection for spread of influence in social networks: Temporal vs. static approach. New Generation Computing 32, 213 -- 235 (2014). 24. Jankowski, J., Michalski, R. & Kazienko, P. Compensatory seeding in networks with varying availability of nodes. In Advances in Social Networks Analysis and Mining (ASONAM), 2013 IEEE/ACM International Confer- ence on, 1242 -- 1249 (IEEE, 2013). 25. Michalski, R., Kazienko, P. & Jankowski, J. Convince a dozen more and succeed -- the influence in multi-layered social networks. In Signal-Image Technology & Internet-Based Systems (SITIS), 2013 International Conference on, 499 -- 505 (IEEE, 2013). 26. Hinz, O., Skiera, B., Barrot, C. & Becker, J. U. Seeding strategies for viral marketing: An empirical comparison. Journal of Marketing 75, 55 -- 71 (2011). 27. Kitsak, M. et al. Identification of influential spreaders in complex networks. Nat. Phys. 6, 888 -- 893 (2010). 28. Libai, B., Muller, E. & Peres, R. Decomposing the value of word-of-mouth seeding programs: Acceleration versus expansion. Journal of Marketing Research 50, 161 -- 176 (2013). 29. Hinz, O., Skiera, B., Barrot, C. & Becker, J. U. Seeding strategies for viral marketing: An empirical comparison. Journal of Marketing 75, 55 -- 71 (2011). 30. Seeman, L. & Singer, Y. Adaptive seeding in social networks. In Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium on, 459 -- 468 (IEEE, 2013). 31. Horel, T. & Singer, Y. Scalable methods for adaptively seeding a social network. In Proceedings of the 24th International Conference on World Wide Web, 441 -- 451 (International World Wide Web Conferences Steering Committee, 2015). 12/13 32. Sela, A., Ben-Gal, I., Pentland, A. & Shmueli, E. Improving information spread through a scheduled seeding approach. In The international conference on Advances in Social Network Analysis and Mining 2015 (2015). 33. Zhang, H., Procaccia, A. D. & Vorobeychik, Y. Dynamic influence maximization under increasing returns In Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems, to scale. 949 -- 957 (International Foundation for Autonomous Agents and Multiagent Systems, 2015). 34. Bulut, E., Wang, Z. & Szymanski, B. Cost-effective multiperiod spraying for routing in delay-tolerant networks. Networking, IEEE/ACM Transactions on 18, 1530 -- 1543 (2010). 35. Wang, C., Chen, W. & Wang, Y. Scalable influence maximization for independent cascade model in large-scale social networks. Data Mining and Knowledge Discovery 25, 545 -- 576 (2012). 36. Zhang, J.X., Chen, D.B. , Dong, Q., Zhao, Z.D.& Wang, Y. Identifying effective multiple spreaders by coloring complex networks. ci. Rep 6, 27823 (2016). 37. Zhao, X. Y., Huang, B., Tang, M., Zhang, H. F. & Chen, D. B. Identifying effective multiple spreaders by coloring complex networks. Europhys. Lett. 108, 68005 (2014). Acknowledgements This work was partially supported by Wrocław University of Science and Technology statutory funds, the Polish National Science Centre, decisions no. 2013/09/B/ST6/02317, 2015/17/D/ST6/04046 and 2016/21/D/ST6/02408; the European Commission under the 7th Framework Programme, Coordination and Support Action, Grant Agree- ment Number 316097, ENGINE - European research centre of Network intelliGence for INnovation Enhancement (http://engine.pwr.edu.pl/); the RENOIR project - Reverse EngiNeering of sOcial Information pRocessing from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 691152, the Army Research Laboratory under Cooperative Agreement Number W911NF-09-2- 0053 and by the Office of Naval Research Grant No. N00014-09-1-0607. Author contributions JJ, PB created an initial concept of sequential seeding; JJ, PB, PK, BKS, RM developed the concept of sequential seeding to its current state; JJ, PB, PK, BKS designed the experiments; JJ, PB executed all experiments and simula- tions; BKS analyzed the proposed method coverage improvement theoretically; JJ, PB, PK, BKS, TK analyzed data and discussed results; JJ, PB, PK, RM drafted the manuscript. All authors critically reviewed the manuscript and approved the final version. 13/13
1703.09693	1	1703	2017-03-28T17:56:51	Fast link prediction for large networks using spectral embedding	[ "cs.SI", "physics.soc-ph" ]	Many link prediction algorithms require the computation of a similarity metric on each vertex pair, which is quadratic in the number of vertices and infeasible for large networks. We develop a class of link prediction algorithms based on a spectral embedding and the k closest pairs algorithm that are scalable to very large networks. We compare the prediction accuracy and runtime of these methods to existing algorithms on several large link prediction tasks. Our methods achieve comparable accuracy to standard algorithms but are significantly faster.	cs.SI	cs	Fast link prediction for large networks using spectral embedding BENJAMIN PACHEV, BENJAMIN WEBB∗ Department of Mathematics, Brigham Young University, Provo, Utah ∗Corresponding author. Email: [email protected] [Received on 9 April 2017] Many link prediction algorithms require the computation of a similarity metric on each vertex pair, which is quadratic in the number of vertices and infeasible for large networks. We develop a class of link prediction algorithms based on a spectral embedding and the k closest pairs algorithm that are scalable to very large networks. We compare the prediction accuracy and runtime of these methods to existing algorithms on several large link prediction tasks. Our methods achieve comparable accuracy to standard algorithms but are significantly faster. Keywords: link prediction; graph embedding; commute time; resistance distance; closest pairs. 1. Introduction The study of networks has become increasingly relevant in our understanding of the technological, natural, and social sciences. This is owing to the fact that many important systems in these areas can be described in terms of networks [1], where vertices represent the system's individual components, e.g. computer routers, neurons, individuals, etc. and where edges represent interactions or relationships between these components. An essential feature of the large majority of these networks is that they have a dynamic topology, i.e. a structure of interactions that evolves over time [2]. The structure of social networks, for instance, change over time as relationships are formed and dissolved. In information networks such as the WWW the network's structure changes as information is created, updated, and linked. Although understanding the mechanisms that govern this structural evolution is fundamental to net- work science, these mechanisms are still poorly understood. Consequently, predicting a network's even- tual structure, function, or whether the network is likely to fail at some point are all currently out of reach for even simple networks. In an attempt to determine which processes cause changes in a network's structure we are lead to the following link prediction problem: Given a network, which of the links, i.e. edges between existing vertices, are likely to form in the near future. Here we adopt the standard convention that links are to be predicted solely on the basis of the network's current topology (see, for instance, [3]). Importantly, the link prediction problem can be used to study more than just which edges will appear in a network. It can also be used to predict which of the non-network edges are, in fact, in the network but currently undetected. Similarly, it can be used to detect which of the current network edges have been falsely determined to be a part of the network. This notion of link prediction is of central importance in numerous applications. Companies such as Facebook, Twitter, and Google need to know the current state and efficiently predict the future structure of the networks they use to accurately sort and organize data [4]. Biologists need to know whether biochemical reactions are caused by specific sets of enzymes to infer causality and so on [5]. 1 2 of 13 BENJAMIN PACHEV AND BENJAMIN WEBB The barrier in determining whether network links truly exist in these and other settings, is that testing and discovering interactions in a network requires significant experimental effort in the laboratory or in the field [6]. Similarly, determining experimentally when and where a new link will form may also be impractical, especially if the precise mechanism for link formation is unknown. For these reasons it is important to develop models for link prediction. At present, there is an ever increasing number of proposed methods for predicting network links [7]. Not surprisingly, certain methods more accurately predict the formation of links in certain networks when compared with others. Additionally, each of these methods has a runtime that scales differently with the size of the network. In our experiments, we discover that a number of link predictors have a runtime that is so high that it effectively prohibits their use on moderately large networks. Here we propose a class of link predicting algorithms that scale to large networks. This method, which we refer to as the approximate resistance distance predictor, integrates a spectral embedding of the network with a known algorithm for efficiently finding the k closest pairs of points in Euclidean space. The spectral embedding aspect of the algorithm is derived as a low-rank approximation of the effective resistance between network vertices, as in [8]. The k closest pairs component of the algorithm is taken from [9] and can be used to predict links based on this embedding. Here we compare the prediction accuracy and runtime of this method against several well-known algorithms on a number of coauthorship networks and a social network consisting of a small subset of Facebook users. We find that our method is achieves the best accuracy on some networks and scales to networks that many other link predictors cannot. The paper is structured as follows. In Section 2 we describe the link prediction problem and outline a number of standard link prediction algorithms. In Section 3 we introduce the method of resistance distance embedding and prove that it is optimal as a low rank approximation of effective resistance (see Proposition 3.1). In Section 4 we describe the experimental setup. Section 5 numerical results comparing the performance of the resistance distance embedding algorithm to other algorithms are given. Section VI concludes with some closing remarks including a number of open questions for future work. 2. The Link Prediction Problem The link prediction problem can be stated as follows. Given a connected graph G = (V, E), and k, the number of predicted nonadjacent links, we seek k pairs of vertices which are most likely to become connected. While the choice of k depends on the application, we adopt the convention that 1 6 k 6 E. The general paradigm for link prediction is to compute a similarity metric score(x, y) on each vertex pair (x, y). The predicted links are then the k (x, y) ∈ V × V − E for which score(x, y) is maximal. By contructing a matrix from the scores, we obtain a graph kernel. We can also go in the other direction. Any real n× n matrix, where n = V, defines a score function on pairs of vertices, and can be used for link prediction. We now give a sampling of existing link prediction algorithms. 2.1 Local Methods A local method for link prediction is an algorithm that uses vertex neighborhoods to compute similarity. Common Neighbors: Common neighbors simply assigns score(x, y) = Γ(x)∩Γ(y), (2.1) where Γ(x) is the neighbor set for x ∈ V . FAST LINK PREDICTION WITH SPECTRAL EMBEDDING 3 of 13 Jaccard's Coefficient: Jaccard's coefficient is a normalized version of common neighbors that takes into account the total number of neighbors for both vertices. It is given by score(x, y) = Γ(x)∩Γ(y) Γ(x)∪Γ(y) . (2.2) Preferential Attachment: Preferential attachement is based on the idea that highly connected nodes are more likely to form links, an observed pattern in coathourship networks [10]. This leads to Adamic-Adar: Resource Allocation: 2.2 Path-based Methods score(x, y) = Γ(x)Γ(y). score(x, y) = ∑ z∈Γ(x)∩Γ(y) 1 logΓ(z) score(x, y) = ∑ z∈Γ(x)∩Γ(y) 1 Γ(z) (2.3) (2.4) (2.5) Path-based methods consider all or a subset of the paths between two vertices to compute similarity. Unlike local similarity measures, they can capture global information about the network. Shortest Path: This link predictor defines score(x, y) as the negated length of the shortest path from x to y. Katz: The Katz metric counts all paths between two nodes, and discounts the longer paths exponen- tially. Define pathℓ x,y to be the set of all paths of length ℓ from x to y. Then given a weight 0 < β < 1, score(x, y) = ∞ ∑ ℓ=1 βℓpathℓ x,y (2.6) A closed form for the associated graph kernel is given by (I − βA)−1 − I = ∑∞ ℓ=1(βA)ℓ, where A is the adjacency matrix of G. 2.3 Random walks A random walk on G starts at some node x and iteratively moves to new nodes with uniform probability. There are a multitude of link predictors based on random walks. These are some of the fundamental ones. Hitting and Commute Time: The hitting time Hx,y is the expected number of steps required to reach y in a random walk starting at x. Commute time is defined as Cx,y = Hx,y + Hy,x. Negated hitting time can be used as a link predictor, but the hitting time is assymetric in general, so we use instead the negated commute time, which is symmetric. The commute time and its variants will be discussed further in Section 3. Rooted Page Rank: A problem with hitting and commute time is that random walks can become lost exploring distant portions of the graph. Rooted Page Rank deals with this problem by introducing random resets. Given a root node x, we consider a random walk starting at x. At each step, with probability α the walk returns back to x. With probability 1−α the walk proceeds to a random neighbor. Given a root node x, for each other node y, score(x, y) is defined as the stationary probability of y under 4 of 13 BENJAMIN PACHEV AND BENJAMIN WEBB the random walk rooted at x. The corresponding graph kernel is given by (1−α)(I −αD−1A)−1, where D is the degree matrix and A is the adjacency matrix. 2.4 Scaling Link Predictors to Large Networks Many link predictors, such as Katz, require the computation of a matrix inverse. This is heinously expensive for large networks, as it is cubic in the number of vertices. One way to circumvent such prob- lems is via a low-rank approximation of the score matrix. We investigate such a low-rank approximation for the commute-time or resistance distance kernel in the next section. Even the simpler local predictors such as common neighbors or preferential attachment face difficul- ties at scale. This is because for sufficiently large networks, it is not possible to compute scores for each pair of vertices and then find the maximal ones. Instead, efficient search techniques must be employed to search only a small subset of the potential links in order to find those of maximal score. In Section 3 we will demonstrate how a class of graph embedding based predictors can efficiently find the k links of maximal score. 3. Spectral Embedding We begin by deriving the approximate resistance distance link predictor as a best low-rank approxima- tion to commute time and show how to evaluate its link prediction scores with a spectral embedding. We then show that this link predictor is part of a family of graph embedding based link predictors that use the k closest pairs algorithm to efficiently find the links of maximal score. Finally, we discuss effi- cient ways to compute the spectral embedding upon which the approximate resistance distance predictor relies. 3.1 Approximating Commute Time Let L = D− A be the Laplacian matrix of a graph G = (V, E), and let n = V. Let L† be the Moore- Penrose inverse of L. Then the commute time is given by x,x + L† y,y − 2L† x,y), (3.1) Cx,y = E(L† y,y − 2L† where the quantity rx,y = (L† x,y) is known as the effective resistance or the resistance dis- tance [8]. Since resistance distance differs from commute-time by a (network-dependant) constant scaling factor, they can be used interchangeably for link prediction. x,x + L† For many networks, G is too large to compute L† exactly, so an approximation must be used. A natu- ral choice is a best rank-d approximation to L† for some fixed dimension d. The resulting approximation of the resistance distances is closely related to distances between points in Euclidean space. PROPOSITION 3.1 Let d be a positive integer and let G = (V, E) be a connected, undirected graph. Then ∃ a best rank-d approximation S of L†, and a map f : V → Rd so that ∀ x, y ∈ V , Sx,x + Sy,y − 2Sx,y = k f (x)− f (y)k2 2. We call this map the resistance distance embedding. Proof. For a connected graph, the Laplacian matrix is positive semidefinite, with eigenvalues 0 = λ1 < λ2 6 ··· 6 λn and corresponding eigenvectors v1, v2, v3, . . . , vn. Then we have the spectral decompositions L = n ∑ i=2 λivivT i FAST LINK PREDICTION WITH SPECTRAL EMBEDDING 5 of 13 and L† = n ∑ i=2 1 λi vivT i . Hence, S = ∑d+1 i=2 1 λi vivT i is a best rank-d approximation to L† in the 2-norm. Then note Sx,x + Sy,y− 2Sx,y = (ex − ey)T S(ex − ey) = d+1 ∑ i=2 1 λi (ex − ey)T vivT i (ex − ey) where f (x) = [ , = d+1 ∑ i=2 1 λi (vi,x − vi,y)2 = k f (x)− f (y)k2 v2,x√λ2 v3,x√λ3 , . . . , vd+1,x pλd+1 2 ]T ∈ Rd (3.2) (3.3) We define the approximate resistance distance link predictor of dimension d by setting score(x, y) = −(Sx,x + Sy,y − 2Sx,y) = −k f (x)− f (y)k2 2 , where S and f are defined as in Proposition 3.1. In the next section, we will see that the approximate resistance distance link predictor is part of a class of link predictors that avoid brute-force search when predicting links. 3.2 Link Prediction with Graph Embeddings The resistance distance embedding is a special case of a graph embedding, which is a map f from V to Rd, d a positive integer. We can use graph embeddings to create link predictors. A natural choice is to set score(x, y) = −k f (x)− f (y)k2, (so maximizing score corresponds to minimizing distance). We refer to this score function as the Euclidean score. If f is the resistance distance embedding, then link prediction with the Euclidean score is equivalent to the approximate resistance distance predictor. Recall that the approximate resistance distance score 2. The k predicted links of maximal score correspond to the k nonadjacent 2 is maximal. These are precisely the k links for which function is −k f (x)− f (y)k2 pairs of vertices (x, y) for which −k f (x)− f (y)k2 k f (x)− f (y)k2 is minimal and are predicted with the Euclidean score. pairs problem is as follows. Given a set of vectors {x1, x2, . . . , xn} ⊂ Rd we seek the k unordered pairs (xi, x j), i 6= j of minimal distance (here we use the Euclidean norm but any Lp norm can be used, 1 6 p 6 ∞). There is an algorithm to solve this problem in Link prediction with the Euclidean score is related to the k closest pairs problem. The closest O(d(n log n + k log n log( n2 k ))) (3.4) [9]. 6 of 13 BENJAMIN PACHEV AND BENJAMIN WEBB We can think of the link prediction problem as the closest pairs problem applied to the set of vectors { f (y), y ∈ V}, with the additional constraint that the best pairs must correspond to non-edges in G. The extra constraint can be handled by finding the E + k closest pairs, then selecting the best k which are non-edges. As there can be no more than E edges, this approach is sure to work. We then have the worst-case complexity bound of O(d(n log n + (E + k)log n log( n2 E + k ))). (3.5) plexity bound can be simplified to Recalling that we require 1 6 k 6 E, and assuming that G is connected so E > n− 1, this com- (3.6) O(d E log2n). For large, sparse networks, E << n2, and this is a tremendous speedup over the O(n2) brute-force approach. Cosine Similarity Score: Another link prediction score function that can be derived from a graph embedding is the cosine similarity score, defined by score(x, y) = < f (x), f (y) > k f (x)kk f (y)k . (3.7) If the cosine similarity score is used, the link prediction problem can still be solved without brute- force search. It is equivalent to the link prediction problem with Euclidean score on a modified graph embedding. The modified embedding is obtained from the original by normalizing the embedding vectors as follows. PROPOSITION 3.2 Given a graph embedding f : V → Rd, the link prediction problem using score(x, y) = < f (x), ( f (y)) > k f (x)kk f (y)k = cosθ is equivalent to the link prediction problem with the Euclidean score function on the modified embedding given by g(y) = f (y) k f (y)k . Proof. Let x, y ∈ V . Note We have < g(x), g(y) >= cosθ = score(x, y). kg(x)− g(y)k2 2 = kg(x)k2 = 2− 2 cosθ = 2− 2score(x, y). 2 +kg(y)k2 2 − 2 < g(x), g(y) > This shows that minimizing Euclidean distance for the modified embedding is the same as maximizing cosine similarity score on the original, so link prediction with Euclidean score on the modified embed- ding is equivalent to link prediction with the cosine similarity score on the original. This section introduced a class of link predictors that avoid a brute-force search when predicting links. These link predictors rely on a precomputed graph embedding. The graph embedding needs to be efficiently computable in order for the overall prediction algorithm to be fast. We are concerned with link predictors that rely on the resistance distance embedding. Consequently, rapid computation of this particular graph embedding is the subject of the next section. FAST LINK PREDICTION WITH SPECTRAL EMBEDDING 7 of 13 3.3 Computing the Resistance Distance Embedding Computing the resistance distance embedding of dimension d requires finding the smallest d nonzero eigenvalues and associated eigenvectors of the Laplacian matrix L. Fortunately, specialized, efficient algorithms exist for this problem which exploit the positive semi-definiteness and sparsity of L. These include TRACEMIN-Fiedler [11] and a multilevel solver MC73 FIEDLER [12]. TRACEMIN-Fiedler is simpler to implement, and is also parallelizable, so we use it in our experiments. 4. Experimental Setup In this section we compare the performance of our link prediction algorithm to others on several large social networks. In a social network, nodes correspond to persons or entities. Edges correspond to an interaction between nodes, such as coauthouring a paper or becoming linked on a social media website. 4.1 The Networks Arxiv High Energy Physics Theory (hep-th): This network is a coauthorship network obtained from the Konect network collection. [13, 14]. Arxiv High Enery Physics Phenomenology (hep-ph): This is another coauthorship network from the Konect network collection [13, 15]. Facebook Friendship (facebook): This social network consists of a small subset of facebook users, where edges represent friendships [16, 17]. Arxiv Condensed Matter Physics (cond-mat): This dataset was obtained from Mark Newman's website [18], and is also a coathourship network. Unlike the other datasets, the edges are not times- tamped. 4.2 Creating Training Graphs In order to perform link prediction, we partition edges into a training set and a test set. Edges in the training set occur before those in the test set and are used to construct a training graph. We run link prediction algorithms on the training graph to predict the contents of the test set. In most cases, edges have timestamps, and we can choose a cutoff time to partition the edges. For one network (cond-mat) the edges are not timestamped. However, there are two versions of the cond-mat network available. One contains all collaborations up to 2003. The second is an updated network with all collaborataions up to 2005. We use the first network as the training graph. The test set consists of all edges in the second network for which both nodes are in the earlier network. Choosing the cutoff between the training and test edges is somewhat arbitrary. If too few edges are used for training, link predictors will struggle. If too few are left for testing, then results may be statistically insignificant. See Table 1 for a comparison of the training networks and original networks. Our spectral embedding based link prediction algorithms require a connected graph. To solve this problem, we reduce each training graph to its largest connected component. For each network we consider, the largest component contains the vast majority of the vertices. 4.3 The Predictors We perform experiments with two spectral embedding based predictors. Each uses the resistance dis- tance embedding of dimension d, with d a parameter to be varied. The first uses the Euclidean score function and is equivalent to the approximate resistance distance predictor of dimension d. The second 8 of 13 BENJAMIN PACHEV AND BENJAMIN WEBB Network cond-mat cond-mat train facebook facebook train hep-ph hep-ph train hep-th hep-th train Nodes 15,803 13,861 63,731 59,416 28,093 26,738 22,908 21,178 Edges Average Degree 60,989 44,619 817,035 731,929 3,148,447 2,114,734 2,444,798 1,787,157 7.7187 6.4381 12.8201 24.6374 112.0723 158.1819 106.7225 168.7749 Table 1. Training network statistics uses the cosine similarity score. We refer to these link predictors as spec euclid and spec cosine respec- tively (spec for spectral). In tables, the dimension of the embedding is indicated by a number after the predictor name. For example, spec euclid8 refers to the spec euclid predictor using an 8-dimensional resistance distance embedding. The other link prediction algorithms used in our experiments are preferential attachment, common neighbors, Adamic Adar, Rooted Page Rank and Katz (with β = .01). Some networks are too large for certain algorithms to handle, so not every algorithm is run on each network. For example, the facebook training graph has 59,416 nodes. Computing the Katz score on this graph requires finding the inverse of a 59,416×59,416 matrix, and is very expensive in time and space, so we do not use the Katz algorithm for the facebook graph. All experiments were performed on the same 4 core machine. The common neighbors, preferential attachment, and Adamic Adar algorithms were implemented in Python and were not parallelized. Our spectral link predictors, Katz, and Rooted Page Rank use the Python library Numpy to parallelize linear algebra operations. All code that was used in the experments in this paper can be found at the git repository bitbucket.org/thorfax/spectral research. For each network, we fix the number of links to be predicted. With the exception of hep-th, this number is equal to 10% of the maximum possible number of correct predictions (i.e the number of new links in the test set). For the hep-th network we discovered that the spec euclid and spec cosine predictors achieve nearly perfect accuracy when predicting 1000 links. As this is not the case for any other network we considered, we report this unusual phenomenon. For all of the networks we consider, the probability of randomly predicting a correct link is very low. Most of the algorithms we consider do much better than the random baseline, but have low raw accuracy since there are few new links compared to the number of possible links. See Table 2 for a summary of the number of links predicted and baseline probability of randomly predicting a correct link. 5. Results On the cond-mat and facebook networks, both the spec euclid and spec cosine predictors performed worse than the simple common neighbors predictor. In addition to the full networks, we also compared predictor accuracy on reduced versions of the hep-th and hep-ph networks, because the full networks are too large for methods like Katz, common neighbors, and Rooted Page Rank to complete in a reasonable amount of time. On our reduced version of the hep-th network, our embedding-based predictors did better than common neighbors but not as well as the Rooted Page Rank predictor. On the reduced hep- FAST LINK PREDICTION WITH SPECTRAL EMBEDDING 9 of 13 Network cond-mat facebook hep-ph reduced hep-ph hep-th reduced hep-th Links 1190 7858 101466 1988 1000 135 Random Accuracy (%) 0.012 0.004 0.286 0.661 0.296 0.084 Table 2. Link Prediction Task Setup ph network, the spec euclid predictor performed significantly better than all other competitors, including our other embedding-based predictor, spec cosine. As Table 3 shows, the best predictors for the cond-mat network were Katz and common neigh- bors. Note that for both spec euclid and spec cosine, the accuracy increases with the dimension of the embedding. Predictor Correct (%) Time (s) katz commonNeighbors prefattach spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec euclid4 spec cosine4 spec euclid8 spec cosine8 spec euclid16 spec cosine16 5.97 5.97 1.93 1.51 0.25 1.51 1.18 1.76 1.34 1.68 1.34 1.68 1.43 62.96 1.55 0.35 2.99 3.35 3.65 3.80 10.54 10.89 11.41 10.73 29.91 32.31 Table 3. Performance of link predictors on the cond-mat network As previously mentioned, the facebook graph was too large to run the Katz predictor on it in a reasonable amount of time. As with the cond-mat network, the simple common neighbors predictor performs best. Our spectral embedding link predictors performed significantly better on the hep-th and hep-ph networks, as Table 5 and Table 6 show. The common neighbors algorithm did not scale to the hep-th and hep-ph networks, unlike the face- book network. Although the facebook network had more nodes, it has a lower average node degree and fewer distance two pairs. The common neighbors algorithm computes intersections of neighbor sets for each distance two pair. Because the average node degree is higher for the hep-th and hep-ph networks, these intersections are more expensive to compute and there are more distance two pairs for which intersections must be computed. In order to compare the performance of our spectral predictors to other predictors on the hep-ph and 10 of 13 BENJAMIN PACHEV AND BENJAMIN WEBB Predictor Correct (%) Time (s) commonNeighbors prefattach spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec euclid4 spec cosine4 spec euclid8 spec cosine8 5.29 0.41 0.42 0.00 0.50 0.42 1.40 1.02 1.95 2.58 151.76 7.00 9.86 47.20 11.52 12.96 24.05 25.96 61.21 62.27 Table 4. Performance of link predictors on the facebook network Predictor Correct (%) Time (s) prefattach spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec cosine4 spec euclid4 spec euclid8 spec cosine8 spec euclid16 spec cosine16 0.00 94.50 1.50 98.70 99.60 100.00 99.90 100.00 100.00 99.90 99.90 9.47 10.88 17.17 17.88 15.97 15.67 18.53 29.81 29.55 96.47 100.93 Table 5. Performance of link predictors on the hep-th network Predictor Correct (%) Time (s) prefattach spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec euclid4 spec cosine4 spec euclid8 spec cosine8 spec euclid16 spec cosine16 0.00 3.93 0.14 9.16 3.44 19.25 13.65 22.90 21.12 24.62 23.97 16.85 18.25 32.37 21.98 23.46 27.04 26.09 46.69 49.62 148.51 135.83 Table 6. Performance of link predictors on the hep-ph network FAST LINK PREDICTION WITH SPECTRAL EMBEDDING 11 of 13 hep-th network data, we conducted another experiment using downsampled versions of these networks. To downsample, we used only the top 10% highest degree nodes. Our spectral predictors performed the best on the reduced hep-ph network (see Table 7), while the Rooted Page Rank algorithm was best for the reduced hep-th network (see Table 8). Predictor prefattach katz commonNeighbors pageRank adamicAdar spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec euclid4 spec cosine4 spec euclid8 spec cosine8 spec euclid16 spec cosine16 Correct (%) Time (s) 0.00 1.16 9.36 11.87 8.85 1.81 0.10 4.73 2.57 13.13 11.12 16.40 9.31 14.13 4.93 1.52 2.75 23.96 2.94 754.40 2.80 6.21 3.73 6.86 5.80 9.01 7.39 7.44 17.32 14.90 Table 7. Performance of link predictors on the reduced hep-ph network Predictor prefattach katz commonNeighbors pageRank adamicAdar spec euclid1 spec cosine1 spec euclid2 spec cosine2 spec euclid4 spec cosine4 spec euclid8 spec cosine8 spec euclid16 spec cosine16 Correct (%) Time (s) 0.00 0.00 2.22 11.11 2.22 0.00 0.00 0.74 0.00 0.74 0.00 2.22 1.48 8.89 5.93 1.28 1.61 16.70 1.97 788.11 2.02 3.77 4.45 4.61 6.84 6.25 10.59 11.02 26.82 24.92 Table 8. Performance of link predictors on the reduced hep-th network 12 of 13 6. Conclusion REFERENCES We present a link prediction framework that can scale to very large networks by avoiding the quadratic costs inherent in methods that exhaustively search all candidate pairs of nonadjacent nodes. We investi- gated the performance of a set of predictors based on this framework and the spectrum and eigenvectors of the graph's Laplacian matrix. These methods achieved high levels of accuracy on certain real-world link prediction tasks, and scaled well to networks with tens of thousands of nodes and millions of edges. We emphasize that there are many other possible graph embeddings to invesitigate. Virtually all the runtime of our spectral link predictors is spent computing the resistance distance embedding. The k closest pairs component of our algorithm is very fast in practice, with nearly linear temporal complexity in the number of edges Replacing the resistance distance embedding with one that is cheaper to com- pute could potentially produce link predictors that can scale to much larger networks than the ones we consider in this paper. Our approximate resistance distance link predictor was derived as a low-rank approximation of resistance distance, an established link prediction score that is expensive to compute. Many other well- known predictors are expensive to compute, such as Katz and Rooted Page Rank. There is much room to explore low-rank approximations of such predictors and investigate whether they can be converted into accurate, scalable, graph embedding based, link predictors of the form we considered. Funding This work was supported by the Defense Threat Reduction Agency [grant number HDTRA1-15-1- 0049]. References [1] NEWMAN, M. (2010a) Networks: An Introduction. Oxford University Press. [2] GROSS, T. & SAYAMA, H. (2000b) Adaptive Networks: Theory, Models and Applications. Springer Publishing Company. [3] LIBEN-NOWELL, D. & KLEINBERG, J. (2001c) The Link-Prediction Problem for Social Net- works. Journal of the American Society for Information Science and Technology, 58(7), 1019– 1031. [4] QUERCIA, D., ASKHAM, H. & CROWCROFT, J. (2012d) TweetLDA: supervised topic classifica- tion and link prediction in Twitter. in Proceedings of the 4th Annual ACM Web Science Conference, pp. 247–250. ACM. [5] BARZEL, B. & BARAB ´ASI, A.-L. (2013e) Network link prediction by global silencing of indirect correlations. Nature biotechnology, 31(8), 720–725. [6] CLAUSET, A., MOORE, C. & NEWMAN, M. E. (2008f) Hierarchical structure and the prediction of missing links in networks. Nature, 453(7191), 98–101. [7] SRINIVAS, V. & MITRA, P. (2016g) Link Prediction in Social Networks-Role of Power Law Dis- tribution. Springer. [8] FOUSS, F., PIROTTE, A., RENDERS, J.-M. & SAERENS, M. (2007h) Random-walk computation of similarities between nodes of a graph with application to collaborative recommendation. IEEE Transactions on knowledge and data engineering, 19(3). REFERENCES 13 of 13 [9] LENHOF, H.-P. (1992i) The k closest pairs problem. http://people.scs.carleton.ca/∼michiel/k-closestnote.pdf, Accessed: 2017-03-11. [10] NEWMAN, M. E. (2001j) Clustering and preferential attachment in growing networks. Physical review E, 64(2), 025102. [11] MANGUOGLU, M., COX, E., SAIED, F. & SAMEH, A. (2010k) TRACEMIN-Fiedler: A parallel algorithm for computing the Fiedler vector. in International Conference on High Performance Computing for Computational Science, pp. 449–455. Springer. [12] HU, Y. & SCOTT, J. (2003l) HSL MC73: A fast multilevel Fiedler and profile reduction code. Rutherford Appleton Laboratory. [13] LESKOVEC, J., KLEINBERG, J. & FALOUTSOS, C. (2007m) Graph Evolution: Densification and Shrinking Diameters. ACM Trans. Knowledge Discovery from Data, 1(1), 1–40. [14] NETWORKS COLLECTION, K. (2016n) arXiv hep-th network dataset – KONECT. http://konect.uni-koblenz.de/networks/ca-cit-HepTh, Accessed: 2017-03-11. [15] [16] dataset http://konect.uni-koblenz.de/networks/ca-cit-HepPh, Acessed: 2017-03-11. network hep-ph (2016o) arXiv – KONECT. (2016p) Facebook friendships network dataset – KONECT. http://konect.uni-koblenz.de/networks/facebook-wosn-links, Accessed: 2017-03-11. [17] VISWANATH, B., MISLOVE, A., CHA, M. & GUMMADI, K. P. (2009q) On the Evolution of User Interaction in Facebook. in Proc. Workshop on Online Social Networks, pp. 37–42. [18] NEWMAN, M. E. (2001r) The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences, 98(2), 404–409.
1709.00714	1	1709	2017-09-03T13:20:16	Home Location Estimation Using Weather Observation Data	[ "cs.SI", "cs.IR" ]	We can extract useful information from social media data by adding the user's home location. However, since the user's home location is generally not publicly available, many researchers have been attempting to develop a more accurate home location estimation. In this study, we propose a method to estimate a Twitter user's home location by using weather observation data from AMeDAS. In our method, we first estimate the weather of the area posted by an estimation target user by using the tweet, Next, we check out the estimated weather against weather observation data, and narrow down the area posted by the user. Finally, the user's home location is estimated as which areas the user frequently posts from. In our experiments, the results indicate that our method functions effectively and also demonstrate that accuracy improves under certain conditions.	cs.SI	cs	Home Location Estimation Using Weather Observation Data Yuki Kondo∗ Masatsugu Hangyo† Mitsuo Yoshida∗ Kyoji Umemura∗ ∗Department of Computer Science and Engeneering, Toyohashi University of Technology, Aichi, Japan [email protected], [email protected], [email protected] †Weathernews Inc., Chiba, Japan [email protected] 7 1 0 2 p e S 3 ] I S . s c [ 1 v 4 1 7 0 0 . 9 0 7 1 : v i X r a Abstract -- We can extract useful information from social media data by adding the user's home location. However, since the user's home location is generally not publicly available, many researchers have been attempting to develop a more accurate home location estimation. In this study, we propose a method to estimate a Twitter user's home location by using weather observation data from AMeDAS. In our method, we first estimate the weather of the area posted by an estimation target user by using the tweet, Next, we check out the estimated weather against weather observation data, and narrow down the area posted by the user. Finally, the user's home location is estimated as which areas the user frequently posts from. In our experiments, the results indicate that our method functions effectively and also demonstrate that accuracy improves under certain conditions. Keywords -- Twitter; AMeDAS; home location estimation; weather observation I. INTRODUCTION Twitter1 is a social media that users can post short texts called tweets. There are many tweets posted on Twitter, and it can be regarded as a social sensor reflecting real world events [1]. Many researchers attempt to develop a more ac- curate user's home location estimation in order to use Twitter as a social sensor associated with location information [2]. As methods of the user's home location estimation, for example, one using user's tweets posted by the user has been proposed. Most of these methods (e.g., [3], [4], [5], [6]) estimate the user's home location by words that appear only in tweets of a few areas. Such words are called local words. If most of the words included in user's tweets consist of local words in a certain area, it can be estimated that the user's home location is likely to be the area. However, it is difficult to estimate the home location for users who have not posted such tweets containing local words. In this paper, we focus on the tweet which contains weather information rather than one which contains local words. For example, in the case of a tweet "It is raining.", the weather at the area the user posted can be regarded as rain at the posted date. We propose a method to estimate the user's home 1https://twitter.com/ (accessed 2017 -- 06 -- 20) 978 -- 1 -- 5386 -- 3001 -- 3/17/$31.00 c(cid:13)2017 IEEE location by using such tweets. This method estimates the weather of the area posted by an estimation target user, then estimates the home location by using the estimated weather and the observed one. In our experiments, the results indicate that our method functions effectively and also demonstrate that accuracy improves in rural areas. II. RELATED WORK Many researchers have been attempting to develop methods in order to detect events and understand trends by considering Twitter as a social sensor [1]. Sakaki et al. detected an earthquake with high probability (96% of earthquakes of JMA seismic intensity scale three or more are detected) merely by monitoring tweets [7]. Vieweg et al. reported situational fea- tures communicated in microblog posts during two concurrent emergency events that took place in North America during Spring 2009 [8]. Aramaki et al. proposed a method to detect influenza epidemics by monitoring tweets and visualize the epidemic state2 [9]. Herfort et al. presented a new approach that relies upon the geographical relations between twitter data and flood phenomena in order to enhance information extraction from social media [10]. Asakura et al. collected posts concerning flood disasters from social media about weather and created a flood disaster corpus [11]. We need the user's home location to use Twitter as social sensor. However, user's home locations are generally not publicly available [3]. Many researches attempt to estimate the user's home location using their information such as tweets [2]. Most of these methods (e.g., [3], [4], [5], [6]) estimate the user's home location by local words included in the tweet. Previous methods use words (local words) in which ap- pearing areas are biased in order to estimate the user's home location. However, there was a lack of geo-location informa- tion that can be used to extract such words. In this paper, we propose a method to estimate the user's home location by using weather information, which can identify the actual area in advance. 2http://mednlp.jp/influ map/ (accessed 2017 -- 06 -- 20) III. PROPOSED METHOD A. Outline of Proposed Method In the proposed method, we first acquire relationships between the tweet and the weather by the following procedure. First, tweets with location information are divided into words and are converted into feature vectors. Next, each tweet is assigned a label as weather (raining or not) where the user posted by using weather observation data from AMeDAS. Finally, we give feature vectors assigned labels to the support vector machine (SVM) and create a classifier for a tweet's weather estimation. We collect tweets posted by an estimation target user and estimate labels using SVM against the tweets. This label represents the weather where the estimation target user posted the tweet. We regard the area where the estimated weather and the weather from AMeDAS data coincide as the user's home location. B. Classifier for Tweet's Weather Estimation First, we convert tweet t to feature vector xt by dividing the tweet into words as follows. (cid:40) xt = (x1, x2, . . . , xW) wi ∈ Wt (otherwise) xi = 1 0 where W is a set of words that occurred more than once in all the tweets, Wt is a set of words included in tweet t, and wi is the i-th word when words included in W are sorted in alphabetical order. In this paper, we regard Wt as all the words included in tweet t under the limit of W . Word selection for a better feature extraction is left for the future study. Next, feature vector xt is assigned a label as the weather where the user posted by using weather observation data from AMeDAS. In label l ∈ {T rue, F alse}, T rue indicates that it was raining at the posted area and F alse indicates that it was not raining. Finally, we give feature vectors assigned labels to the SVM as a state-of-the-art binary classifier. In this paper, we used a linear kernel and set the cost parameter C = 0.025 which sample value is shown in the scikit-learn document3. We will plan to search for optimal parameters in the future study. C. Location Estimation by using Weather Observation In order to estimate a user's home location by using weather observation data, we collect tweets posted by an estimation target user and estimate the labels (weather) of the tweets. Thereafter, the user's home location is estimated based on the weather indicated by the estimated labels and the observed one. In this paper, we hypothesized that a user would always posts tweets from the same area. Let T = {t1, t2, . . . , tn} be a set of tweets posted by an estimation target user and let date(tj) denote the posted date 3http://scikit-learn.org/stable/auto examples/classification/plot classifier comparison.html (accessed 2017 -- 07 -- 28) of the tweet tj ∈ T . Also, let w-est(tj) be an estimated weather of the tweet tj. In estimating weather, we estimate label l ∈ {T rue, F alse} by using feature vector xtj converted from tweet tj. Then w-est(tj) ∈ {T rue, F alse}. Let A = {a1, a2, . . . , am} be a set of all areas, and we assume that a user's home location is one of A. m is the total number of areas, and each area corresponds one-to-one with the location of Meteorological Office where AMeDAS is installed. Let w-obs(a, k) be the weather at area a and date k, based on weather observation data from AMeDAS, then w-obs(a, k) ∈ {T rue, F alse}. User's home location aest is estimated by following expression. a∈A (cid:88) (cid:40) t∈T 0 1 aest = arg min count(T, a) count(T, a) = diff (l1, l2) = diff(w-est(t), w-obs(a, date(t))) l1 = l2 l1 (cid:54)= l2 IV. EXPERIMENT DATA A. Weather Observation Data AMeDAS (Automated Meteorological Data Acquisition System) is an automatic meteorological observing system operated by Japan Meteorological Agency (JMA). AMeDAS is installed in about 1,300 locations throughout Japan and ob- serves weather information such as precipitation, temperature and so on. These weather observation data are published on the JMA's website. However, the observation data of precipitation is recorded in units of 0.5 mm, and it is not possible to distinguish between precipitation less than 0.5 mm and no precipitation. In other words, we cannot determine exactly whether it was raining or not. In this paper, we use detailed rainfall observation data measured by a precipitation detector and label whether it was raining or not. A precipitation detector is a facility that detects slight precipitation of less than 0.5 mm, and is installed in 157 meteorological offices throughout Japan. As a result, it is possible to distinguish between precipitation less than 0.5 mm and no precipitation in the data observed at the meteorological office among the observation data from AMeDAS. B. Area and Tweet Data Among the 157 meteorological offices listed on the JMA's website4, we use 154 locations (areas) other than Showa Sta., Mt. Fuji and Mt. Aso. Then m = 154. We assume that both locations where a tweet is posted and user's home locations are one of these areas. We used the data on the latitude and longitude of the center of each area which is published on the JMA's website5. The following pre-processing applies to tweets with location information posted in 2016. 4http://www.data.jma.go.jp/obd/stats/data/mdrr/chiten/sindex2.html (accessed 2017 -- 04 -- 20) (accessed 2017 -- 04 -- 20) 5http://www.jma.go.jp/jma/kishou/know/amedas/kaisetsu.html name ("source"). 3) Find the nearest AMeDAS within a 10 km radius based on the location information, and assign the area ID as the nearest AMeDAS to the tweet (remove the tweet if an area ID cannot be assigned). 4) Collate the weather observation data by the area ID and posted date, and assign a label as to whether it was raining or not to the tweet (remove the tweet if a label cannot be assigned). 5) Remove hashtags, URLs, screen names from the tweet. 6) Divide the tweet into words by using MeCab6 if the tweet is in Japanese. 1) Remove tweets that have no location information con- sisting of latitude and longitude ("coordinates") given by GPS. 2) Remove tweets posted by Bots judged based on client 7) Remove tweets that contain none of the weather related words shown in Sec. IV-C. After the above pre-processing has been applied to the tweets, we assign the home location to users remaining over 10 tweets. If 90% or more of tweets are posted in the same area, we regard the area as the home location. Tweets are removed for those users who cannot be assigned a home location. In this paper, we pre-processed 171,614,139 tweets with location information posted in 2016, and eventually use 4,613 tweets posted by 153 users for our experiments. C. Collecting Weather-related Words Our method firstly estimates the weather of the area posted by an estimation target user, then estimates the home loca- tion by using the estimated weather and the observed one. However, the weather of the tweet cannot be estimated if the tweet contains no weather information. In this paper, we remove tweets that do not contain weather information before the experiment. In order to remove tweets that do not contain weather information, we focused on words used to express weather in everyday conversation. We use words such as "sunny" and "good weather" when expressing to "It is sunny weather." in everyday conversation. Therefore, if a tweet contains such words used to express a weather, it is considered that the tweet includes weather information at the posted area. In this paper, we collect words used to express weather from tweets, and use only tweets containing collected words for our experiments. In some cases, words used to express the weather appear unevenly for a day when the weather is either sunny or rainy. Therefore, these words are considered to be biased to either T rue or F alse for label l assigned to the tweet where the word appeared. Let Wt be a set of words included in tweet t. Label l is assigned to tweet t, although we assume that label l is assigned to all words included in Wt. We calculate the self- mutual information (PMI) of the following two events for each word and evaluate a bias of the word. 6http://taku910.github.io/mecab/ (accessed 2017 -- 06 -- 20). ipadic-2.7.0. • The word appeared. • The label assigned to the word is T rue (or F alse). PMI(w, l) is the ease of attaching label l to word w, this is defined as follows. PMI(w, l) = log P (w, l) P (w)P (l) P (w, l) = P (w) = P (l) = C(w, l) N C(w) N C(l) N where C(w, l) is the number of times the label of word w was l, C(w) is the total frequency of word w, C(l) is the total frequency of label l, and N is the sum of all word frequencies. For the tweets set up to pre-process (5) in Sec. IV-B, PMI(w, l) is calculated for the case of l = True and l = False for the word that appears more than 10 times, and we collected 2,000 words with higher values. However, there was a problem that low frequency words were included. If the words are used by few users and the frequency is low, P (w, T rue) and P (w, F alse) values are greatly different and the PMI tends to be excessively high. Therefore, we calculated the number of users using each word and removed words that ware used less than 10 times. From the remaining words, we subjectively picked up weather-related words, then we manually selected 50 words where the label tends to be T rue and 18 words where the label tends to be F alse. V. EXPERIMENT SETTING A. Baseline Methods To verify the effectiveness of our method, we compare our method to baseline methods that use local words. In order to compare in different viewpoints, we prepare two methods as the baseline methods. Cheng et al. [3] propose a method for user's home location estimation by using the word occurrence probability of each area. This method estimates the occurrence probability of words included in the user's tweets in each area, and consider the area where the sum of occurrence probabilities of words is maximum as the user's home location. Let this method be the baseline (A). Our method consists of two steps, estimating weathers of user's tweets and estimating user's home location by using estimated weathers. Therefore, we compare the estimation method that consists of two steps similar to our method. For a set of words included in the tweet, the location where the tweet was posted can be estimated by the method similar to location estimation method of Cheng et al. [3]. We construct the baseline (B) as follows: estimate the locations where tweets were posted by an estimation target user, and estimate the user's home location by majority vote from the estimated tweet locations. Although filtering and smoothing of words can be applied to both baseline methods, we do not apply them in our experiments. B. Evaluation Metrics In this experiment, we use Precision@k to evaluate the estimation performance. This is the ratio of users where their actual home location is contained within the top k estimated home locations. We use the correct distance to evaluate whether an estimated home location is an actual home location. This metric considers that the correct home location was estimated if the distance between an estimated home location and an actual home location is less than the threshold called correct distance. The following expression defines Precision@k(U, d) where estimating home location for user set U at correct distance d. Precision@k(U, d) = {uu ∈ U ∧ At(u, d) (cid:54)= φ}/U At(u, d) = {aa ∈ Ak(u) ∧ dist(a, at(u)) < d} where at(u) is the actual home location of user u, Ak(u) is a set consisting of the top k estimated home locations for user u, and dist(a1, a2) is the distance between area a1 and a2. Precision@k is also used for the evaluation of estimation performance by prefecture. It differs in that user set U is grouped by prefecture in which the actual user's home location belongs, and the precision is evaluated for each prefecture. User set Up whose prefecture p contains the user's actual home location is expressed as follows. Up = {uu ∈ U ∧ pref (at(u)) = p} where pref (a) is a prefecture that contains area a. Let P be a set of all prefectures in Japan, for each p ∈ P , Precision@(cid:48)(Up, d) is calculated by the following formula. Precision@k(cid:48)(Up, d) = {uu ∈ Up ∧ A(cid:48) t(u, d) (cid:54)= φ}/U t(u, d) = {aa ∈ Ak(u)∧ A(cid:48) dist(a, at(u)) < d ∧ pref (a) = pref (at(u))} VI. EXPERIMENT RESULT A. Evaluate Precision by Correct Distance Fig. 1, 2 and 3 show Precision@k when k = 1, 3, 5 and the correct distance d is changed from 10 km to 160 km respectively. In all cases of k = 1, 3, 5, the baseline (A) exceeds the baseline (B). Baseline (B) is a method of estimating the posted location of tweets and estimating the home location by majority vote from the estimated tweet locations. An error in posted location estimation of tweets is considered to have led to a reduction in Precision@k. Our method exceeds the precision of the baseline (A) when the correct distance is larger than about 60 km to 80 km. This is because our method can estimate correct locations for prefectures with less population as shown in Sec. VI-B. Fig. 1. Precision of Varied Correct Distance (k = 1). Fig. 2. Precision of Varied Correct Distance (k = 3). Fig. 3. Precision of Varied Correct Distance (k = 5). 020406080100120140160DPSSFDUEJTUBODF<LN>0.00.20.40.60.81.0QSFDJTJPO#BTFMJOF " #BTFMJOF # 1SPQPTFENFUIPE020406080100120140160DPSSFDUEJTUBODF<LN>0.00.20.40.60.81.0QSFDJTJPO#BTFMJOF " #BTFMJOF # 1SPQPTFENFUIPE020406080100120140160DPSSFDUEJTUBODF<LN>0.00.20.40.60.81.0QSFDJTJPO#BTFMJOF " #BTFMJOF # 1SPQPTFENFUIPEFig. 4. Precision for Each Prefecture: Prefecture names are sorted in descending order of their population. The baseline methods can estimate correct locations for only a few prefectures that have a large population, whereas in our method it can estimate correct locations for prefectures with less population. Also, the gradient of the graph is steeper than that of the baseline methods, and the improvement of the precision when the correct distance is larger is clearer. In other words, it is indicated that our method estimated the area closer to the user's actual home location compared to the baseline methods even if it is incorrect. B. Evaluate Precision by Prefecture Fig. 4 shows Precision@k(cid:48) evaluated by each prefecture with k = 1 and correct distance d = 10. If there is no user in a prefecture, then the prefecture is not shown in Fig. 4. In addition, prefecture names are sorted in descending order of their population in 20167. From this result, it is found that the baseline methods are correct only for prefectures with a large population. In our method, accurate estimation is made in more prefectures than in the baseline methods. This is clear from the macro average, and our method was 0.242, the baseline (A) was 0.0689, and the baseline (B) was 0.0458. The macro average of our method is the largest. Most location estimation methods use word frequencies for each area and estimate the occurrence probability of a word. Some words have high occurrence probabilities in some areas, and such words are called local words. Most location estimation methods use local words. In Japan, however, it is known that about 40% of tweets with location information were posted from the urban areas surrounding Tokyo, tweets are concentrated in urban areas. Because the number of posted tweets depends on the area, words with equal occurrence probability in any area is estimated to have higher occurrence probability in areas with a large population. Therefore, users who did not post tweets including local words are estimated to be that their home location is in an area with a large population. According to experimental results, at the baseline methods, many users were estimated to be that their home location is in an area with a large population. We used only tweets contain- ing weather-related words for the experiments. Therefore, this experimental setup is regarded as users that barely post tweets 7https://www.e-stat.go.jp/SG1/estat/GL08020103.do? toGL08020103 & listID=000001154737 (accessed 2017 -- 05 -- 15) containing local words. From the above discussion, our method is more effective than the baseline methods when users post many tweets that contain weather-related words not including local words. VII. CONCLUSION In this paper, we have proposed a method to estimate the user's home location by using tweets without local words. This method estimates the weather of the area posted by an estimation target user, then estimates the home location by using the estimated weather and the observed one. We compared our method to the baseline methods that use local words. As a result, under the condition of estimating three areas or more and the correct distance 70 km or more, the estimated precision of our method exceeded baseline methods. In addition, we analyzed the estimated precision for each prefecture in order to find the factor that our method achieves more precision. The baseline methods can estimate correct locations for only a few prefectures that have a large population, whereas in our method it can estimate correct locations for prefectures with less population. From these experimental results, we found that in some cases baseline methods cannot estimate home location whereas our method can. REFERENCES [1] F. Atefeh and W. Khreich, "A survey of techniques for event detection in twitter," Computational Intelligence, vol. 31, no. 1, pp. 132 -- 164, 2015. [2] X. Zheng, J. Han, and A. Sun, "A Survey of Location Prediction on Twitter," 2017, arXiv:1705.03172v1. [3] Z. Cheng, J. Caverlee, and K. Lee, "You are where you tweet: A content-based approach to geo-locating twitter users," in Proceedings of the 19th ACM International Conference on Information and Knowledge Management, 2010, pp. 759 -- 768. [4] S. Kinsella, V. Murdock, and N. O'Hare, ""I'm Eating a Sandwich in Glasgow": Modeling Locations with Tweets," in Proceedings of the 3rd International Workshop on Search and Mining User-Generated Contents, 2011, pp. 61 -- 68. [5] B. Han, P. Cook, and T. Baldwin, "Geolocation Prediction in Social Media Data by Finding Location Indicative Words," in Proceedings of the International Conference on Computational Linguistics 2012, 2012, pp. 1045 -- 1062. 00.20.40.60.81TokyoKanagawaOsakaAichiSaitamaChibaHyogoHokkaidoFukuokaShizuokaIbarakiHiroshimaKyotoMiyagiNiigataNaganoGifuGummaTochigiFukushimaOkayamaKagoshimaOkinawaEhimeNagasakiIwateIshikawaYamagataMiyazakiToyamaAkitaKagawaWakayamaYamanashiSagaFukuiKochiTottoriprecisionproposed methodbaseline(A)baseline(B)[6] S. Roller, M. Speriosu, S. Rallapalli, B. Wing, and J. Baldridge, "Super- vised Text-based Geolocation Using Language Models on an Adaptive Grid," in Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, 2012, pp. 1500 -- 1510. [7] T. Sakaki, M. Okazaki, and Y. Matsuo, "Earthquake shakes twitter users: Real-time event detection by social sensors," in Proceedings of the 19th International Conference on World Wide Web, 2010, pp. 851 -- 860. [8] S. Vieweg, A. L. Hughes, K. Starbird, and L. Palen, "Microblogging During Two Natural Hazards Events: What Twitter May Contribute to Situational Awareness," in Proceedings of the 28th international conference on Human factors in computing systems, 2010, pp. 1079 -- 1088. [9] E. Aramaki, S. Maskawa, and M. Morita, "Twitter catches the flu: Detecting influenza epidemics using twitter," in Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, 2011, pp. 1568 -- 1576. [10] B. Herfort, J. P. de Albuquerque, S.-J. Schelhorn, and A. Zipf, "Ex- ploring the Geographical Relations Between Social Media and Flood Phenomena to Improve Situational Awareness," in Connecting a Digital Europe Through Location and Place, 2014, pp. 55 -- 71. [11] Y. Asakura, M. Hangyo, and M. Komachi, "Disaster Analysis using User-Generated Weather Report," in Proceedings of the 2nd Workshop on Noisy User-generated Text, 2016, pp. 24 -- 32.
1803.05084	1	1803	2018-03-14T00:25:37	Local Partition in Rich Graphs	[ "cs.SI", "cs.DS", "physics.soc-ph" ]	Local graph partitioning is a key graph mining tool that allows researchers to identify small groups of interrelated nodes (e.g. people) and their connective edges (e.g. interactions). Because local graph partitioning is primarily focused on the network structure of the graph (vertices and edges), it often fails to consider the additional information contained in the attributes. In this paper we propose---(i) a scalable algorithm to improve local graph partitioning by taking into account both the network structure of the graph and the attribute data and (ii) an application of the proposed local graph partitioning algorithm (AttriPart) to predict the evolution of local communities (LocalForecasting). Experimental results show that our proposed AttriPart algorithm finds up to 1.6$\times$ denser local partitions, while running approximately 43$\times$ faster than traditional local partitioning techniques (PageRank-Nibble). In addition, our LocalForecasting algorithm shows a significant improvement in the number of nodes and edges correctly predicted over baseline methods.	cs.SI	cs	Local Partition in Rich Graphs Scott Freitas Arizona State University Tempe, Arizona [email protected] Hanghang Tong Arizona State University Tempe, Arizona [email protected] 8 1 0 2 r a M 4 1 ] I S . s c [ 1 v 4 8 0 5 0 . 3 0 8 1 : v i X r a Nan Cao Tongji University Shanghai, China [email protected] Yinglong Xia Huawei Santa Clara, California [email protected] ABSTRACT Local graph partitioning is a key graph mining tool that allows researchers to identify small groups of interrelated nodes (e.g. peo- ple) and their connective edges (e.g. interactions). Because local graph partitioning is primarily focused on the network structure of the graph (vertices and edges), it often fails to consider the ad- ditional information contained in the attributes. In this paper we propose-(i) a scalable algorithm to improve local graph partition- ing by taking into account both the network structure of the graph and the attribute data and (ii) an application of the proposed local graph partitioning algorithm (AttriPart) to predict the evolution of local communities (LocalForecasting). Experimental results show that our proposed AttriPart algorithm finds up to 1.6× denser local partitions, while running approximately 43× faster than traditional local partitioning techniques (PageRank-Nibble [1]). In addition, our LocalForecasting algorithm shows a sig- nificant improvement in the number of nodes and edges correctly predicted over baseline methods. ACM Reference Format: Scott Freitas, Hanghang Tong, Nan Cao, and Yinglong Xia. 2018. Local Partition in Rich Graphs. In Proceedings of ACM KDD (KDD'18). ACM, New York, NY, USA, Article 4, 9 pages. https://doi.org/10.475/123_4 1 INTRODUCTION Motivation. With the rise of the big data era an exponential amount of network data is being generated at an unprecedented rate across many disciplines. One of the critical challenges before us is the translation of this large-scale network data into meaningful infor- mation. A key task in this translation is the identification of local communities with respect to a given seed node 1. In practical terms, the information discovered in these local communities can be uti- lized in a wide range of high-impact areas-from the micro (protein interaction networks [13] [26]) to the macro (social [21] [4] and transportation networks). 1we interchangeably refer to local community as a local partition Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). KDD'18, August 2018, London, United Kingdom © 2018 Copyright held by the owner/author(s). ACM ISBN 123-4567-24-567/08/06...$15.00 https://doi.org/10.475/123_4 Problem Overview. How can we quickly determine the local graph partition around a given seed node? This problem is tradi- tionally solved using an algorithm like Nibble [19], which identifies a small cluster in time proportional to the size of the cluster, or PageRank-Nibble, [1] which improves the running time and ap- proximation ratio of Nibble with a smaller polylog time complexity. While both of these methods provide powerful techniques in the analysis of network structure, they fail to take into account the attribute information contained in many real-world graphs. Other techniques to find improved rank vectors, such as attributed PageR- ank [10], lack a generalized conductance metric for measuring cluster "goodness" containing attribute information. In this paper, we propose a novel method that combines the network structure and attribute information contained in graphs-to better identify local partitions using a generalized conductance metric. Applications. Local graph partition plays a central role in many application scenarios. For example, a common problem in recom- mender systems is that of social media networks and determining how a local community will evolve over time. The proposed Local- Forecasting algorithm can be used to determine the evolution of local communities, which can then assist in user recommendations. Another example utilizing social media networks is ego-centric network identification, where the goal is to identify the locally im- portant neighbors relative to a given person. To this end, we can use our AttriPart algorithm to identify better ego-centric networks using the graph's network structure and attribute information. Fi- nally, newly arrived nodes (i.e., cold-start nodes) often contain few connections to their surrounding neighbors, making it difficult to ascertain their grouping to various communities. The proposed LocalForecasting algorithm mitigates this problem by introduc- ing additional attribute edges (link prediction), which can assist in determining which local partitions the cold start nodes will belong to in the future. Contributions. Our primary contributions are three-fold: • The formulation of a graph model and generalized conduc- tance metric that incorporates both attribute and network structure edges. • The design and analysis of local clustering algorithm At- triPart and local community prediction algorithm Local- Forecasting. Both algorithms utilize the proposed graph model, modified conductance metric and novel subgraph identification technique. KDD'18, August 2018, London, United Kingdom S. Freitas et al. • The evaluation of the proposed algorithms on three real- world datasets-demonstrating the ability to rapidly identify denser local partitions compared to traditional techniques. Deployment. The local partitioning algorithm AttriPart is currently deployed to the PathFinder [7] web platform (www. path-finder.io), with the goal of assisting users in mining local net- work connectivity from large networks. The design and deployment challenges were wide ranging, including-(i) the integration of four different programming languages, (ii) obtaining real-time perfor- mance with low cost hardware and (iii) implementation of a visually appealing and easy to use interface. We note that the AttriPart algorithm, deployed to the web platform, has performance nearly identical to the results presented in section 4. Figure 1: Close-up of the AttriPart algorithm on the PathFinder web platform. This paper is organized as follows-Section 2 defines the prob- lem of local partitioning in rich graphs; Section 3 introduces our proposed model and algorithms; Section 4 presents our experimen- tal results on multiple real-world datasets; Section 5 reviews the related literature; and Section 6 concludes the paper. 2 PROBLEM DEFINITION In this paper we consider three graphs-(1) an undirected, un- weighted structure graph G = (V , E), (2) an undirected, weighted attribute graph A = (V , E) and (3) a combined graph consisting of both G and A that is undirected and weighted B = (V , E). In each graph, V is the set of vertices, E is the set of edges, n is the number of vertices and m is the number of edges (i.e. G, H and B contain the same number of vertices and edges by default). In order to denote the degree centrality we say δ(v) is the degree of vertex v. We use bold uppercase letters to denote matrices (e.g. G) and bold lowercase letters to denote vectors (e.g. v). For the ease of description, we define terms that are interchange- ably used throughout the literature and this paper-(a) we refer to network as a graph, (b) node is synonymous with vertex, (c) local partition is referred to as a local cluster, (d) seed node is equivalent to query and start vertex, (e) topological edges of the graph refers to the network structure of the graph, (f) a rich graph is a graph with attributes on the nodes and or edges. Having outlined the notation, we define the problem of local partitioning in rich graphs as follows: Problem 1. Local Partitioning in Rich Graphs Given: (1) an undirected, unweighted graph G = (V , E), (2) a seed node q ∈ V and (3) attribute information for each node v ∈ V containing a k-dimensional attribute vector xi -with an attribute matrix X = [x1, x2, ..., xn] ∈ Rk×n representing the attribute vector for each node v. Output: a subset of vertices S ⊂ V such that S best represents the local partition around seed node q in graph B. Symbol G, A, B n, m me p, mp s, q, ϕo Table 1: Symbols and Definition Definition network, attribute & combined graphs number of nodes & edges in graphs G, A, B number of edges in B after LocalForecasting number of nodes & edges in T preference vector, seed node & target conductance lazy random walk transition matrix set of vertices representing local partition rank truncation and iteration thresholds rank vector iterations; number of vertices to sweep AttriPart & LocalProximity teleport values subgraph relevance threshold & number of walks subgraph of B; walk count dictionary & list W S ϵ, ϵt tm, ns αn, αr ts, nw T ; D, L µ(L), σ(L) mean and standard deviation of L te edge addition threshold 3 METHODOLOGY This section first describes the preliminaries for our proposed algo- rithms, including the graph model and modified conductance metric. Next, we introduce each proposed algorithm-(1) LocalProximity, (2) AttriPart and (3) LocalForecasting. Finally, we provide an analysis of the proposed algorithms in terms of effectiveness and efficiency. 3.1 Preliminaries Graph Model. Topological network G represents the network structure of the graph and is formally defined in Eq. (1). Attribute network A represents the attribute structure of the graph and is computed based on the similarity for every edge (u, v) ∈ E in G. In order to determine the similarity between the two nodes, we use Jaccard Similarity J(u, v). A is formally defined in Eq. (2) where 0.05 is the default attribute similarity between an edge (u, v) ∈ E in G if J(xu , xv) = 0. In addition, te is the similarity threshold for the addition of edges not in G where 0 < te ≤ 1. Combined Network B represents the combined graph of G and A and is formally defined in Eq. (3). Formally, we define each of the three graph models G, A and B in Eq. (1), Eq. (2) and Eq. (3). Figure 2 presents an illustrative example. (cid:40)1, G(u, v) = if (u, v) ∈ E and u(cid:44)v 0, otherwise (1) Local Partition in Rich Graphs KDD'18, August 2018, London, United Kingdom By definition, B can be split into its representative components, G and A. We also note a few key properties of the parallel cut metric below: Figure 2: Example of the three graph models: (a) graph G is the network structure with nodes {1, 2, 3, 4} and correspond- ing attribute set {x1, x2, x3, x4} given as input. (b) Graph A is the attribute network with the same set of edges as G with each edge (u, v) assigned a positive similarity weight suv . (c) Graph B is a linear combination of the each respective edge (u, v) from G and A.  A(u, v) = B(u, v) = J(u,v), 0.05, J(u,v), 0, 1+A(u,v), A(u,v), 0, (2) (3) if (u, v) ∈ E, u(cid:44)v and J(u, v) > 0 if (u, v) ∈ E, u(cid:44)v and J(u, v) = 0 if (u, v) (cid:60) E, u(cid:44)v and J(u, v) > te otherwise if (u, v) ∈ E and (u, v) ∈ A if (u, v) (cid:60) E and (u, v) ∈ A otherwise Conductance. Conductance is a standard metric for determin- ing how tight knit a set of vertices are in a graph [12]. The traditional conductance metric is defined in Eq. (4), where S is the set of ver- tices representing the local partition. The lower the conductance value ϕ(S), where 0 ≤ ϕ(S) ≤ 1, the more likely S represents a good partition of the graph. cut(S) δ(v). v ∈S ϕ(S) = (4) Where the cut is Cut(S) = {(u, v) ∈ Eu ∈ S, v (cid:60) S}, and the volume min(vol(S), vol( ¯S)) is vol(S) = This definition of conductance will serve as the benchmark to compare the results of our parallel conductance metric. Parallel Conductance. We propose a parallel conductance metric which takes into account both the attribute and topological edges in the graph. Instead of simply adding the cut of each vertex v ∈ S, we want to determine whether v is more similar to the vertices in S or ¯S. The new cut and conductance metric is formally defined in Eq. (5) and Eq. (6), respectively. The key idea behind the parallel conductance metric is to determine whether each vertex in S is more similar to S or ¯S using the additional information provided by the attribute links. parallel_cut(S) = iϵS jϵS jϵS B(i, j) B(i, j) = iϵS (cid:2)A(i, j) + G(i, j)(cid:3) (cid:2)A(i, j) + G(i, j)(cid:3) jϵS jϵS (5) (1) Parallel_cut = 1 means that the vertices in S have connec- tions of equal weighting between S and ¯S. (2) Parallel_cut < 1 means that the vertices in S have only a few strong connections to ¯S. (3) Parallel_cut > 1 means that the vertices in S are more strongly connected to ¯S than S. Eq. (6) uses the cut as defined in Eq. (5) and the volume as defined above with the modification that δ(v) is a sum of it's components in G and A. ϕ(S) = parallel_cut(S) vol(S) (6) We note that the parallel conductance metric has a different scale compared to the traditional conductance metric. For example, a conductance of 0.3 in the traditional conductance doesn't have the same meaning as a conductance of 0.3 in the parallel definition. We also bound the volume of S to vol(S) < 1/2vol(B). This allows us to reduce the min(vol(S), vol( ¯S)) computation to vol(S). Figure 3: A toy example calculating the parallel cut and con- ductance with local partition S containing vertices {1, 2, 3, 4}. Parallel cut(V1) = 1.05/2.1 = 0.5, parallel cut(V2) = 0, paral- lel cut(V3) = 1.05/2.2 = 0.477, parallel cut(V4) = 0, parallel cut(Total) = 0.5 + 0.477 = 0.977. Volume(S) = 12. Parallel conductance(S) = 0.977/12 = 0.0814. 3.2 Algorithms We propose three algorithms in this subsection, including (1) Lo- calProximity (2) AttriPart and (3) LocalForecasting. First, we introduce the LocalProximity algorithm as a key building block for speeding-up the AttriPart and LocalForecasting algorithms by finding a subgraph containing only the nodes and edges rele- vant to the given seed node. Based on LocalProximity, we further propose the AttriPart algorithm to find a local partition around a seed node by minimizing the parallel conductance metric. Finally, we propose the LocalForecasting algorithm, which builds upon AttriPart, to predict a local community's evolution. KDD'18, August 2018, London, United Kingdom S. Freitas et al. LocalProximity. There are two primary purposes for the Lo- calProximity algorithm-(i) the requisite computations for the LocalForecasting algorithm require a pairwise similarity calcu- lation of all nodes, which is intractable for large graphs due to the quadratic run time. To make this computation feasible, we use the LocalProximity algorithm to determine a small subgraph of relevant vertices around a given seed node q. (ii) We experimentally found that the PageRank vector utilized in the AttriPart algo- rithm is significantly faster to compute after running the proposed LocalProximity algorithm. Algorithm Details. The goal is to find a subgraph T around seed node q, such that T contains only nodes and edges likely to be reached in nw trials of random walk with restart. We base the im- portance of a vertex v ∈ V on the theory that random walks can measure the importance of nodes and edges in a graph [5][17]. This is done by defining node relevance proportional to the frequency of times a random walk with restart walks on a vertex in nw tri- als (nodes walked on more than once in a walk will still count as one). Instead of using a simple threshold parameter to determine node/edge relevance as in [5], we utilize the mean and standard deviation of the walk distribution in order for the results to remain insensitive of nw given that nw is sufficiently large. In conjunction with the mean and standard deviation, we introduce ts as a rele- vance threshold parameter to determine the size of the resulting subgraph T . See section 3.3 for more details. Algorithm Description. The LocalProximity algorithm takes a graph B, a seed node q ∈ B, a teleport value αr , the number of walks to simulate nw , a relevance threshold ts-and returns a subgraph T containing the relevant vertices in relation to q. This algorithm can be viewed in three major steps: (1) Compute the walk distribution around seed node q in graph B using random walk with restart (line 2). We omit the Ran- dom Walk algorithm due to space constraints, however, the technique is described above. (2) Determine the number of vertices to include in the subgraph T based on the relevance threshold parameter ts, mean of the walk distribution list µ(L) and the standard deviation of the walk distribution list σ(L) (lines 4-6). (3) Create a subgraph based on the included vertices (line 8). walks to simulate nw , relevance threshold ts Algorithm 1: Local Proximity Input: Graph B, seed node q, teleport value αr , number of Result: Subgraph T 1 subgraph_nodes = []; 2 D = RandomWalk(q, αr , nw , B); 3 L = D.values; 4 for vertex u in B do 5 6 7 end 8 T = B.subgraph(subgraph_nodes); 9 return T ; if D[u] > µ(L) + σ(L) / ts then subgraph_nodes.append(u); AttriPart. Armed with the LocalProximity algorithm, we further propose an algorithm AttriPart, which takes into account the network structure and attribute information contained in graph to find denser local partitions than can be found using the net- work structure alone. The foundation of this algorithm is based on [19][1][30] with subtle modifications on lines 1, 4 and 9. These modifications incorporate the addition of a combined graph model, approximate PageRank computation using the LocalProximity algorithm, and the parallel cut and conductance metric. In addition, AttriPart doesn't depend on reaching a target conductance in order to return a local partition-instead it returns the best local partition found within sweeping ns vertices of the sorted PageRank vector. Algorithm Description. Given a graph B, seed node q ∈ V , tar- get conductance ϕo, rank truncation threshold ϵ, the number of iterations to compute the rank vector tlast , teleport value αn, rank iteration threshold ϵt and number of nodes to sweep ns-AttriPart will find a local partition S around q within ns iterations of sweep- ing. This algorithm can be viewed in five steps: (1) Set values for ϵ and tlast as seen in Eq. (7) and Eq. (9) respec- tively. We experimentally set b = 1+loд(m) and ϵt to 0.01. For additional detail on parameters ϵ, tlast and b see [19]. For all other parameter values see Section 4. (2) Run LocalProximity around seed node q in order to reduce 2 the run time of the PageRank computations (line 1). (3) Compute the PageRank vector using a lazy random tran- sition with personalized restart-with preference vector s containing all the probability on seed node q. At each it- eration truncate a vertex's rank if it's degree normalized PageRank score is less than ϵ (lines 2-7). (4) Divide each vertex in the PageRank vector by its correspond- ing degree centrality and order the rank vector in descending order (line 8). (5) Sweep over the PageRank vector for the first ns vertices, returning the best local partition S found (lines 9-10). The sweep works by taking the re-organized rank vector and creating a set of vertices S by iterating through each vertex in the rank vector one at a time, each time adding the next vertex in the rank vector to S and computing ϕ(S). ϵ = 1/(1800(l + 2)tlast 2b) (7) (8) l = ⌈loд2(2m/2)⌉ ϕ2 ln(c1(l + 2)(cid:112)2m/2)⌉ tlast = (l + 1)⌈ 2 (9) LocalForecasting. As a natural application of the AttriPart algorithm, we introduce a method to predict how local communities will evolve over time. This method is based on the AttriPart algorithm with two significant modifications-(i) required use of the LocalProximity algorithm to create a subgraph around the seed node and (ii) the use of the ExpandedNeighborhood algorithm to predict links between nodes in the subgraph. The idea behind using the ExpandedNeighborhood algorithm is that nodes are often missing many connections they will make in the future, which in turn affects the grouping of nodes into communities. To aid in Local Partition in Rich Graphs KDD'18, August 2018, London, United Kingdom Algorithm 2: AttriPart Input: Graph B, seed node q, target conductance ϕo, truncation threshold ϵ, iterations tlast , teleport value αn, iteration threshold ϵt , vertices to sweep ns 2(I + D−1T); qt = (1 − α)qt−1W + αs; rt(i) = qt(i) if qt(i)/d(i) > ϵ, else 0; Result: Local partition S 1 T = Local_Proximity(B, q, αr , nw , ts); 2 Di,i = δ(vi); 3 W = 1 4 for t = 1 to tlast and sum(qt ) - sum(qt−1) < ϵt do 5 6 7 end 8 Order i from large to small based on rt(i)/d(i); 9 Sweep Parallel_Conductance ϕ(S{i = 1..j}) while i < ns; 10 If there is j : ϕ(Sj) < ϕo, return S; predicting future edge connections we use Jaccard Similarity [14] to predict the likelihood of each vertex connecting to the others-with edges added if the similarity between two nodes is greater than threshold te. Algorithm Description. Given a graph B, a seed node q ∈ V , a target conductance ϕo, a rank truncation threshold ϵ, the number of iterations to compute the rank vector tlast , a teleport value αn, rank iteration threshold ϵt , similarity threshold te and number of nodes to sweep ns-this algorithm will find a predicted local partition around q within ns iterations of sweeping. As the Local- Forecasting algorithm is similar to AttriPart, we highlight the three primary steps: (1) Determine the subgraph around a given seed node using the LocalProximity algorithm (line 1). (2) Determine the pairwise similarity between all nodes in the subgraph using Jaccard Similarity, adding edges that are above a given similarity threshold (line 2). (3) Run the AttriPart algorithm to find the predicted local partition around the seed node (line 3). Algorithm 3: Local Forecasting Input: Graph B, seed node q, target conductance ϕo, truncation threshold ϵ, iterations tlast , teleport value αn, iteration threshold ϵt , similarity threshold te, vertices to sweep ns Result: Predicted local partition S 1 T = Local_Proximity(B, q); 2 T = Expanded_Neighborhood(T , te) ; 3 S = AttriPart(T , q, ϕo, ϵ, tlast , αn, ϵt , ns) ; 4 return S ; Algorithm 4: Expanded Neighborhood Input: Subgraph T , edge addition threshold te Result: Subgraph T with predicted edges for v in T and v not u do u_attr = T[u]; v_attr = T[v]; similarity_score = JaccardSimilarity(u_attr, v_attr); if similarity_score > te and not T[u][v] then T[u][v] = similarity_score; 1 for u in T do 2 3 4 5 6 7 8 end 9 return T ; end few links while a handful encompasses the majority. In Figure 4, we found that after running nw trials of random walk with restart, a scale-free like distribution formed-with a large majority of the nodes containing a small number of 'hits', while a few nodes con- stituted the bulk. As the number of ran- dom walks nw is in- creased, the scale-free like distribution is main- tained since each node is proportionally walked with the same distribu- tion. We therefore need only some minimum value for nw , which we set to 10,000. We use this skewed scale-free like distribution in combina- tion with Eq. (10) below to ensure the extraction of relevant nodes in rela- tion to a query vertex. Figure 4: Random walk w/ restart-distribution of node walk counts. nw = 10,000, αr = 0.15; dataset: wikipedia, start vertex: 'ewok', y-axis: right; dataset: Aminer, start vertex: 364298, y-axis: left. We omit nodes walked zero times in the graph, however, they're used in calculating µ(L), σ(L). Mathematically we de- fine node relevance based on Eq. (10), where D is a dictionary containing the walk count of each vertex and D(v) represents the number of times vertex v is walked in nw trials of the random walk with restart. L is a list of each node's walk count in the graph, µ(L) is the average number of times all of the nodes in the graph are walked and σ(L) is the standard deviation of the number of times all of the nodes in the graph are walked. In section 4 we discuss values of ts that have been shown to be empirically effective. 3.3 Analysis Effectiveness. LocalProximity (Algorithm 1). The objective is to ensure that all relevant nodes in proximity to seed node q are included. We use the fact that many real-world graphs follow a scale-free distribution [3] [6], with many nodes containing only a D(v) > µ(L) + σ(L)/ts (10) After determining the relevant nodes we create a subgraph T from a portion of the long-tail curve as defined by threshold param- eter ts in conjunction with µ(L) and σ(L). We say that subgraph T contains p ≪ n nodes-with p increasing nearly independently of the graph size (depending on threshold ts). As seen in Figure 4 KDD'18, August 2018, London, United Kingdom S. Freitas et al. the number of nodes with r walks converges independent of graph size. Efficiency. All algorithms use the same data structure for stor- ing the graph information. If a compressed sparse row (CSR) format is used, the space complexity is O(2m + n + 1). Alternatively, we note that with minor modification to the algorithms above we can use an adjacency list format with O(n + m) space. Lemma 3.1 (Time Complexity). LocalProximity has a time com- plexity of O(n + mp + nw) while AttriPart has a time complexity of O(p2 + pmp + n + nw) and LocalForecasting a time complexity of O(p2 + pme + n + nw). Proof. LocalProximity: There are three major components to this algorithm: (1) nw random walks with walk length l for a time complexity of O(nw) (line 2). (2) Linear iteration through the number of nodes taking O(n) (lines 4-7). (3) Subgraph T creation based on the number of included vertices p with node set Vt - requiring iteration through every edge of node v ∈ Vt for mp total edges. Iterating through every edge is linear in the number of edges for a time complexity of O(mp) (line 8). This leads to a total time complexity of O(n + mp + nw) AttriPart: There are six major steps to this algorithm: (1) calling LocalProximity which returns a subgraph T containing p nodes and mp edges for a time complexity of O(n + mp + nw) (line 1). (2) Creating a diagonal degree matrix by iterating through each node in T with time complexity O(p) (line 2). (3) Creating the lazy random walk transition matrix W , which requires O(mp) from multiplying the corresponding matrix entries (line 3). (4) In lines 4-7 we iterate for tlast iterations, with each iteration (i) updating the rank vector by multiplying the corresponding edges in the transition matrix W , with the rank vector q for a time complexity of O(mp) and (ii) truncating every vertex with rank qt(i)/d(i) ≤ ϵ for a time complexity linear in the number of nodes in the rank vector O(p). (5) Sort the rank vector which will be upper bounded by O(ploдp) (line 8). (6) Compute the parallel conductance, which takes O(p2 + pmp) time (lines 9-10). Combining each step leads to a total time complexity of O(p2 + pmp + n + nw). LocalForecasting: This algorithm has three major steps: (1) run the LocalProximity algorithm, which has a time complexity of O(n + mp + nw). (2) Perform the ExpandedNeighborhood algo- rithm, which densifies T by adding predicted edges for a total of me edges in T . This algorithm has a time complexity of O(p2) due to the nested for loops. (3) Run the AttriPart algorithm, which has a time complexity of O(p2 + pme + n + nw) with the modification of mp to me for the additional edges. This leads to an overall time complexity of O(p2 + pme + n + nw). □ While AttriPart and LocalForecasting both scale quadrati- cally with respect to p, we note that in practice these algorithms are very fast since p ≪ n and p scales nearly independent of graph size as shown in section 3.3. 4 EXPERIMENTS In this section, we demonstrate the effectiveness and efficiency of the proposed algorithms on three real-world network datasets of varying scale. 4.1 Experiment setup Datasets. We evaluate the performance of the proposed algorithms on three datasets-(1) the Aminer co-authorship network [27], (2) a Musician network mined from DBpedia and (3) a subset of Wikipedia entries in DBpedia containing both abstracts and links. All three networks are undirected with detailed information on each below: • Aminer. Nodes represents an author, with each author con- taining a set of topic keywords, and an edge representing a co-authorship. To form the attribute network, we com- pute attribute edges based on the similarity between two authors for every network edge, using Jaccard Similarity on the corresponding authors's topic set. • Musician. Nodes represent a Musician, with each Musician containing a set of music genres, and an edge representing two Musicians who have played in the same band. To form the attribute network, we compute attribute edges based on the similarity between two Musicians for every network edge, using Jaccard Similarity on the corresponding artist's music genre set. • Wikipedia. Nodes represent an entity, place or concept from Wikipedia which we will jointly refer to as an item. Each item contains a set of defining key words; with edges representing a link between the two items. The dataset orig- inates from DBpedia as a directed graph with links between Wikipedia entries. We modify the graph to be undirected for use with our algorithms-which we believe to be a reasonable as each edge denotes a relationship between two items. In addition, this dataset uses only a portion of the Wikipedia en- tries containing both abstracts and links to other Wikipedia pages found in DBpedia. To form the attribute network, we compute attribute edges based on the similarity between two items for every network edge using Jaccard Similarity on the corresponding item's key word set. Category Aminer Musician Wikipedia Network Co-Author Co-Musician Link Nodes 1,560,640 6,006 237,588 Edges 4,258,946 8,690 1,130,846 Table 2: Network Statistics Metrics. (1) To benchmark the LocalProximity algorithm's effectiveness and efficiency, we compare (i) the difference between local partition created with and without the LocalProximity algo- rithm on AttriPartand (ii) the run time and difference between the top 20 PageRank vector entries with and without the Local- Proximity algorithm. (2) To benchmark the AttriPart algorithm's effectiveness and efficiency we compare the triangle count, node count, local partition density and run time to PageRank-Nibble. Normally, PageRank-Nibble does not return a local partition if the target conductance is not met, however, we modify it to return the best local partition found-even if the target conductance is not met. This modification allows for more comparable results to AttriPart. (3) To provide a baseline for the LocalForecasting Local Partition in Rich Graphs KDD'18, August 2018, London, United Kingdom algorithm's effectiveness, we compare the local partition results to AttriPart on two graph missing 15% of their edges. Repeatability. All data and source code used in this research will be made publicly available. The Aminer co-authorship net- work can be found on the Aminer website 2; the Musician and Wikipedia datasets used in the experiments will be released on the author's website. All algorithms and experiments were conducted in a Windows environment using Python. 4.2 Effectiveness LocalProximity. In Figure 5 parts (a)-(c), we can see that the pro- posed LocalProximity algorithm significantly reduces the compu- tational run time, while maintaining high levels of accuracy across both metrics. Parts (a)-(b) demonstrate to what extent the accuracy of the results are dependent upon the parameter values. In particu- lar, a low value of αr (random walk alpha) and a high value of ts (relevance threshold) are critical to providing high accuracy results. In Figure 5 part (a), we measure accuracy as the number of vertices that differ between the local partitions w/ and w/o the LocalProximity algorithm on AttriPart. A small partition differ- ence indicates that the LocalProximity algorithm finds a relevant subgraph around the given seed node and that the full graph is unnecessary for accurate results. In part (b), we define the accuracy of the results to be the difference between the set of top 20 entries in the PageRank vectors for the full graph and subgraph using the LocalProximity algorithm. Overall, the results from part (b) cor- relate well to (a)-showing that for low values of αr (random walk alpha) and high values of ts (relevance threshold), their is negligible difference between the results computed on the full graph and the subgraph found using the LocalProximity algorithm. AttriPart. In Figure 6, we see that AttriPart finds signifi- cantly denser local partitions than PageRank-Nibble-with local partition densities approximately 1.6×, 1.3× and 1.1× higher in At- triPart than PageRank-Nibble in the Aminer, Wikipedia and Mu- 2m sician datasets respectively. Density is measured as n(n−1) where m is the number of edges and n is the number of nodes. In Figure 6, we observe that the triangle count of the Attri- Part algorithm is lower than PageRank-Nibble in the Musician and Aminer datasets. We attribute this to the fact that AttriPart is finding smaller partitions (as measured by node count) and, there- fore, there are less possible triangles. We also note that each triangle is counted three times, once for each node in the triangle. While no sweeps across algorithm parameters were performed, we believe that the gathered results provide an effective baseline for parameter selection. LocalForecasting. In order to measure the effectiveness of the LocalForecasting algorithm we setup the following experi- ment with three local partition calculations: (1) calculate the local partition using AttriPart, (2) calculate the local partition using At- triPart with 15% of the edges randomly removed from the graph and (3) calculate the local partition using the LocalForecasting algorithm with 15% of the edges randomly removed from the graph. We treat (1) as the baseline local community and want to test if (3) finds better local partitions than (2). The idea behind randomly removing 15% of the edges in the graph is to simulate the evolution 2https://Aminer.org/data (a) Y-axis represents the difference in vertices between the local par- tition calculated w/ and w/o the LocalProximity algorithm. (b) Y-axis represents the # of vertices differing between the top 20 rank vector entries w/ and w/o the LocalProximity algorithm. (c) Y-axis represents the difference in run time between the PageR- ank calculation w/ and w/o the LocalProximity algorithm. Figure 5: Each data point averages 10 randomly sampled ver- tices in both the Aminer and Musician datasets. Default pa- rameters (unless sweeped across): αn = 0.2, αr = 0.15, ϕo = 0.2, ts = 2, nw = 10,000, ns = 200. Parameter ranges: αr , αn and ϕo [0.1-0.7] in 0.1 intervals; ts [1-5] in 0.5 intervals. of the graph over time and test if the LocalForecasting algorithm can predict better local communities in the future. Ideally, we would have ground-truth local community data for a rich graph with time series snapshots, however, in its absence we use the above method. In Figure 7, each data point is generated in three steps-(i) tak- ing the difference between the set of vertices and edges in local partitions (1) and (3), (ii) taking the difference between the set of vertices and edges in local partitions (1) and (2) and (iii) by taking the difference between (ii) and (i). Step (i) tells us how far off the LocalForecasting algorithm is from the baseline, step (ii) tells us how far off the local partition would be from the baseline if no prediction techniques were used and step (iii) tells us the difference between the local partitions with and without the LocalForecast- ing algorithm (which is what we see graphed in Figure 7). KDD'18, August 2018, London, United Kingdom S. Freitas et al. (a) Scalability: Each data point represents the Aminer dataset in 1/10th intervals, with each point averaged over 3 randomly sampled vertices. Parameters: αn = 0.2, αr = 0.15, ϕo = 0.2, ts = 2, nw = 10,000, ns = 200. (b) Figure 6: Effectiveness: results are averaged over 20 and 100 randomly sampled vertices in the Aminer/Wikipedia and Musician datasets, respectively. Parameters: αn = 0.2, αr = 0.15, ϕo = 0.05, ts = 2, nw = 10,000, ns = 200. In Figure 7, we see that the local partition prediction accuracy, for both the edges and vertices, is above the baseline calculations in the Aminer dataset for a majority of edge similarity threshold values (te). The best results were obtained when te is 0.6, with an average of 1.4 vertices and 2.75 edges predicted over the baseline using the LocalForecasting algorithm. This number, while relatively small, is an average of 20 randomly sampled vertices-with one result reaching up to 14 vertices and 26 edges over baseline. In addition, we can see that the Musician dataset does not perform as well as the Aminer dataset, with most of the prediction results performing worse than the baseline (as indicated by the negative difference). We believe that this result on the Musician dataset is due to the different nature of each dataset's network structure- with the Musician dataset being significantly more sparse (no giant connected component) than the Aminer dataset. 4.3 Efficiency For both the proposed and baseline algorithms, the efficiency re- sults represent only the time taken to run the algorithm (e.g. not including loading data into memory). LocalProximity. Across a majority of the parameters the run time for the full graph PageR- ank computation is approximately 450 seconds longer compared to computing the PageRank vector based on the LocalProximity sugraph. AttriPart. In Figure 6, we see that the AttriPart al- gorithm finds local partitions 43× faster than PageRank-Nibble. Figure 7: Each data point averages 20 randomly sampled ver- tices in the Aminer and Musician datasets. Default parame- ters (unless sweeped across): αn = 0.2, αr = 0.15, ϕo = 0.2, ts = 5, te = 0.7, nw = 10,000, ns = 200. Parameter ranges: te [0.1-0.9] in 0.1 intervals, ϕo [0.1-0.6] in 0.1 intervals. LocalForecasting. This algorithm has an expected run time nearly identical to AttriPart, we therefore refer the reader to Figure 6 for run time results. 5 RELATED WORK We provide a high level review of both local and global community detection methods, with a focus on the research that pertains to the algorithms we propose in this paper. A - Local Community Detection. Given an undirected graph, start vertex and a target conductance-the goal of Nibble is to find a subset of vertices that has conductance less than the target con- ductance [19]. This algorithm has strong theoretical properties with a run time of O(2b(loд6m)/ϕ4), where b is a user defined con- stant, ϕ is the target conductance and m is the number of edges. PageRank-Nibble builds on the work of Nibble by introducing the use of personalized PageRank [9, 22], in addition to an algorithm for the computation of approximate PageRank vectors [1]. Since PageRank-Nibble and Nibble run on undirected graphs, they use truncated random walks in order to prevent the stationary distribu- tion from becoming proportional to the degree centrality of each node [8]. There are also many alternative techniques for local com- munity detection. To name a few, the paper by Bagrow and Bollt [2] introduces a method of local community identification that utilizes an l-shell spreading outward from a start vertex. However, their algorithm requires knowledge of the entire graph and is therefore not truly local. The research by J. Chen et. al. [4] proposes a method for local community identification in social networks that avoids the use of hard to obtain parameters and improves the accuracy of identified communities by introducing a new metric. In addition, the work by [29] and [25] introduces two methods of local com- munity identification that take into account high-order network structure information. In [29], the authors provide mathematical guarantees of the optimality and scalability of their algorithms, in addition to the generalization of it to various network types (e.g. signed and multi-partite networks). B - Global Community Detection. The basic idea behind the Walktrap algorithm is that random walks on a graph tend to get "trapped" in densely connected parts that correspond to communi- ties [18]. Utilizing the properties of random walks on graphs, they define a measurement of structural similarity between vertices and Local Partition in Rich Graphs KDD'18, August 2018, London, United Kingdom between communities, creating a distance metric. The algorithm itself has an upper bound of O(mn2). Another popular choice for global community detection is spectral analysis. In the paper by M. Newman [15] it is shown that the problems of community detection by modularity maximization, community detection by statistical inference and normalized-cut graph partitioning when tackled us- ing spectral methods, are in fact, the same problem. The work by S. White et. al. in [24] attempts to find communities in graphs using spectral clustering. They achieve this by using an objective func- tion for graph clustering [16] and reformulating it as a spectral relaxation problem, for which they propose two algorithms to solve it. A systematic introduction to spectral clustering techniques can be found in [23]. There also exists many alternative techniques for global community detection. Among others, two interesting techniques relevant to this work are [11] [20]. In [11], the authors propose a community detection algorithm that uses the information in both the network structure and the node attributes, while in [20] the authors use network feature extraction to predict the evolution of communities. A detailed review of various community detection algorithms can be found in [28]. 6 CONCLUSION This paper proposes new algorithms for attributed graphs, with the goal of (i) computing denser local graph partitions and (ii) pre- dicting the evolution of local communities. We believe that the proposed algorithms will be of particular interest to data mining researchers given the computational speed-up and enhanced dense local partition identification. The proposed local partitioning al- gorithm AttriPart has already deployed to the web platform PathFinder (www.path-finder.io) [7] and allow users to interac- tively explore all three datasets presented in the paper. In addition, the source code and datasets will be made publicly available by the conference date. REFERENCES [1] R. Andersen, F. Chung, and K. Lang. 2006. Local Graph Partitioning using PageR- ank Vectors. In 2006 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS'06). 475–486. https://doi.org/10.1109/FOCS.2006.44 [2] James P. Bagrow and Erik M. Bollt. 2005. Local method for detecting communities. Phys. Rev. E 72 (Oct 2005), 046108. Issue 4. https://doi.org/10.1103/PhysRevE.72. 046108 [3] Albert-László Barabási and Réka Albert. 1999. Emergence of Scaling in Random Networks. Science 286, 5439 (1999), 509–512. https://doi.org/10.1126/science.286. 5439.509 arXiv:http://science.sciencemag.org/content/286/5439/509.full.pdf [4] J. Chen, O. ZaÃŕane, and R. Goebel. 2009. Local Community Identification in Social Networks. In 2009 International Conference on Advances in Social Network Analysis and Mining. 237–242. https://doi.org/10.1109/ASONAM.2009.14 [5] Pierre Dupont, J Callut, G Dooms, J N. Monette, and Yves Deville. 2017. Relevant subgraph extraction from random walks in a graph. (12 2017). [6] Michalis Faloutsos, Petros Faloutsos, and Christos Faloutsos. 1999. On Power-law Relationships of the Internet Topology. In Proceedings of the Conference on Appli- cations, Technologies, Architectures, and Protocols for Computer Communication (SIGCOMM '99). ACM, New York, NY, USA, 251–262. https://doi.org/10.1145/ 316188.316229 [7] Scott Freitas, Hanghang Tong, Nan Cao, and Yinglong Xia. 2017. Rapid Analysis of Network Connectivity. In Proceedings of the 2017 ACM on Conference on In- formation and Knowledge Management (CIKM '17). ACM, New York, NY, USA, 2463–2466. https://doi.org/10.1145/3132847.3133170 [8] Vince Grolmusz. 2015. A Note on the PageRank of Undirected Graphs. Inf. Process. Lett. 115, 6 (June 2015), 633–634. https://doi.org/10.1016/j.ipl.2015.02.015 [9] T. H. Haveliwala. 2003. Topic-sensitive PageRank: a context-sensitive ranking algorithm for Web search. IEEE Transactions on Knowledge and Data Engineering 15, 4 (July 2003), 784–796. https://doi.org/10.1109/TKDE.2003.1208999 [10] Chin-Chi Hsu, Yi-An Lai, Wen-Hao Chen, Ming-Han Feng, and Shou-De Lin. 2017. Unsupervised Ranking Using Graph Structures and Node Attributes. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining (WSDM '17). ACM, New York, NY, USA, 771–779. https://doi.org/10.1145/ 3018661.3018668 [11] Jure Leskovec Jaewon yang, Julian McAuley. 2013. Community Detection in Networks with Node Attributes. ICDM (2013). [12] Ravi Kannan, Santosh Vempala, and Adrian Vetta. 2004. On Clusterings: Good, Bad and Spectral. J. ACM 51, 3 (May 2004), 497–515. https://doi.org/10.1145/ 990308.990313 [13] Songsong Liu Lazaros G. Papageorgiou Sophia Tsoka Laura Bennett, Aris- totelis Kittas. 2014. Community Structure Detection for Overlapping Modules through Mathematical Programming in Protein Interaction Networks. PLOS ONE (2014). https://doi.org/10.1371/journal.pone.0112821 [14] David Liben-Nowell and Jon Kleinberg. 2007. The Link-prediction Problem for Social Networks. J. Am. Soc. Inf. Sci. Technol. 58, 7 (May 2007), 1019–1031. https://doi.org/10.1002/asi.v58:7 [15] Mark E. J. Newman. 2013. Spectral methods for network community detection and graph partitioning. CoRR abs/1307.7729 (2013). [16] M. E. J. Newman and M. Girvan. 2004. Finding and evaluating community structure in networks. Physical Review E 69, 026113 (2004). [17] M.E. J. Newman. 2005. A measure of betweenness centrality based on random walks. Social Networks 27, 1 (2005), 39 – 54. https://doi.org/10.1016/j.socnet.2004. 11.009 [18] Pascal Pons and Matthieu Latapy. 2005. Computing Communities in Large Net- works Using Random Walks. In Proceedings of the 20th International Conference on Computer and Information Sciences (ISCIS'05). Springer-Verlag, Berlin, Heidelberg, 284–293. https://doi.org/10.1007/11569596_31 [19] Daniel A. Spielman and Shang-Hua Teng. 2013. A Local Clustering Algorithm for Massive Graphs and Its Application to Nearly Linear Time Graph Parti- tioning. SIAM J. Comput. 42, 1 (2013), 1–26. https://doi.org/10.1137/080744888 arXiv:https://doi.org/10.1137/080744888 [20] M. Takaffoli, R. Rabbany, and O. R. ZaÃŕane. 2014. Community evolution pre- diction in dynamic social networks. In 2014 IEEE/ACM International Confer- ence on Advances in Social Networks Analysis and Mining (ASONAM 2014). 9–16. https://doi.org/10.1109/ASONAM.2014.6921553 [21] Chayant Tantipathananandh, Tanya Berger-Wolf, and David Kempe. 2007. A Framework for Community Identification in Dynamic Social Networks. In Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '07). ACM, New York, NY, USA, 717–726. https://doi.org/10.1145/1281192.1281269 [22] Hanghang Tong, Jingrui He, Mingjing Li, Wei-Ying Ma, Hong-Jiang Zhang, and Changshui Zhang. 2006. Manifold-Ranking-Based Keyword Propagation for Image Retrieval. EURASIP Journal on Applied Signal Processing 2006 (2006), Article ID 79412, 10 pages. doi:10.1155/ASP/2006/79412. [23] Ulrike von Luxburg. 2007. A tutorial on spectral clustering. Statistics and Com- puting 17, 4 (01 Dec 2007), 395–416. https://doi.org/10.1007/s11222-007-9033-z [24] Scott White and Padhraic Smyth. [n. d.]. A Spectral Clustering Approach To Finding Communities in Graphs. 274–285. https://doi.org/10.1137/1.9781611972757.25 arXiv:http://epubs.siam.org/doi/pdf/10.1137/1.9781611972757.25 [25] Hao Yin, Austin R. Benson, Jure Leskovec, and David F. Gleich. 2017. Local Higher- Order Graph Clustering. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '17). ACM, New York, NY, USA, 555–564. https://doi.org/10.1145/3097983.3098069 [26] Sune Lehmann Yong-Yeol Ahn, James P. Bagrow. 2010. Link communities reveal multiscale complexity in networks. Nature (August 2010), 761–764. https://doi. org/doi:10.1038/nature09182 [27] Jing Zhang, Jie Tang, Cong Ma, Hanghang Tong, Yu Jing, Juanzi Li, Walter Luyten, and Marie-Francine Moens. 2017. Fast and Flexible Top-k Similarity Search on Large Networks. ACM Trans. Inf. Syst. 36, 2, Article 13 (Aug. 2017), 30 pages. https://doi.org/10.1145/3086695 [28] Claudio J. Tessone Zhao Yang, RenÃľ Algesheimer. 2016. A Comparative Analysis of Community Detection Algorithms on Artificial Networks. Scientific Reports (2016). https://doi.org/doi:10.1038/srep30750 [29] Dawei Zhou, Si Zhang, Mehmet Yigit Yildirim, Scott Alcorn, Hanghang Tong, Hasan Davulcu, and Jingrui He. 2017. A Local Algorithm for Structure-Preserving Graph Cut. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '17). ACM, New York, NY, USA, 655–664. https://doi.org/10.1145/3097983.3098015 [30] Leonid Zhukov. [n. d.]. works. PLriUvS7IljvkBLqU4nPOZtAkp7rgpxjg1&index=11 ([n. d.]). Structural Analysis and Visualization of Net- https://www.youtube.com/watch?v=jIS5pZ8doH8&list=
1803.07509	2	1803	2018-03-26T02:04:30	Workforce migration and its economic implications: A perspective from social media in China	[ "cs.SI", "physics.soc-ph" ]	The workforce remains the most basic element of social production, even in modern societies. Its migration, especially for developing economies such as China, plays a strong role in the reallocation of productive resources and offers a promising proxy for understanding socio-economic issues. Nevertheless, due to long cycle, expensive cost and coarse granularity, conventional surveys face challenges in comprehensively profiling it. With the permeation of smart and mobile devices in recent decades, booming social media has essentially broken spatio-temporal constraints and enabled the continuous sensing of the real-time mobility of massive numbers of individuals. In this study, we demonstrate that similar to a natural shock, the Spring Festival culturally drives workforce travel between workplaces and hometowns, and the trajectory footprints from social media therefore open a window with unparalleled richness and fine granularity to explore laws in national-level workforce migration. To understand the core driving forces of workforce migration flux between cities in China, various indicators reflecting the benefits and costs of migration are introduced into our prediction model. We find that urban GDP (gross domestic product) and travel time are two excellent indicators to help make predictions. Diverse migration patterns are then revealed by clustering the trajectories, which give important clues to help understand the different roles of Chinese cities in their own development and in regional economic development. These patterns are further explained as a joint effect of the subjective will to seek personal benefits and the capacity requirements of local labour markets. Our study implies that the non-negligible entanglement between social media and macroeconomic behaviours can be insightful for policymaking in social-economic issues.	cs.SI	cs	Workforce migration and its economic implications: A perspective from social media in China Xiaoqian Hu1, Junjie Wu1,2 and Jichang Zhao1,2,(cid:63) 1School of Economics and Management, Beihang University 2Beijing Advanced Innovation Center for Big Data and Brain Computing, Beihang University (cid:63)Corresponding author: [email protected] Abstract The workforce remains the most basic element of social production, even in modern societies. Its migration, especially for developing economies such as China, plays a strong role in the re- allocation of productive resources and offers a promising proxy for understanding socio-economic issues. Nevertheless, due to long cycle, expensive cost and coarse granularity, conventional surveys face challenges in comprehensively profiling it. With the permeation of smart and mobile devices in recent decades, booming social media has essentially broken spatio-temporal constraints and enabled the continuous sensing of the real-time mobility of massive numbers of individuals. In this study, we demonstrate that similar to a natural shock, the Spring Festival culturally drives workforce travel between workplaces and hometowns, and the trajectory footprints from social media therefore open a window with unparalleled richness and fine granularity to explore laws in national-level workforce migration. To understand the core driving forces of workforce migration flux between cities in China, various indicators reflecting the benefits and costs of migration are in- troduced into our prediction model. We find that urban GDP (gross domestic product) and travel time are two excellent indicators to help make predictions. Diverse migration patterns are then revealed by clustering the trajectories, which give important clues to help understand the different roles of Chinese cities in their own development and in regional economic development. These patterns are further explained as a joint effect of the subjective will to seek personal benefits and the capacity requirements of local labour markets. Our study implies that the non-negligible entan- glement between social media and macroeconomic behaviours can be insightful for policymaking in social-economic issues. 8 1 0 2 r a M 6 2 ] I S . s c [ 2 v 9 0 5 7 0 . 3 0 8 1 : v i X r a 1 1. Introduction Given its importance, the workforce remains one of the most basic elements of social production in modern societies, even in the age of the knowledge economy. The workforce's intercity migration within a country, usually between hometowns and workplaces, essentially reflects the reallocation of production resources that is profoundly responsible for the normal operation of the national economy and the promotion of production and operating activities. Accordingly, workforce migration exerts a strong influence on social-economic issues [1, 33], to name a few examples, including the balanced development of the regional economy, national transportation design, urban infrastructure planning, policymaking in household registration systems and climate change mitigation [32]. Understanding the laws of workforce migration at the national scale will therefore help to solve these issues. Although individual movement is stochastic, the collective behaviour of migration is not random but closely linked to extensive social, economic and even political factors [6]. Laws or patterns underlying group travel can emerge as the population size increases [16, 26], suggesting that the laws of workforce migration can be probed and modelled. In fact, tremendous efforts have been devoted to profiling and forecasting human migration flux in recent years [3, 23, 34, 44–46]. In addition, diverse patterns of human mobility are revealed [4, 8, 10, 11, 13, 20, 26, 28, 31, 38], and their driving forces are also extensively explored [5, 7, 9, 18, 22, 29, 35–37, 41]. However, most existing studies assume that human migration behaviours are determined only by passively accepting an attraction from cities, neglecting people's subjective will to seek benefits and the capability requirement of distinct labour markets that indeed pro- foundly affect workforce movement. In fact, from the viewpoint of economics, the driving force behind workforce migration is a higher economic benefit or the Pareto optimality of the economy [5, 7, 18, 29, 36]. The development of science and technology also leads to a major labour market concern that tends to focus more on technical capacity than on the size of the workforce. Moreover, from the cognition perspective, sensitivity regarding influence factors varies across different individuals [14]. Hence, the cost of migration should not always be represented by the spherical distance, as in previous studies. Richer information should be included as indicators to demonstrate that the workforce possesses different cognitive sensitivities to indicators when making behavioural decisions [47]. 2 Apart from the above limitations, the data adopted in previous studies have adverse ef- fects on the ability to carry out an in-depth study of workforce migration. Specifically, survey samples from questionnaires or government censuses constitute the major data sources for previous studies on the workforce [4, 11, 46]. However, it is well known that limited samples, biased sample selection, and the uncontrollable quality of investigations might lead to un- reliable survey data. Although government censuses supply valid nationwide records, they involve an enormous expense and are quite time-consuming. Indeed, population censuses usually have a very long survey circle (once per decade, as designated by the government of China), and the resulting non-current investigations are not helpful for flexible and effective policy formulation. Moreover, it is difficult to obtain a fine-grained survey in geographical terms. For instance, the population census in China supplies only the migration flux between provinces (see http://www.stats.gov.cn/tjsj/pcsj/), while fine-grained inter-city trajectories are necessary to investigate workforce migration on the city level. In summary, traditional investigations cannot offer the desired data in terms of cost, granularity and reliability. Online media, such as Twitter and its variant Weibo in China, have boomed in recent decades and offer an unprecedented opportunity to understand human movements. Massive numbers of users leave their digital footprints as messages on social media, including both spatial and temporal information. The continuous penetration of smart and mobile devices further boosts the scale and granularity of these footprints, facilitating the investigation of human movements on national and systemic levels. In order to distinguish the workforce migration behaviours from the movements for other purposes, we adopt the Spring Festi- val scenario. Being the most important traditional custom in China, the Spring Festival culturally drives people working in other places back to their hometowns to reunite with their families and celebrate the Chinese New Year. It can be treated as a natural shock in workforce movement on a national scale, leading to the full-scale migration with high traffic peaks before and after the Lunar New Year, namely, the Spring Festival travel rush. Accordingly, the resulting trajectories form an ideal data source for modelling workforce migration on a national scale. In this paper, we collect massive trajectories of movements from Weibo during the 2017 Spring Festival travel rush and establish a nearly complete picture of workforce migration in China at the city level. To explore the core driving force of this systematic migration, the most widespread in the world, we introduce economic factors and alternative cost measures 3 to the well-known gravity model (GM) framework [15, 17, 48] to make high-quality pre- dictions. By evaluating the forecasting performance for real-life data, urban GDP and the travel time between cities in flux are two key factors that lead to the best performance. We then investigate the diverse patterns of workforce migration by clustering trajectories using the K-means algorithm, and the prediction accuracy of the models improves accordingly. More importantly, two dominant patterns are highlighted, i.e., the migration flows from small neighbourhoods to local core cities and the flows between cities in good economic statuses, especially in the eastern coastal region. We further provide a rigorous analysis of gaming between interest articulation and cost elusion in different patterns, and these transfer modes also embody the filter effect of the labour market when the workforce makes migration decisions. The results indeed testify to the interest-driven intrinsic motivation of the workforce and the extrinsic limiting condition of market requirements. Finally, an evalu- ation approach inspired by the dividing modes is proposed to value regional development in China. Some practical problems, such as the deficiency in regional leading roles for Beijing and Chongqing, are exposed, and a thorough discussion of corresponding strategies is also provided. Our models and findings could shed light on related social-economic issues such as urban planning, workforce management and regional development. 2. Results 2.1. Sensing the trajectories of inter-urban workforce migration An increasing number of people in China choose to leave their homes in search of job opportunities because of the relaxation of the policy that had restricted inter-city population movements since the early 1980s. Additionally, a great number of migrating workers travel back to their hometowns during the Spring Festival in what is known as the Spring Festival travel rush [21, 40, 42] due to its unique cultural and traditional meaning. Lasting 40 days surrounding the Spring Festival, the great travel rush defined by the National Development and Reform Commission begins on the 15th day of the twelfth month of the lunar year and ends on the 25th day of the first lunar month of the following year. We argue that the Spring Festival travel rush can be regarded as a natural experiment on the national scale, and it provides us with reliable crowdsourcing migration trajectories to probe workforce movement 4 patterns [25, 43]. Combined with the spread of the online social media and mobile devices, we find that it is possible to break through the limitation of traditional methods in order to gain a nearly complete profile on a nationwide scale for such a long period. Note that there must be noise in the collected trajectories, but fortunately, those few noisy movements by non-workforce participants have no pronounced influence on our findings. Specifically, messages in Weibo contain the locations where they are posted, and thus, a 4-tuple (m, u, t, g) can be used to denote a message m posted by user u in timestamp t at location g of city-granularity. For a user u, an array L(u) = {(tu,1, gu,1), (tu,2, gu,2), . . . , (tu,m, gu,m)} is defined to denote all locations u experienced, where tu,i represents the time at which u posts a message (tu,1 < tu,2 < . . . < tu,m) and gu,i is the corresponding posting city. Then, we filter the same locations in which people post messages in subsequent iterations. If {(tu,i, gu,i), (tu,i+1, gu,i+1), . . . , (tu,i+k, gt,i+k)} ⊆ L(u), gu,i = gu,i+1 = . . . = gu,i+k and gu,i−1 (cid:54)= gu,i , gu,i+k+1 (cid:54)= gu,i+k, only (tu,i+k, gu,i+k) is reserved. Finally, we obtain L(u) = {(tu,1, gu,1), (tu,2, gu,2), . . . , (tu,n, gu,n)}, where gu,i (cid:54)= gu,i+1, i ∈ [1, n − 1]. Additionally, the trajectory of u is cut as (gu,1, gu,2), (gu,2, gu,3), . . . , (gu,n−1, gu,n). Then, we can gain the transition set by summing all users' trajectories as {(g1, g2, f1,2), . . . , (gi, gj, fi,j), . . .}, where fi,j denotes the migration flux from city gi to gj. In the same way, the undirected transition set can be obtained as {(g1, g2, F1,2), . . . , (gi, gj, Fi,j), . . .}, where Fi,j represents the number of people in the population transferring between cities gi and gj, suggesting that Fi,j = fi,j +fj,i and Fi,j = Fj,i. More particularly, 371 cities are contained in the transition set, and there are 61,759 city-pairs in undirected trajectories and 120,361 city-pairs in directed trajectories, totaling 41,454,268 migration traces (see Methods). To prove the rationality of using data from Weibo to investigate workforce migration problems, we collect national railway line data, 5,878 trains in total, and count the number of trains through city pairs. It shows that the flow of workforce migration has a positive correlation with the number of trains between city pairs (see Fig. S1). Therefore, workforce migration flux acquired from social media is representative and indeed embodies population movement in real life from the perspective of trains. Furthermore, it cannot be neglected that only approximately one-fifth of city pairs extracted from Weibo data can be explained by railway line data, specifically, 12,382 city pairs. According to the comparison above, Weibo reflects richer information about workforce migration than the national railway line does. Therefore, the rationality and completeness of migration data gained from Weibo 5 builds a foundation to carry out the following in-depth investigations. 2.2. Modelling inter-urban workforce migration In the classical gravity model, human migration is assumed to be dependent on the at- tractiveness and accessibility of locations. The former is always estimated as the population size, and the latter is calculated according to the geographical distance between two loca- tions. In detail, the migration flux between cities i and j is defined as Fij = a Pi·Pj denotes the number of people residing in city i, Fij represents the migration flow between , where Pi dγ ij i and j, dij is the metric of distance between the two cities, a is a constant to adjust the forecast value, and γ is usually assumed to be positive. This model implies that the migra- tion flow increases with the population size and decreases with the distance between the two locations. The motivation for workforce migration has been extensively investigated in previous ef- forts, and it has been proven that migrating behaviour is not completely blind or disorderly but rather abides by certain laws. Based on the microeconomic perspective, profit maxi- mization, utility optimization and intuitive revenue enhancement [5, 7, 36, 37] are treated as the target of supplying labour to discuss the specific influence factors when workers make migration decisions. In addition, from the sociological perspective, human migration aims at improving living conditions and is the result of push and pull dynamics [18, 29]. More specifically, immigrant areas with abundant job opportunities, higher wage levels and other preferential treatment conditions produce a "pull" force. Additionally, harsh living condi- tions in the original place of residence would "push" human emigration. Following the idea of a subjective measure for the pursuit of profit maximization, we introduce economic factors rather than demographic factors into our models. Based on this, we propose an extension of GM as follows. Fij = a i · Eβ Eα C γ ij j , (1) where Ei stands for the economic indicator of city i, and Cij represents the cost of migration between city i and city j. There is an assumption that the labour force would prefer to migrate to cities with a high economic status and more job opportunities, and these cities could provide more potential for the workforce to meet labour market demands, meaning 6 that α and β should be positive. Additionally, γ is still assumed to be positive, indicating that the workforce prefers to migrate to cities with a low cost of migration. With regard to cities' economic indicators, GDP is one of the most important indices of national economic accounting, containing abundant information about social and economic situations, and it is almost always available, especially in undeveloped areas. As shown in Fig. 1, the GDP of different provinces is positively correlated with per capita disposable income, implying that the workforce is more likely to have better revenue in a province with high GDP. Furthermore, the investment of funds in research and development (R&D) has a strong positive correlation with provinces' GDP. Intuitively, this effect is due to greater investment in R&D and more advanced degrees in the market. Additionally, the ratios of practitioners in high-technology industries, such as the information technology industry, the financial industry, the real estate industry, and the scientific and technical services in- dustry, all show significant positive correlations with GDP in Fig. 1. This result means that provinces with better economic statuses have higher technology-related workforce re- quirements and that the market in these provinces could offer more highly skilled labour. As discussed above, GDP can simultaneously reflect a region's income level and required level of labour skills. It is natural to select GDP as an economic indicator in our model, specifically called the G-GM model. Considering the demographic information included in GDP, we select per capita GDP as another economic indicator because it can be better compared to the classical GM model with a population reflecting cities' attraction; we call this model the aveG-GM model. To determine the relationship between the origin and the destination in detail, we introduce a variant of G-GM, named dirG-GM. Unlike the three models introduced above, Ei and Ej in dirG-GM denote the total GDP of the origin and the destination, and Fij in Eq. 1 is changed to fij, the migration flux from city i to j. Because of the undirected migration flux between cities in G-GM and aveG-GM, we simplify the parameters α and β to 1. With respect to the index for migration cost, the migration cost between two cities can be intuitively measured in terms of the geographical distance, or spherical distance; for example, in GM, the spherical distance is calculated using the longitudes and latitudes of city centers. However, due to the complexity of China's geographical circumstances, the same geographical distance does not suggest the same accessibility between cities. For instance, the reachability between two cities in a mountainous area is worse than that in 7 plains regions with the same geographical distance. In this case, direct geographical distance might provide less information to the workforce than metrics such as travel time. Therefore, the wealth of information included in metrics shows great differences. In the meantime, from the perspective of decision making, the sensitivity of the labour force to indicators is diverse. Hence, alternative measures such as the shortest travel distance and travel time, which are both obtained from the navigation system of the Baidu Map, enrich the distance metrics in all models. The forecasting ability of the four models is evaluated in terms of the fitness between predicted flux and realistic flux, as shown in Fig. 2, where the metric of distance is defined as the travel time (results with alternative distance metrics can be found in Fig. S2 and Fig. S3). As can be seen, independent of distance measures, extensions of the G-GM and dirG-GM models demonstrate better accuracy than GM. To further quantify the forecasting ability in four models with different distance metrics, an indicator named SSI based on the Sørensen index [19, 24] is introduced to evaluate the model performance directly. In detail, , where Fij and Fij denote the actual flux and the forecast value of the size of the population migrating between city i and the measurement index is defined as SSI = j min(Fij , Fij ) Fij (cid:80) (cid:80) 2(cid:80) (cid:80) (cid:80) j Fij +(cid:80) i i i j city j, respectively. The range value for SSI is 0, with a complete difference between the actual value and the predictive value of 1 with complete equality. A larger SSI means a better predictive capability of models and vice versa. From Fig. 3, it can be seen that the predictive performance of G-GM with travel time being a cost metric is the best approach among four models, and the SSI index of GM and aveG-GM is obviously worse. This finding further illustrates that GDP demonstrates adequate information about the city's ability to promote labour mobility, in contrast to per capita GDP, which is conventionally employed to measure demographic and economic influences simultaneously. Because of G-GM's stronger performance, the subjective consciousness of the workforce, e.g., pursuit of profit, is indeed profound in helping understand migration. Consistent with the previous study [47], the appearance of travel time as the most appropriate distance metric also suggests that travel time weighs heavily in choosing migration destinations, instead of the geographical distance. In addition, as reported in Table 1, the exponents for the distances between cities and eco- nomic levels of origin and destination cities are all positive, suggesting that our assumption of more migration flows between close and well-developed cities may be justified. Addition- ally, the positiveness of γ in all models clearly indicates that distance blocks labour mobility. 8 However, aveG-GM has the worst performance, which suggests that GDP per capita is not a good choice. Note that α and β are close to each other in dirG-GM (0.9278 and 0.9351 separately), implying that economic levels (GDP) of the origin and destination exert an approximately equivalent effect on the migrant population of the workforce, although the destination has a slightly stronger influence than the origin. Therefore, the direction of flows can reasonably be ignored, and dirG-GM can be reduced to G-GM. Therefore, combined with predictive performance and the simplicity of parameters, G-GM with travel time is the best solution to predict inter-urban migration. 2.3. Mining inter-urban migration patterns Although the G-GM model takes economic status as a driving force in workforce migra- tion, the sensitivity of individuals to travel costs and economic benefits is diversified because of different backgrounds. Thus, this sensitivity is significant for understanding workforce migration behaviours from a micro-perspective. We argue that instead of exploring this phenomenon at the collective level, focusing the analysis by mining workforce migration patterns of different sensitivities might suggest a new way to visualize the big picture of workforce migration. To effectively detect workforce migration patterns, an unsupervised approach of K-means is employed to cluster the migration trajectories of massive numbers of individuals (see Methods). As for identifying the optimal number of clusters, two methods, the silhouette coefficient [30] and elbow method [39], are both used. We find that excellent clusterings can be achieved with three, four, or five clusters (see Fig. S4-S5). To obtain the best choice, dif- ferent partitions are compared thoroughly, and the partition with four clusters demonstrates more patterns than that with three clusters and patterns with greater differences than that with five clusters (see Fig. S6-S7). Thus, the workforce is grouped into four explicable clusters, as shown in Fig. 4. As can be seen, trajectories in four clusters have significantly different patterns from the perspectives of geography (see Fig. 4), the product of GDP (see Fig. S8), travel time (see Fig. S9) and migration flux (see Fig. S10). All the patterns are depicted and explained as follows. Pattern I Migration between local core cities and their surroundings. As 9 demonstrated in Fig. 4(a), divergent, star-like patterns of trajectories indicate local yet core cities such as Guangzhou, Zhengzhou, Changsha, Wuhan, and Hefei, which are capital cities of provinces and which have a great economic influence within their neighbourhoods. The workforce in this group spends the least travel time and dominates the migration flux more than other groups. This finding implies that the essential component of migration flows between capital cities and their adjacent areas and travel time has an intensive effect on migration. The radiation to labour forces within relatively small areas of local core cities, especially capital cities, is reflected in this pattern. Pattern II Migration between developed cities. As can be seen in Fig. 4(b), city pairs show the best economic statuses in this pattern, for instance, the most developed cities, such as Beijing, Shanghai, Chongqing, and Suzhou, are key destinations for the workforce of this group. This finding suggests that this pattern of workforce migration is more sensitive to cities' economic levels. Apart from that, the city pairs reflect that the workforce capacity in cities with high GDP match the requirements of the labour market of similarly developed cities and provide the workforce with more opportunities to obtain a lucrative job. As compared to Pattern I, migration flows in this pattern involve the second-lowest amount of travel time. Additionally, the mutual promotion of city pairs in this group, which are better developed, is fully reflected. Pattern III Migration between undeveloped cities. From the point of view of regional distribution, there are distinct differences between Pattern III and the former two patterns, and the bright region of this pattern concentrates mainly in central China, such as Jinzhong, Sanmenxia, and Xianning, among other cities. Cities experiencing this migration pattern are less extrusive than those in the former two groups, as can be seen in Fig. S8, and the product of GDP in these trajectories is relatively small. Based on the above discussion, we know that a city's GDP may not only have potential benefits for the labour market of this city but also represent technology capability requirements (see Fig. 1). Therefore, the workforce in this pattern, lacking adequate technical skills, cannot meet the capacity requirements of other labour markets in cities with high GDP, as discussed in Pattern II, and must select other cities with lower GDP as the migration destination. In comparison to Pattern II, the filtering effect of markets is highlighted, and the intrinsic motivation for benefit appeals is limited by objective constraints. Pattern IV Migration due to emotions. This group's trajectories geographically 10 cover all areas of China, particularly peripheral provinces such as Xinjiang, Tibet, Hei- longjiang, and Hainan and other remote districts located on the country's borders, including the cities of Tonghua, Haikou, Mudangjiang, Hulun Buir, and Qiannanzhou. City pairs in this pattern possess the lowest product of GDP but the longest travel time compared to other groups. Meanwhile, the average volume of migration between city pairs in this pat- tern is the smallest, and the total flux in this pattern accounts for a small proportion of all migration volume, implying that these few "emotional" migrations may be not completely profit-driven. Although some city pairs reflect labour migration to distant cities to escape harsh living environments, such as migration from Tibet and Xinjiang, much migration is closely related to tourism and sightseeing and protection from winter; the key cities in this pattern, Mudanjiang and Hulun Buir, are located in the cold region of northeast China, and Haikou and Qiannanzhou have a warmer southern climate. The dominant occupation of the workforce in Pattern I indeed illustrates the agglom- eration effects of core cities to their surroundings in local districts. More surprisingly, these surroundings are not less economically developed cities, as intuitively expected, but demon- strate relatively better economic statuses within their provinces (GDP product ranks second in four groups; see Fig. S8). This pattern can also be further justified by the similar values for α and β in dirG-GM (see Table 1), suggesting the identical roles of origins and desti- nations in economic attraction. In fact, as super-cities or mega-cities, most capital cities expand continuously, according to the development report of the Chinese floating population in 2016, and our results fully confirm this point. To some extent, this finding implies that the organization pattern of urban groups within provinces in China is gradually formed. In addition, considering city-pairs with good economic development and located a short dis- tance apart in Pattern II, the changes in labour demand are particularly apparent. With the development of these prosperous regions, labour with knowledge and techniques is more needed than simple manual labour, explaining the fact that few people in the workforce would like to move between less affluent cities located a long distance apart, as reflected in Patterns III & IV. It is increasingly difficult to gain employment positions in developed areas because of workers' lack of skills. Thus, a workforce that pursues profit maximization not only needs a developed destination that provides many job opportunities but also needs an origin featuring a satisfactory economic level. We further use G-GM to model the trajectories of different groups. As shown in Fig. S11 11 and Fig. S12, the prediction performance of grouped trajectories are all better than that of all trajectories, especially for the first two groups, which occupy the largest proportion (see also Fig. S10). The significantly enhanced forecasting ability suggests the rationality of clustering trajectories, and the workforce possesses different migration preferences (see also Fig. S11 and Fig. S12). In fact, individuals' sensitivity to migration costs can be sufficiently reflected by the gaming between economic benefits and migration costs, and the variations in γ across groups are decent proxies to probe this phenomenon. As listed in Table 2, γ varies for different patterns and is clearly distinguished from the parameter of all trajectories, further implying the existence of workforce sensitivity to migration costs. Specifically, the small γ (0.1910) for the first group indicates that labour is more sensitive to economic attraction and cares little about the travel time. With the incremental increase in travel time between cities, the sensitivity to travel costs is significantly improved; for example, in the second group, γ rises to 0.3868, suggesting the priority of moving to well- developed neighbouring cities. Similarly, as economic statuses worsen in Pattern III, γ jumps to 0.4186, reflecting the greatest sensitivity to travel time. However, the γ for the last group actually falls with the longest travel time and the worst economic status compared to that of the other groups. Under this circumstance, the migration followed by the trajectories in Pattern IV are not affected by the economic benefit and the travel time, e.g., emotional motivation instead of profit-seeking motivation. Considering the specificity of the Spring Festival travel rush as a long holiday in winter, we conjecture that migration for nonprofit reasons, such as tourism sightseeing, protection from winter or visits with relatives, is filtered out and isolated in this cluster. For instance, as is well known, an increasing number of people from northeast areas have selected Hainan province as a place to seek refuge from the cold and to celebrate the Spring Festival in recent years. From Fig. 4, we also find the trajectories between cities in Hainan are indeed prominent in Pattern IV compared to other groups; to a certain extent, this pattern confirms our speculation about escaping cold winters. In fact, Pattern I exactly embodies the development model for local core cities, e.g., their leading impacts on their surroundings. Pattern II further illustrates the mutual promotions among these similarly developed cities. In comparison to the other two patterns, the development models reflected in Patterns I & II are more strongly expected to sustain growth. Therefore, the flux ratio of cities in Patterns I & II reflects the sustainability 12 of city development models, and we define this ratio as an indicator DI to evaluate cities' development patterns. From Fig. 5, we find that there are obvious variances in DI, and the cities with high development levels are scattered across the country. Additionally, from the view of the geographical distribution, the value of DI in the country as a whole appears to significantly and progressively increase in a hierarchy from the western regions to the eastern areas, reflecting the disproportion and irrationality of development models in many regions, especially the western and northeastern areas. While unable to evaluate individual cities, this indicator is also applied to assess regional development models on a larger scale. Hence, the flux ratio in different patterns of provinces is used to further value the province development in Fig. 6. There are significant differences among provinces in terms of the ratio of different patterns. The provinces with high ratios of Patterns III & IV, as we well understand, are all undeveloped areas, such as Tibet, Xinjiang, Yunnan and etc. 2.4. Regional development models inspired by migration patterns It is well known that urban agglomeration is the organization of space in a mature stage of city development. On one hand, because of their small-scale industries and relatively simple structures, small cities are not closely associated with greater production and operation networks and are considered to lack development vitality and momentum. There is an urgent need to optimize the allocation of a wider range of resources. On the other hand, from the perspective of big cities, many economic activities are no longer limited to a certain city with the growth of the urban population, the extension of industrial chains and the improvement of transportation infrastructures. Hence, it is necessary to strengthen ties among different cities in certain regions and create a collection of relatively win-win city pairs through appropriate regional development models. Meanwhile, urban diseases could be relieved by spatial expansion and the optimization of urban functions. Therefore, it is significant to promote regional development by making local cities play a leading role for neighbouring cities. More specifically, we hope for more Pattern I migration between city pairs. To understand the relation between cities and their surroundings in detail, we compute the cumulative ratio of flux from proximal to distant cities. This method is effective, uni- versal and easy to execute in order to learn the present condition of the city development 13 model. From Fig. 7, we find that most cities are affected by their locality to some degree. In other words, for a certain city, more workforce members migrate from or to a closer city. Under these circumstances, local core cities could play a stronger role in radiating to peripheral cities, i.e., undeveloped cities could enhance their economic development by stimulating effects in nearby developed cities. At the same time, cities with unexpected development models are also detected; however, they lack locality and contribute little to regional development. In terms of cities with good statuses, we find that most cities with high GDP have more workforce migration with neighbouring cities, as the red lines in Fig. 7 show. Nevertheless, unexpectedly, two developed cities, Beijing and Chongqing, are very different from other cities with high GDP because of their lack of locality. In particular, compared to Guangzhou, Shanghai and Shenzhen, the most important cities in China, Beijing and Chongqing, make an extremely low contribution to boost the development of their local regions. Our finding indicates that Hebei province, which is close to Beijing, is not similar to the Yangtze Delta Metropolitan and Pearl River Delta regions, whose rapid economic development is led by Shanghai and Guangzhou, respectively. Beijing, as the capital of China, has unique advan- tages in economic production resources and should play an significant lead role in radiating region development. As stated by a report on the government's work in 2014, achieving the coordinated development of Beijing, Tianjin and Hebei is a major national strategy. To some degree, our findings support the government strategy from the perspective of workforce migration. Furthermore, we analyze the influence factors for why Beijing lacks regional promotion relative to Guangzhou, Shanghai and Shenzhen. In Fig. 8, by comparing the GDP of core cities and their surroundings, we find that the matching extent of economic status between Beijing, Chongqing and its surroundings is worse than that of Shanghai, Shenzhen and Guangzhou. For Beijing, there is just one city, Tianjin, which has a matching economic status around the periphery. Additionally, the GDP of cities close to Chongqing are less than almost one-tenth of that city's GDP. Combined with the discussion above, due to the disparity between the technical level of the workforce in the cities' surroundings and the skill requirements of the labour markets in both Beijing and Chongqing, the labour mobility between these city pairs is discouraged, and hence, their radiation effects are essentially undermined. To promote regional development in a coordinated manner, enhancing the 14 technical level of the workforce in undeveloped areas is a substantial requirement, except in policy-oriented economic cooperation. Almost all cities with a low GDP and a lack of locality, specifically, those ranking lower than Beijing, are in remote and undeveloped areas, for instance, Tibet, Xinjiang, Yunnan, and Guizhou. Combined with the discussion about Patterns III & IV above, the workforce in these areas makes migration decisions under the threshold filtering in markets, migrating to other cities with similarly low GDP, or for "emotional" reasons, neglecting the cost of the long-distance migration in order to escape their harsh life environments. To shift workforce migration from Patterns III & IV to Patterns I & II, a large, high-quality labour force is the key. From Fig. 9, we observe that the ratio of the investment of funds in R&D to GDP in provinces and the ratio of flux in Patterns III & IV is negatively correlated (r=-0.83). This finding once again confirms the need for cities and provinces to vigorously expand scientific research and advanced education. 3. Discussion The social media boom indeed offers an opportunity to understanding human behaviours in unparalleled richness and fine granularity. To the best of our knowledge, this is the first study to systemically explore the workforce migration from the perspective of social media. Our study overcomes the limitations of traditional survey-based approaches and justifies the feasibility of probing workforce migration both collectively and individually using social media. More importantly, the present study confirms the possibility of quasi real-time policy making regarding social-economic issues, which are generally greatly constrained by the long census cycle and lead to the problem of being "out-of-sync", which negatively influences the flexibility of policy formulation. Unlike previous models of human movement, in this study, we systematically discuss how the workforce migrates between cities within a country, considering the core driving force, interest articulation and the threshold filtering of the labour market. By introducing GDP as an economic indicator and travel time into the classical GM model, our G-GM model shows better performance than GM and justifies our assumptions. Additionally, due the diverse backgrounds of the workforce, the different migration patterns in terms of individual will and objective market requirements are mined and revealed. Meanwhile, we expound the gaming 15 behaviour on benefits sought and travel cost evasion in exact patterns. Based on the patterns we find, a simple and effective method to evaluate city development models is proposed and provides support to find and solve realistic problems to enhance the sustainability of regional development. Our results also imply that human behaviors like collective movement can be essentailly entangled with economic development, and more inportantly, this entanglement can be sensed and prifiled through social media only. In addition to regional economic development, a demonstrated application discussed pre- cisely in the paper, our study on workforce migration provides insightful ideas to solve a good deal of critical social-economic issues. The prediction model of migration flows offers a quantitative perspective to evaluate and design policies on household registration sys- tems, floating population management and inter-city transportation planning in real time. Meanwhile, the disclosed patterns of workforce migration not only have implications for the government in policy formulation and evaluation but also inspire enterprises and individuals to solve real issues. For example, the gaming between the benefit and distance costs of making migration strategies could offer suggestions regarding where to release recruitment information to ease or solve the problems of labour shortages, especially in the southeast coastal areas of China. 4. Conclusion In this study, geographical tweets posted on Weibo during the Spring Festival travel rush in 2017 are considered as a natural experiment on the national scale to understand the rules of workforce migration. Intrinsic motivation for the decision to migrate, such as the pursuit of interest and cost avoidance, are discussed systematically by introducing the economic indicators and travel costs into gravity laws. Additionally, a model called G-GM is presented to predict the migration flux. We find that the product of GDP and travel time between city pairs can be excellent indicators of workforce prediction; in particular, GDP reflects cities' potential benefits and the technical requirements of their labour markets. Under the combined effect of intrinsic motivation and external restrictions on the labour market, workforce migration in China presents four typical patterns. Moreover, individuals' sensitivity to migration costs can be well detected in terms of parameter fluctuation across different patterns. The migration flux ratio in divergent patterns provides us with a new 16 bellwether for development models for cities or provinces. As an application, the practical problems and causes of regional development – for instance, Beijing's and Chongqing's lack of a leading role in regional radiation and the mismatch between labour market skills in surrounding areas and labour market requirements in these two cities – are detected from workforce migration patterns and inspire related policy formulation and evaluation. 5. Methods 5.1. Data description Migration trajectories in Weibo. Tweets on Weibo during the Spring Festival travel rush, from January 13 to February 21, 2017, announced by the National Development and Reform Commission (see http://zizhan.mot.gov.cn/zfxxgk/) , are thoroughly collected. Weibo is a Twitter-like service in China and its active users are 392 million, more than 330 million that of Twitter. Our dataset is a complete snapshot of Weibo during the travel rush. A 4-tuple (m, u, t, g) is defined to represent a message with both temporal and spatial features, in which m is the message number, u is the unique identifier for an anonymous user, t is the timestamp of the post and g is the location, including the country, province and city information. Specifically, the posting city for each tweet can be further employed to establish the workforce migration network at the city level. Our dataset is publicly available and can be downloaded freely through https://doi.org/10.6084/m9.figshare.5513620.v2. The national railway line data. We collect national railway line data on a total of 5,878 trains from a train schedule. For each city pair, we count the number of trains through two cities. The dataset is publicly available and can be downloaded through https://doi.org/10.6084/m9.figshare.5513620.v2. Distance measures between cities. To measure inter-city accessibility, we use three metrics, the actual geographical distance, the travel distance and the travel time in route planning. The former is the distance computed by the longitude and latitude of cities on Google maps. More specifically, for city i, with longitude and latitude (xi, yi), and city j, with (xj, yj), the distance is defined as (cid:113) dij = D × arcsin( sin2(a) + cos(rxi) × cos(ryj) × sin2(b)), (2) where ryi = yi×π 180 , a = yi−yj 2 and rxi = xi×π 180 , b = xi−xj 2 , D is the diameter of the earth. The 17 travel distance and travel time between cities are crawled from the travel path planning in Baidu Map API. All distance measures between different city pairs are also publicly available and can be downloaded freely through https://doi.org/10.6084/m9.figshare.5513620.v2. Demographic and economic surveys. To introduce the economic driving force of labour migration in our model, we collect the GDP and per capita GDP of all cities from a statis- tical yearbook of all provinces in China. In order to compare our model with the classical GM model, the population size in cities is also collected from a statistical yearbook of all provinces in China. Because of the restrictions in our household registration system, there is a significant difference between the population size by households registered and that by permanent residents. In our study, we choose the number of permanent residents to de- note cities' population size. Because of the delay in publishing government statistics and the future forecasts using past data, we collect GDP, per capita GDP and the number of permanent residents in China in 2015 from the 2016 statistical yearbook. Furthermore, the ratio of practitioners in high-technology industries, the disposable income per capita and the investment of funds for research and development (R&D) are all collected from the 2016 statistical yearbook. The datasets are publicly available and can be downloaded through https://doi.org/10.6084/m9.figshare.5513620.v2. 5.1.1. Estimation of model parameters To obtain the parameters in the proposed models, we use the least squares polynomial i ·Gβ Gα dγ ij j fit. As an example, the process of solving the model dirG-GM fij = a will be discussed in more detail as follows. Because the flux between cities is greater than zero, we take its log and convert the model to a polynomial, as log fij = log a + α log Gi + β log Gj − γ log dij. To i,j p(i, j) − yij2, where fit the parameters, the solution minimizes the squared error E =(cid:80) p(i, j) = α log Gi + β log Gj − γ log dij + log a and yij = log fij. 5.1.2. Clustering migration trajectories Considering the diversity of human behaviour, we categorize workforce mobility into different kinds of movements by the clustering algorithm of K-means [2, 12, 27]. In workforce trajectory set D = {p1, p2, . . . , pm}, the trajectory between city i and city j is defined as 18 a 6-tuple pk = (i, j, Gi, Gj, Tij, Fij), where Gi and Gj denote the GDP of city i and city j, respectively; Tij is the travel time in driving route plans between two locations, and Fij represents the population flow between city i and j. The feature vector of trajectory pk is defined as vk = ( Gij, Tij, Fij), where Gij is the normalization of the product of Gi and Gj, Gij = Gi·Gj−min(G·G) max(F )−min(F ). The labour trajectory set can be divided into n clusters C = {C1, C2, . . . , Cn} by minimizing E = dist(v, mk)2, where Ck is the kth cluster, mk denotes the center of cluster Ck. In max(G·G)−min(G·G); Tij = Tij−min(T ) (cid:80) max(T )−min(T ) and Fij = Fij−min(F ) (cid:80) v∈Ck k=1 (cid:80)n and the optimization function is converted to maximize E = (cid:80)n detail, mk = 1Ck between two workforce trajectories, dist(v, mk) = v·mk v. In this paper, we employ cosine similarity to measure the distance ij + F 2 ij, sim(v, u). (cid:113) (cid:113)(cid:80) (cid:107)v(cid:107)(cid:107)mk(cid:107), where v = ij + T 2 v∈Ck G2 k=1 v,u∈Ck The CLUTO (http://glaros.dtc.umn.edu/gkhome/views/cluto) package is employed in our study to conduct the clustering. Competing interests We have no competing interests. Author contributions H.X. and W.J. conceived the research. H.X. and Z.J. conducted the experiments and anal- ysed the results. H.X. and Z.J. wrote the manuscript. All authors reviewed the manuscript. Acknowledgments Dr.Junjie Wu was supported by the National Natural Science Foundation of China (NSFC) (71531001, 71725002, U1636210, 71471009, 71490723) and the Fundamental Re- search Funds for Central Universities. Z.J. thanks the support from NSFC(71501005). 19 [1] Guy J Abel and Nikola Sander. Quantifying global international migration flows. Science, 343(6178):1520–1522, 2014. [2] Michael R. Anderberg. Cluster analysis for applications. Probability & Mathematical Statistics New York Academic Press, pages 347–353, 1973. [3] Daniele Barchiesi, Tobias Preis, Steven R Bishop, and Helen Susannah Moat. Modelling human mobility patterns using photographic data shared online. Royal Society Open Science, 2(8):150046–150046, 2015. [4] Joshua Evan Blumenstock. Inferring patterns of internal migration from mobile phone call records: evidence from Rwanda. Information Technology for Development, 18(2):107–125, 2012. [5] William J Carrington, Enrica Detragiache, and Tara Vishwanath. Migration with endogenous moving costs. The American Economic Review, 86(4):909–930, 1996. [6] Heide Castaneda. Migration is part of the human experience but is far from natural. NATURE HUMAN BEHAVIOUR, 1:0147 EP –, 2017. [7] Jorg Decressin and Antonio Fatas. Regional labor market dynamics in Europe. European Economic Review, 39(9):1627–1655, 1995. [8] Giorgio Fagiolo and Marina Mastrorillo. International migration network: topology and mod- eling. Physical Review E, 88(1):012812, 2013. [9] C Cindy Fan. Interprovincial migration, population redistribution, and regional development in China: 1990 and 2000 census comparisons. The Professional Geographer, 57(2):295–311, 2005. [10] A. Ghosh, V. Berg, K. Bhattacharya, D. Monsivais, J. Kertesz, K. Kaski, and A. Rotkirch. Migration patterns across the life course of families: Gender differences and proximity with parents and siblings in Finland. arXiv preprint arXiv:1708.02432, August 2017. [11] Bartosz Hawelka, Izabela Sitko, Euro Beinat, Stanislav Sobolevsky, Pavlos Kazakopoulos, and Carlo Ratti. Geo-located twitter as the proxy for global mobility patterns. Cartography and Geographic Information Science, 41(3):260–271, 2014. [12] Anil K Jain and Richard C Dubes. Algorithms for clustering data. Technometrics, 32(2):227– 229, 1988. 20 [13] David Karemera, Victor Iwuagwu Oguledo, and Bobby Davis. A gravity model analysis of international migration to North America. Applied Economics, 32(13):1745–1755, 2000. [14] M Khasin, E Khain, and L. M. Sander. Fast migration and emergent population dynamics. Physical Review Letters, 109(24):248102, 2012. [15] Gautier Krings, Francesco Calabrese, Carlo Ratti, and Vincent D Blondel. Urban gravity: a model for inter-city telecommunication flows. Journal of Statistical Mechanics: Theory and Experiment, 2009(07), 2009. [16] Caterina La Porta, Oleksandr Chepizhko, Costanza Giampietro, Eleonora Mastrapasqua, Mehdi Nourazar, Miriam Ascagni, Michela Sugni, Umberto Fascio, Livio Leggio, and Chiara Malinverno. Bursts of activity in collective cell migration. Proceedings of the National Academy of Sciences of the United States of America, 113(41):11408, 2016. [17] Renaud Lambiotte, Vincent D Blondel, Cristobald De Kerchove, Etienne Huens, Christophe Prieur, Zbigniew Smoreda, and Paul Van Dooren. Geographical dispersal of mobile commu- nication networks. Physica A-statistical Mechanics and Its Applications, 387(21):5317–5325, 2008. [18] Everett S Lee. A theory of migration. Demography, 3(1):47–57, 1966. [19] Maxime Lenormand, Sylvie Huet, Floriana Gargiulo, and Guillaume Deffuant. A universal model of commuting networks. Plos One, 7(10):e45985, 2012. [20] Moshe Levy. Scale-free human migration and the geography of social networks. Physica A-statistical Mechanics and Its Applications, 389(21):4913–4917, 2010. [21] Jiwei Li, Qingqing Ye, Xuankai Deng, Yaolin Liu, and Yanfang Liu. Spatial-temporal anal- ysis on Spring Festival travel rush in China based on multisource big data. Sustainability, 8(11):1184, 2016. [22] Xiaomeng Li, Hongzhong Xu, Jiawei Chen, Qinghua Chen, Jiang Zhang, and Zengru Di. Characterizing the international migration barriers with a probabilistic multilateral migration model. Scientific Reports, 6(1):32522, 2016. [23] Xiao Liang, Jichang Zhao, Lixia Dong, and Ke Xu. Unraveling the origin of exponential law in intra-urban human mobility. Scientific Reports, 3(1):2983–2983, 2013. [24] Xiao Liang, Jichang Zhao, and Ke Xu. A general law of human mobility. Science in China Series F: Information Sciences, 58(10):1–14, 2015. 21 [25] Yu Liu, Xi Liu, Song Gao, Li Gong, Chaogui Kang, Ye Zhi, Guanghua Chi, and Li Shi. Social sensing: A new approach to understanding our socioeconomic environments. Annals of the Association of American Geographers, 105(3):512–530, 2015. [26] Yu Liu, Zhengwei Sui, Chaogui Kang, and Yong Gao. Uncovering patterns of inter-urban trip and spatial interaction from social media check-in data. PLOS ONE, 9(1), 2014. [27] J Macqueen. Some methods for classification and analysis of multivariate observations. In Proc. of Berkeley Symposium on Mathematical Statistics and Probability, pages 281–297, 1967. [28] Anastasios Noulas, Salvatore Scellato, Renaud Lambiotte, Massimiliano Pontil, and Cecilia Mascolo. A tale of many cities: universal patterns in human urban mobility. PLOS ONE, 7(5), 2012. [29] Ernest George Ravenstein. The laws of migration. Journal of the statistical society of London, 48(2):167–235, 1885. [30] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20(1):53–65, 1987. [31] Hoon Lee Sang, Robyn Ffrancon, Daniel M. Abrams, Beom Jun Kim, and Mason A. Porter. Matchmaker, matchmaker, make me a match: Migration of populations via marriages in the past. Physical Review X, 4(4), 2013. [32] Soheil Shayegh. Outward migration may alter population dynamics and income inequality. Nature Clim. Change, advance online publication, 2017. [33] Jianfa Shen. Changing patterns and determinants of interprovincial migration in China 1985- 2000. Population Space and Place, 18(3):384–402, 2012. [34] Filippo Simini, Marta C Gonzalez, Amos Maritan, and Albertlaszlo Barab´asi. A universal model for mobility and migration patterns. Nature, 484(7392):96–100, 2012. [35] O Stark. The migration of labor. Cambridge Massachusetts/oxford England Basil Blackwell, 26(4), 1991. [36] Edward J Taylor. The new economics of labour migration and the role of remittances in the migration process. International Migration, 37(1):63–88, 1999. [37] J Edward Taylor, Scott Rozelle, and Alan De Brauw. Migration and incomes in source communities: A new economics of migration perspective from China. Economic Development and Cultural Change, 52(1):75–101, 2003. 22 [38] Christian Thiemann, Fabian J Theis, Daniel Grady, Rafael Brune, and Dirk Brockmann. The structure of borders in a small world. PLOS ONE, 5(11), 2010. [39] Robert L Thorndike. Who belongs in the family. Psychometrika, 18(4):267–276, 1953. [40] Lei Wang, Qian Zhang, Yun Cai, Jianlin Zhang, and Qingguo Ma. Simulation study of pedestrian flow in a station hall during the Spring Festival travel rush. Physica A-statistical Mechanics and Its Applications, 392(10):2470–2478, 2013. [41] Lizhe Wang and Lajiao Chen. The impact of new transportation modes on population distri- bution in jing-jin-ji region of china. Scientific Data, 5:170204, 2018. [42] Xianwen Wang, Chen Liu, Wenli Mao, Zhigang Hu, and Li Gu. Tracing the largest seasonal migration on earth. arXiv preprint arXiv:1411.0983, 2014. [43] Jun Xu, Aoyong Li, Dong Li, Yu Liu, Yunyan Du, Tao Pei, Ting Ma, and Chenghu Zhou. Difference of urban development in China from the perspective of passenger transport around Spring Festival. Applied Geography, 87:85–96, 2017. [44] Xiaoyong Yan, Wenxu Wang, Ziyou Gao, and Yingcheng Lai. Universal model of individual and population mobility on diverse spatial scales. Nature Communications, 8(1):1639, 2017. [45] Xiaoyong Yan, Chen Zhao, Ying Fan, Zengru Di, and Wenxu Wang. Universal predictability of mobility patterns in cities. Journal of the Royal Society Interface, 11(100):20140834–20140834, 2014. [46] Emilio Zagheni, Venkata Rama Kiran Garimella, Ingmar Weber, et al. Inferring international and internal migration patterns from twitter data. In Proceedings of the 23rd International Conference on World Wide Web, pages 439–444. ACM, 2014. [47] Jichang Zhao, Ruiwen Li, Xiao Liang, and Ke Xu. Second ring is the most jammed road in Beijing: A view from taxis and Weibo. In International Conference on Service Systems and Service Management, pages 1–6, 2016. [48] George Kingsley Zipf. The P1P2/D hypothesis: On the intercity movement of persons. Amer- ican Sociological Review, 11(6):677–686, 1946. 23 Distance metric Geographical distance Travel distance Travel time - γ β α Model GM 0.3137∗∗∗ aveG-GM 0.7574∗∗∗ G-GM 0.2872∗∗∗ dirG-GM 0.3066∗∗∗ 0.9278∗∗∗ 0.9351∗∗∗ - - - - - - - GM 0.3910∗∗∗ aveG-GM 0.8608∗∗∗ G-GM 0.3409∗∗∗ dirG-GM 0.3373∗∗∗ 0.9018∗∗∗ 0.9088∗∗∗ - - - - - - GM 0.3901∗∗∗ aveG-GM 0.7307∗∗∗ G-GM 0.3165∗∗∗ dirG-GM 0.3628∗∗∗ 0.9167∗∗∗ 0.9239∗∗∗ - - - - Prob(F-statistic) ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ TABLE 1: The fitting results for different models. In this table, γ is the exponent for the distance measurements, and α and β are exponents for the GDP of origins and destinations, respectively. Additionally, ∗∗ stands for p < 0.01, and ∗∗∗ represents p < 0.001, all parameters in the models have been examined by significance tests. Trajectory set γ Prob(F-statistic) All Group I Group II Group III Group IV 0.3165∗∗∗ 0.1910∗∗ 0.3868∗∗∗ 0.4186∗∗∗ 0.1961∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ TABLE 2: γ varies in different migration patterns and all trajectories. ∗∗ stands for p < 0.01, and ∗∗∗ represents p < 0.001. 24 FIG. 1: Relations between GDP and per capita disposable income, R&D, and talent distribution. There is a high correlation between GDP and the per capita disposable income in different provinces. The investment of funds in research and development (R&D) is positively correlated with GDP in provinces. Meanwhile, the ratios of practitioners in the high-technology industries, such as the information technology industry, financial industry, real estate industry and scientific research and technical services industry, are all significantly positively related to GDP. This finding shows that GDP, as an indicator, has information on potential income, technical workforce levels and labour market requirements. The city with high GDP would have more job opportunities with higher yields, a higher-skilled workforce and high labour requirements. The outliers are marked by orange circles; in detail, Beijing and Shanghai are unusual points in the first subgraphs, and the outliers of the other subgraph denote Beijing. 25 0123456Per Capita Disposable Income×104r=0.630.00.51.01.52.02.5R&D×103r=0.890.000.050.100.150.20Ratio of PractitionerInformation Technology Industryr=0.88Financial Industryr=0.9101234567GDP×10120.000.050.100.150.20Ratio of PractitionerReal Estate Industryr=0.8501234567GDP×1012Scientific Research andTechnical Services Industryr=0.84FIG. 2: Modelling flux predictions for four models, with the distance metric being travel time. The grey points are a scatter plot of fitness between the actual labour migration flows and forecast flows between each pair of cities. The red lines of y = x indicate the best prediction, and the red points represent the means of the predicted migration flows. Additionally, the whiskers of the box plots show the 5th and 95th percentiles, respectively. 26 FIG. 3: Comparison of four models based on SSI. The bars with different colors show the predictive capability of models with different distance metrics. Higher SSI reflects better accuracy in flux prediction. 27 GMaveG-GMG-GMdirG-GM0.250.300.350.400.450.500.55SSIGeographical DistanceTravel DistanceTravel TimeFIG. 4: Migration flows of grouped trajectories. The bright lines denote the migration flows between cities. (a) Pattern I, (b) Pattern II, (c) Pattern III and (d) Pattern IV. A brighter area denotes that there are more city pairs with migration flows. The average migration flows between city pairs is 18522, 3182, 1045 and 195, respectively. Each group accounts for 29%, 28%, 20% and 23% of the total flux, respectively. The representative cities in each pattern are marked by red triangles and are obtained by filtering the flux ratios of more than 50% in Patterns I & II & IV(40% in Pattern III), sorting each city's migration flux and selected cities. 28 FIG. 5: Evaluation of cities' development models. The ratio flux in Patterns I & II is defined as the index DI to evaluate the development models of cities, , where E(1) is the set of city pairs in Pattern I, E(2) is the set (cid:80) (cid:80) DI(i) = (i,j)∈E(1)∪E(2) Fij (i,j)∈E Fij of city pairs in Pattern II, E is the set of city pairs and fij is the amount of workforce migration flux between city i and j. A deeper red color indicates the city with a better development model. The value of DI ranges from 0 to 1. 29 FIG. 6: Evaluation of provinces' development models. We define the average flux ratio in each pattern of cities in a province to evaluate the development model of the (cid:80) (cid:80) (cid:80) (cid:80) province. To be specific, p(k, P ) = number, P is the province and i ∈ P means that city i is under the jurisdiction of province P . T represents all city pairs, and T (k) is the set of city pairs belonging to Pattern k. , in which k stands for the pattern i∈P i∈P (i,j)∈T (k) Fij (i,j)∈T Fij 30 0.00.20.40.60.81.0Ratio of FluxShanghaiBeijingTianjinChongqingJiangsuShandongHebeiZhejiangHenanFujianGuangxiHunanGuangdongJiangxiAnhuiHubeiShaanxiShanxiLiaoningSichuanInner MongoliaGuizhouJilinHeilongjiangGansuQinghaiNingxiaHainanYunnanXinjiangTibetPattern IPattern IIPattern IIIPattern IVFIG. 7: Locality of cities. The cumulative ratio of the migration flux of the neighbouring cities from the near to the distant, defined as c(i, r), is computed to show the (cid:80) (cid:80) (i,j)∈Ti(r) Fij (i,j)∈Ti Fij , where Ti(r) is locality characteristics of the target city. Specifically, c(i, r) = the set of city pairs between city i and top r cities, sorted by travel time. Ti is the set of city pairs between city i and other cities. The color of lines stands for the GDP of city i, and a deeper red color means better economic status. 31 FIG. 8: Extent of GDP matching. To understanding the extent of GDP matching between target cities and their neighbouring cities, the average ratio of GDP is defined as (cid:80) l(r, i) = (i,j)∈Ti(r) Ej rEi , where Ti(r) represents the set of pairs between city i and top r cities, sorted by travel time, and Ei is the GDP of city i. Because Guangzhou has a higher GDP than Shenzhen, A is an abnormal point in the line of Shenzhen. B is the highest point in the line of Beijing, and the inflection point results from Tianjin, which is the only city with a high GDP that is similar to that of Beijing. 32 100101102Rank0.00.20.40.60.81.0Average Ratio of GDPABShanghaiShenzhenChongqingBeijingGuangzhouFIG. 9: R&D and the development models of cities. The ratio of flux in Patterns III & IV is negatively correlated with the ratio of the investment of funds in R&D to provinces' GDP. 33 0.00.20.40.60.81.01.2Ratio of Flux in Pattern III and IV0.020.000.020.040.060.08Ratio of R&D to GDPr=-0.83FIG. S1: Relation between migration flux and the number of trains between city pairs. The migration flux is positively correlated with the number of trains that pass through city pairs, and the coefficient of correlation is approximately 0.37. 34 100101102103104105106Migration Flux100101102103Number of Trainsr=0.37FIG. S2: Modelling migration flows with the distance metric of geographical distance. The comparison between actual and predicted labour migration flows in four models. 35 FIG. S3: Modelling migration flows with the distance metric of travel distance. The comparison between actual and predicted migration flows in four models. 36 FIG. S4: Comparison of different cluster numbers based on the silhouette coefficient. The silhouette coefficient for each sample i is calculated as s(i) = b(i)−a(i) where a(i) reflects the average dissimilarity between i and all other samples within the max(a(i),b(i)), same cluster and b(i) is the lowest average dissimilarity between i and any other cluster. In this paper, the cosine distance is defined as 1 minus the cosine similarity. Additionally, the average s(i) over all samples is a metric to measure how appropriately the data have been clustered, with a value ranging from -1 to 1. A higher value denotes a better clustering result. It can be seen that the best effect (silhouette coefficient is 0.93) is achieved when the number of clusters is 3. 37 12345678910Number of Clusters0.700.750.800.850.900.95Silhouette CoefficientFIG. S5: Comparison of different cluster numbers based on the elbow method. The elbow method is used to find the number of clusters when an obvious inflection point is obtained in the cost function. In this paper, the cost function maximizes the cosine similarity of samples within the same cluster. As can be seen, the best number of clusters is chosen at this point, where the number of clusters is set to 5. 38 12345678910Number of Clusters6.106.116.126.136.146.156.166.17Cost Function×104FIG. S6: Migration flows of grouped trajectories with three clusters. (a) Pattern I; (b) Pattern II; and (c) Pattern III. Compared to the distinct patterns of four clusters, the partition with three clusters is not abundant enough. Although Pattern I and Pattern II can correspond to the first two patterns of the four-cluster case, Pattern III gives us an intangible result, including the brighter area in the midlands and the transition relationship between distant city pairs that are divided into two patterns in the partition of four clusters. Therefore, it is better to divide trajectories into four clusters because it offers more information. 39 FIG. S7: Migration flows of grouped trajectories with five clusters. (a) Pattern I; (b) Pattern II; (c) Pattern III; (d) Pattern IV; (e) Pattern V. There is no obvious distinction among Patterns II and III. Compared with this partition result, the patterns of the four clusters are obviously more different from each other. Accordingly, the partition with four clusters is a better choice than that with five clusters. 40 FIG. S8: GDP product distributions in grouped trajectories. (a) Pattern I; (b) Pattern II; (3) Pattern III; (4) Pattern IV. The highest probability of migration occurs between two cities with excellent economic levels in the second cluster, and the first subgroup and the third subgroup follow, with the second- and third-highest probabilities. In comparison, almost all city pairs in the fourth group have low levels of economic development. 41 10-1210-1110-1010-910-810-710-6Probability Density(a)(b)012345Product of GDP×10810-1210-1110-1010-910-810-710-6Probability Density(c)012345Product of GDP×108(d)FIG. S9: Distributions of travel time between cities in subgroups. (a) Pattern I; (b) Pattern II; (c) Pattern III; (d) Pattern IV. It is apparent that the city pairs in the first group have the shortest travel time. Additionally, there is an increase from the first group to the fourth group, where travel time reaches the maximum. 42 10-910-810-710-610-510-4Probability Density(a)(b)0123456789Travel Time×10510-910-810-710-610-510-4Probability Density(c)0123456789Travel Time×105(d)FIG. S10: Distributions of migration flux between cities in grouped results. (a) Pattern I; (b) Pattern II; (c) Pattern III; (d) Pattern IV. There is a dramatic decline in the average migration flows between city pairs from the first group to the fourth group (18522, 3182, 1045 and 195 in detail). Additionally, the proportion of total flux in each group is 29%, 28%, 20% and 23% separately. Patterns I, II and III indeed explain the extremely weighty types of workforce migration behaviours. The scarce flow in each city pair and a small percentage of total flux in Pattern IV imply that the migration in this cluster can be regarded as noisy movements to some extent. 43 10-910-810-710-610-510-410-3Probability Density(a)(b)0.00.51.01.52.0Migration Flux×10510-910-810-710-610-510-410-3Probability Density(c)0.00.51.01.52.0Migration Flux×105(d)FIG. S11: SSI of G-GM in grouped trajectories. The dashed line represents the SSI for all the trajectories. 44 Group IGroup IIGroup IIIGroup IVAll0.400.450.500.550.600.650.70SSIFIG. S12: Modelling grouped trajectories. The comparison of the prediction capability of the grouped trajectories based on G-GM. (a) Pattern I; (b) Pattern II; (c) Pattern III; (d) Pattern IV. 45
1902.00171	1	1902	2019-02-01T04:15:44	Social Network Based Substance Abuse Prevention via Network Modification (A Preliminary Study)	[ "cs.SI" ]	Substance use and abuse is a significant public health problem in the United States. Group-based intervention programs offer a promising means of preventing and reducing substance abuse. While effective, unfortunately, inappropriate intervention groups can result in an increase in deviant behaviors among participants, a process known as deviancy training. This paper investigates the problem of optimizing the social influence related to the deviant behavior via careful construction of the intervention groups. We propose a Mixed Integer Optimization formulation that decides on the intervention groups, captures the impact of the groups on the structure of the social network, and models the impact of these changes on behavior propagation. In addition, we propose a scalable hybrid meta-heuristic algorithm that combines Mixed Integer Programming and Large Neighborhood Search to find near-optimal network partitions. Our algorithm is packaged in the form of GUIDE, an AI-based decision aid that recommends intervention groups. Being the first quantitative decision aid of this kind, GUIDE is able to assist practitioners, in particular social workers, in three key areas: (a) GUIDE proposes near-optimal solutions that are shown, via extensive simulations, to significantly improve over the traditional qualitative practices for forming intervention groups; (b) GUIDE is able to identify circumstances when an intervention will lead to deviancy training, thus saving time, money, and effort; (c) GUIDE can evaluate current strategies of group formation and discard strategies that will lead to deviancy training. In developing GUIDE, we are primarily interested in substance use interventions among homeless youth as a high risk and vulnerable population. GUIDE is developed in collaboration with Urban Peak, a homeless-youth serving organization in Denver, CO, and is under preparation for deployment.	cs.SI	cs	Social Network Based Substance Abuse Prevention via Network Modification (A Preliminary Study) Aida Rahmattalabi University of Southern California [email protected] Milind Tambe University of Southern California [email protected] Anamika Barman Adhikari University of Denver [email protected] Eric Rice University of Southern California [email protected]@usc.edu Phebe Vayanos University of Southern California [email protected] Robin Baker Urban Peak Organization [email protected] 9 1 0 2 b e F 1 ] I S . s c [ 1 v 1 7 1 0 0 . 2 0 9 1 : v i X r a ABSTRACT Substance use and abuse is a significant public health problem in the United States. Group-based intervention programs offer a promising means of preventing and reducing substance abuse. While effective, unfortunately, inappropriate intervention groups can result in an increase in deviant behaviors among participants, a process known as deviancy training. This paper investigates the problem of optimiz- ing the social influence related to the deviant behavior via careful construction of the intervention groups. We propose a Mixed In- teger Optimization formulation that decides on the intervention groups to be formed, captures the impact of the intervention groups on the structure of the social network, and models the impact of these changes on behavior propagation. In addition, we propose a scalable hybrid meta-heuristic algorithm that combines Mixed Integer Programming and Large Neighborhood Search to find near- optimal network partitions. Our algorithm is packaged in the form of GUIDE, an AI-based decision aid that recommends interven- tion groups. Being the first quantitative decision aid of this kind, GUIDE is able to assist practitioners, in particular social workers, in three key areas: (a) GUIDE proposes near-optimal solutions that are shown, via extensive simulations, to significantly improve over the traditional qualitative practices for forming intervention groups; (b) GUIDE is able to identify circumstances when an intervention will lead to deviancy training, thus saving time, money, and effort; (c) GUIDE can evaluate current strategies of group formation and discard strategies that will lead to deviancy training. In developing GUIDE, we are primarily interested in substance use interventions among homeless youth as a high risk and vulnerable population. GUIDE is developed in collaboration with Urban Peak, a homeless- youth serving organization in Denver, CO, and is under preparation for deployment. KEYWORDS Social Networks; Substance Abuse Intervention, Optimization 1 INTRODUCTION Peers have a direct influence in adolescents' risk behaviours Sub- stance use and abuse is a significant public health problem in the Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 2019, Montreal, Canada © 2019 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. https://doi.org/doi Figure 1: Urban Peak site. We are collaborating with Urban Peak organization for conducting the group-based interven- tions United States, particularly among youth. According to the Monitor- ing the Future study [12], around 54 percent of high school students have tried at least one illicit substance. Homeless youth, in particu- lar, are shown to be disproportionately affected, with substantially higher levels of substance use compared to the housed youth [14]. Notably, interventions attempting to reduce substance use have successfully utilized social networks to disseminate and reinforce behavioral norms supportive of protective behaviors related to substance use (e.g., [19]). The way the social network is utilized is often through formation of subgroups where the individuals can talk, share experiences and engage in various constructive activities. Such social network-based approaches to substance abuse prevention are considered more promising because young people learn better from one another, find each other more credible and have a shared environment and culture [6]. Unfortunately, these social network-based efforts may inadver- tently increase the chances of youth being exposed to negative social influences, as they do not explicitly structure the interven- tion groups. This can result in an effect known as deviancy training. Deviancy training occurs when high-risk youth are aggregated together and reinforce negative behaviors and attitudes. Social network-based interventions have typically grouped par- ticipants into intervention groups in three ways [19]: (1) random assignment: participants are randomly assigned to the groups; (2) network based assignment: participants are assigned based on their own nominations; (3) teacher nominated assignment: groups are created based on teacher nominations, often corresponding to an assignment based on the participants' behaviors, i.e., even distribu- tion of high-risk individuals. The random assignment and teacher nominated methods are less effective because they are not designed to leverage naturally occurring ties in participants' social networks. The network-based method on the other hand does leverage these natural forms of influence. However, it is shown that it can have unintended effect of exacerbating the problem behavior [19]. AAMAS'19, May 2019, Montreal, Canada Aida Rahmattalabi, Anamika Barman Adhikari, Phebe Vayanos, Milind Tambe, Eric Rice, and Robin Baker Recently, successful applications of AI based decision aids such as [21] and [5] have encouraged efforts to address such complicated social problems using techniques in AI and optimization [15]. In the present work, we aim to tackle the problem of deviancy training in substance abuse intervention by structuring more effective groups, one that effectively partitions the participants' social network. Building this decision aid, however, raises major challenges. First, it has been observed that in the process of being a part of the network-based intervention, the sub-networks (peer-groups) un- dergo a transformation, whereby network ties change in strength [17]. Such changes will consequently change the influence spread pat- terns and therefore constitute an important component of the model. Second challenge is understanding the influence process. In this regard, we propose an influence spread model which is based on the competitive Linear Threshold (LT) Model in [3] to explain both pos- itive and negative social influences within interventions. The third major challenge is to strategically choose the right peer-groups from the larger social network to mitigate the challenge of deviancy training. In the following, we will refer to those engaged in sub- stance use as "users", that we wish to have socially influenced by "non-users" peers in the smaller groups, so as to reduce/prevent such substance use and abuse. We view this problem as an influence maximization problem, with the goal of maximizing/minimizing the total positive/negative influence by careful construction of the intervention groups, consid- ering the consequent changes in the network topology. Influence maximization in networks has been widely studied in the past decade and several works have tried to address the problem of mod- ifying the network topology in order to maximize positive influence spread or curtail undesirable behaviors. In particular, [9, 10] study the problem of edge deletion/addition under the LT model. Our problem however is different: we do not directly decide on edges to add/delete. Instead, we can only select how participants are parti- tioned into intervention groups. Depending on this choice, there is a disciplined process that controls how social network changes. More precisely, we will cluster the given social network into interven- tion groups, which causes further changes in the network topology as a byproduct, and subsequently controls how the positive and negative influences propagate. Social influence optimization and network dynamics distinguish this work from the existing literature on network clustering. For ex- ample, [2] studies clustering graphs with both positive and negative edges, but the structure of the network is static. Further, unlike [7] that considers the network dynamic, we are interested in optimizing a stochastic influence function, plus, the evolution of the network topology is not just a function of time but a function of the deci- sions variables or clusters. To best of our knowledge no work has addressed such influence-based clustering of networks to this date. To address these challenges, we propose an AI-based decision aid, called GUIDE (GroUp-based Intervention DEcision aid). GUIDE assists interventionists in substance abuse prevention by giving rec- ommendations regarding the intervention groups. In what follows we detail our main contributions: (I) First, we introduce a novel problem to AI researchers; We propose a mathematical model for this type of intervention, which includes the key aspects of the influence spread and network dynamics. The model enables us to predict, both the expected success of the intervention, measured as the expected number of "non-users" at the end of the interven- tion, and the possibility of harm, or deviancy training. (II) We show that the problem of finding the optimal partition to minimize the expected substance use is NP-hard. (III) Therefore, we propose both a Mixed Integer Linear Program (MILP) to solve this prob- lem and a hybrid meta-heuristic algorithm that combines MIP and Large Neighborhood Search (LNS) to find near-optimal network partitions. (IV) We provide extensive analysis of the model, using both real-world social network data and synthetic graphs and we show that our proposed partitions can significantly outperform all common practices such as random assignment of individuals or participants choice. 2 PROBLEM STATEMENT Our goal is to create an assistant that will aid an interventionist for substance abuse prevention. The intervention that we consider is a 6-week program, in which participants are placed into different groups. The size of each group is bounded between 4 and 7. An interventionist interacts with each group. Pre-intervention, the par- ticipants are asked to report both their hard-drug using behaviors, and their network relationships (i.e., who they know) and what the strength of that relationship is (whether a person is a strong or a weak tie). Over the course of the intervention and based on the groups assignments, these social ties will change: some may become weaker, while others may grow stronger. We aim to maxi- mize the number of "non-users" one-month after the intervention is complete. We view the problem of group configuration for substance abuse prevention as a graph partitioning problem. Given a social network, the goal is to find a partition such that the expected number of "non-users," at the end of the intervention, is maximized. More formally, let graph G = (V, E) be a directed graph representing the given social network, with V as the set of all nodes (individuals in the social network), and E ⊂ V × V as the set of all edges (social ties). Arc (i, j) ∈ E indicates existence of an arc pointing from i to j, signifying that j has reported i as a friend pre-intervention. Associated with each node i ∈ V is a node behavior indicating that node i is a substance "user" or a "non-user." The scalar bi ∈ {0, 1} encodes the node behavior with bi = 1 if and only if node i is a "user." Additionally, associated with each edge e = (i, j) ∈ E is the edge strength se ∈ {0, 1}, where se = 1 (resp. 0), indicates that j considers i as strongly (resp. weakly) influential for him. We model the intervention as a partition of V into S subsets Ps, s = 1, . . . , S, such that ∩S s =1Ps = V, where each subset must consist of at least C and at most C nodes. Note that the number of subsets S is a decision variable, but C and C are pre-specified. Dur- ing the intervention and based on the group configurations, the ties undergo changes. Some ties are cut or weakened and some will be re- inforced. We assume that, given a choice of partition P := {Ps}S s =1, the post-intervention structure of the network is known determin- istically. We let G+(P) = (V, E+(P)) denote the post-intervention graph when groups are formed according to partition P, with E+(P) corresponding to the new edge set. Thus, (i, j) ∈ E+(P) if and only if j considers i as a friend post-intervention. Accordingly, for e ∈ E+(P), we denote by s e (P) ∈ {0, 1} the strength of edge e post-intervention. s =1Ps = ∅, ∪S + Social Network Based Substance Abuse Prevention via Network Modification (A Preliminary Study)AAMAS'19, May 2019, Montreal, Canada Same Group (user, user) (non-user, non-user) (non-user, user) (user, non-user) Separate Groups (user, user) (non-user, non-user) (non-user, user) (user, non-user) no-tie weak strong strong strong strong weak weak weak weak no-tie weak none none weak none none none none none strong strong strong strong strong strong strong strong weak weak Table 1: Changes in tie strength post-intervention. The ex- isting relationships, and the behavior of the individuals as well as their assignment to groups impacts the changes. The behavior of node i ∈ V post-intervention is random. We let Bi(P) denote the random variable that represents the behavior of node i post-intervention, i.e., Bi(P) = 1 if and only if node i is a "user." This is a complicated stochastic function of the partition that depends on both the link formation and breakage rules and the influence model assumed. Figure 2 shows an example network pre- and post-intervention. Mathematical Formulation. Mathematically, the problem of se- lecting the optimal partition P that maximizes the expected number of "non-users" in the network post-intervention can be formulated as: Where E[·] denotes the expectation operator with respect to the distribution of {Bi}i∈V. The objective counts the expected number of "non-users" in G+(P). Note that we are considering the inter- vention period as one time step and therefore, we are assuming a single-stage influence process. The first three constraints ensure that P forms a partition of the node set. The last constraint enforces capacity constraints on each group. Problem (1) is a combinatorial optimization problem which is hard to solve as formalized by Theorem 2.2. In order to solve this problem, both the random behavior B, as well as the network struc- tural changes must be modeled. In the following two sections, we propose models for network and influence dynamics that are sup- ported by the social work literature. Tie Formation and Breakage. While these youth have pre- existing strong or weak ties, especially centered around their sub- stance use, the strength of these relationships would weaken or strengthen during the intervention [4]. There is empirical evidence maximize subject to (cid:35) (1 − Bi(P)) (cid:34) E i∈V S ∈ N+ S Ps ⊂ V ∀s ∈ {1, . . . , S} S s =1 C ≤ Ps ≤ C ∀s ∈ {1, . . . , S}, Ps = ∅ Ps = V s =1 (1) Figure 2: An example network pre- (left) and post- (right) intervention. The black (resp. white) circles indicate "user" (resp. "non-user") nodes. Weak (resp. strong) links are de- noted by thin (resp. thick) arrows. The ellipsoids represent the two groups that are formed for the intervention. As seen, new edges are created within the groups, while some edges are cut across the groups. partitions to suggest that the more similar two individuals are, the stronger their ties are [1]. This is explained by their behavioral homophily. Homophily refers to the tendency of people to associate with peo- ple who are like themselves and these ties also often facilitate more communication and influence [13]. Therefore, if two individuals are both "users" or both "non-users," we assume that they would develop a strong tie over time. Based on this rationale, we also model the interventionist who will exert a positive influence as an additional "non-user" who develops a weak tie to the "users" and a strong tie to the "non-users," during the intervention period. An important component of the proposed intervention is to de- velop skills that would facilitate bonding with pro-social peers and discourage interactions with youth who have high-risk behaviors [18]. Therefore, if two individuals are separated and at least one of them has "user" behavior, their tie weakens or is cut. Table 1 summarizes how we expect strength of ties to change in response to graph partitioning. In this table, the rows indicate the behavior of the pair of nodes under consideration, and the columns show their pre-intervention tie. The changes are defined as a function of the ties pre-intervention (E), the behavior of the end nodes (bi), and whether they are assigned to the same group or not. Example 2.1. Consider the example network in Figure 2. In this network, black circles indicate "user" individuals and white circles represent "non-user" people. Initially, individuals are connected with either weak or strong ties, where weak ties are shown via light arrows, and strong ties are depicted via bold arrows. Given this social network as input, we need to group the indi- viduals into two intervention groups. Based on this decision, new connections will form. Consider the grouping as depicted in Fig- ure 2. Post-intervention, the ties undergo changes according to table 1. Therefore, new ties will be formed between individuals that did not have any ties before intervention and are placed in the same group. Also, some of the weak existing ties become stronger before the intervention. Also, some of the weak ties will become stronger as a result of the grouping. Finally, the existing ties that are across groups will be broken or weakened. AAMAS'19, May 2019, Montreal, Canada Aida Rahmattalabi, Anamika Barman Adhikari, Phebe Vayanos, Milind Tambe, Eric Rice, and Robin Baker Substance Abuse Prevention Influence Spread Model. Fol- lowing the intervention, and depending on how the network evolves, we evaluate the influence to predict the changes in the nodes' be- haviors. We use a variant of the popular Linear Threshold model [8] proposed in [3], in which they study the competitive influence processes in social networks. Our problem is similar to [3] as there are two concurrent influences, a positive influence which originates from the "non-users" and a negative influence from "users." How- ever, it differs from their model in that the nodes in our network have already adopted one of these two behaviors. Consider the post-intervention graph as G+(P) = (V, E+(P)). For notational convenience, we henceforth eliminate the dependence of G+ and E+ on P. For each (i, j) ∈ V × V, a weight value wij ∈ R is assigned, which characterizes the amount of influence from node i to node j. These weights are such that the sum of the weights of the incoming edges to a certain node is bounded. Specifically, 0 ≤ i∈V wij ≤ 1, ∀j ∈ V, Mathematically, if for example, bj = 1 and Initially, each node i ∈ V is assigned a threshold value, denoted by T i, corresponding to the influence threshold for switching behavior (from "user" to "non-user" and vice-versa). We assume these thresh- olds are uniformly distributed in the range [0, 1]. A necessary con- dition for a node to change his behavior is for the sum of his neigh- bors' weights with opposite behavior to exceed the node's threshold. w(i, j) ≥ T j, node j has the necessary condition to become a "non-user." Following the model from [3], if the threshold of a "user" (resp. "non-user") is exceeded, the node behavior will switch with probability Ωun (resp. Ωnu). Thus, for any node i, we can express their post-intervention behavior distribution as: i:bi =0 P(Bj = 1) = ΩunP(cid:169)(cid:173)(cid:171) P(Bj = 0) = ΩnuP(cid:169)(cid:173)(cid:171) i:bi =0 i:bi =1 wij ≥ T j(cid:170)(cid:174)(cid:172) , if (bj = 0) wij ≥ T j(cid:170)(cid:174)(cid:172) , if (bj = 1). Theorem 2.2. Problem (1) is NP-Hard, even under the presented influence model. Proof. We construct a reduction from Set Cover problem with N items to be covered by M subsets. The reduction can be explained as follows: For any item in the set cover problem, a "non-user" node, and for any set, a "user" node will be created. There is a tie from a "user" to a "non-user", if and only if the item is included in a set. In the drug prevention problem, we aim to partition the network into S sub- graphs. However, several assumptions are made. First, all ties across partitions are cut. Every pair of "user" in the same group form a tie. Every member of the network is connected to a "non-user" (the interventionist). All the ties have the same weight. The threshold values for "non-users" are set to Number of "user" connections Number of "user" connections + 1. "User" thresholds are set equal to 1 2 + ϵ (ϵ > 0 , small positive number.) The capacity of each group is N + M. If the optimal solution is at least (N + S + 1), there is at most one group with more than two "users". This is because, if there are at least N +S +1 "non-user" at the end of the intervention, at least S "users" must have flipped to "non-user", and the way the thresholds are set requires that the "users" are not put together in a group, in more that one group. Also, it is best to separate the "users" from "non-user". Therefore, the optimal S "user" to isolate will be equivalent to the solution of the Set Cover. Also, a solution to the Set Cover results in at least N + S + 1 "non-user". □ 3 MIXED-INTEGER PROGRAMMING FORMULATION s ij ∅ ij ∅ ij w ij w ij , s We present a MILP formulation of Problem (1). First, we define s as indicators for whether the pre-intervention tie from ij , s s i to j is strong, weak, or there is no tie, respectively. The values for these variables are known from the initial data on the network := I((i, j) ∈ := I((i, j) (cid:60) E), s structure and are determined as s E and s(i, j) = 1), and s := I((i, j) ∈ E and s(i, j) = 0), for each (i, j) ∈ V × V, where I(·) is the indicator function. Let us first introduce the decision variables. We define zij as a binary variable which is equal to 1 if nodes i and j are in the same be binary variables to encode the group. Also, let x post-intervention tie strength. Finally, wij is the normalized post- intervention weight of the tie between i and j. Next, we introduce the constraints in this problem. s+ ij , x The groups must satisfy the capacity constraints, that is the sum of all the nodes that are in the same group must meet the following constraint,i.e., w+ ij ∅+ ij , x C ≤ i∈V zij ≤ C ∀j ∈ V. we require that if i and j are in the same group, and so are j and k, i and k must be in the same group as well. This can be expressed by saying that: zij + zjk − 1 ≤ zki ∀i, j, k ∈ V. The post-intervention link strengths can be defined dependent on the pre-intervention edge strengths, the group assignments, and the behavior of the nodes. Therefore, according to Table 1, the link from i to j will become strong if for example, nodes i and j belong to the same group and either have similar behaviors, or they already have a strong tie, which is encoded as part of the following equation ij)(cid:105) w (cid:104) s+ ij x = zij s ∅ ij + bi + bj − 1(s s ij + s +(1 − zij)bi + bj − 1s s ij , where the term bi + bj − 1 is 1 iff i and j have the same behavior. Similarly, (cid:104)bi − bj(s (cid:104) ij)(cid:105) w w+ ij x ∅ ij + s = zij ij(1 − max(bi , bj)) + (ss w s + (1 − zij) ij)(bi − bj)(cid:105) . determines whether the new tie from i to j is weak. Once the post- interventions network is determined, the normalized weights are Social Network Based Substance Abuse Prevention via Network Modification (A Preliminary Study)AAMAS'19, May 2019, Montreal, Canada calculated according to equation: s+ Ws x ij i′∈V Ws x wij = w+ ij + Ww x s+ i′j + Ww x , w+ i′j where Ws and Ww ,(Ww < Ws) are numerical values for strong and weak ties, respectively. In this equation, the numerator is the strength of the tie from i to j, before normalization which is equal to Ws if it is a strong tie, or Ww if it is a weak tie. The denominator is the sum of all the incoming weights. While the above constraint consists of nonlinear terms, they can be linearized using standard techniques which require introducing a new decision variable for each product term. Finally, the objective can be defined as follows: (Ωnu wij)(cid:1) + (cid:0)1 − (Ωun wij). j:bj =0 i:bi =1 j:bj =1 i:bi =0 This equation gives the closed form objective to our problem, which is the expected number of "non-users" post-intervention. The first term in the above equation is the expected number of "non-users" that remain "non-users." The second term is the expected number of "users" that change. This equation is equivalent to the objective in Problem (1). Note that since in the influence model, the thresholds are drawn from a uniform distribution, the probability that a node changes its behavior equals the normalized incoming edge weights from neighbors with that behavior. 4 LARGE NEIGHBORHOOD SEARCH (LNS) Local search algorithms target a set of feasible candidate solutions (neighborhood) and iteratively improve a solution by searching within that limited space. Small neighborhoods are faster to ex- plore, but can make escaping a local optimum much harder. Large neighborhood search which was first proposed by [16], can alleviate this problem by exploring a carefully chosen large neighborhood. LNS iteratively improves a solution by alternatively destroying and repairing a solution. Our solution approach is based on a mixture of LNS and MILP optimization which has been proven successful in [11]. Starting from an initial feasible solution, each loop starts by destroying the current solution by selecting two groups at random. It then repairs the solution by optimally deciding how to partition people between these two groups. Since this MILP subproblem is much smaller than the original MILP, it yields a significant improvement in the run time. It is important to note that in the destroy step, determining the amount of destruction is crucial. If too little is destroyed, the effect of large neighborhood is lost, whereas large destructions will result in a repeated re-optimization. We experimented with different neighborhoods, for example destroying the group assignment of random nodes or three groups at the same time (instead of two). In these cases, the results were not performing as well as the two- group neighborhood. 5 EXPERIMENTAL SETUP Network Data. Social Network data primarily comes from the Youthnet study in which homeless youth were recruited from a drop-in center in Los Angeles County, CA. Clearly, our approach is independent of the input network data. Therefore, we use this Figure 3: Comparison between small and large neighbor- hood search dataset for our preliminary study. We also plan to deploy our ap- proach in Denver and in collaboration with Urban Peak organiza- tion. In order to explore different network structures, we use randomly generated Watts Strogatz (WS) graphs [20] with parameters (p = 0.25, k = 4). WS graph models have properties such as short average path lengths which makes them close to real-world networks. Baselines. For evaluation, we compare MILP and LNS against three different baselines. These baselines are based on the three approaches that were detailed in the Introduction section (random assignment, network-based or by participants choice, and teacher nominated assignments). Model Parameters. In the following experiments, and based on input from domain experts we choose the model parameters as: Ωun = 1.0, Ωnu = 0.8, C = 3, C = 8, Ws = 3, and Ww = 1. We also explore the impact of variations of some of these parameters. For the LNS, 50 trials are performed, each starting from a different random initial solution and we report the solution with highest expected success rate. We also report averages over 25 different graph samples, both for the real and WS graphs. success = i∈V (1 − bi) The numerator is the expected number of youth that have become E( i∈V (1 − Bi(P))) − max(cid:0)E( i∈V (1 − Bi(P))) − 6 RESULTS AND DISCUSSION Solution Quality Metrics. Different solution strategies are com- pared based on a success metric, which we define as: i∈V (1 − bi)(cid:1) "non-users" as the result of the intervention. The term E( corresponds to the objective function of Problem (1). The denom- inator is its maximum possible value which is bounded above by bi -- it corresponds to the case where all "users" thresholds Ωnu are exceeded, and thus the maximum expected number of "users" that will change. Neighborhood Search: Small vs. Large. We first compare the success rate of the two local search based algorithms. In each step of the small neighborhood search, a randomly chosen pair of nodes switch groups. If the resulting solution has a higher success rate, the solution is accepted and this process continues until a locally optimal solution is found. As explained previously, the LNS im- i∈V i∈V (1 − Bi(P))) 00.20.40.60.811.2Success rateSmall NeighborhoodLarge NeighborhoodAAMAS'19, May 2019, Montreal, Canada Aida Rahmattalabi, Anamika Barman Adhikari, Phebe Vayanos, Milind Tambe, Eric Rice, and Robin Baker Figure 4: Solution quality of MIP and LNS in GUIDE, against three baselines commonly employed by practitioners (Syn- thetic Graphs) Figure 6: Success vs. Ωun/Ωnu Figure 5: Solution quality of MIP and LNS in GUIDE, against three baselines commonly employed by practitioners (Real Graphs) plementation is based on re-optimization of the randomly chosen groups. For both algorithms, each run is repeated for different initial random solutions until a time limit is reached. In this experiment, we use 10 different randomly generated WS graphs of size 30. The time limit is set to 1 hour. Figure 3 shows that the small neighbor- hood search can perform poorly relative to LNS under a variety of instances. This results indicates that between MILP, which is not easily scalable to larger networks, and pure small neighborhood local search which is favorable due its lower computational demand, LNS can achieve a balance between quality and speed and it is able to outperform local search, given the same time budget. Solution Quality. Figures 4 and 5 compare the success rate of the optimization techniques, MIP and LNS against our baselines across 4 different network sizes, for both real network samples and WS samples. The MIP solver is given a cutoff time equal to the solution time of LNS (summed over the 50 trials). These results indicate that the solutions of both MIP and LNS are significantly better than any of the traditional methods for forming these groups, both statistically (p < 0.01) and practically. From the practical point of view, prevention science cares about effect size of interventions in addition to the statistical significance, and we observed in the results that we were able to provide solutions that outperform the traditional methods up to (20 - 41)%. Figure 5 also shows that in the real network samples (sizes 20 and 30), the baselines have a neg- ative success rate. Negative success rate corresponds to the deviancy training effect, which means the expected number of "non-users" Figure 7: LNS solution time analysis. This plot indicates how the solution evolves with respect to the number of runs. post-intervention is fewer than pre-intervention. This observation can be explained by the structure of the network as well as the behavior of the nodes. In fact, in these samples, a significant ratio of the individuals are "users" (68%), hence the negative influence dom- inates. Additionally, many of these "users" are strongly connected with their "user" peers, making it hard for them to be positively influenced. We should note that the optimized strategies are always strictly positive, though the success rate is limited due to the cer- tain network structure. It can also be observed that the common intuition of evenly distributing the "users" across groups is in fact performing relatively well as compared to the other baselines. The importance of this result is twofold; First, it validates our model by reflecting the intuition of the social work partners that such a strategy is helpful. Secondly, we can clearly see that there is room for further improvements over these baselines. There are computational lessons learned as well. For example, MIP is guaranteed to find the optimal solution by searching the entire solution space, but as shown here, it is not a practical solu- tion due to time constraints. In fact, LNS outperforms MIP solution given the same time budget. Figure 6 compares the intervention success with respect to different Ωun ratios. This experiment is Ωnu performed on synthetic graphs of size 40. It illustrates that the increase of this ratio will result in a decline in expected success of our intervention. Scalability. Figure 7 shows the LNS solution quality in a 40-node real-world graphs against the number of runs. Recall that LNS is run for 50 20304080Number of nodes00.20.40.60.8Success rateMIPLNSRandom-BaselineFriendship-based BaselineDistribute-users Baseline20304080Number of nodes-0.200.20.40.6Success rateMIPLNSRandom-BaselineFriendship-based BaselineDistribute-users Baseline00.40.811.251.662.5Ωun/Ωnu00.20.40.60.8Success rate050100Runs0.60.650.70.75Success40-Node Real NetworksSocial Network Based Substance Abuse Prevention via Network Modification (A Preliminary Study)AAMAS'19, May 2019, Montreal, Canada Figure 8: LNS solution time analysis. This plot shows how the time required to find a locally optimal solution changes with respect to the network size. times, each starting from a different random initial solution. Figure 7 (a) shows that almost after 20 runs, the solution does not change significantly. Figure 8 shows the time needed for the LNS to find a locally optimal solution for different network sizes. This plot shows that as the size of the network increases, the LNS time increases polynomially. These results suggest that LNS is an efficient choice for our purpose, mainly because a decision aid, such as GUIDE, needs to be adjustable to unanticipated occurrences. For example in cases a person does not show up, or refuses to accept his assignment, GUIDE needs to recalculate a new solution relatively quickly based on the imposed constraints. Therefore we will utilize LNS as the core optimization algorithm in GUIDE. 7 DEPLOYMENT GUIDE is developed in collaboration with Urban Peak, a homeless- youth serving organization in Denver, CO, and it is under prepara- tion for deployment. We have developed an application that encap- sulates the data collection phase as well as the implementation of the LNS, the optimization algorithm. This application has a user- friendly interface that facilitates the use of our algorithm for the practitioners. In the first phase, detailed data about the social network and substance use behavior of each volunteer is collected. Then, a per- sonalized ID is assigned to each individual and the data is stored on the local machine for future use. See Figure 9 (a). Next, the data is fed into the LNS algorithm to generate a suggestion for intervention groups. The practitioners can easily upload or choose an existing social network data file. After the optimization is performed, the results will appear on the screen, See Figure 9 (b). The results in- clude a group assignment for each individual which is formatted as a table. Finally, the output can be printed and saved to a file for future references. 8 CONCLUSION Substance abuse is a very significant public health and social prob- lem in the United States. We have identified a niche in this wicked social problem where AI can make a contribution, illustrated the many modeling complexities, and provided algorithmic improve- ments over current practice. The result is GUIDE, an AI-based (a) A snapshot of the data collection form (b) A snapshot of the GUI for the group recommendation Figure 9: Two snapshots of the GUIDE application. (a) First, detailed information about the substance-use behavior of each individual is collected and saved in a data base. (b) Sec- ond, the practitioners will query the application for a group recommendation. decision aid which leverages social network to structure interven- tion groups. We evaluated our approach against different traditional methods ranging from those based on intuitions such as distributing "users" to purely random strategies with no effort to understand the social circle of these youth. We showed that these traditional strate- gies significantly under-perform relative to our proposed method, emphasizing the importance of social influence in composition of these groups. In fact, here our focus is on the life choices of a extremely vulnerable section of the population. The result, as we discussed, is a model that is of tremendous value to prevention sci- entists, providing them with a quantitative means of understanding phenomena for which until now only qualitative observations were feasible. In particular, as these interventions are very costly and time consuming, the ability to forecast the likely impact of different strategies is very helpful. REFERENCES [1] Sinan Aral and Dylan Walker. 2014. Tie strength, embeddedness, and social influence: A large-scale networked experiment. Management Science 60, 6 (2014), 1352 -- 1370. [2] Nikhil Bansal, Avrim Blum, and Shuchi Chawla. 2004. Correlation clustering. Machine Learning 56, 1-3 (2004), 89 -- 113. [3] Allan Borodin, Yuval Filmus, and Joel Oren. 2010. Threshold models for competi- tive influence in social networks. In Proceedings of the 6th international conference on Internet and network economics. Springer-Verlag, 539 -- 550. [4] Damon Centola and Michael Macy. 2007. Complex contagions and the weakness of long ties. American journal of Sociology 113, 3 (2007), 702 -- 734. [5] John P Dickerson, Ariel D Procaccia, and Tuomas Sandholm. 2012. Optimizing kidney exchange with transplant chains: Theory and reality. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems- Volume 2. International Foundation for Autonomous Agents and Multiagent Systems, 711 -- 718. [6] J Georgie, Harrison Sean, M Deborah, Hickman Matthew, Campbell Rona, et al. 2016. Peer-led interventions to prevent tobacco, alcohol and/or drug use among young people aged 11 -- 21 years: a systematic review and meta-analysis. Addiction 111, 3 (2016), 391 -- 407. [7] Robert Görke, Tanja Hartmann, and Dorothea Wagner. 2009. Dynamic Graph Clustering Using Minimum-Cut Trees. In Proceedings of the 11th International Symposium on Algorithms and Data Structures. Springer-Verlag, 339 -- 350. [8] David Kempe, Jon Kleinberg, and Éva Tardos. 2003. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 137 -- 146. [9] Elias B. Khalil, Bistra Dilkina, and Le Song. 2014. Scalable Diffusion-Aware Optimization of Network Topology. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 15 20 30 40 10-1100101102LNS Time (s)AAMAS'19, May 2019, Montreal, Canada Aida Rahmattalabi, Anamika Barman Adhikari, Phebe Vayanos, Milind Tambe, Eric Rice, and Robin Baker [10] Masahiro Kimura, Kazumi Saito, and Hiroshi Motoda. 2008. Solving the con- tamination minimization problem on networks for the linear threshold model. PRICAI 2008: Trends in Artificial Intelligence (2008), 977 -- 984. [11] Boon Ping Lim, Menkes Van Den Briel, Sylvie Thiébaux, Russell Bent, and Scott Backhaus. 2015. Large neighborhood search for energy aware meeting sched- uling in smart buildings. In International Conference on AI and OR Techniques in Constriant Programming for Combinatorial Optimization Problems. Springer, 240 -- 254. [12] Sean Esteban McCabe, Brady T West, Philip Veliz, Kenneth A Frank, and Carol J Boyd. 2014. Social contexts of substance use among US high school seniors: a multicohort national study. Journal of Adolescent Health 55, 6 (2014), 842 -- 844. [13] Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather: Homophily in social networks. Annual review of sociology 27, 1 (2001), 415 -- 444. [14] Adeline Nyamathi, Angela Hudson, Barbara Greengold, and Barbara Leake. 2012. Characteristics of homeless youth who use cocaine and methamphetamine. The American journal on addictions 21, 3 (2012), 243 -- 249. [15] Aida Rahmattalabi, Phebe Vayanos, and Milind Tambe. 2018. A Robust Opti- mization Approach to Designing Near-Optimal Strategies for Constant-Sum Monitoring Games. In Decision and Game Theory for Security, Linda Bushnell, Radha Poovendran, and Tamer Başar (Eds.). Springer International Publishing, Cham, 603 -- 622. [16] Paul Shaw. 1998. Using constraint programming and local search methods to solve vehicle routing problems. In International Conference on Principles and Practice of Constraint Programming. Springer, 417 -- 431. [17] Jennifer R Spoor and Janice R Kelly. 2004. The evolutionary significance of affect in groups: Communication and group bonding. Group processes & intergroup relations 7, 4 (2004), 398 -- 412. [18] Steve Sussman, Clyde W Dent, Alan W Stacy, and Sande Craig. 1998. One-year outcomes of project towards no drug abuse. Preventive Medicine 27, 4 (1998), 632 -- 642. [19] Thomas W Valente, Beth R Hoffman, Annamara Ritt-Olson, Kara Lichtman, and C Anderson Johnson. 2003. Effects of a social-network method for group assign- ment strategies on peer-led tobacco prevention programs in schools. American journal of public health 93, 11 (2003), 1837 -- 1843. [20] Duncan J Watts and Steven H Strogatz. 1998. Collective dynamics of'small- world'networks. nature 393, 6684 (1998), 440. [21] Amulya Yadav, Hau Chan, Albert Xin Jiang, Haifeng Xu, Eric Rice, and Milind Tambe. 2016. Using social networks to aid homeless shelters: Dynamic influence maximization under uncertainty. In Proceedings of the 2016 International Confer- ence on Autonomous Agents & Multiagent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 740 -- 748.
1108.0027	1	1108	2011-07-29T23:00:59	Revisiting Degree Distribution Models for Social Graph Analysis	[ "cs.SI", "physics.soc-ph" ]	Degree distribution models are incredibly important tools for analyzing and understanding the structure and formation of social networks, and can help guide the design of efficient graph algorithms. In particular, the Power-law degree distribution has long been used to model the structure of online social networks, and is the basis for algorithms and heuristics in graph applications such as influence maximization and social search. Along with recent measurement results, our interest in this topic was sparked by our own experimental results on social graphs that deviated significantly from those predicted by a Power-law model. In this work, we seek a deeper understanding of these deviations, and propose an alternative model with significant implications on graph algorithms and applications. We start by quantifying this artifact using a variety of real social graphs, and show that their structures cannot be accurately modeled using elementary distributions including the Power-law. Instead, we propose the Pareto-Lognormal (PLN) model, verify its goodness-of-fit using graphical and statistical methods, and present an analytical study of its asymptotical differences with the Power-law. To demonstrate the quantitative benefits of the PLN model, we compare the results of three wide-ranging graph applications on real social graphs against those on synthetic graphs generated using the PLN and Power-law models. We show that synthetic graphs generated using PLN are much better predictors of degree distributions in real graphs, and produce experimental results with errors that are orders-of-magnitude smaller than those produced by the Power-law model.	cs.SI	cs	Revisiting Degree Distribution Models for Social Graph Analysis Alessandra Sala§, Sabrina Gaito†, Gian Paolo Rossi‡, Haitao Zheng§ and Ben Y. Zhao§ §UC Santa Barbara, †Università degli Studi di Milano {alessandra, htzheng, ravenben}@cs.ucsb.edu, [email protected], [email protected] 1 1 0 2 l u J 9 2 ] I S . s c [ 1 v 7 2 0 0 . 8 0 1 1 : v i X r a ABSTRACT Degree distribution models are incredibly important tools for analyzing and understanding the structure and formation of social networks, and can help guide the design of efficient graph algorithms. In particular, the Power-law degree distri- bution has long been used to model the structure of online social networks, and is the basis for algorithms and heuris- tics in graph applications such as influence maximization and social search. Along with recent measurement results, our interest in this topic was sparked by our own experimental results on so- cial graphs that deviated significantly from those predicted by a Power-law model. In this work, we seek a deeper un- derstanding of these deviations, and propose an alternative model with significant implications on graph algorithms and applications. We start by quantifying this artifact using a variety of real social graphs, and show that their structures cannot be accurately modeled using elementary distributions including the Power-law. Instead, we propose the Pareto- Lognormal (PLN) model, verify its goodness-of-fit using graph- ical and statistical methods, and present an analytical study of its asymptotical differences with the Power-law. To demon- strate the quantitative benefits of the PLN model, we com- pare the results of three wide-ranging graph applications on real social graphs against those on synthetic graphs gener- ated using the PLN and Power-law models. We show that synthetic graphs generated using PLN are much better pre- dictors of degree distributions in real graphs, and produce experimental results with errors that are orders-of-magnitude smaller than those produced by the Power-law model. 1. INTRODUCTION Graph degree distributions are fundamental tools used in the study of complex networks such as online social net- works. Not only do they reveal insights into the structure and formation of these networks, but they also lay the foun- dation for modeling network dynamics and help guide the design of graph algorithms and applications. In particular, it is widely believed that the Power-law distribution accu- rately captures node degrees in complex networks such as online social networks, i.e. their degree distribution follows a Power-law function f (x) = cx−α, for some normaliza- tion constant c and exponent α. As a result, the Power-law model has already played a significant role in guiding the design of algorithms on social network problems such as in- fluence maximization [9], landmark selection [28], and link privacy protection [17]. Deviating from Power-law. Our interest in verifying the Power-law model for social graphs was first sparked by our own efforts to partition social graphs. In a recent project, we searched for an efficient way to partition and divide large social graphs across distributed machines for parallel graph computation. The goal is to minimize edges between parti- tions, thereby reducing data dependencies between machines when resolving graph queries. Given the known density of social graphs, it was not sur- prising when popular partitioning algorithms such as Metis produced poor partitions (those with low modularity) from our Facebook graphs [39]. Our next step was to leverage the known fact that Power-law graphs are vulnerable to targeted attacks, i.e. they quickly fragment when their "supernodes" are removed [4]. We evaluated a new partitioning approach where a small portion of nodes with the highest degree are replicated across all machines. This allows us to essentially "remove" the supernodes from the graph, which should frag- ment the graph into numerous disconnected or weakly con- nected subgraphs, which are easily partitioned. Surprisingly, our results showed that unlike prior results on peer-to-peer networks [34], social graphs were extremely resilient to this approach. On one of our typical Facebook graphs with 500K nodes, cutting the top 10% (50K) supern- odes had no impact on the connected graph. Nearly all of the remaining nodes (∼440K nodes) were still connected in a strongly connected component. Revisiting degree models. These surprising results led us to re-examine how well social graphs obey the Power-law degree model. Recent measurements of popular online so- cial networks have shed light on their internal structure, and also produced a number of real social graphs suitable for model validation [39, 23, 38, 24]. Despite using the Power- law parameter as a popular graph metric, it is increasingly evident that the Power-law only captured a portion of the degree distribution curve. While never explained, several re- sults consistently show that the Power-law distribution over- 1 predicts the number of "high-degree nodes" in social net- works [2, 14]. Similar divergence has also been observed in Internet topology and web graphs or mobile cell graphs [27, 35]. In this paper, we revisit the problem of modeling node degree distributions in online social networks. We question existing assumptions, propose the use of an alternative dis- tribution model, and show that choosing an appropriate de- gree distribution model has dramatic impact on a wide-range of OSN and graph applications. We are primarily interested in answer three key questions. First, how significant is the fitting error when modeling degree distributions using the Power-law and other elementary distributions? Second, can we propose an alternative distribution that provides a fit with significantly better accuracy? Finally, what is the real impact of switching to this alternative distribution from Power-law, and how does it impact real graph applications? Results and Contributions. In addressing the questions outlined above, our work makes several contributions to the problem of developing a more accurate degree distribution model for online social networks. First, we assert that the seminal Power-law distribution does not accurately model degree distributions in OSNs. We verify this hypothesis using a number of measured social graphs from the Facebook and Orkut OSNs, and show that other elementary distribution also perform poorly. Second, we search for a more accurate model by examining complex distributions, and propose the use of the Pareto-Lognormal (PLN) distribution. Intuitively, the Power-law distribution forms when each new node joins the network via a bootstrap node bi, and builds edges using the "rich-get-richer" model in some graph neighborhood around bi. In contrast, the PLN distribution forms when each new node joins multiple com- munities and uses a stochastic process to drive connections within each community. To support PLN's use in analysis, we analytically describe its probability distribution function, cumulative distribution function, and maximum likelihood function. Using three different error measures, we com- pare the accuracy of PLN, Power-law and 5 other models on 7 different OSN social graphs ranging in size from 740K edges and 14K nodes, to 118 million edges and 1.6 million nodes. Our results confirm that PLN provides the most ac- curate edge distribution model. Third, we highlight the differences between the PLN and Power-law models, quantifying the asymptotic differences between them when predicting high degree nodes in the net- work. For both models, we derive a close form to bound the lowest degree of a node for any percentile of high-degree nodes in the network. We then make predictions of the car- dinality of nodes in that percentile. Using our social graph data, we validate the predictions from both models, and find that PLN generally produces errors that are at least two- orders-of-magnitude smaller than those of Power-law. Fourth, we examine the end-to-end impact of degree dis- tribution models on multiple graph applications, including graph partitioning, influence maximization [9], and attacks on social graph link privacy [17]. Prior studies of these ap- plications used assumptions of Power-law graphs to drive their algorithm design. We implement each application, and show that running them on synthetic Power-law graphs pro- duced dramatically different results from those on real so- cial graphs . In contrast, we show that running applications on synthetic PLN graphs produces results nearly identical to those produced using real social graphs. Finally, we give some preliminary intuition towards the ongoing design of a generative model that both captures the temporal evolution of social networks and produces degree distributions match- ing our PLN model. We draw insights from a series of daily snapshots of a Facebook social graph that capture dynamic growth over a period of a month. Social Graph Datasets. To evaluate our models, we use 7 real social graphs gathered from recent measurements of popular online social networks. The majority of our datasets come from Facebook, the most popular OSN today with more than 500 million users. We use traces gathered through crawls of the Facebook network in 2009, when Facebook was structurally organized into geographical/regional net- works. We had crawled and analyzed over 10 million users (∼15% of Facebook in 2008), as part of an earlier measure- ment study [39]. For this paper, we utilize 6 anonymized social graphs representing a range of network sizes, from a small Monterey, CA graph (13K nodes, 704K edges) to a large London graph (1.6 million nodes, 118 million edges). We also include in our study the "Manhattan Random Walk" Facebook graph from [14], which has been proven to be a representative uniform random sampling of the total Face- book graph. We use this graph to validate the representa- tiveness of our Facebook results. Finally, we also include a public social graph from the Orkut OSN [23]. With 3 million nodes, 111 million edges, it has more nodes but less edges than our largest Facebook graph (London). Our datasets and their key statistics are summarized in Table 1. Roadmap. We begin in Section 2 by examining how well elementary functions fit degree distributions from real so- cial graphs. Next, in Section 3, we describe the PLN model through its PDF, CDF and maximum likelihood functions, and show the accuracy of this model through statistical anal- ysis. Then, we quantify asymptotic differences between the models in Section 4, present experimental application-level results in Section 5, and discuss intuition for a generative model in Section 6. Finally, we discuss related work in Sec- tion 7 and conclude in Section 8. 2. ELEMENTARY DISTRIBUTIONS We start by examining how well elementary distribution models fit real social graphs from deployed OSNs. We in- clude in our analysis the three elementary distributions: Power- law, Lognormal and Exponential distributions [25, 11]. We leave out the formal introduction of these models, and in- 2 stead focus on presenting the results of our experimental analysis. Fitting Models to Real Datasets. We fit each distribu- tion model to our real OSN datasets, using an optimal esti- mator to derive the best model parameters [11]. For each considered distribution model, we analyze the quality of the fit on our real data through three probability plots: the probability distribution, the complementary cumu- lative distribution (CCDF), and the "quantile-quantile" plot. The probability distribution function is an important met- ric for understanding the portion of nodes that have a certain degree value. For example, the seminal Power-law model predicts that a large number of nodes have a high degree. The CCDF quantifies how often a node's degree is above a value. It is particularly useful for identifying the general slope of OSN connectivity. Finally, the Quantile-Quantile analysis can graphically compare different distributions, in our case a theoretical model and a real social graph. The plot shows the discrepancy between degree values that corre- spond to the same quantile in both distributions. The greater the distance from the reference line, the stronger the evi- dence that the dataset follows a different distribution. Experimental Results. We perform experiments on our 6 Facebook datasets, the random walk Facebook graph, and the Orkut dataset. Results across all graphs are highly con- sistent, and we only show 3 graphs for brevity, Santa Barbara Los Angeles and London Facebook networks. These social graphs range in size from 13K to 1.6M nodes. Figure 1 plots the CCDF of the fitted models on the three datasets. The CCDF presents a detailed view on the tail of these distributions. We see that none of the presented mod- els accurately captures the decay slope shown by the real datasets. Both the Power-law and the Lognormal distribu- tions overestimate the tail. Take London for example (Fig- ure 1). The Power-law model overestimates the highest de- gree nodes by up to 5 orders of magnitude compared to the real graph. In contrast, the Exponential distribution decays sharply, and significantly underestimates the number of high degree nodes. On the London graph, this error reaches 4 orders of magnitude for nodes with degrees of 2000. The quantile-quantile plot formally quantifies the distance between the real data and the model. Figure 2 shows that both the Power-law and the Lognormal model underestimate nearly 90% of nodes (i.e. Figure 2, x-axis ∈ [1, 500]) and largely overestimate the very high degree nodes. In con- trast, the Exponential distribution model exhibits a reason- ably good behavior along the lower tail and the body of the distribution compared to the dataset, although a strongly di- verging slope is displayed on the high degree nodes. We do not include plots on the comparison of probability distribution functions because of space constraints. How- ever, they lead to the same conclusion: the high density of nodes with low degree are not well predicted by the Power- law and Lognormal models, and all models fail to accurately predict the number of high degree nodes. Finally, while this section only reports results from graphical assessment of goodness-of-fit, we confirm our observations using statis- tical error measures later in Section 3.3. 3. COMPLEX DISTRIBUTION MODELS Our experimental results show that none of the elementary distributions, including the Power-law distribution, provide a satisfactory fit for degree distributions in today's OSNs. These results, however, do lead to an intuition where low de- grees are distributed following a Pareto model while higher degrees can be modeled with an asymptotically faster decay- ing distribution. We believe a combination of two models will offer a better representation of the complex phenomena observed on the OSNs. Similar approach has been success- fully used by other fields, for example, modeling actuarial data in economics [29]. In this section, we test our intuitions by evaluating four distribution models with a Pareto component, each with vary- ing degrees of accuracy and fitting complexity. Ultimately, we confirm via both graphical and statistical analysis that the beginning of these distributions is characterizable as a Pareto and the upper tail decays as a Lognormal, and that the Pareto-Lognormal distribution provides the best combi- nation of accuracy and fitting overhead. 3.1 Pareto with Exponential Cutoff Prior work has proposed modeling human behavior using the Pareto with Exponential cutoff distribution [36], where the beginning of the distribution follows a Pareto and the tail exponentially decays. This formulation is particularly interesting to our study given the decay observed in the tail of our data (see Figure 1). We study this distribution using the derivations from [15]. Fitting to real datasets. Figure 3 shows how well the Pareto with Exponential cutoff distribution fit the Santa Bar- bara, Los Angeles and London datasets. We plot the CCDF to graphically show how the tail of this distribution decays compared to real datasets. While this model is able to closely fit the low degree nodes of the real data, it fails short in cap- turing the upper tail by underestimating the density of nodes with high degrees. We conclude that the exponential cutoff is too sharp to properly capture OSN connectivity of central (high-degree) nodes. Intuitively, we are looking for a more gradual decay slope. 3.2 Pareto-Lognormal Distribution Given the low accuracy in modeling the upper tail of our data with the Pareto with exponential cutoff, we turn our at- tention to a family of distributions that mixes Pareto with Lognormal distributions. These models, compared to the Pareto with exponential cutoff model, provide a smother de- cay on the upper tail that provides a better match for our real data. We start with the Double Pareto-Lognormal (DPLN) distribution introduced by Reed in [32]. The DPLN is a 3 F D C C 1 0.1 0.01 0.001 0.0001 1 Santa Barbara Power Law Lognormal Exponential 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1 Los Angeles Power Law Lognormal Exponential 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 London Power Law Lognormal Exponential 1 10 100 1000 10000 Node Negree Figure 1: Complementary cumulative distribution function of elementary models fitted on Santa Barbara, Los Angeles and London Facebook graphs. s e l i t n a u Q a a D t 10000 1000 100 10 1 1 Power-law London 10 100 1000 10000 Theoretical Quantiles s e l i t n a u Q a a D t 10000 1000 100 10 1 1 Lognormal London 10 100 1000 10000 Theoretical Quantiles s e l i t n a u Q a a D t 10000 1000 100 10 1 1 Exponential London 10 100 1000 10000 Theoretical Quantiles Figure 2: Quantile-Quantile plot of elementary distributions fitted on the London dataset to graphically show the similarity of real social network data to a particular distribution model. F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1 Santa Barbara Pareto Exp. Cutoff 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 1e-07 Los Angeles Pareto Exp. Cutoff 1 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 1e-07 London Pareto Exp. Cutoff 1 10 100 1000 10000 Node Degree Figure 3: The CCDF of the Pareto with Exponential Cutoff fitted on Santa Barbara, Los Angeles and London networks. complex model with the ability to fuse two Pareto and a Log- normal distribution. It includes four parameters: two Pareto exponents α and β that identify the slope of the upper and lower tails of the distribution, and µ and τ that describe the Lognormal parameters connecting the two Pareto tails. The DPLN also gives rise to two other distributions [32]: the Pareto-Lognormal (PLN) and the Lognormal-Pareto (LNP). Both are expressed by three parameters: Pareto exponent β, and Lognormal components µ and τ. Next, we derive the precise formulation of the PLN dis- tribution in terms of the PDF, CDF and the likelihood func- tion. We omit the detailed derivation on DPLN and LNP for brevity. In Section 3.2 and 3.3, we use both graphical and statistical assessments of goodness-of-fit to compare these three distributions. Pareto-Lognormal PDF and CDF Derivation. The PLN is expressed as the combination of two probability dis- tributions. We derived the correct formula of this distribution (which is a limit form of the DPLN) using results on the DPLN [32]. 4 The Pareto-Lognormal probability distribution function is: f (x) = βxβ−1e(−βµ+ β2 τ 2 2 )Φc(cid:18) log x − µ + βτ 2 τ (cid:19) (1) where β characterizes the slope of the lower tail of this distri- bution which follows a Pareto behavior, and µ and τ charac- terize the body and the upper tail of this distribution, which approximate a Lognormal decline. We also derived the Pareto- Lognormal cumulative distribution function, formalized as follows: 2 τ τ (cid:19) + xβ e(−βµ+ β2τ 2 β and V AR[X] = τ 2 + 1 )Φc(cid:18) log x − µ + βτ 2 F (x) = Φ(cid:18) log x − µ (cid:19) (2) with E[X] = µ− 1 β2 . In Section 4, we will analyze the CDF and formally prove that the Pareto component describes the low values of the distribution and the Lognormal one dominates the high values. Pareto-Lognormal Likelihood Derivation The likeli- hood function estimates the parameters of a distribution func- tion in light of the observed data. The likelihood L is a function of parameters θ of the distribution and is defined d o o h i l e k L i g o L 9.72e+06 9.7e+06 9.68e+06 9.66e+06 9.64e+06 9.62e+06 9.6e+06 9.58e+06 9.56e+06 0.7 0.8 0.9 Pareto-Lognormal 1 1.1 1.2 1.3 1.4 1.5 Beta 10000 1000 100 10 s e l i t n a u Q a t a D 1 1 Pareto-Lognormal London 10 100 1000 10000 Theoretical Quantiles Figure 4: The β value corresponding to the minimum reverse log likelihood value maximizes the likelihood of the Pareto-Lognormal dis- tribution to fit the London sample. Figure 6: The Pareto-Lognormal Quantile-Quantile plot on London shows that this model almost perfectly approximates the real data. as follows: L(θ) = n Yi=1 f (xiθ) Using the definition of the Pareto Lognormal distribution in Equation 1, the likelihood becomes: L(β, µ, τ ) = n Yi=1 βxβ−1 i e(−βµ+ β2 τ 2 2 )Φc(cid:18) log xi − µ + βτ 2 τ (cid:19) The likelihood is defined as a product, and maximizing a product is usually more difficult than maximizing a sum. In- stead, we use a monotonously increasing conversion func- tion to transform the function L(β, µ, τ ) into a new func- tion L′(β, µ, τ ), such that L(β, µ, τ ) and L′(β, µ, τ ) have their maximum values for the same β, µ and τ values. The monotonous transformation we use is the logarithmic func- tion, which turns the maximization of a product into an eas- ier maximization of a sum. For simplicity, let A0 be −βµ + β2τ 2 2 , then log L = log n Yi=1 f (xi) = n Xi=1 log f (xi) = n Xi=1 log β+ n Xi=1 log xβ−1 i + n Xi=1 log e(A0) n n log β +(β−1) n Xi=1 log xi +nA0 + n Xi=1 Φc (cid:18) log xi − µ + βτ 2 Xi=1 Φc (cid:18) log xi − µ + βτ 2 τ τ (cid:19) (3) (cid:19) = We use (3) to estimate the parameters β, µ and τ in order to fit the Pareto-Lognormal model to our real OSN graphs. We also reverse the sign of (3) such that the likelihood to fit the data is maximized by minimizing −L′(β, µ, τ ). Parameter estimation. The Pareto-Lognormal distri- bution has three parameters (i.e. β, µ and τ) to estimate in order to fit real data. We perform a grid search, i.e. a multi- dimensional numerical search over the parameters space, to identify the best triplet of values to fit a particular dataset, as in [35]. 5 We bound the search of the parameters µ and τ using the Moment Method Estimation to determine the initial values, and then refine the search around those values to identify the best ones. While this methodology could lead to suboptimal results that lower the performance of our models when com- pared to those using optimal parameter configurations, we choose it because of its computational efficiency. We use the likelihood metric as the objective function in our parameter search [6], i.e. our goal is to minimize the reverse log likelihood "−L(β, µ, τ )" when searching for the optimal (β, µ, τ) [6]. To show that there is a clear concave trend around the minima, we plot in Figure 4 the values of −L(β, µ, τ ) for the London dataset as a function of β. We have identified similar trends in all our datasets. Fitting to real datasets. Figure 5 examines the results of fitting the PLN, DPLN and LNP distributions to our graphs as CCDF plots. Among these three models, LNP overes- timates the data on the upper tail, which suggests that the decay of the right tail follows more of a Lognormal model than a Pareto one. The flexibility of the DLPN model, with its four parameters, allows for a very good fit. However, it comes at the cost of a higher fitting complexity which grows exponentially in the number of parameters. On the other hand, the PLN distribution achieves the accuracy of the DPLN, and also has the reduced fitting complexity of LNP (i.e. 3 parameters instead of 4). Figure 5 clearly shows that on all three datasets (i.e. Santa Barbara, Los Angeles and London), PLN and DPLN overlap along the entire dis- tribution. In addition, we use the Quantile-Quantile plot to evaluate the fitting accuracy of PLN, and show the results for London in Figure 6. This plot is representative and others are omitted for brevity. It shows a near perfect fitting from the Pareto-Lognormal to the real London graph. 3.3 Statistical Analysis The graphical assessments in Section 2 (i.e. the PDF, CCDF and the Quantile-Quantile plot) are the first step to- wards a complete characterization of these distributions. We now look at statistical measures to quantify how well each model fits real social graphs [10]. F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1 Santa Barbara Lognormal Pareto DPLN Pareto Lognormal 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 Los Angeles Lognormal Pareto DPLN Pareto Lognormal 1 10 100 1000 10000 Node Degree F D C C 1 0.1 0.01 0.001 0.0001 1e-05 1e-06 1e-07 London Lognormal Pareto DPLN Pareto Lognormal 1 10 100 1000 10000 Node Degree Figure 5: The CCDF of the DPLN and its two limit forms, i.e. Pareto-Lognormal (PLN) and Lognormal Pareto (LNP) fitted on Santa Barbara, Los Angeles and London networks. The PLN provides the same accurate fit as the DPLN but with much lower fitting complexity. Real Graph Monterey Egypt nodes 13.843 edges ≈ 704K Santa Barbara nodes 27.140 edges ≈ 2M nodes 283K edges ≈ 11M Los Angeles nodes 572K edges ≈ 43M New York nodes 855K edges ≈ 66M Manhattan R.W. nodes 957K edges ≈ 80M nodes 1.6M edges ≈ 118M London Orkut nodes 3 M edges ≈ 111 M Statistical Methods Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Log L. AIC RSS Distribution Models Power-law 8.75338e+04 1.75069e+05 3.05267e+02 1.80247e+05 3.60497e+05 4.38954e+02 1.62073e+06 3.24148e+06 1.17928e+03 3.81640e+06 7.63281e+06 1.74060e+03 5.66194e+06 1.13239e+07 1.81230e+03 6.76810e+06 1.35362e+07 1.31910e+03 1.07131e+07 2.14262e+07 1.79927e+03 1.81346e+07 3.62692e+07 1.94280e+03 Lognormal 7.83008e+04 1.56605e+05 2.07443e+00 1.63328e+05 3.26660e+05 5.81415e+00 1.53376e+06 3.06754e+06 7.68335e+00 3.45975e+06 6.91950e+06 8.60777e+00 5.23512e+06 1.04702e+07 1.65956e+01 5.92193e+06 1.18438e+07 4.01862e+00 9.75799e+06 1.95159e+07 1.24657e+01 1.63179e+07 3.26358e+07 2.41064e+00 Exponential 7.78214e+04 1.55644e+05 4.86935e-01 1.61209e+05 3.22421e+05 7.31640e-01 1.53832e+06 3.07664e+06 1.69934e+00 3.44290e+06 6.88581e+06 1.50800e+00 5.17802e+06 1.03560e+07 1.87038e+00 5.86365e+06 1.17273e+07 9.62000e-02 9.60413e+06 1.92082e+07 7.33091e-01 1.62442e+07 3.24885e+07 3.75274e-01 Pareto-Exp. 7.76411e+04 1.55188e+05 3.43766e-01 1.60516e+05 3.21036e+05 3.37774e-01 1.50289e+06 3.00579e+06 3.43834e-01 3.42677e+06 6.85354e+06 8.55362e-01 5.12808e+06 1.02561e+07 5.88628e-01 6.02838e+06 1.20567e+07 8.07448e+00 9.65522e+06 1.92104e+07 2.30840e-01 1.61573e+07 3.23147e+07 2.95699e-01 LNP 7.76411e+04 1.55288e+05 2.59320e-01 1.61152e+05 3.22311e+05 6.10471e-01 1.49892e+06 2.99785e+06 2.13671e-01 3.41280e+06 6.82561e+06 5.02215e-01 5.13177e+06 1.02635e+07 7.90849e-01 5.89459e+06 1.92208e+07 6.29269e-01 9.61043e+06 1.17892e+07 1.26502e+00 1.61822e+07 3.23645e+07 4.16978e-01 PLN 7.74057e+04 1.54817e+05 2.26463e-02 1.60448e+05 3.20902e+05 7.44256e-02 1.49855e+06 2.99710e+06 1.58090e-01 3.40372e+06 6.80745e+06 4.82168e-02 5.11728e+06 1.02345e+07 1.70891e-01 5.85632e+06 1.17126e+07 2.44948e-02 9.56562e+06 1.91312e+07 1.99847e-01 1.60868e+07 3.21738e+07 1.60096e-02 DPLN 7.74089e+04 1.54825e+05 2.32905e-02 1.60455e+05 3.20919e+05 7.74719e-02 1.49830e+06 2.99661e+06 2.03417e-01 3.40379e+06 6.80759e+06 4.75582e-02 5.11753e+06 1.02350e+07 1.65433e-01 5.85679e+06 1.17135e+07 2.34597e-02 9.56626e+06 1.91325e+07 2.12646e-01 1.60829e+07 3.21659e+07 1.01767e-02 Table 1: Quantifying the "Goodness of the fit" of 7 distribution models via statistical methods on 6 Facebook datasets crawled in 2009, and on the Orkut dataset. Goodness-of-Fit Analysis. We consider three measures to evaluate the goodness of the proposed statistical models, including the likelihood function, the Akaike's Information Criterion and the Residual Sum of Squares. The likelihood function is used to quantify the likelihood that a particular model fits a given dataset. Section 3.2 ex- plained how to estimate this function for PLN by computing the minimum of the reverse log likelihood. We compute the minimum of the reverse of the log likelihood function for each model such that the best fit models generate the lowest values. This matches the two other measures, where the best models also generate the smallest fitting errors. Akaike's Information Criterion (AIC) [3] is a measure of the quality of the fit that is capable of capturing the tradeoff between accuracy and fitting costs. The AIC value is com- puted as: AIC = 2k − 2 log L where k is the number of parameters in the statistical model, and L is the maximized value of the likelihood function for the estimated model. The value k in the AIC test is used to tradeoff the accuracy of a model with its complexity. The model that shows the lowest AIC is considered the one that best fits the data. Residual Sum of Squares(RSS) [35] is a statistical method that computes the sum of squares of residuals between the empirical distribution and the data sample. It measures the discrepancy as euclidean distance between the data and the estimation model. A small RSS indicates a tight fit of the model to the data, and it can be formally expressed as: RSS = P (yi − f (xi))2, where yi is the empirical evaluation and f (xi) is the estimate value of the statistical model. Error Measures on Real Traces. We now compare mod- els based on the numerical values of these three statistical methods computed on our OSN datasets. We explore sam- ple size variation to prove that the datasets manifest the same trend across different sample sizes. In Table 1 we report a set of statistical values to quantify 6 the goodness-of-the-fit for each of the seven analyzed mod- els. We highlight in bold the smallest values (i.e. those that identify the best model) for each metric. Across all datasets, the Power-law model consistently per- forms the worst. Its values of Log Likelihood and AIC test are the highest and the RSS values are up to 4 orders of magnitude higher than the best model. The second worst model on our measured datasets is the Lognormal. Its RSS values are up to 3 order of magnitude larger than the best model, due to the high imprecision in estimating the high degree nodes. Exponential, Pareto with Exponential cutoff and LNP provide reasonable accuracy in the sense their val- ues are with 1 order of magnitude from the best. The two best models are DPLN and PLN. The RSSs for some datasets identify DPLN as the best model, with a very small difference separating it from PLN (on the second or third decimal point). On the other side, the AIC test slightly penalizes the model with more parameters, such as DPLN. Based on the results of both the likelihood and the AIC, we see that PLN does consistently well on all our Facebook datasets. Finally, analysis of the Orkut graph produces re- sults consistent to our Facebook graphs. Overall we see a consistent trend: Power-law does not produce accurate results, and the best models are PLN and DPLN. Only in a small number of cases, DPLN is slightly more accurate than PLN, but differences are exceptionally small compared to other models. Given the significant in- crease in fitting complexity for DPLN (i.e. deriving 4 rather than 3 parameters), the PLN model clearly produces the best combination of fitting accuracy and complexity. 4. IMPLICATIONS OF THE PLN MODEL We have shown that the PLN model provides a much more accurate fit to today's OSN graphs. But are these differences large enough to really matter? In this section, we answer this question by quantifying the magnitude of error intro- duced by the Power-law model, and in the process, show that social algorithms and protocols based on the Power-law assumption must be re-evaluated (and some re-designed) us- ing the PLN model. The remainder of this section includes three parts. First, we analyze PLN using its CDF to characterize the asymptot- ical slope of its tail, which we use later to derive more com- plete bounds, in order to better understand the high degree nodes of these networks. Second, for both PLN and Power- law model, we analytically determine a close form to bound the lowest degree of high degree nodes. By comparing these bounds, we formally quantify the divergence of Power-law from PLN. We also validate our analytical results from em- pirical analysis on our Facebook and Orkut graphs. Finally, for both models, we approximate the cardinality of high de- gree nodes (those with degrees in the top 10% of the net- work), and evaluate the discrepancy between the two. Again, we validate our analytical results using our real datasets to understand the actual prediction errors from both models. 4.1 Modeling High Degree Nodes with PLN The cardinality of high degree nodes is a key factor in de- signing social applications and protocols. To characterize this distribution in the PLN model, it is necessary to under- stand the limit behavior on its CDF (or F (x)), defined by (2). We aim to bound the limit behavior so that we can di- rectly derive the expected number of high degree nodes and their connectivity. Let z = log x−µ ), we have , A = e(−βµ+ β2τ 2 τ 2 F (x) = Φ(z) + xβ A Φc(z + βτ ) (4) Leveraging the erf function (i.e. the Gauss error function, which is a special function of sigmoid shape), we can ex- press the standard normal cumulative distribution Φ(x) and its complementary form Φc(x) respectively as follows: Φ(x) = 1 2 [1 + erf ( x √2 )] = 1 2 erf c(− x √2 ) Φc(x) = 1 − 1 2 [1 + erf ( x √2 )] = 1 2 erf c( x √2 ) (5) (6) with x ∈ ℜ. Note that by definition, erf (−x) = −erf (x) and erf c(x) = 1 − erf (x). We can now reformulate (4) through the use of Gauss error functions as: √2 (cid:19) (7) erf c(cid:18) z + βτ √2(cid:19) + xβA erf c(cid:18)− F (x) = 1 2 1 2 z Next we use the asymptotical expansion of the Gauss error function to further expand F (x). For large x, the erf c() function can be approximated with the following series: erf c(x) ≈ e−x2 x√π +∞ Xn=0 (−1)n 2n! n!(2x)2n (8) Without loss of generality, in the remainder of this analysis, we will only use the first term of the series in (8). This is be- cause it achieves enough accuracy to accomplish the goals of this investigation. While we omit a formal error analysis of the loss in the approximation, we will show via experimental analysis that the loss is negligible, and the results are suffi- ciently accurate. Thus the erf c() function is approximated as follows: erf c(x) ≈ e−x2 x√π (9) LEMMA 1. [32] For a sufficiently large x, 1 − (Φ(z) + xβAΦc(z + βτ )) goes to Φc(z). The claim states that for large x the PLN distribution has a Lognormal form with the parameters of the PLN model. We will use this Lemma extensively in the following discus- sions. 4.2 Quantile Analysis: a Degree Threshold Next, we quantify how well the PLN and the Power-law models predict the degree threshold that defines the top γ 7 subset of the highest degree nodes in the network. In other words, we wish to predict the minimum degree ξγ, such that nodes with degree > ξγ are in the top γ portion of all nodes sorted by degree. Formally, ξγ is the γ-th quantile of the complementary cumulative distribution. As a concrete ex- ample, we will use γ = 0.10, i.e. top 10%. We compute ξγ for both models. For the Power-law model, the degree ξγ can be expressed as: ( 1 α . To quantify the γ ) same quantile for the PLN distribution, we again leverage its limit behavior. We then use the asymptotic expansion of the complementary Gauss error function to obtain a tight upper bound in Lemma 2. 1 LEMMA 2. A tight upper bound of the γ-th quantile of the PLN distribution is ξγ = eµ+q−2τ 2 log ( √2πγ τ ). PROOF. Given the result in Lemma 1, we consider 1 − F (x) ≈ Φc( log x−µ ) for large values of x. Since we are using the Pareto-Lognormal parameters, we can reformu- late the complementary cumulative distribution function for large x = ξγ values as: τ e−(log ξγ −µ)2 /(2τ 2) log ξγ −µ√2τ 2 . In order to derive the minimum de- 1 2√π gree ξγ, which characterizes the high degree nodes, we first quantify the γ% of high degree nodes in the network, and then derive the required degree value. We leverage the tail of the complementary cumulative distribution from Lemma 1, and thus, the following holds: 1 2√π e−(log ξγ−µ)2/(2τ 2) log ξγ−µ √2τ 2 = γ (10) e−y2 /(2τ 2) y τ√2π Let y = log ξγ − µ, then (10) becomes = γ. Applying the logarithmic function and approximation with the quadratic term, it reduces to y = q−2τ 2 log ( γ√2π ). By substituting y, we have log ξγ − µ = q−2τ 2 log ( γ√2π or ξγ = eµ+q−2τ 2 log ( of the high degree nodes in the γ-th quantile. ) which is the minimum degree ) √2πγ τ τ τ Next, we approximate the differences among the ξγ quan- tiles between the PLN and the Power-law distributions. Note that we are not approximating through a limit formulation but we are estimating the difference of the two, i.e. PLN and Power-law, analytical quantile values when γ ∈ [0.01, 0.1]. In order to compute the difference of the ξγ estimated by the Power-law and the PLN, we approximate the PLN's ξγ as eµ, since q−2τ 2 log ( ) is negligible compared to µ when γ is around its typical value 0.05. √2πγ τ The Power-law overestimation can be computed as the ra- tio of the ξγ quantiles of the Power-law and the PLN. Since µ and 1 α are approximated by the mean value, ν, of the sample logarithms, the following theorem applies: THEOREM 1. For γ ∈ [0.01, 0.1], the ratio of the ξγ quantiles of the Power-law and PLN is ≈ ( 1 eγ )ν. Real Graphs Monterey Santa Barbara Egypt Los Angeles New York Manhattan R.W. London Orkut (see Lemma 2) Power-law PLN Real Data ξγ 8906 160362 29987 170457 152937 357804 176742 41377 582 845 775 1101 1099 797 939 369 246 345 206 363 395 583 364 155 Table 2: Comparing the min degree of the top 10% high-degree nodes on our datasets, against the values predicted by the theoretical bounds. Both the PLN and Power-law estimates will express the ξγ degree as an exponential function, but they have different bases. For the Power-law, the base of the exponent is 1 γ = 10; for PLN it is e. This discrepancy accounts for the large difference in the predictions made by the two models. As we will show using experimental validation on real datasets, this difference can be as large as two order of magnitude. Experimental Validation on Real Data. Now we aim to validate the theoretical result shown in the previous sec- tion using experimental analysis on our datasets. For each dataset presented in Table 1, we compute its ξγ value (i.e. the estimated minimum node degree within the 10% of the highest degree nodes), and compare them against their ana- lytical bounds from the PLN (see Lemma 2) and Power-law models. We list the results in Table 2. Clearly, our theoretical ap- proximation of ξγ values using PLN are consistently more accurate than those from the Power-law model, which over- estimates the real data up to 3 orders of magnitude. For in- stance, on the New York graph, the ξγ value from the real sample is 395; our theoretical approximation from PLN is 1099; but the Power-law predicts 152937. 4.3 Cardinality of High-degree Nodes We next use the distribution models to predict the cardi- nality of high-degree nodes in OSN graphs. The cardinality of high-degree nodes is commonly used to design social al- gorithms and protocols, and to evaluate their performance and complexity. In the following, we first analytically quan- tify this metric using both the Power-law and PLN mod- els, and then empirically validate our predictions using real datasets. Power-law model. Using the Power-law density func- tion, we can derive the number of high degree nodes by com- puting the integral of the upper tail of the distribution. Let ξ be the minimum node degree among the high degree nodes, then the number of high degree nodes is approximated as: N Z ∞ ξ cx−αdx = cN 1 α − 1 ξ−α+1 ≈ N ξα−1 (11) where c is the normalization constant, approximated by (α− 8 Real Graphs Monterey Santa Barbara Egypt Los Angeles New York Manhattan R.W. London Orkut Power-law Real Data # of nodes with degree ≥ ξγ 1385 1843 6763 2724 28319 61170 57182 141980 85581 204540 72854 76133 398529 160050 307240 761970 PLN 1387 2442 30387 54894 77879 76133 156850 334900 Table 3: Number of nodes with degree higher than ξγ, with ξγ esti- mated from the 10% of the nodes, computed on the fitted models and the real data. x > e e r g e D h t i w s e d o N 1e+06 100000 10000 1000 100 10 Power Law Pareto Lognormal 1000 Node Degree 10000 Figure 7: When predicting # of high degree nodes, results from the Power-law and PLN models diverge up to 5 orders of magnitude. 0 1)dα−1 , and d0 is the minimum node degree in the network. PLN model. As mentioned before, a high degree node has a degree at least ξ, where ξ can be computed as in Sec- tion 4.2. Thus, the next lemma follows: LEMMA 3. Let ξ be the minimum degree for a high de- gree node, then the number of high degree nodes ≈ N 2π −( e log ξ−µ√2τ 2 log ξ−µ√2τ 2 )2 . Using Lemma 1, the number of high degree nodes de- scribed by the PLN distribution can be studied from the Φc(x) component of the CCDF of the PLN, where x is log ξ−µ . Thus this number becomes: N Φc(cid:16) log ξ−µ (cid:17). Using the ap- proximation in (8), we approximate the number of nodes in τ τ the network with degree no less than ξ by N 2π )2 −( e log ξ−µ√2τ 2 log ξ−µ√2τ 2 . Comparing High-degree Node Cardinality and Connec- tivity. We use two experiments to validate the quality of predic- tions on high-degree nodes from PLN and Power-law. First, we study the number of high-degree nodes in the measured network datasets, and compare this number with that esti- mated from the fitted models in Table 3. Specifically, we study the total number of high-degree nodes, defined as the nodes with degree higher than ξγ generated from the top 10% of the network as proven in Section 4.2. Table 3 shows the cardinality of high-degree nodes in the 9 s e d o N e e r g e D h g H i f o # e v i t a u m u C l 1e+08 1e+07 1e+06 100000 Power Law Pareto Lognormal 1000 Node Degree 10000 Figure 8: Comparing the cumulative degree of high degree nodes be- tween the Pareto-Lognormal model and the Power-law reveals different values up to 3 orders of magnitude. measured networks, as well as those predicted by the PLN and Power-law models. Clearly, the PLN model provides a tight approximation of the real dataset values, while predic- tions made by the Power-law model generally lead to more than 100% estimation error. Empirical results like these can sometimes be biased be- cause of specific distributions in the real data. To eliminate any possible bias, we generate pure sample networks from the Pareto-Lognormal and the Power-law models, then com- pute and compare each model's predicted number of high degree nodes and their connectivity. The pure sample networks are generated as follows. For the Power-law, we take the parameters of a fitted model on a particular real sample, and generate a pure sample by uni- formly extracting N numbers between 0 and 1. Then for each extraction x ∈ N , a node with the following degree is generated: degreex = 1 u(1/α) , where α has the value of the Power-law exponent fitted on a real dataset. For the Pareto- Lognormal, we generate a pure sample as follows: let x be a random number uniformly extracted between 1 and N , and let y be a random number from the exponential distribution with mean parameter 1/β. Then degreexy = eµ+τ∗x−y, where β , µ and τ are the parameters of the Pareto-Lognormal fitted on the real dataset. In Figures 7 and 8, we compare the number of high degree nodes and the sum of all node degrees respectively. We do this for N = 80K, and use the average of the fitted model values as the parameters of the Power-law and the Pareto Lognormal. Figure 7 shows the divergence of these two models in estimating the high degree nodes. In particular for nodes with degree > 2000, Power-law gener- ates 2 orders of magnitude more nodes than PLN. 5. IMPACT ON SOCIAL APPLICATIONS In the last section, we argued using both analytical predic- tions and empirical data that the Pareto-Lognormal model provides a much more accurate representation of node de- gree distributions in OSNs. In this section, we seek to bet- ter understand how the choice of degree distribution model impacts the performance of applications on social graphs. C C S n i s e d o N g n n a m e R % i i 100 90 80 70 60 50 40 30 20 10 0 Real Graph Pareto Lognormal Graph Power-law Graph 1 2 3 4 5 6 7 8 9 10 % of High Degree Nodes Removed (a) Santa Barbara, CA C C S n i s e d o N g n n a m e R % i i 100 90 80 70 60 50 40 30 20 10 0 Real Graph Pareto Lognormal Graph Power-law Graph 1 2 3 4 5 6 7 8 9 10 % of High Degree Nodes Removed (b) Los Angeles, CA C C S n i s e d o N g n n a m e R % i i 100 90 80 70 60 50 40 30 20 10 0 Real Graph Pareto Lognormal Graph Power-law Graph 1 2 3 4 5 6 7 8 9 10 % of High Degree Nodes Removed (c) London, UK Figure 9: Number of users remaining connected after deleting x% of the highest degree nodes from the graph. The reported results compare partitioning effects on the synthetic PLN and Power-law graphs and a real graph. The original graphs are Facebook graphs representing users in Santa Barbara CA, Los Angeles CA and London UK. More precisely, for applications that run on social graphs, we wish to quantify just how much the choice of a degree distribution model alters their experimental results. When actual measured social graphs are not available, so- cial applications generally use synthetic graphs generated using statistics extracted from real graphs [9]. To highlight the differences when applications use either the Pareto Log- normal or the Power-law degree model, we take each of our real Facebook social graphs, and generate two synthetic ver- sions of them: one assuming a PLN model for degree dis- tribution, and one assuming Power-law. We generate these synthetic graphs using prior works from [26, 7], which can generate a synthetic graph with no self loops, given a speci- fied degree distribution and a given network size. We present results from implementations of three impor- tant applications on social graphs. They include: a) a graph partition approach that replicates high degree nodes across partitions (introduced in Section 1), b) influence maximiza- tion on social graphs using three different information spread models [9], and c) a link-privacy attack on anonymized so- cial graphs [17]. All results are computed on a local cluster of Dell Xeon servers with 24–32GBs of main memory. Memory constraints limited the size of graphs we used in our results. Using Ama- zon EC2 large memory machines, we verified that our ob- served trends hold for larger graphs such as the Orkut graph, but such computations were too slow and costly in resources for us to generate comprehensive graphs. 5.1 Partitioning via Supernode Replication Efficiently answering queries on large graphs is a difficult challenge faced by companies dealing with large network datasets, e.g. Facebook, Zynga. Scalable solutions require splitting the graph data across machines in computing clus- ters. While some systems [30, 12] distribute nodes randomly across a cluster, an ideal solution would find a way to dis- tribute the graph across the cluster as subgraphs with min- imal edges between them. The best solutions would min- imize edges between partitions, which minimizes data de- pendencies between machines and maximizes parallel query processing. Unfortunately, graph partition is known to be NP-Complete [5], and social graphs are known to be very dense graphs that do not partition well. Fortunately, Power-law networks are known to be vulnerable to "targeted attacks," i.e. they quickly fragment into disconnected subgraphs when nodes with the highest degree are removed [4]. Since social graphs are widely accepted as Power-law graphs, we hypothesize we can easily partition a social graph by first fragmenting it into subgraphs, through the removal of a small number of supern- odes. These supernodes (and their edges) can be replicated to every node in a cluster. In addition to potentially parti- tioning highly connected graphs, this approach is attractive because it does not require the entire graph to be in memory, and can thus efficiently run on extremely large social graphs. We now look at the impact of running this partition scheme both on real social graphs, and on synthetic graphs gener- ated using the Power-law and PLN synthetic models. We take three of our social graphs, the Facebook regional graphs from Santa Barbara, CA, USA and London, UK [39], and the Orkut social graph [23], and create synthetic graphs that match each of them in size and node degrees, based on either the Power-law or the PLN model for degree distributions. We generate synthetic graphs from degree distributions us- ing Newman's approach [26]. For each of these graphs, we test the effectiveness of our graph partitioning strategy, by incrementally removing nodes from the graph, starting with those nodes with the highest de- gree. After these super nodes are removed, we examine the remaining graph, and measure the size of the largest con- nected component as a function of the original full graph. We plot the results in Figure 9 for each of the original graphs. The results are strongly consistent among our datasets. Regardless of the original graph in question, synthetic graphs produced using a Power-law distribution are quickly frag- mented. Removing just the top 3% of the highest degree nodes reduces the largest connected component to only 40% of the original graph. In contrast, the real social graphs and synthetic graphs from the Pareto-Lognormal distribution are highly resistant to the fragmentation. Even after removing 10% of the highest degree nodes, the graph remains largely connected, and the biggest connected component still con- 10 s r e s U d e c n e u l f n I f o # 1000 100 10 Power-law Graph Real Graph Pareto Lognormal Graph 0 5 10 15 20 25 30 35 40 45 50 # of Seed Nodes (a) Independent Cascade s r e s U d e c n e u l f n I f o # 10000 1000 100 Power-law Graph Real Graph Pareto Lognormal Graph 0 5 10 15 20 25 30 35 40 45 50 # of Seed Nodes (b) Weighted Cascade s r e s U d e c n e u l f n I f o # 10000 1000 100 Power-law Graph Real Graph Pareto Lognormal Graph 0 5 10 15 20 25 30 35 40 45 50 # of Seed Nodes (c) Linear Threshold Figure 10: Number of users influenced by a given number of seed nodes under three different influential dissemination models. Seed nodes are sorted by degree in decreasing order. tains 80% of all nodes. In this case, relying on an inaccu- rate degree distribution model produces results fundamen- tally different from the original graph. 5.2 Influence Maximization Social networks have proven to be exceptionally useful tools for information dissemination and marketing, and are used by companies and individuals to promote their ideas, opinions and products. One critical problem of interest is that of influence maximization, or how to identify an initial set of users who can influence the most number of users. This is known as the influence maximization problem. We examine the impact of degree distribution models on the influence maximization problem. Prior works have pro- vided algorithms that use statistical methods to model infor- mation dissemination over social links [9, 16]. We consider three different models to spread information: independent cascade model, weighted cascade model and linear thresh- old. These models differ in how they compute the proba- bility of influencing a node in the graph. For example, in- dependent cascade assumes a node is influenced indepen- dently from nearby nodes, while the probability of a node in weighted cascade being influenced is a function of its de- gree. In each case, heuristics provided by [9] allow us to compute the number of users influenced by an initial set of "seed" users. We seek to understand how the spread of in- fluence in these models is impacted by the structure of the network. We take the Santa Barbara Facebook graph, its two synthetic variants based on Power-law and PLN, and com- pute how many users are influenced as the number of seed nodes increases for each of the three influence models men- tioned above. We focus on results from the Santa Barbara network, since it is the largest graph we can execute given the significant memory footprint of code from [9]. Prior work has shown this graph to be a representative social graph in all graph metrics [33]. Figure 10 shows the spread of information for each of the three different models on the Santa Barbara network. For each model, we compare the results on Santa Barbara with results on its Power-law and PLN synthetic graphs. Our re- sults are consistent across all three influential models. While the exact result differs for each particular influential model, s e d o N d e r e v o C f o % 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 Power-law Real Graph Pareto Lognormal 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Fraction of Seeds Figure 11: The capability to spread an attack through the network is highly dependent on assumptions of the underlying topology. The attack is much more effective on a Power-law network than both the real graph and a Pareto-Lognormal network. results from the PLN graph results always closely follow re- sults from the real graph. In contrast, results on the Power- law graphs always overestimate the number of influenced users on the real graph, sometimes overestimating by an or- der of magnitude. These results are further confirmation that using an inaccurate degree distribution model can introduce dramatic errors in application-level experimental results. 5.3 Attacks on Link Privacy For our third application study, we focus on the problem of link privacy in social graphs. Prior studies such as [17] used experiments to quantify the impact of attacks to dis- close the presence of connections between social network users. This example differs from our other application stud- ies in that the application results are simulated based on ana- lytical derivations that integrate a degree distribution model. [17] presents different approaches to disclosure network information by attacking particular nodes in the network. The most effective attack strategy is "Highest," which tries to compromise nodes by spreading the attack from high degree nodes. Intuitively, because high degree nodes have many in- cident edges, they are more likely to successfully spread the attack across other nodes. The authors provide a mathemati- cal form to quantify the effect of this attack by using a prob- abilistic approach that leverages the probability distribution 11 t h w o r G e t l u o s b A 100 50 0 -50 1 Evolution SF1st-SF30th day 10 100 Starting Node Degree Figure 12: The evolution process affects nodes independently of their node degree. function of the network1. This leads to a simple question: how much does the effectiveness of this attack depend on the choice of network topology model? Quantifying the strength of this attack relies a notion of node coverage. A node is covered if and only if its 1-hop neighbors are unequivocally known. This definition allows [17] to present results where the strength of this attack is quantified as the fraction f of nodes whose entire immediate neighborhood is known. Let us begin by introducing the required variables to re- peat the experiment: N , m and d0 are respectively the total number of nodes in the networks, the sum of the node de- grees and the minimum degree. The variable D = Pk identifies the sum of the degree of the k nodes from where the attack originates. Theorem 3 in [17] states that if D = − ln ǫ0 2m, then the disclosed nodes after the attack are: N (1− ǫ − o(1)). ǫ is a variable that we can compute as follows: let ǫ0 = e(− i=1 d(xi)), which is needed to estimate the fraction of covered nodes. We can derive ǫ as: 2m Pk i=1 d(xi) d0 d0 ǫ = kmax Xx=d0 e−( x 2m Pk j=1 d(xj))f (x) where kmax is the maximum node degree, d(x) is the de- gree of x and f (x) is the density function of the degree distribution. If we assume a Power-law degree distribution, we substitute f (x) with cx−α and in the case of the Pareto- Lognormal we use f (x) as defined in (1). Figure 11 shows that results on a Power-law network sig- nificantly overpredicts the impact of attacks using the "High- est" strategy. In contrast, results on the real social graph closely match those from a synthetic graph following the Pareto-Lognormal degree model. Clearly, applications that use the Power-law model in their analytic derivations and simulations are also significantly affected by their choice of node degree distribution models. 6. TOWARDS A GENERATIVE MODEL In this section, we present preliminary results towards a 1 This use of network topology models is also commonly used in problems involving information propagation or dissemination in social networks. 12 generative model able to reproduce a Pareto Lognormal graph. We also show using evidence in our datasets that the for- mation of these graphs diverges from the hypothesized pref- erential attachment-based scheme presented in prior stud- ies [37, 21]. Our goal is to provide an intuition of an algorithm that captures a lognormal multiplicative process, and accurately models the temporal evolution of online social networks. Unlike generative models that focus on reproducing a single snapshot of a network [21], we focus on modeling the evolu- tionary process that captures the network growth. While the Power-law curve is generated by an iterative process follow- ing the preferential attachment rule, our PLN model may be explained by the following process: "A node joins the net- work through an introduction node, and builds connections within the local community; then it completes its growth by joining and growing in multiple other communities." To re- alize a generative model based on this intuition, we need a stochastic process derived from the PLN distribution that balances the Pareto and the Lognormal process to drive the growth rate of each node. Dynamic Graph Snapshots. To understand the growth of social networks, we perform 30 daily crawls of the San Francisco Facebook network during the month of November 2009. We use the same crawling strategy used to capture other Facebook datasets [39]. We use these daily snapshot graphs to help us understand the rate of network growth and how new edges form. Network Generative Models. Since our PLN model is itself a combination of a Pareto distribution and a Log- normal distribution, our search for an iterative process should naturally integrate two theories. Starting with "Preferential Attachment" [37], which produces a Power-law distribution, we add the "Law of Proportional Effects" [13], where the growth of the degree of a node, at discrete time t + 1, is multiplicative and independent from its actual size Xt. This contributes to a Lognormal distribution. We study the cor- relation of growth and current degree using our 30 snapshot graphs of the San Francisco network. The results in Fig- ure 12 confirm that the degree growth of different nodes is not influenced by their starting degree, meaning that the growth is a constant independent of node degrees. A Two-Phase Iterative Algorithm. We envision a two- phase iterative algorithm that integrates fundamental proper- ties from the "Law of Proportional Effects" [13] and "Prefer- ential Attachment" [37]. The bimodal connotation is meant to highlight the ability of our algorithm to integrate the Power- law and Lognormal probability models. Our two-phase algorithm alternates between adding new nodes to the network using a preferential attachment model, and growing the connectivity among nodes using the law of proportionate effects. Its two phases are driven by a prob- ability parameter p. We do not claim to provide a formal derivation of the algorithm or its proof of correctness. Both 0.1 F D P 0.001 1e-05 1 0.1 F D P 0.001 Santa Barbara Pareto Lognormal 10 Node Degree 100 1e-05 1 Synthetic SB Pareto Lognormal 10 Node Degree 100 Figure 13: The comparison between the PDFs of the real and syn- thetic Santa Barbara network shows that our generative model pro- vides an accurate reproduction of the original network. are beyond the scope of this paper. Here we limit ourselves to preliminary results to validate our underlying intuition. Initial Validation. To evaluate our intuition about the generative method, we implement a simple instance of a "Two-Phased Iterative Algorithm" summarized above, and run a structural comparison between the generated graphs and the real social network graphs. We fit the parameters of this algorithm by following the approach presented in [33]. Our goal is to show that our intuitive algorithm for gen- erating PLN distributions can generate synthetic graphs that accurately reproduce the degree distribution of our real data. In Figure 13, we show the similarity between the pdf of node degrees from the real data, versus those from the synthetic graph produced by our algorithm. The results are consis- tent across all seven of our datasets, and we only show re- sults only on Santa Barbara Network for brevity2. Given these initial results, we are continuing to work on formal- izing the mathematical formulation of our generative model and a more thorough analysis via larger, more complete syn- thetic datasets. 7. RELATED WORK Social Networks. Historically, both online and offline social networks have been explained through the seminal Power-law model. Power-law is often described with the "rich gets richer" paradigm which has been proven to hold in real datasets across multiple disciplines, including Inter- net router topology graphs, biological graphs [11, 31], hu- man mobility traces [8], etc. 2These synthetic graphs also show similar structural characteristics to the original graphs, including clustering coefficient, Knn, sepa- ration metrics, and betweenness centrality. 13 One of the first OSN study was conducted on Club Nexus website [1]. Later analytical studies attracted attention for their large scale, including CyWorld, MySpace and Orkut [2], YouTube, Flickr, LiveJournal in [23], and the most recent studies of Facebook [39] and Twitter[19]. More recently, re- searchers have begun to investigate the temporal properties of OSNs [18, 24, 22]. Preliminary analysis of OSN structures in these and other studies has shown that the degree distribution does not fol- low a pure Power Law distribution. As a result, followup work proposed to segment these distributions and fit the seg- mented pieces with distinct Power-law settings [14, 2]. Social Applications and Systems. We have shown that our proposed PLN is statistically more accurate in describ- ing OSNs than the seminal Power-law model. We believe that many social applications and protocols designed based on the Power-law assumption need to be re-evaluated, es- pecially algorithms and protocols that rely on the popula- tion of high degree nodes or their connectivity. Examples include distributed resource replication strategies to mini- mize routing delay and social search, epidemics dissemina- tion strategies to maximize information spread [9], landmark selection strategies to accurately predicts shortest paths in graphs [28], community detection to improve social recom- mendation systems, and social attack strategies [17]. 8. CONCLUSION Degree distributions are incredibly important tools for study- ing and understanding the structure and formation of social networks. They give us insight into network structure, and are the foundations of generative models that model growth and network dynamics [36, 20, 21]. Finally, they are key tools in the design and analysis of efficient algorithms for a number of challenging graph problems. Our work sheds light on an existing discrepancy between the commonly used Power-law model and real measurement data from today's online social networks. While most prior studies have ignored this error, we show that it is consis- tent across different communities and propose the Pareto- Lognormal distribution (PLN) as a more accurate alterna- tive. Our analysis shows that PLN significantly outperforms existing elementary models, and is the most accurate and efficient of all complex distribution models we studied. Fi- nally, we analytically quantify the magnitude of error reduc- tion achieved by moving from the Power-law to the PLN model, and confirm our analysis with empirical data. 9. REFERENCES [1] ADAMIC, L. A., BUYUKKOKTEN, O., AND ADAR, E. A social network caught in the web. First Monday 8, 6 (2003). [2] AHN, Y.-Y., ET AL. Analysis of topological characteristics of huge online social networking services. In Proc. of WWW (2007). [3] AKAIKE, H. A new look at the statistical model identification. IEEE Transactions on Automatic Control 19, 6 (1974), 716 – 723. [4] ALBERT, R., JEONG, H., AND BARABASI, A.-L. Error and attack tolerance of complex networks. Nature 406 (July 2000). [34] SAROIU, S., GUMMADI, K., AND GRIBBLE, S. D. A measurement study of peer-to-peer file sharing systems. In Proc. of MMCN (2002). [35] SESHADRI, M., MACHIRAJU, S., SRIDHARAN, A., BOLOT, J., FALOUTSOS, C., AND LESKOVEC, J. Mobile call graphs: beyond power-law and lognormal distributions. In Proc. of KDD (2008). [36] TREVOR, F., MARK, L., AND GEORGE, L. A stochastic evolutionary model exhibiting power-law behaviour with an exponential cutoff. Physica A Statistical Mechanics and its Applications (2005), 641–656. [37] VAZQUEZ, A. Growing network with local rules: Preferential attachment, clustering hierarchy, and degree correlations. Physical Review E 67-056104 (2003). [38] VISWANATH, B., MISLOVE, A., CHA, M., AND GUMMADI, K. P. On the evolution of user interaction in Facebook. In Proc. of WOSN (2009). [39] WILSON, C., BOE, B., SALA, A., PUTTASWAMY, K. P. N., AND ZHAO, B. Y. User interactions in social networks and their implications. In Proc. of EuroSys (April 2009). [5] ANDREEV, K., AND RÄCKE, H. Balanced graph partitioning. In Proc. of SPAA (2004). [6] ARNOLD, B. C. Pareto distribution. Encyclopedia of Statistical Sciences. John Wiley & Sons, Inc. (2004). [7] BRITTON, T., DEIJFEN, M., AND MARTIN-LoF, A. Generating simple random graphs with prescribed degree distribution. Journal of Statistical Physics 124, 1377–1397. [8] CHAINTREAU, A., ET AL. Impact of human mobility on opportunistic forwarding algorithms. IEEE Transactions on Mobile Computing 6, 6 (2007), 606–620. [9] CHEN, W., WANG, Y., AND YANG, S. Efficient influence maximization in social networks. In Proc. of KDD (2009). [10] CLAESKENS, G., AND HJORT, N. L. Model selection and model averaging. Cambridge University Press (October 2008). [11] CLAUSET, A., ET AL. Power-law distributions in empirical data. SIAM Review 51, (4) (2009), 661–703. [12] DEAN, J., AND GHEMAWAT, S. Mapreduce: Simplified data processing on large clusters. In Proc. of OSDI (December 2004). [13] GIBRAT, R. La loi de l'effet proportionnel. Compte-Rendus de l'Acadamie des Sciences (1932), 843–845. [14] GJOKA, M., KURANT, M., BUTTS, C. T., AND MARKOPOULOU, A. Walking in Facebook: A case study of unbiased sampling of OSNs. In Proc. of INFOCOM (2010). [15] KAGAN, Y. Y., AND SCHOENBERG, F. Estimation of the upper cutoff parameter for the tapered pareto distribution. Journal of Applied Probability, Vol. 38 (2001), 158–175. [16] KEMPE, D., KLEINBERG, J. M., AND TARDOS, E. Maximizing the spread of influence through a social network. In Proc. of KDD (2003). [17] KOROLOVA, A., ET AL. Link privacy in social networks. In Proc. of CIKM (2008). [18] KUMAR, R., ET AL. Structure and evolution of online social networks. In Proc. of KDD (2006). [19] KWAK, H., LEE, C., PARK, H., AND MOON, S. What is twitter, a social network or a news media? In Proc. of WWW (2010). [20] LESKOVEC, J., ET AL. Microscopic evolution of social networks. In Proc. of KDD (2008). [21] LESKOVEC, J., AND FALOUTSOS, C. Scalable modeling of real graphs using kronecker multiplication. In Proc. of ICML (2007). [22] LESKOVEC, J., KLEINBERG, J., AND FALOUTSOS, C. Graphs over time: Densification laws, shrinking diameters and possible explanations. In Proc. of KDD (2005). [23] MISLOVE, A., ET AL. Measurement and analysis of online social networks. In Proc. of IMC (Oct 2007). [24] MISLOVE, A., ET AL. Growth of the flickr social network. In Proc. of WOSN (Seattle, WA, August 2008). [25] MITZENMACHER, M. A brief history of generative models for power law and lognormal distributions. Internet Mathematics 1, 2 (2003). [26] NEWMAN, M. E. J., STROGATZ, S. H., AND WATTS, D. J. Random graphs with arbitrary degree distributions and their applications. Phys. Rev. E 64, 2 (2001), 026118. [27] PENNOCK, D. M., FLAKE, G. W., LAWRENCE, S., GLOVER, E. J., AND GILES, C. L. Winners don't take all: Characterizing the competition for links on the web. In PNAS (2002), pp. 5207–5211. [28] POTAMIAS, M., ET AL. Fast shortest path distance estimation in large networks. In Proc. of CIKM (2009). [29] PREDA, V., AND CIUMARA, R. On composite models: Weibull-pareto and lognormal-pareto. - a comparative study -. Journal for Economic Forecasting 3, 2 (2006), 32–46. [30] PUJOL, J. M., ET AL. The little engine(s) that could: scaling online social networks. In Proc. of SIGCOMM (2010). [31] QIAN, J., N, M. L., AND GERSTEIN, M. Protein family and fold occurrence in genomes: power-law behavior and evolutionary model. Journal of Molecular Biology 313, 4 (2001), 673 – 681. [32] REED, W. J., AND JORGENSEN, M. The double pareto-lognormal distribution - a new parametric model for size distribution. Com. Stats – Theory & Methods 33, 8 (2004), 1733–1753. [33] SALA, A., CAO, L., WILSON, C., ZHENG, H., AND ZHAO, B. Y. Measurement-calibrated graph models for social network experiments. In Proc. of WWW (Raleigh, NC, April 2010). 14
1502.06775	2	1502	2015-06-09T11:32:26	Limitations in the spectral method for graph partitioning: detectability threshold and localization of eigenvectors	[ "cs.SI", "cond-mat.stat-mech", "physics.soc-ph" ]	Investigating the performance of different methods is a fundamental problem in graph partitioning. In this paper, we estimate the so-called detectability threshold for the spectral method with both unnormalized and normalized Laplacians in sparse graphs. The detectability threshold is the critical point at which the result of the spectral method is completely uncorrelated to the planted partition. We also analyze whether the localization of eigenvectors affects the partitioning performance in the detectable region. We use the replica method, which is often used in the field of spin-glass theory, and focus on the case of bisection. We show that the gap between the estimated threshold for the spectral method and the threshold obtained from Bayesian inference is considerable in sparse graphs, even without eigenvector localization. This gap closes in a dense limit.	cs.SI	cs	Limitations in the spectral method for graph partitioning: Detectability threshold and localization of eigenvectors Tatsuro Kawamoto1 and Yoshiyuki Kabashima1 1Department of Computational Intelligence and Systems Science, Tokyo Institute of Technology, 4259-G5-22, Nagatsuta-cho, Midori-ku, Yokohama, Kanagawa 226-8502, Japan (Dated: August 22, 2018) Investigating the performance of different methods is a fundamental problem in graph partitioning. In this paper, we estimate the so-called detectability threshold for the spectral method with both unnormalized and normalized Laplacians in sparse graphs. The detectability threshold is the critical point at which the result of the spectral method is completely uncorrelated to the planted partition. We also analyze whether the localization of eigenvectors affects the partitioning performance in the detectable region. We use the replica method, which is often used in the field of spin-glass theory, and focus on the case of bisection. We show that the gap between the estimated threshold for the spectral method and the threshold obtained from Bayesian inference is considerable in sparse graphs, even without eigenvector localization. This gap closes in a dense limit. I. INTRODUCTION Over recent decades, significant attention has been paid to the clustering problem on graphs or networks [1]. Although clustering is sometimes considered as simply an optimization problem, e.g., finding the most efficient par- titioning in parallel computing, it can also be used to find physically meaningful modules, or communities, among which each vertex or edge shares a common attribute. The latter technique is usually called community detection, and is often formulated as a discrete optimization problem. Many methods and algorithms have been developed for clustering graphs [1], and their performance has been investigated both theoretically [2 -- 15] and experimentally [16 -- 23]. One important theoretical problem concerns the so-called detectability threshold [5, 8]. Suppose that we apply a clustering method to a set of random graphs produced by a generative model with a planted block structure, and examine the degree of correlation between the planted partition and the partition obtained by the method. The standard generative model is called the stochastic block model or planted partition model. As the connection between modules gets stronger, i.e., as the block structure gets weaker, the partition given by the clustering method will be less correlated with the planted partition. The detectability threshold is the critical point of a parameter that indicates the strength of the block structure. Below this threshold, the partition obtained by the method is completely uncorrelated to the planted partition. This transition is especially problematic in sparse graphs. For dense graphs, it has been shown that such a transition does not exist and we are always able infer the planted structure in the large size limit [24 -- 26]; as we increase the total number of nodes, the average degree also increases in dense graphs, and therefore we obtain more information about the module assignment of each node. Unfortunately, this is not the case for sparse graphs because the average degree remains constant independent of the graph size. Needless to say, the detectability threshold in sparse graphs is a significant problem, as many real networks are sparse. Note that, although we specify the algorithm, i.e., the spectral method, the existence of the detectability threshold itself is not an algorithmic problem, but the theoretical limit of an objective function. Although there are many levels of graph clustering, e.g., hierarchical clustering and clustering with overlaps, we focus on undirected graphs without hierarchical structures and consider the graph partitioning problem, i.e., partitioning of a graph into non overlapping modules. In addition, we focus on the case of bisection in the large-size limit. As mentioned above, the planted partition is given in the analysis of the detectability threshold. In practice, however, not only do we not know the model parameters a priori (such as the module sizes and the fraction of edges between modules), we are not even sure whether it is appropriate to assume that the graph was generated by the same mechanism as the stochastic block model in the first place. Therefore, knowing the detectability of a method does not readily contribute to its practical use. It does, however, provide an important clue for categorizing the detection method: It reveals the similarity of outcomes between methods through the stochastic block model. Thus, for this purpose, the stochastic block model should be regarded as a model that gives a measure for comparison. For example, although block structures exist in many senses [1], applying all the existing methods to the given data obviously entails a huge cost and is also redundant. If we know which methods tend to give a similar partitioning, we could reduce the cost significantly. Whereas performance comparison studies are often experimental, it is promising that theoretical analyses can provide deeper insights. In this paper, we discuss the detectability threshold of the spectral method. The spectral method can be employed with three major discrete objective functions: ratio-cut (RatioCut), normalized-cuts (Ncut), and modularity. The 2 spectral method solves the continuous relaxation of these objective functions as an eigenvalue/eigenvector problem of the corresponding matrix. The unnormalized Laplacian L corresponds to RatioCut, the normalized Laplacian L corresponds to Ncut, and the so-called modularity matrix B corresponds to modularity. Nadakuditi and Newman [8] predicted that the detectability threshold of the spectral method with the modularity function coincides with the threshold given by Bayesian inference [5, 6]. However, this assumes that the average degree c is sufficiently high. It was argued in [27] that the estimate in [8] may not be precise for sparse graphs and that there exists a gap between the spectral method with modularity and Bayesian inference; indeed, it was confirmed numerically that the spectral method does not detect the planted partition all the way down to the detectability threshold. It was later discovered that the spectral method with a non-backtracking matrix [28] was a promising means of filling this gap. This approach provides a formalism that avoids the emergence of localized eigenvectors, known to be a drawback of the spectral method. A localized eigenvector is one in which the weight of its elements is concentrated on a few characteristic vertices. Once the eigenvector used for partitioning becomes localized, the information of the block structure will be washed out. Thus, preventing this effect will enhance our ability to detect the planted structure. To the best of our knowledge, however, the true detectability threshold for the spectral method in sparse graphs remains unknown. Therefore, we do not know to what extent the gap actually exists. Furthermore, although the non-backtracking matrix approach improves the detectability by avoiding the eigenvector localization problem, the logical and quantitative connection to detectability seems to be incomplete. That is, it is not known whether the gap appears because of the localization or if it exists even when the localization is absent. When both are present, evaluating the relative degree of their effects is an important problem. Our results show that, in sparse graphs, a considerable gap exists even when localization is absent and that the effect of localization may be significant when the degree fluctuates considerably. Using the so-called replica method, which is often used in spin-glass theory, we derive estimates for the detectability threshold of the spectral method with both unnormalized and normalized Laplacians. Note that, as pointed out in [11], the spectral method with the normalized Laplacian L and the modularity matrix B are equivalent for the bisection problem, as long as the continuous relaxation gives a good estimate of the original discrete problem. We compare our analytical estimates with the results of numerical experiments for the two-block random graphs with uniform, bimodal, and Poisson degree distributions. Although our estimates contain some approximations, they agree quite well with the results of the numerical experiment, as long as localization does not occur. For the analysis of localized eigenvectors, we show that our estimate is fairly accurate for graphs with bimodal degree distributions. The rest of this paper is organized as follows. We first introduce a more precise definition of the stochastic block model (Sec. II) and the spectral method in graph partitioning (Sec. III). Then we derive an estimate of the detectability threshold in two-block random regular graphs in Sec. IV. Note that there is no distinction between unnormalized and normalized Laplacians in this case. We analyze the effect of degree fluctuation for the unnormalized Laplacian L in Secs. V and VI. In Sec. V we present a formal solution for estimating the detectability threshold and analyze the case of graphs with bimodal degree distributions, and in Sec. VI we estimate the localized eigenvector and its eigenvalue for a graph with a bimodal degree distribution. A similar analysis is done for the normalized Laplacian L in Secs. VII and VIII. For the normalized Laplacian L of graphs with an arbitrary degree distribution, the resulting estimate of the detectability threshold resembles that of random regular graphs. Finally, we discuss the case of stochastic block models, i.e., two-block random graphs with Poisson degree distributions, with the normalized Laplacian L in Sec. IX. We summarize our results in Sec. X. II. STOCHASTIC BLOCK MODEL AND THE DETECTABILITY THRESHOLD The stochastic block model [29] is a generative model of random graphs with a block structure, and is commonly used for analyzing the performance of clustering methods. While many variants have been proposed [30 -- 35], the model is fundamentally a generalization of the Erdos-R´enyi random graph. In the stochastic block model, the number of modules q, size of each module, and probability prs that vertices in modules r and s are connected are specified as inputs. With these parameters, the graphs are constructed as follows. Each vertex has a preassigned module index σi = r (r ∈ {1, . . . , q}) to which the vertex belongs. Based on this block structure, edges are generated between pairs of vertices at random, i.e., vertices i ∈ r and j ∈ s are connected with probability prs. In the case of sparse graphs of N vertices, we set prs = crs/N , where crs remains constant in the limit N → ∞. To construct an assortative block structure, we typically choose prr = cin/N and prs = cout/N for r 6= s, where cin and cout are constants that satisfy cin > cout. In the case of two modules of equal size, we have the average degree c = (cin + cout)/2. Note that, in this case, unlike the dense case, the fluctuation of the degree of each vertex does not vanish in the limit N → ∞. According to [5, 6], the Bayesian inference method has a detectability threshold at cin − cout = 2√c. 3 (1) That is, even when the generative model has the assortative property cin > cout, it is impossible to retrieve that information with any detection algorithm, unless the difference cin − cout is greater than 2√c (for other detectability analyses with sufficiently large average degree, see [12, 13, 36, 37]). In the present paper, we parametrize the stochastic block model differently. As mentioned above, we restrict ourselves to the two-block model. Instead of setting cin and cout as the model parameters, we set the average degree c and the average number of edges γ from one module to the other, or, in other words, the total number of edges γN between two modules. The parameters γ controls the strength of block structure; the larger is the value of γ, the weaker is the block structure. It is related to cout and cin by 2 (cid:19)2 γN =(cid:18) N 2 − γN = 2(cid:18) N 2(cid:19) = 2 cN pout = N 4 cout, N 4 cin, (2) (3) in the limit N → ∞. Although the total number of edges between modules fluctuates in the standard formulation when the graphs are finite, we let every realization have exactly γN and cN/2− γN for the number of edges between modules and the total number of edges within modules, respectively. This is called the "microcanonical" formulation of the stochastic block model [34]. III. SPECTRAL METHOD IN GRAPH BISECTION Graph partitioning is often formulated as a discrete optimization problem for some objective function that is computationally difficult. The spectral method constitutes a continuous relaxation of the original problem using eigenvectors of a proper matrix. The unnormalized Laplacian L, which is used for RatioCut, and the normalized Laplacian L, which is used for Ncut, are defined as L = D − A, L = D−1/2LD−1/2. (4) (5) The matrix A is the adjacency matrix, i.e., Aij = 1 if vertices i and j are connected and Aij = 0 otherwise. The matrix D is a diagonal matrix with degree ci of vertex i on the diagonal element, i.e., Dij = ciδij . Although the details differ depending on the objective function, the basic procedure of spectral bisection is the same and quite simple [11, 38]. We denote the total degree of the graph as K. For a graph with a set of vertices V partitioned into V1 and V2, the objective function of RatioCut is defined as fRatioCut(V1, V2) = E(V1, V2) N1N2 , (6) where the cut size E(V1, V2) is the number of edges between modules V1 and V2, and we denote by N1 and N2 the number of vertices in each module. Similarly, the objective function of Ncut is defined as fNcut(V1, V2) = E(V1, V2) K1K2 , (7) where K1 and K2 are the total degrees of each modules. Using the unnormalized Laplacian L and the normalized Laplacian L, minimizing RatioCut and Ncut are equivalent to and TLx, x min x min x x TLx, subject to xi =(pN2/N1 −pN1/N2 subject to xi =(pK2/K1 −pK1/K2 i ∈ V1, i ∈ V2, i ∈ V1, i ∈ V2. (8) (9) 4 FIG. 1: (Color online) A realization of two-block 4-random regular graphs. Finally, allowing xi to take an arbitrary real number, we obtain the relaxed versions of the above discrete optimization problems. For RatioCut, and for Ncut, min x∈RN x TLx subject to x ⊥ 1, x T x = N, min x∈RN x TLx subject to x ⊥ D1/21, x T x = K. (10) (11) TLx and x TLx are achieved when the x are the eigenvectors corresponding to the smallest The smallest values of x eigenvalue of L and L, respectively. Note, however, that L is positive semi-definite and 1 is the eigenvector of L corresponding to the zero eigenvalue (i.e., the smallest eigenvalue). Because of the constraint that x must be TLx in Eq. (10) is achieved when we select the eigenvector corresponding perpendicular to 1, the smallest value of x to the second-smallest eigenvalue of L. Similarly, L is also positive semi-definite, and D1/21 is the eigenvector corresponding to the zero eigenvalue. Hence, the eigenvector corresponding to the second-smallest eigenvalue of L TLx in Eq. (11). As the sign of xi indicates which module vertex i belongs to in Eqs. (8) gives the smallest value of x and (9), we retrieve information about the optimal partition from the solutions of the relaxed problems by referring to the sign of each element in the eigenvector, i.e., vertices with the same sign belong to the same module. It is known that this prescription works well when the module sizes are not very different. Of course, it is not obvious whether the optimal partition of the relaxed problem coincides with that of the unrelaxed problem. However, this is beyond the scope of the present paper, and we concentrate on the relaxed problem, i.e., the spectral method. In the following sections, we analyze how the optimal partitions in the spectral method are correlated to the planted partitions in various random graphs. IV. DETECTABILITY THRESHOLD IN RANDOM REGULAR GRAPHS In this section, we analyze the detectability of the c-random regular graph with a two-block structure. A realization of such graphs is shown in Fig. 1. As c is a constant and does not increase as a function of the graph size N , this is regarded as a sparse random graph that has the property of dense graphs that the degree fluctuation is negligible. This is worth investigating, because the results we show in Sec. VII for the normalized Laplacian L are analogous to those we obtain for random regular graphs in this section. Although we analyze the unnormalized Laplacian L here, as the degree is the same for every vertex, there is no distinction between the unnormalized and normalized Laplacians. We calculate the average of the second-smallest eigenvalue of the unnormalized Laplacian [λ2]L, where [. . . ]L represents the average with respect to L of each realization of a random graph. As we increase the fraction γ of edges between modules, the value of [λ2]L increases until it reaches the edge of the spectral band, above which [λ2]L becomes constant irrespective of γ. As we show in the following, by calculating [λ2]L, we can obtain the distribution of the elements in the corresponding eigenvector. The basic methodology here runs parallel to that in [39]. To calculate the second-smallest eigenvalue, we first introduce the Hamiltonian H(xL), partition function Z(βL), and free energy f (βL), 1 2 TLx, x H(xL) = Z(βL) =Z dx e−βH(xL)δ(x2 − N )δ(1T f (βL) = − ln Z(βL), 1 N β x), 5 (12) (13) (14) where x is an N -dimensional vector, 1 is the vector in which each element is 1, and T denotes the transpose. The Hamiltonian corresponds to the objective function to be minimized in Eq. (10); the factor 1/2 in Eq. (12) is purely conventional. The δ functions in Eq. (13) impose the constraints in Eq. (10). The crucial aspect of this formulation is that, in the limit β → ∞, in conjunction with the operation of δ(1T x), the contribution from the second-smallest eigenvalue is dominant in the integral in Eq. (13). Thus, Eq. (13) actually evaluates exp [−N βλ2/2]. Hence, the second-smallest eigenvalue λ2 is given by We then take the average over all realizations of random graphs. However, the direct calculation of this average is not tractable. Therefore, we recast [λ2]L as λ2 = 2 lim β→∞ f (βL). (15) [λ2]L = −2 lim β→∞ 1 N β = −2 lim β→∞ lim n→0 [ln Z(βL)]L 1 N β ∂ ∂n ln [Z n(βL)]L . (16) The assessment of [Z n(βL)]L is also difficult for a general real number n. However, when n is a positive integer, n, 2, . . . , x [Z n(βL)]L can be expressed as a high-dimensional integral with respect to n replicated variables x which can be analytically evaluated by the saddle-point method as N tends to infinity. In addition, the resulting expression of N −1 ln[Z n(βL)]L is shown to be a function of n that can be extended to real values of n under a certain ansatz concerning the permutation symmetry among the replica indices a = 1, 2, . . . , n. Therefore, we employ such an expression to calculate the right-hand side of Eq. (16). This procedure is often termed the replica method. Although the mathematical validity of the replica method has not yet been proved, we see that our assessment based on the simplest permutation symmetry for the replica indices offers a fairly accurate prediction of the experimental results. 1, x For n ∈ N, we can write [Z n(βL)]L as [Z n(βL)]L =Z n Ya=1 and the exponential factor is given by dxaδ(xa2 − N )δ(1T xa)!"exp − x T a Lxa!#L , β 2 Xa (17) (18) exp − β 2 Xa = exp − x T a Lxa!   Xij∈V1 Xa=1 n β 4 uij(xia − xja)2 + Xij∈V2 uij(xia − xja)2 + 2Xi∈V1 Xj∈V2 wij(xia − xja)2    , where uij = {Aiji ∈ r, j ∈ r} and wij = {Aiji ∈ V1, j ∈ V2}. For the average over the random graphs, we assume that each realization occurs with equal probability. The condition of being a regular graph requires Xl∈V1 Xl∈V2 uil + Xk∈V2 ujl + Xk∈V1 wik = c (for i ∈ V1), wjk = c (for j ∈ V2), and the number of edges between modules is wik = γN. Xi∈V1 Xk∈V2 (19) (20) (21) Therefore, we have "exp − = T x 1 β a Lxa!#L 2 Xa NG X{uij }{wij }(Yi∈V1 δ Xl∈V1  × exp  Xij∈V1 Xa=1 uij(xia − xja)2 + Xij∈V2 − uil + Xk∈V2 β 4 n wik − c! Yj∈V2 δ Xl∈V2 ujl + Xk∈V1 uij(xia − xja)2 + 2Xi∈V1 Xj∈V2 wjk − c! δ Xi∈V1 Xk∈V2 wik − γN! wij (xia − xja)2  ),  where NG is the number of random regular graphs with two modules (see Appendix A for the count of NG in random regular graphs). With the technique in Appendix A, Eq. (22) can be written as 6 (22) "exp − β 2 Xa x T a Lxa!#L = 1 β 2π z−(1+c) j dzi 2π dzj 2π z−(1+c) i I Yj∈V2 NG I Yi∈V1 × Yi<j∈V1 Xuij ={0,1} zizj exp"− 2 Xa × Yi<j∈V2 Xuij ={0,1} zizj exp"− 2 Xa × Yi∈V1 Yj∈V2 Xwij ={0,1} zizj exp"−η − eηγN Z dη (xia − xja)2#!uij (xia − xja)2#!uij 2 Xa β β (xia − xja)2#!wij When N ≫ 1, an element in Eq. (23) can be approximated as β Yi<j∈Vr Xuij ={0,1} zizj exp"− 2 Xa zizj exp − 2 Xa Z a Qr(µ ≈ exp  Xi<j∈Vr ≈ exp" (p1N )2 (xia − xja)2#!uij (xia − xja)2!  (r)) exp − (r))Qr(ν a dν(r) dµ(r) β 2 n Ya=1 (µ(r) a − ν(r) a )2!# , β 2 Xa where we have introduced the order parameter functions for r = 1, 2. Then, inserting the identity (r)) = Qr(µ 1 prN Xi∈Vr ziYa δ(xia − µ(r) a ), (r))δ 1 1 =Z dQr(µ = prNZ dQr(µ prN Xi∈Vr (r))d Qr(µ 2π ziYa δ(xia − µ(r) exp" Qr(µ (r)) (r)) Xi∈Vr a ) − Qr(µ (r))! ziYa δ(xia − µ(r) a ) − prNQr(µ (r))!# for each µ (r) and replacing the δ functions as x2 δ N ia − N! =Z βdφa Xi=1 δ N xia! =Z βdψa Xi=1 4π 4π e− β 2 φa(Pi x2 ia−N ), e− β 2 ψa(Pi xia), . (23) (24) (25) (26) (27) (28) we can recast Eq. (17) as [Z n(βL)]L = p1p2N 2Z Yr=1,2 dQr(µ (r))d Qr(µ 2π × exp"N 2KI(Qr, Qr) + N(cid:18) β (r)) n βdφa 4π Z Ya φa − Xr=1,2 2π βdψa 4π Z dη KIIr(Qr, Qr) + 1 N Xr=1,2 ln KIII,r( Qr,{φa},{ψa}) + ηγ − 1 N where 2 Xa ln NG − ln c!(cid:19)#, prps (r)dν (s) Qr(µ 2 Z dµ KI(Qr, Qr) = Xr,s=1,2 KIIr(Qr, Qr) = prZ dµ (r) Qr(µ (r))Qr(µ dxia Yi∈Vr Qc KIIIr( Qr,{φa},{ψa}) =Z Yi∈Vr Ya=1 (r)), n (r))Qs(ν (s))e−(1−δrs)η− β 2 Pa(µ(r) a −ν(s) a )2 , r(xi) exp"− β 2 Xa (cid:0)φax2 ia + ψaxia(cid:1)#! . Now we evaluate limN→∞(ln[Z n(βL)]L)/N with the saddle-point method and calculate the second-smallest eigen- value according to Eq. (16). For this, we assume that the functional forms of Qr(µ) and Qr(µ) are invariant under any permutations of replica indices a ∈ {1, 2, . . . , n}, which is often termed the replica-symmetric ansatz. Further, the Gaussian nature of the current problem allows us to assume that Qr(µ) and Qr(µ) are mixtures of Gaussian functions; this originates from the fact that the effective Hamiltonian yielded by appropriate exponentiations of the δ functions is composed of quadratic forms. These restrict the functional forms of Qr(µ) and Qr(µ) as 2π(cid:19) Qr(µ) = TrZ dAdH qr(A, H)(cid:18) βA Qr(µ) = TrZ d Ad H qr( A, H) exp" β 2 n 2 n exp"− Xa=1(cid:16) Aµ2 n βA 2 Xa=1(cid:18)µa − a + 2 Hµa(cid:17)# , H A(cid:19)2# , (33) (34) i.e., some superpositions of Gaussian functions with weights qr(A, H) and qr( A, H), where A and A denote the variances and H and H denote the means of each Gaussian distribution, respectively. When n = 0, the order parameters Qr(µ) and Qr(µ) coincide with Eqs. (A9) and (A10) in Appendix A and thus the normalization factors Tr and Tr are equal to Eqs. (A19) and (A20). With Eqs. (33) and (34), (30) -- (32) become functions of n that are extendable to real values of n. Inserting these expressions into the identity N −1 [ln Z(βL)]L = limn→0(∂/∂n)N −1 ln [Z n(βL)]L, we have 7 (29) (30) (31) (32) (35) (36) (37) [λ2]L = − extr cp1p2 p2 qr,qr ,φ,ψ(Z dAdHZ dA′dH ′ Ξ(A, H, A′, H ′) 2 (cid:18)(cid:18) p1 + Γ(cid:19) q1(A, H)q1(A′, H ′) +(cid:18) p2 × prZ dAdHZ d Ad H qr(A, H)qr( A, H) (H + H)2 + φ − c Xr=1,2 A − A − Hg(cid:17)2 Yg=1(cid:16)d Agd Hg qr( Ag, Hg)(cid:17)(cid:16)ψ/2 −Pg Ag ), prZ + Xr=1,2 φ −Pg p1 c where we set + Γ(cid:19) q2(A, H)q2(A′, H ′) + 2 (1 − Γ) q1(A, H)q2(A′, H ′)(cid:19) H 2 A ! Γ = 1 − γ cp1p2 , Ξ(A, H, A′, H ′) = (1 + A′)H 2 + (1 + A)H ′2 + 2HH ′ (1 + A)(1 + A′) − 1 H 2 A − H ′2 A′ . − In the above calculation, we have assumed the replica symmetry for φa and ψa, i.e., φa = φ and ψa = ψ, respectively, for a = 1, 2, . . . , n. The saddle-point conditions yield the following set of integral equations: 8 q1( A, H) =Z dA′dH ′ (cid:20)(cid:18)1 + q2( A, H) =Z dA′dH ′ (cid:20)(cid:18)1 + p2 p1 p1 p2 Γ(cid:19) p1q1(A′, H ′) + (1 − Γ) p2q2(A′, H ′)(cid:21) δ(cid:18) A + Γ(cid:19) p2q2(A′, H ′) + (1 − Γ) p1q1(A′, H ′)(cid:21) δ(cid:18) A + and A′ 1 + A′(cid:19) δ(cid:18) H − 1 + A′(cid:19) δ(cid:18) H − A′ H ′ 1 + A′(cid:19) , 1 + A′(cid:19) , H ′ qr(A, H) =Z c−1 Yg=1(cid:16)d Agd Hg qr( Ag, Hg)(cid:17) δ H + ψ/2 − c−1 Xg=1 Hg! δ A − φ + Ag! . c−1 Xg=1 (38) (39) (40) Note that we set c > 2 here. To obtain nontrivial random regular graphs, the degree c of each vertex needs to be greater than 2. Furthermore, the saddle-point conditions with respect to the auxiliary parameters ψ and φ give = 0, H A prZ dAdH Qr(A, H) A(cid:19)2 prZ dAdH Qr(A, H)(cid:18) H Xr Xr = 1, respectively, where we have defined Qr(A, H) =Z c Yg=1(cid:16)d Agd Hg qr( Ag, Hg)(cid:17) δ H + ψ/2 − c Xg=1 Hg! δ A − φ + Ag! , c Xg=1 (41) (42) (43) which corresponds to the complete marginal of Eq. (40). The distribution of A and H in module r, Qr(A, H), can be obtained by iteratively updating the saddle-point equations (38) -- (40), while keeping the constraints (41) and (42). i /N = 1 (see Eq. (27)). Then, Eq. (42) should indicate the same restriction, as it incorporates the average over the random graphs at the same time. Hence, Eq. (43) indicates δ(x− xi)/Nr of the elements of the second-smallest eigenvector of each module. As shown in Fig. 2, the distributions obtained by iterating the saddle-point equations agree very well with the numerical experiments. Recall that φ leads to the normalization condition Pi x2 that the distribution of H/A gives the distribution Pr(x) =Pi∈Vr Eqs. (33) and (34) have the same values a and a, i.e., q(A) = R dHq(A, H) = δ(a − A) and q( A) = R d H q( A, H) = δ(a − A). Then, Eq. (35) becomes To obtain an analytical expression, we further restrict the form of the solution. We assume that the variances in [λ2]L = −extr( 11(cid:3) + c p2(cid:2)(1 + a)m22 + m2 12(cid:3) − γ (m11 − m12)2(cid:19) 1 1 − a(a + 2)(cid:18)c p1(cid:2)(1 + a)m21 + m2 a − a Xr=1,2 φ − c a Xr=1,2 pr(cid:18) ψ2 + 1 c pr (m2r + 2m1r m1r + m2r) 4 − c ψ m1r + c m2r + c(c − 1) m2 1r(cid:19) + φ), where we have denoted the moments of H and H as mnr =R dHH nqr(H) and mnr =R d H H n qr( H). Equation (44) has solutions with m1r 6= 0 and m1r = 0. For the solution with m1r 6= 0, after taking the saddle point, we have 1 + a = 1 1 + a , a = φ − (c − 1)a, cp1 (cid:18)1 − m2 11 = p2 1 (c − 1)2Γ2(cid:19)(cid:0)(c − 1)Γ2 − 1(cid:1) , [λ2]L = (1 − Γ)(cid:18)c − 1 − 1 Γ(cid:19) . (44) (45) (46) (47) (48) (a) (b) 9 FIG. 2: (Color online) Distributions of the elements of the second-smallest eigenvector of the two-block random regular graphs. The dots in each plot represent the numerical results with N = 104 vertices, in which the average is over 100 samples. The solid lines in each plot represent the results from the saddle-point equations (38) -- (40) with parameters (a) c = 3, p1 = 0.7, p2 = 0.3, γ = 0.1, and (b) c = 4, p1 = 0.6, p2 = 0.4, γ = 0.1. The requirement that m2 11 ≥ 0 implies that the above solution is valid for or, in terms of γ, 1 √c − 1 ≤ Γ, γ ≤ cf (c)p1p2, (cid:18)f (c) = 1 − 1 √c − 1(cid:19) . (49) (50) The point at which the equality holds in Eq. (50) is the detectability threshold of the random regular graphs. Above this point, we have the solution with m1k = 0. When m1k = 0, we have [λ2]L = −φ = c − 2√c − 1, (51) which matches Eq. (48) at the boundary of (50). In both cases, we have ψ = 0, which comes from the symmetry property in which the problem is invariant under conversion from x to −x. Equations (48) and (51) are plotted in Fig. 3(a), together with the results of the numerical experiments. Again, the agreement is quite good. To compare our results with those reported in the literature, we recast Eq. (49) in terms of cin = pinN and cout = poutN , and set the module sizes to p1 = p2 = 0.5. First, we recast Eq. (50) in terms of the total degree within a module Kin and the total degree from one module to the other Kout. With these quantities, we have Kout = γN and K = cN = 2(Kin + Kout) for the total degree of the whole graph K. Then, Eq. (50) reads In the limit N → ∞, Kin(= cN/2 − γN ) and Kout(= γN ) are related to cin and cout as Eqs. (2) and (3), respectively. Therefore, the detectable region (52) is Kin − Kout ≥ N 2 c √c − 1 . (52) cin − cout ≥ 2 c √c − 1 . (53) In the dense limit c → ∞, Eq. (53) converges to the result in [8]. The implication of Eq. (53) is that, even if the effect of localization is absent, the spectral method for sparse graphs never reaches the ultimate limit 2√c [5, 7, 8]. Although the distribution Pr(x) is not of Gaussian form, even for the case of equal size modules [see Figs. 4(a) and 4(b) for the apparent cases], it is expected to be somewhat close to the Gaussian distribution, especially when (a) (b) 10 FIG. 3: (Color online) Average second-smallest eigenvalues of the unnormalized Laplacian L of the random regular graphs as a function of (a) γ and (b) cin − cout. The solid lines represent the analytical solutions [(48) and (51)] and the dots represent the numerical results. The numerical experiments used N = 104 vertices, and each dot represents the average over 100 samples. the peaks are not well-separated. Estimating the mean m(x) and variance s2(x) from the replica analysis above, we can calculate the mean value of the fraction of correctly classified vertices as [1 + erf(m(x)/p2s2(x))]/2, under the Gaussian approximation (see Appendix B for details). Figure 5 illustrates this estimate, together with the results of numerical experiments and the populations obtained by iterating the saddle-point equations. The Gaussian fitting appears to give a fairly good approximation, especially around the detectability threshold. We can also observe some convergence of the numerical results to our estimate, which is for the infinite-size limit. Note that there exists a finite fraction of misclassified vertices, even in the limit N → ∞. Although the poor performance of optimization algorithms may have crucial effects in practice, misclassification occurs in principle because the planted partition is not necessarily the partition that optimizes the objective function. Unless the block structure of the planted partition is sufficiently strong, the random graph is likely to have a partition that is better, in the sense of the objective function, than the planted partition. We now discuss the behavior of the detectability threshold (50). We consider γ/c, the fraction of connections between modules normalized by the degree. The upper bound of this quantity is min{p1, p2}, which is achieved when all edges in a smaller module are connected to the other module, i.e., the case where the small module has a bipartite structure. For a uniform random graph, i.e., a graph with no block structures, the expected value of γ/c is p1p2. This is because, for each stub or half-edge in the first module in cN1, the probability of being connected to the second module is N2/N . Therefore, we have γN = cN1 × N2/N . This can also be obtained as the value of γ with η = 0 in (A15). Note that the parameter 1 − Γ is the ratio of γ to cp1p2. Because 0 < f (c) < 1, the detectability threshold γc/c falls somewhere in the range (0, p1p2) (see Fig. 6). Let us consider the value of γ/c such that the total degree within a module is greater than the number of edges between modules. As the number of edges within a smaller module is cN min{p1, p2} − γN , we have γwd c < 1 2 min{p1, p2}. (54) This corresponds to the region where the weak definition of a community [40] is satisfied with respect to the smaller module. While the detectability threshold (50) is always in the region stated in (54) for equal-size modules, this may not be the case for unequal-size modules, because the larger module may possess a strong block structure even if the smaller one does not. Such a situation is achieved when 2f (c) max{p1, p2} > 1. In other words, because f (5) = 1/2, the spectral method always loses all information about the planted solution when the weak definition is not satisfied in any module size for c ≤ 5. Note that this is for bisection using the sign of the eigenvector with the second-smallest eigenvalue. Although this is the standard approach, as we mentioned at the end of Sec. III, its performance is not reliable in practice when the module sizes are very different. (a) (b) 11 FIG. 4: (Color online) Gaussian fitting of the distributions of elements for the second-smallest eigenvector of the two-block 3-random regular graphs. The values of γ are (a) 0.1 and (b) 0.05, and the equal module sizes are considered in both plots. The dots in each plot represent the results of the numerical experiments with N = 104 vertices, in which the average is over 100 samples. The solid lines in each plot represent the Gaussian distributions with the same mean and variance as the values in the numerical experiments. FIG. 5: (Color online) Estimate of the fraction of correctly classified vertices in the two-block 3-random regular graphs with equal module sizes and the numerical results. The crosses represent the estimates using the replica method, and the solid line represents their Gaussian approximation. The numerical experiments were conducted for various graph sizes. V. DETECTABILITY THRESHOLD IN RANDOM GRAPHS WITH DEGREE FLUCTUATION: THE UNNORMALIZED LAPLACIAN We now analyze the case where the degree fluctuates in the unnormalized Laplacian L. Within Nr = prN vertices, t=1 bt = 1) vertices have degree ct. In the limit N → ∞, the we consider a graph in which btNr (t ∈ {1, 2, . . . , T}, PT 12 weak definition 0 0.5 FIG. 6: (Color online) Parameter region of the detectability threshold for random regular graphs, f (c)p1p2, in the space of γ/c. This is in the range (0, p1p2), whereas the region in which the weak definition of a community is satisfied is below min{p1, p2}/2. replica-symmetric solution of the second eigenvalue is given by [λ2]L = − extr 1 qr,qr ,φ,ψ(Z dAdHZ dA′dH ′ Ξ(A, H, A′, H ′) 2(cid:18)(c p1 − γ)q1(A, H)q1(A′, H ′) + (c p2 − γ)q2(A, H)q2(A′, H ′) + 2γq1(A, H)q2(A′, H ′)(cid:19) × prZ dAdHZ d Ad H qr(A, H)qr( A, H) (H + H)2 + φ − c Xr=1,2 A − A − Hg(cid:17)2 Ag ). btZ + Xr=1,2 prXt Yg=1(cid:16)d Agd Hg qr( Ag, Hg)(cid:17)(cid:16)ψ/2 −Pg φ −Pg A ! H 2 ct The saddle-point equations derived from (55) are q1( A, H) =Z dA′dH ′ (cid:20)(cid:18)1 − q2( A, H) =Z dA′dH ′ (cid:20)(cid:18)1 − and γ c p1(cid:19) q1(A′, H ′) + c p2(cid:19) q2(A′, H ′) + γ γ c p1 γ c p2 q2(A′, H ′)(cid:21) δ(cid:18) A + q1(A′, H ′)(cid:21) δ(cid:18) A + A′ 1 + A′(cid:19) δ(cid:18) H − 1 + A′(cid:19) δ(cid:18) H − A′ H ′ 1 + A′(cid:19) , 1 + A′(cid:19) , H ′ qr(A, H) =Xt btct c Z ct−1 Yg=1 (cid:16)d Agd Hg qr( Ag, Hg)(cid:17) δ H + ψ 2 − ct−1 Xg=1 Hg! δ A − φ + Ag! . ct−1 Xg=1 (55) (56) (57) (58) Compared with the saddle-point equations for the random regular graphs, we now have a step in (58) to draw a degree ct from the excess degree distribution btct/c at every iteration. Note that, in the case of the random regular graphs, we had a set of solutions qr(A, H) and qr(A, H) of the saddle-point equations (38) -- (40), with A and A fixed at certain values a and a. Although these give exact solutions with random regular graphs, functions of this form cannot solve Eqs. (56) -- (58) because of the degree fluctuation. Although the saddle-point equations (56) -- (58) no longer have solutions of simple form, it is still important to obtain an analytical expression. For this purpose, we again employ qr(A) and qr( A) of the form qr(A) = δ(a − A) and qr(a − A), respectively. This is called the effective medium approximation (EMA) [39, 41]. We can then recast Eq. (55) as [λ2]L = −extr(φ + 1 a(a + 2) cXr − c Xk + Xr=1,2 pr a − a prXt bt pr(cid:2)(1 + a)m2r + m2 1r(cid:3) − γ(m11 − m12)2! (m2r + 2m1r m1r + m2r) φ − cta(cid:18) ψ2 4 + ct( m2r − ψ m1r) + ct(ct − 1) m2 1r(cid:19)). (59) Introducing Rn =Xt Sn =Xt , btcn t φ − cta btcn t (φ − cta)2 , and provided that a and φ are obtained at the saddle point, we have m2 11 = [λ2]L = where we set X1 − S2 (S2 − S1)X1X2 + (S1S3 − S2 R2 − R1 p1 (1 + a)2 Γ p2 a(a + 2) Γ − m2 c 2 )X2 − S1X3 2(cid:21) − φ, 11(cid:20) γ , cp1p2 (a2 + 2a + 2), Γ = 1 − R2 1 c p1 X1 = X2 = We have m11 = 0 when S2 ≥ X1, i.e., X3 = 2 p2 (cid:18) Γ 1 + a(cid:19) R2 1 c p1 p2 Γ 1 + a 2 , . (1 + a)2 ≤ cS2 1 − 1. R2 , 13 (60) (61) (62) (63) (64) (65) (66) (67) (68) Equation (63) gives the formal solution for the second-smallest eigenvalue, and the equality condition in (68) is our estimate of the detectability threshold with the EMA. When the graph has a bimodal distribution, i.e., {b1, b2} for c1 and c2, we can solve for a and φ at the saddle point analytically. In this case, the saddle-point conditions give φ = c1c2a 1 + a (cid:18) 1 + a Γ + 1 − c(cid:19)(cid:20)c2 − c(cid:18)1 + 1 + a Γ (cid:19)(cid:21)−1 , where we have defined c2 = b1c2 1 + b2c2 2, and a is the solution of the following quadratic equation: Γ (cid:19)2 c(cid:18) 1 + a +(cid:18)c − c2 + c1c2 Γ − 1(cid:19) 1 + a Γ + c1c2(1 − c) Γ − 1 = 0. (69) (70) We take the smaller value for the solution of Eq. (70), which gives a nonnegative value for m2 11. Figures 7(a) and 7(b) show the eigenvalues of the unnormalized Laplacian with the EMA and those of the localized eigenvectors, together with the results of the numerical experiments. The estimate of the localized eigenvectors and their eigenvalues is discussed in Sec. VI, and we describe the resulting behavior here. As γ increases, the eigenvector possessing information about the modules will eventually have a higher eigenvalue than that of the localized eigenvector; the standard spectral method fails in such a region. Our estimate agrees very well with the numerical result, as long as localization does not occur. Although we do not know which g should be chosen a priori (see Sec. VI for the meaning of g), if we choose a value that is consistent with the result of the numerical experiment, our estimate of the localization transition is close to the point where the result given by the replica method with the EMA starts to deviate from numerical result. Eigenvector localization is expected to result from the existence of a few vertices with irregular degrees, which we call the defects. In Figs. 7(a) and 7(b), we regard the vertices with the lower population as the defects. As we show in Secs. VI and VIII, localization tends to occur in both unnormalized and normalized Laplacians when the defects have a lower degree. Therefore, when we have equal populations, i.e., b1 = b2 = 0.5, we regard vertices with lower degree as defects. Indeed, when vertices with lower degree are dominant, localization does not seem to occur, or produces only a negligible effect. (a) (b) 14 FIG. 7: (Color online) Average second-smallest eigenvalues of the unnormalized Laplacian L of the two-block random graphs with bimodal degree distributions, as a function of γ. The degree sets are (a) {c1, c2} = {3, 6} and (b) {c1, c2} = {3, 9}. The ratio of module sizes is set to p1 = 0.6 (p2 = 0.4) in both cases. The estimates given by the replica method with the EMA and the estimated eigenvalues for the localized eigenvectors are represented by the solid lines (see Sec. VI for details of the localized eigenvectors). When the latter gives a lower eigenvalue, the former is indicated by a dashed line. The dots represent the numerical results with N = 104. The average is over ten samples. In each plot, results are plotted, from top to bottom, for b1 = 0.1, 0.5, 0.9 (b2 = 0.9, 0.5, 0.1). FIG. 8: (Color online) A tree with the defects aggregated around its root. VI. LOCALIZED EIGENVECTOR OF THE UNNORMALIZED LAPLACIAN WITH BIMODAL DEGREE DISTRIBUTIONS The behavior of the eigenvalues of localized eigenvectors for the unnormalized Laplacian L was described in Sec. V. This section is devoted to their derivation and an analysis of some specific examples. As mentioned in the previous section, localized eigenvectors emerge because of degree fluctuations. They are the vectors in which the weight of their elements is concentrated around a few defects, the vertices with characteristic degrees. Here, we analyze the localization of an eigenvector for the unnormalized Laplacian L and consider this process for the normalized Laplacian L in Sec. VIII. We focus on sparse graphs with bimodal degree distributions. The graphs have two types of degree, cD and cB, with populations bD and bB (bD + bB = 1), respectively. We let bD < bB and refer to the vertices with degree cD as the defects. As frequently analyzed for sparse graphs in the literature [39, 41 -- 43] (see Fig. 8), we consider a tree with defects aggregated around its root, up to a distance g from the root. Hereafter, we denote a localized eigenvector as v. We now simplify the graph by letting all vertices at distance d > g have a uniform degree cB. Then, the symmetry of the graph implies that, for all vertices located at distance d from the root, vi = Vd. From the definition of the eigenvector Lv = λv, for g ≥ 1, we have 15 (cD − λ)V0 − cDV1 = 0, (cD − 1)Vd+1 − (cD − λ)Vd + Vd−1 = 0 (cB − 1)Vd+1 − (cB − λ)Vd + Vd−1 = 0 (0 < d ≤ g), (d ≥ g + 1). (71) To be a localized eigenvector, the element Vd needs to vanish at d → ∞. Thus, we choose a solution of the form Vd+1 = κVd with κ < 1 for d > g, where κ satisfies (72) In addition, we have the constraint that the norm of the eigenvector needs to be finite, i.e., v2 < ∞. Let us consider the case g = 0, i.e., only the vertex at the root is defective. From Eq. (71), λ must satisfy (cB − 1)κ2 − (cB − λ)κ + 1 = 0. cD − λ cD = V1 V0 = V2 V1 = κ(λ), (73) where κ(λ) is a function of λ determined by Eq. (72). The resulting non-zero eigenvalue is λ = cD(cB − cD − 2)/(cB − cD − 1), and the corresponding damping factor κ is κ = 1/(cB − cD − 1). In addition, the constraint finite norm requires cD < cB − 1 − √cB − 1. Interestingly, the defect must have a lower degree than the other vertices, whereas, in the case of the adjacency matrix, it was hubs that caused localization. Note, however, that this is the condition for a tree with uniform degree at d > 0 and does not hold in general. Let us now consider using the results given by the replica method with the EMA from the previous section in the case where the vertex degrees at d > g are not necessarily uniform. Recall that [λ2]L = −φ when m11 = 0, and that a satisfies Eqs. (45) and (46) for random regular graphs. From these relations, we see that the factor 1 + a obeys the same characteristic equation as (72). This connection can also be seen from the saddle-point equation; Eqs. (38) -- (40) yield ∂H/∂Hg = (1 + a)−1 = 1 + a. This implies that, in tree-like graphs, the response function ∂Vi/∂Hj generally damps as O((1 + a)D(i,j)), where D(i, j) is the distance between sites i and j. In particular, taking vertex j to the root (g = 0) and comparing the relation with the solution form Vd = const κd for d > g, where const is proportional to V0, we conclude κ = 1 + a. Hence, one way to estimate the localized eigenvector and its eigenvalue is to replace the damping factor κ with 1 + a∗(λ), where a∗(λ) is the value of a obtained via the EMA as a function of λ. Then, analogously to the example above, we numerically compute the eigenvalue λ so that it is consistent with the equations for d ≤ g in Eq. (71), i.e., Vg+1 Vg = Vg+2 Vg+1 = 1 + a∗(λ). (74) If this eigenvalue is smaller than the second-smallest eigenvalue calculated in the previous section, we can conclude that the localization transition has occurred. VII. DETECTABILITY THRESHOLD IN RANDOM GRAPHS WITH DEGREE FLUCTUATION: THE NORMALIZED LAPLACIAN We now analyze the spectral method with the normalized Laplacian L and consider the two-block random graph with degree fluctuations. With y := D−1/2 x, the Hamiltonian can be written as and, noting that the total degree reads K = cN , we define the partition function in terms of y as H(xL) = 1 2 x TLx = 1 2 TLy, y Z(βL) =Z dy e−βH(yL)δ(y TDy − cN )δ(1TDy), (75) (76) where we have omitted the constant factor obtained by defining the partition function in terms of x. The free energy defined by this partition function is related to the second-smallest eigenvalue according to 2 lim β→∞ [f (βL)]L = −2 lim β→∞ 1 N β [ln Z(βL)]L = c [λ2]L . (77) Note that the vector y is not the eigenvector of L, and therefore the distribution P (y) does not give the distribution of the second-smallest eigenvector. However, as the sign of each vector element is not changed by the conversion from y to x, the fraction of correctly classified vertices is readily obtained from the distribution P (y). In the limit N → ∞ (as in Secs. IV and V), the replica-symmetric solution can be obtained as c [λ2]L = − extr qr,qr ,φ,ψ(Z dAdHZ dA′dH ′ Ξ(A, H, A′, H ′) 16 1 2(cid:18)(c p1 − γ)q1(A, H)q1(A′, H ′) + (c p2 − γ)q2(A, H)q2(A′, H ′) + 2γq1(A, H)q2(A′, H ′)(cid:19) × prZ dAdHZ d Ad H qr(A, H)qr( A, H) (H + H)2 + c φ − c Xr=1,2 A − A − Hg(cid:17)2 Yg=1(cid:16)d Agd Hg qr( Ag, Hg)(cid:17)(cid:16)ctψ/2 −Pg Ag ). btZ + Xr=1,2 prXt ctφ −Pg A ! H 2 ct (78) We have the same saddle-point equations as (56) and (57) for qr( A, H). For qr(A, H), we have an analogous equation to (58), but with ψ and φ replaced with ctψ and ctφ, i.e., qr(A, H) =Xt btct c Z ct−1 Yg=1 (cid:16)d Agd Hg qr( Ag, Hg)(cid:17) δ H + ctψ 2 − ct−1 Xg=1 Hg! δ A − ctφ + Ag! . ct−1 Xg=1 (79) With the EMA, Eq. (78) is approximated as pk(cid:2)(1 + a)m2k + m2 1k(cid:3) − γ c (m11 − m12)2! pk (m2k + 2m1k m1k + m2k) + Xk=1,2 1k(cid:1) ). pk(cid:0) m2k − cψ m1k + (c − 1) m2  [λ2]L = −extr(φ + − 1 1 a(a + 2) Xk a − aXk c(φ − a)  cψ2 4 1 + From the saddle-point conditions, for the solution with m1k 6= 0, we have 1 + a = 1 1 + a , a = c φ − (c − 1)a, cp1 1 − 11 = (c − 1)2 p2 m2 c (cid:18)c − 1 − 1 − Γ [λ2]L = −φ = 2!(cid:16)(c − 1)Γ (c − 1)2Γ Γ(cid:19) , 1 1 2 − 1(cid:17) , and in the region where m1k = 0, we have Again, ψ = 0 in both cases. The condition for the solution with m1k 6= 0 to exist is [λ2]L = −φ = (cid:0)√c − 1 − 1(cid:1)2 c . 1 √c − 1 ≤ Γ, (80) (81) (82) (83) (84) (85) (86) and the detectability threshold is the condition when the equality holds. Figures 9(a) and 9(b) plot Eqs. (84) and (85), together with the results of the numerical experiments for bimodal distributions. These figures are plotted in the same way as in Figs. 7(a) and 7(b). Again, our estimate gives a fairly (a) (b) 17 FIG. 9: (Color online) Average second-smallest eigenvalues of the normalized Laplacian L of two-block random graphs with bimodal degree distributions as functions of γ. The degree sets are (a) {c1, c2} = {3, 6} and (b) {c1, c2} = {3, 9}. The ratio of module sizes is set to p1 = 0.6 (p2 = 0.4) in both cases. The estimates given by the replica method with the EMA and the estimated eigenvalues of the localized eigenvectors are represented by solid lines (see Sec. VIII for details of the localized eigenvector). When the latter gives a lower eigenvalue, the former is indicated by a dashed line. The dots represent the numerical results with N = 104. The average is over 10 samples. In each plot, results are plotted, from top to bottom, for b1 = 0.1, 0.5, 0.9 (b2 = 0.9, 0.5, 0.1). accurate prediction of the numerical results [44]. As conjectured in Sec. VIII, a comparison of Figs. 7(a) and 7(b) and Figs. 9(a) and 9(b) shows that localization is less likely to occur with the normalized Laplacian L than with the unnormalized Laplacian L. Figures 10(a) and 10(b) show the fraction of correctly classified vertices, ignoring the effect of localization. Although our estimates are close to the numerical results when the effect of localization is negligible, they differ significantly in the region where localization is present. Note that the difference between the crosses and the solid line in each figure is not totally due to the Gaussian approximation of the distribution of the eigenvector elements (see Appendix C for details). In this case, the Gaussian fitting (solid line) whose mean and variance are estimated by the EMA of the free energy seems to be a better approximation. VIII. LOCALIZED EIGENVECTOR OF THE NORMALIZED LAPLACIAN WITH BIMODAL DEGREE DISTRIBUTIONS In this section, we analyze the localized eigenvectors of the normalized Laplacian L. The behavior of the eigenvalues of these localized eigenvectors for the normalized Laplacian L were described in Sec. VII, and this section is devoted to their derivation and an analysis of some specific examples. We consider the tree in which the defects are aggregated around its root up to a distance of g. The analysis here is completely analogous to that for the unnormalized Laplacian L in Sec. VI. The equations corresponding to (71) for g ≥ 1 are V1 − (1 − λ)V0 = 0, (cBc2)−1/2(c2 − 1)Vg+1 − (1 − λ)Vg + c−1 2 Vg−1 = 0, c−1 B (cB − 1)Vg+2 − (1 − λ)Vg+1 + (cBc2)−1/2Vg = 0, B Vd−1 = 0 c−1 B (cB − 1)Vd+1 − (1 − λ)Vd + c−1 (d ≥ g + 2). (87) As in Sec. VI, we consider the case where the vertices at d > g have a uniform degree cB. Interestingly, when g = 0, an analogous calculation as for the unnormalized Laplacian L in Sec. VI yields that the eigenvector with a non-zero eigenvalue of the form Vd+1 = κVd for d > 1 has κ = −1, i.e., we never have a localized eigenvector. Moreover, when g = 1, the condition 2cD < cB must be satisfied to give a localized eigenvector with a finite norm; again, the defects need to have a lower degree. Note also that the results obtained here are more severe than the conditions in Sec. VI (a) (b) 18 EMA (moment) EMA (distribution) EMA (moment) EMA (distribution) FIG. 10: (Color online) Fraction of correctly classified vertices with the spectral method for two-block random graphs with bimodal degree distributions as functions of γ. The normalized Laplacian L is used, and the degree sets are (a) {c1, c2} = {3, 6} and (b) {c1, c2} = {3, 9}. The population of each degree and the ratio of module sizes are b1 = 0.5 (b2 = 0.5) and p1 = 0.5 (p2 = 0.5) in both cases. The dots represent the numerical results with various graph sizes. The crosses represent the results obtained by iterating the saddle-point equations (56), (57), and (79). The solid lines represent the results obtained by fitting the distribution of eigenvector elements to a Gaussian distribution whose mean and variance are estimated by the saddle point of the free energy with the EMA (see Appendix B for details). for the unnormalized Laplacian L in the sparse case. This implies that localization tends to be suppressed in the normalized Laplacian L. In fact, this tendency is consistent with the analysis in data clustering [38, 45]. In the case where the vertex degrees at d > g are not necessarily uniform, a localized eigenvector can be estimated in almost the same way as for the unnormalized Laplacian L in Sec. VI, utilizing the results of the replica method with the EMA in Sec. VII. Note that, unlike the case of the unnormalized Laplacian L, the value of cB is needed to determine the ratio Vg+2/Vg+1 in Eq. (87). Thus, when we solve for the consistent eigenvalue by Vg+2/Vg+1 = Vg+3/Vg+2 = 1+a∗(λ), we replace cB in the ratio Vg+2/Vg+1 with the average degree c. IX. DETECTABILITY THRESHOLD OF THE STOCHASTIC BLOCK MODEL WITH THE NORMALIZED LAPLACIAN Finally, we consider the eigenvalues and detectability threshold in the stochastic block model, i.e., the random graph with a Poisson degree distribution. To compare with the literature, we recast our result in terms of cin and cout and set the module sizes to p1 = p2 = 0.5. With these quantities, the detectability threshold in the normalized Laplacian, Eq. (86), reads cin − cout = 2c √c − 1 . (88) Compared with the threshold obtained in [8], we have a correction factor of √c/√c − 1. The phase diagrams of these two thresholds are shown in Fig. 11. While the difference between them is negligible when the average degree is sufficiently large, considering the fact that the upper bound of the parameter cin − cout is 2c, this gap is indeed considerable in sparse graphs. We can compare the average estimate of the second-smallest eigenvalue [λ2]L with the numerical results. In terms of cin and cout, Eq. (84) reads [λ2]L = 1 − (c − 1) 2c2 (cin − cout) − 2 cin − cout . (89) Figures 12(a) and 13(a) show that the estimated eigenvalues with the EMA agree excellently with the numerical results, as long as the localized eigenstate does not occupy the second-smallest eigenvalue. We measured the localization i )2 for a vector x, and have plotted this in Figs. 12(b) and 13(b). The IPR grows rapidly below the point at which the estimates start to deviate from the numerical results in Figs. 12(a) and 13(a). Similarly, as shown in Figs. 12(c) and 13(c), our estimates for the fraction of correctly classified vertices start to deviate from the numerical results at that point. strength with the inverse participation ratio (IPR), defined as PN i /(PN i=1 x4 i=1 x2 19 FIG. 11: (Color online) Phase diagram of the detectable and undetectable regions given by the replica method with the EMA (solid line) and the ultimate threshold [5, 7, 8] (dashed line). The model cannot take parameter values with cin − cout > 2c (invalid region) because of the condition that cout ≥ 0. For the stochastic block model with c = 6, the localized eigenvector appears in the region significantly above the detectability threshold, i.e., in the detectable region, and its effect is not negligible. However, for the stochastic block model with c = 8, although the estimate with the EMA is still not precise, the error due to localization seems to be much smaller for the graph sizes we tested. Note that as the average degree c increases the degree fluctuation of each vertex decreases because of the law of large numbers. Therefore, the effect of localization is expected to eventually disappear. That is, the point at which localization occurs will finally become buried in the undetectable region. As mentioned in Sec. VII, the solid lines and crosses behave differently in Figs. 12(c) and 13(c). Moreover, the results of the saddle-point equation with the EMA seem to converge to the result in [8] around the detectability threshold. It should be noted, however, that this coincidence is due to the property of the Poisson degree distribution in the stochastic block model. As described in Appendix C, the difference between these results with the EMA generally depends on the form of the degree distribution. X. SUMMARY In summary, we have analyzed the limitations of the spectral method for graph partitioning, known as the de- tectability threshold, and the localization of eigenvectors. We derived estimates for the detectability thresholds of the spectral method with the unnormalized Laplacian L [Eq. (68)] and the normalized Laplacian L [Eq. (86)] for sparse graphs. The detectability threshold with the normalized Laplacian L can generally be written as (88) (for equal size modules), which is analogous to the threshold for random regular graphs (53). This converges to the result in [8] in the dense limit c → ∞. For the condition where a localized eigenvector emerges, although it is difficult for a graph with an arbitrary degree distribution, our estimates give a fairly good prediction for two-block random graphs with bimodal degree distributions. Overall, our estimates with the replica method agree with the numerical results quite well, as long as localization is absent, for the graph sizes we tested. It should be noted, however, that the localization of eigenvectors is expected to be sensitive to rare events such as the emergence of vertices of irregular degree. Therefore, when the support of the degree distribution is infinite, e.g., a Poisson distribution, we must be careful as the finite size effect may not be negligible. We revealed that the spectral method with Laplacians does not detect modules all the way down to the ultimate detectability threshold 2√c in any sparse graph. In fact, the estimated gap between the precise detectability threshold and the ultimate one is already considerable for very sparse graphs even in the case where eigenvector localization is absent or negligible (see Fig. 11). Another way of viewing this result is that the method with the non-backtracking matrix closed the gap of detectability in two ways. Furthermore, for the graph sizes we tested, the effect of localization was relatively weak in the stochastic block model when the average degree was not very low. Finally, we comment that we must be careful when we compare the performance between the normalized Laplacian and modularity. Although (a) (b) EMA (moment) 20 (c) Nadakuditi-Newman EMA (moment) EMA (distribution) FIG. 12: (Color online) (a) Average second-smallest eigenvalue of the method with the normalized Laplacian L, (b) the IPR of its eigenvector, and (c) the fraction of correctly classified vertices with the spectral method in the stochastic block model. We set the average degree c = 6. In each plot, the dots represent the numerical results with various graph sizes. In (a), the estimate of the eigenvalue with the EMA is represented by a solid line. The dotted line shows the estimated detectability threshold and the dashed line shows the estimate in [8]. In (c), the crosses represent the fraction of correctly classified vertices with the EMA and the solid line represents their Gaussian approximation (see Appendix B for details). The dashed line is, again, the estimate in [8]. their spectral methods become equivalent for a certain choice of normalization [11], there is no guarantee that our results precisely coincide with the detectability threshold of the modularity matrix, i.e., the method with the spherical normalization [46]. This work was supported by JSPS KAKENHI No. 26011023 (TK) and No. 25120013 (YK) and the JSPS Core-to- Core Program "Non-equilibrium dynamics of soft matter and information." Acknowledgements Appendix A: Number of random graphs with two modules In this section, we calculate NG, the number of possible graph realizations. Although the value of NG does not appear in the final result in the main text, the technique used here is essential for calculating the moment of the partition function. Let us first consider the two-block c-random regular graphs. We sum all the connection patterns (a) (b) EMA (moment) 21 (c) Nadakuditi-Newman EMA (moment) EMA (distribution) FIG. 13: weaker, and the estimate with the EMA is more precise. (Color online) Same plots as in Fig. 12, but with the average degree c = 8. In this case, the effect of localization is ({uij},{wij}) that satisfy the constraint of constant degree. That is, NG = X{uij }{wij } Yi∈V1 δ Xl∈V1 uil + Xk∈V2 wik − c! Yj∈V2 δ Xl∈V2 ujl + Xk∈V1 wjk − c! δ Xi∈V1 Xk∈V2 wik − γN! . Using the relations zx−1, 2π δ(x) =I dz δ(x) =Z +i∞ −i∞ dη 2π e−ηx, (A1) (A2) (A3) we have NG = X{uij }{wij } Yi∈V1I dzi 2π (Pl∈V1 i z uil+Pk∈V2 wik−c−1) Yj∈V2I dzj 2π (Pl∈V2 j z ujl+Pk∈V1 wjk −c−1) 2π Z dη =I Yi∈V1 dzi 2π dzi 2π =I Yi∈V1 × Yi<l∈V1 2π dzj 2π z−(1+c) i z−(1+c) j wik − γN )(cid:1) exp(cid:0)−η(Xi∈V1 Xk∈V2 Z dη I Yj∈V2 (zizl)uil Yj<l∈V2 Xujl ={0,1} Z dη I Yj∈V2 (1 + zizl) Yj<l∈V2 z−(1+c) j z−(1+c) i dzj 2π 2π eηγN eηγN (1 + zjzl) Yi∈V1 Yk∈V2(cid:0)1 + zizke−η(cid:1) . × Yi<l∈V1 Xuil={0,1} (zjzl)ujl Yi∈V1 Yk∈V2 Xwik={0,1} (zizke−η)wik Setting the contours with respect to zi and zj to be small, we can approximate the last factors as (1 + zizl) = exp Xi<l∈V1 zi Xl∈V1 zk! , where we have neglected the diagonal term in (A7) because N ≫ 1. Introducing the order parameters Yi<l∈V1 Yi∈V1 Yk∈V2(cid:0)1 + zizke−η(cid:1) = exp Xi∈V1 Xk∈V2 zizl! ≈ exp 1 ln(1 + zizke−η)! ≈ exp e−η Xi∈V1 ln(1 + zizl)! ≈ exp Xi<l∈V1 2 Xi∈V1 zi Xk∈V2 zl! , q1 = q2 = we can recast (A6) as 1 p1N Xi∈V1 p2N Xj∈V2 1 zi, zj, eηγN 2π dzi 2π dzj 2π z−(1+c) i z−(1+c) j Z dη × δ p1N q1 − Xi∈V1 × exp(cid:20) 1 NG = p1p2N 2Z dq1Z dq2I Yi∈V1 zi! δ (p1N q1)2(cid:21) exp(cid:20) 1 Z dq2dq2 × exp"−q1 p1N q1 − Xi∈V1 × exp(cid:20) 1 I Yj∈V2 zj p2N q2 − Xj∈V2  (p2N q2)2(cid:21) exp(cid:2)e−ηp1p2N 2q1q2(cid:3) I Yi∈V1 I Yj∈V2 zj zi!# exp p2N q2 − Xj∈V2  (p2N q2)2 + e−ηp1p2N 2q1q2 + ηγN(cid:21) . = p1p2N 2Z dq1dq1 −q2 (p1N q1)2 + z−(1+c) i dzj 2π dzi 2π 1 2 2π 2π 2 2 2 z−(1+c) j 2π Z dη   22 (A4) (A5) (A6) (A7) (A8) (A9) (A10) (A11) (A12) Since dzi 2π I Yi∈Vr ezi qr z−(1+c) i =I Yi∈Vr dzi 2π Xm (zi qr)m m! z−(1+c) i r =(cid:18) qc c!(cid:19)N pr (r = 1, 2), (A13) Eq. (A12) becomes NG = p1p2N 2Z dq1dq1 Z dq2dq2 + N (ηγ − p1 q1q1 − p2 q2q2 + c p1 ln q1 + c p2 ln q2 − ln c!)#. In the limit N → ∞, the saddle point of the integrand gives NG. The saddle-point conditions yield 2 + 2e−ηp1p2q1q2(cid:1) 2 (cid:0)p2 Z dη 1 + p2 2q2 1q2 2π 2π 2π exp" N 2 γ = N e−ηp1p2q1q2, N p1q1 + N p2e−ηq2 − q1 = 0, N p2q2 + N p1e−ηq1 − q2 = 0, q1 q1 = q2 q2 = c. We then have N p2 1 q1 =s c p1 − γ q2 =s c p2 − γ N p2 2 , . Inserting the values at the saddle point, we obtain the number of graphs for N ≫ 1: NG ≃ exp"N(cid:18) c − 2 (ln N − 1) − ln c! − γ ln γ + cp1 ln (cp1) + cp2 ln (cp2) (cp2 − γ) ln (cp2 − γ)(cid:19)#. (cp1 − γ) ln (cp1 − γ) − 1 2 1 2 23 (A14) (A15) (A16) (A17) (A18) (A19) (A20) (A21) t=1 be the sequence of degrees, each of which has probability {bt}T A completely analogous result holds for the number of possible two-block random graphs with a given degree sequence, i.e., the two-block configuration model or the degree-corrected ensembles with "hard" constraints [34]. We let {ct}T t=1; i.e., the number of nodes with degree ct is N bt. The number of graphs NG can then be written as NG = X{uij }{wij } wik − ct! Yj∈V(2,t) wjk − ct! wik − γN! , ujl + Xk∈V1 uil + Xk∈V2 δ Xl∈V1 δ Xl∈V2 Yt=1 T  δ Xi∈V1 Xk∈V2   Yi∈V(1,t) (A22) where, as in the main text, we denote the set of vertices in module r with degree ct as V(r,t). A similar calculation to the single-degree case yields NG = p1p2N 2Z dη 2π 2π Z dq1dq1 × exp"N(cid:18) N 2 Z dq2dq2 2π Yt   Yi∈V(1,t) 1 ct! Yj∈V(2,t) 1 ct!  p2 1q2 1 + p2 2q2 2 + N e−ηp1p2q1q2 N 2 where, as in the main text, c =Pbt + ηγ − p1 q1q1 − p2 q2q2 + c (p1 ln q1 + p2 ln q2)(cid:19)#, (A23) btct is the average degree. The saddle-point conditions yield analogous results, γ = N p1q1p2q2e−η, N p1q1 + N p2q2e−η = q1, N p1q1e−η + N p2q2 = q2, q1 q1 = q2 q2 = c, (A24) (A25) (A26) (A27) and Finally, we have where c! =Pt btct!. NG ≃ exp"N(cid:18) c − 2 N p2 1 q1 =s cp1 − γ q2 =s cp2 − γ N p2 2 , . (ln N − 1) − ln c! − γ ln γ + cp1 ln (cp1) + cp2 ln (cp2) (cp2 − γ) ln (cp2 − γ)(cid:19)#, (cp1 − γ) ln (cp1 − γ) − 1 2 1 2 24 (A28) (A29) (A30) Appendix B: Gaussian approximation of the fraction of correctly classified vertices We consider the distribution of elements of the second-smallest eigenvector belonging to the rth module, averaged over the realization of the unnormalized Laplacian L. This is defined as Pr(x) = 1 Nr Xi∈Vr [δ(x − xi)]L , and, as mentioned in the main text, it can be expressed in terms of H and A as Pr(x) =Z dAdH Qr(A, H) δ(cid:18)x − H A(cid:19) . (B1) (B2) For the Gaussian fitting of the distribution Pr(x), we solve for the mean and variance in (B2). To obtain an analytical expression, we fix the distribution of A, i.e., Qr(A) = R dHQr(A, H) = δ(afull − A). Note that afull here is different from a in the saddle-point equations. From the cavity interpretation of Eqs. (38) -- (40), Qr(A, H) can be regarded as the complete marginal distribution corresponding to qr(A, H). Therefore, instead of (46), afull can be determined as afull = φ − ca = (c − 1)Γ − 1 (c − 1)Γ , (B3) where we have inserted the values of φ and a at the saddle point. Then, in the detectable region, the mean can be approximated as mr(x) = = = 1 1 afull Z dH HQr(H) afull Z Yg=1 c(c − 1)Γ m1r (c − 1)2Γ2 − 1 , c d Hg qr( Hg) − where we have used the fact that ψ = 0 at the saddle point. Similarly, for the second moment, we have (cid:10)x2(cid:11)r = c d Hg qr( Hg) c Xg=1 1 a2 Yg=1 =(cid:18) (c − 1)Γ full Z (c − 1)2Γ2 − 1(cid:19)2 ψ 2 + Hg! c Xg=1 Hg!2 (cid:0)c m2r + c(c − 1) m2 1r(cid:1) . (B4) (B5) Then, the variance reads 25 s2 (c − 1)2Γ2 − 1(cid:19)2 r(x) =(cid:10)x2(cid:11)r −(cid:10)x2(cid:11)r = c(cid:18) (c − 1)Γ r(x))i /2 gives the fraction of correctly classified vertices in module r. For 2(x)(=: s2(x)) by symmetry. The total fraction (cid:0) m2r − m2 1r(cid:1) . (B6) With mr(x) and sr(x), h1 + erf(mr(x)/p2s2 the equal-size modules, m1(x) = m2(x)(=: m(x)) and s2 of correctly classified vertices is then h1 + erf(m(x)/p2s2(x))i /2. For the normalized Laplacian L with the EMA, the analogous calculation for the detectable region gives 1(x) = s2 mr(x) = 2 c2 c (c − 1)Γ m1r (c − 1)2Γ − 1 r(x) = (c − 1)Γ (c − 1)2Γ 2 s2 , − 1!2  c2 c m2r +  c3 − c2 c − c2 1r c !2  m2  , (B7) (B8) where cn =Pt btcn t . Appendix C: Approximations of the saddle-point equations and the free energy As shown in Figs. 10(a), 10(b), 12(c), and 13(c), the results obtained by the EMA of the saddle-point equation (crosses in the figures) and the Gaussian fitting of the distribution of the eigenvector elements (solid lines in the figures), whose mean and variance are estimated by the EMA of the free energy, are different. One may expect that the former is simply more accurate, as long as the stationary state is achieved by a sufficient number of iterations, because the latter contains a Gaussian approximation. However, this is not correct. In fact, the EMA of the saddle- point equations is not equivalent to the EMA of the free energy; in the latter, the approximation is applied before taking the saddle point. For example, the relation between the first moment of H, m1r, and its conjugate m1r, derived by the saddle-point equation (79) with the EMA, is m1r = c2 c − 1! m1r, (C1) whereas we have m1r = (c − 1) m1r from the saddle point of the free energy with the EMA. Hence, the difference depends on the ratio of the mean and the variance, or the Fano factor. In the case of the Poisson degree distribution, Eq. (C1) reads m1r = c m1r, which corresponds to the dense limit of the saddle point for the free energy. As neither quantity is exact, it is not readily obvious which offers the better estimate in general. [1] S. Fortunato, Phys. Rep. 486, 75 (2010). [2] S. Fortunato and M. Barth´elemy, Proc. Natl. Acad. Sci. U.S.A. 104, 36 (2007). [3] T. Kawamoto and M. Rosvall, Phys. Rev. E 91, 012809 (2015). [4] J. Reichardt and M. Leone, Phys. Rev. Lett. 101, 078701 (2008). [5] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, Phys. Rev. Lett. 107, 065701 (2011). [6] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, Phys. Rev. E 84, 066106 (2011). [7] E. Mossel, J. Neeman, and A. Sly, Probab. Theory Relat. Fields pp. 1 -- 31 (2014). [8] R. R. Nadakuditi and M. E. J. Newman, Phys. Rev. Lett. 108, 188701 (2012). [9] G. V. Steeg, C. Moore, A. Galstyan, and A. Allahverdyan, Eur. Phys. Lett. 106, 48004 (2014). [10] T. P. Peixoto, Phys. Rev. Lett. 110, 148701 (2013). [11] M. E. J. Newman, Phys. Rev. E 88, 042822 (2013). [12] F. Radicchi, Phys. Rev. E 88, 010801 (2013). [13] F. Radicchi, Eur. Phys. Lett. 106, 38001 (2014). [14] P. Ronhovde, D. Hu, and Z. Nussinov, Eur. Phys. Lett. 99, 38006 (2012). [15] R. K. Darst, D. R. Reichman, P. Ronhovde, and Z. Nussinov, Journal of Complex Networks (2014). 26 [16] D. Hu, P. Ronhovde, and Z. Nussinov, Philos. Mag. 92, 406 (2012). [17] J. Leskovec, K. J. Lang, D. Anirban, and M. W. Mahoney, Internet Mathematics 6, 29 (2009). [18] A. Lancichinetti and S. Fortunato, Phys. Rev. E 80, 056117 (2009). [19] A. Lancichinetti and S. Fortunato, Phys. Rev. E 84, 066122 (2011). [20] B. H. Good, Y.-A. de Montjoye, and A. Clauset, Phys. Rev. E 81, 046106 (2010). [21] R. Aldecoa and I. Mar´ın, Sci. Rep. 3, 2216 (2013). [22] D. Hric, R. K. Darst, and S. Fortunato, Phys. Rev. E 90, 062805 (2014). [23] R. K. Darst, Z. Nussinov, and S. Fortunato, Phys. Rev. E 89, 032809 (2014). [24] A. Condon and R. M. Karp, Random Struct. Algorithms 18, 116 (2001). [25] M. Onsjo and O. Watanabe, in Algorithms and Computation (Springer, 2006), pp. 507 -- 516. [26] P. J. Bickel and A. Chen, Proc. Natl. Acad. Sci. U.S.A. 106, 21068 (2009). [27] P. Zhang, F. Krzakala, J. Reichardt, and L. Zdeborov´a, J. Stat. Mech. 2012, P12021 (2012). [28] F. Krzakala, C. Moore, E. Mossel, J. Neeman, A. Sly, L. Zdeborov´a, and P. Zhang, Proc. Natl. Acad. Sci. U.S.A. 110, 20935 (2013). [29] P. W. Holland, K. B. Laskey, and S. Leinhardt, Soc. Networks 5, 109 (1983). [30] E. M. Airoldi, D. M. Blei, S. E. Fienberg, and E. P. Xing, J. Mach. Learn. Res. 9, 1981 (2008). [31] A. Clauset, C. Moore, and M. E. J. Newman, Nature 453, 98 (2008). [32] B. Karrer and M. E. J. Newman, Phys. Rev. E 83, 016107 (2011). [33] B. Ball, B. Karrer, and M. E. J. Newman, Phys. Rev. E 84, 036103 (2011). [34] T. P. Peixoto, Phys. Rev. E 85, 056122 (2012). [35] D. B. Larremore, A. Clauset, and A. Z. Jacobs, Phys. Rev. E 90, 012805 (2014). [36] X. Zhang, R. R. Nadakuditi, and M. E. J. Newman, Phys. Rev. E 89, 042816 (2014). [37] P.-Y. Chen and A. O. Hero III, arXiv:1409.3207 (2014). [38] U. Luxburg, Statistics and Computing 17, 395 (2007). [39] Y. Kabashima and H. Takahashi, J. Phys. A 45, 325001 (2012). [40] F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi, Proc. Natl. Acad. Sci. U.S.A. 101, 2658 (2004). [41] G. Biroli and R. Monasson, J. Phys. A 32, L255 (1999). [42] G. Semerjian and L. F. Cugliandolo, Journal of Physics A: Mathematical and General 35, 4837 (2002). [43] F. L. Metz, I. Neri, and D. Boll´e, Phys. Rev. E 82, 031135 (2010). [44] In Fig. 8(a), the localization transition might have occurred right before the detectability threshold. However, we could not find an eigenvalue close to the numerial experiment within g ≤ 3. [45] U. Von Luxburg, M. Belkin, and O. Bousquet, Ann. Statist. pp. 555 -- 586 (2008). [46] M. E. J. Newman, Phys. Rev. E 74, 036104 (2006).
1708.08136	1	1708	2017-07-24T23:59:17	An Ensemble Framework for Detecting Community Changes in Dynamic Networks	[ "cs.SI", "cs.LG" ]	Dynamic networks, especially those representing social networks, undergo constant evolution of their community structure over time. Nodes can migrate between different communities, communities can split into multiple new communities, communities can merge together, etc. In order to represent dynamic networks with evolving communities it is essential to use a dynamic model rather than a static one. Here we use a dynamic stochastic block model where the underlying block model is different at different times. In order to represent the structural changes expressed by this dynamic model the network will be split into discrete time segments and a clustering algorithm will assign block memberships for each segment. In this paper we show that using an ensemble of clustering assignments accommodates for the variance in scalable clustering algorithms and produces superior results in terms of pairwise-precision and pairwise-recall. We also demonstrate that the dynamic clustering produced by the ensemble can be visualized as a flowchart which encapsulates the community evolution succinctly.	cs.SI	cs	An Ensemble Framework for Detecting Community Changes in Dynamic Networks Timothy La Fond, Geoffrey Sanders, Christine Klymko, and Van Emden Henson Lawrence Livermore National Laboratory Livermore, California 94550 7 1 0 2 l u J 4 2 ] I S . s c [ 1 v 6 3 1 8 0 . 8 0 7 1 : v i X r a Abstract-Dynamic networks, especially those representing so- cial networks, undergo constant evolution of their community structure over time. Nodes can migrate between different communities, communities can split into multiple new com- munities, communities can merge together, etc. In order to represent dynamic networks with evolving communities it is essential to use a dynamic model rather than a static one. Here we use a dynamic stochastic block model where the underlying block model is different at different times. In order to represent the structural changes expressed by this dynamic model the network will be split into discrete time segments and a clustering algorithm will assign block memberships for each segment. In this paper we show that using an ensemble of clustering assignments accommodates for the variance in scalable clustering algorithms and produces superior results in terms of pairwise-precision and pairwise-recall. We also demonstrate that the dynamic clustering produced by the ensemble can be visualized as a flowchart which encapsulates the community evolution succinctly. 1. Introduction Over the past decade or so, network analysis has emerged as an important area of mathematics and computer science, as well as in many other disciplines (see [1], [2], [3] among many others). One of the significant questions in network analysis is how to cluster the nodes in the network to identify underlying communities [4], [5]. Clustering static graphs reveals important community structure that can be leveraged for many important analysis tasks involving rela- tional datasets. Several fast clustering techniques have been defined for detecting communities that have high internal connectivity relative to their external connectivity [6], [7], [8], [9]. Analysis of community structure in dynamic graphs, where each edge is associated with a timestamp, is vastly more complex. Many real-world networks experience con- cept drift as the networks evolve; such evolutions cannot be accurately represented with a static model. Ignoring the tem- poral information and applying static clustering techniques entire dataset risks missing important structural changes; im- portant but transient community development could remain undiscovered. Several types of changes in community structure are of high analytic interest, including (i) the birth/death of a community [10], (ii) the growth/decay of a community [10], (iii) the changes of internal connectivity of a community Figure 1: Schematic of the dynamic community detection frame- work employed. For each timestep, t, a representative partitioning is determined. Representative partitionings from adjacent timesteps (t− 1) and t + 1 are compared for tracking communities and their dynamic structural changes such as community death (dark orange) or community splitting (green into blue and yellow). (such as a dense community becoming increasingly bipartite [11]), (iv) groups of communities splitting/merging [10], [12], and (v) member migration between groups of two or more communities (as well as the identification of nodes that do not migrate [13]). Clearly, there is a need for both the ability to perform dynamic community detection to find the community behavior at any given time and to represent community evolution to display the kinds of temporal effects that occur in the network. In this paper, we design a framework for taking a static community detection algorithm and employing it to find the evolving community structure of dynamic networks. Our framework also provides a interpretable visualization of the major changes to the community structure. Some dynamic community detection algorithms, for ex- ample, GraphScope, have memory and will continue to use the discovered data model from prior time steps as long as that model passes some performance metrics [11]. This memory effect can sometimes be detrimental if the algorithm fails to update the model during a change in the underlying structure. Other algorithms have no memory at all and merely connect the results of independent community detection passes run at different times in the network [10]. We will describe how to apply our framework in both a memory and memoryless fashion, where the version with memory attempts to smooth the temporal noise of the data at the cost of sometimes lagging behind immediate changes to the underlying model. Often edge timestamps are essentially continuous, where the resolution of timestamps is on much smaller intervals than the span of time required to resolve communities, both statically and during their evolution. For example, we might have timestamps that have resolution in the seconds while a few hours of connectivity data are necessary to resolve most community structure present in the dataset. A large class of dynamic graph analysis techniques discretize the time dimension into a number of uniform non-overlapping time segments. The selection of an appropriate timestep width which provides sufficient resolution to detect temporal events re- mains an open problem, with time segments typically being defined to start and end at a natural frequency determined by the data source, e.g. daily or weekly. We will also restrict ourselves to community dynamics (i) and (iv) (birth, death, splitting, and merging). In this paper we: • • Define a dynamic stochastic block model capable of producing data with community evolution events. Introduce a Dynamic Clustering Ensemble Frame- work which, given a baseline static clustering algo- rithm, finds the dynamic community behavior of a network and reports any evolution in the community structure. • Demonstrate that the ensemble framework has su- perior pairwise-precision and pairwise-recall perfor- mance in recovering block assignments than than the baseline algorithm. We apply this ensemble to synthetic data, drawn from both static and dynamic models, as well as to semi-synthetic data. • Create a visualization of the community evolution in a real world dataset and illustrate the types of community dynamics found (Figure 7). λij(t) = θiθjΣbi,bj (t) every edge Aijt is sampled according to a Poisson distribu- tion with a time-dependent rate λij: Aijt ∼ P oisson(λij(t)) where θi controls node i's expected degree, bi is i's block assignment, and Σbi,bj (t) is the interaction strength between blocks bi and bj at time t. As the block model is a function of time, by gradually changing the edge probability between collections of nodes over time we can introduce smooth splits and merges into the community structure. For exam- ple, if the value of Σbi,bj (t) grows linearly with increasing t we effectively cause bi and bj to merge into a single block as their interconnectivity grows. 2.2. Ensemble Framework Given a dynamic graph as defined above and an a priori division of the graph stream into discrete time segments, the algorithm presented clusters the graph at every segment, using as a baseline the Markov chain Monte Carlo (MCMC) algorithm described in the works of Peixoto [7], [8], [9]. This algorithm uses an ergodic Markov chain to randomly modify the block membership of each node and probabilis- tically accepts or rejects each modification. Although the algorithm will converge given a sufficiently long mixing time, the required mixing times are often prohibitively long and, in practice, approximations to the MCMC process are used. These approximations, combined with the non- determinism of the MCMC process, lead to variability in the produced clustering given different runs of the MCMC algorithm on the same data. 2. Approach 2.1. Data Model To aid in the development of temporal algorithms, gen- erative dynamic graph models are often used. These models support: (a) building fully synthetic examples with vari- ous community dynamics, allowing algorithm validation on examples with strong ground-truth; (b) injecting synthetic community dynamics into real world dynamic graphs, al- lowing algorithm validation in the face of realistic noise; and (c) gauging the statistical significance of various community events that have been detected. Degree-corrected stochastic block models whose model parameters are functions of continuous time are able to exhibit the kinds of phenomenon described in (i)-(v) above, and we will be using a version for our synthetic and semi-synthetic data analysis. In order to create the type of community dynamics we expect to see in real networks we introduce a dynamic stochastic block model as a generative graph model where Figure 2: The clusterings determined by various runs of the MCMC clustering algorithm are used to form a representative clustering of the nodes. We address this variability using an ensemble approach which performs multiple runs of the MCMC clustering algorithm at each time segment. This accommodates for variance in the clustering algorithm by applying the al- gorithm multiple times and obtaining a cloud of possible partitions. These clouds are then resolved to produce a best overall cluster assignment for the time segment, referred to as the representative clustering, which corresponds to the most typical partitions seen in the cloud. A small example of this can be seen in Figure 2. Let H(X, Y ) be any similarity function across two sets of nodes. Given K runs of the MCMC clustering algorithm at time t, leading to K different partitions of the form ,} for k = 1 to K, the algorithm reports {Bk,t the partition Bt = {Bt c} that satisfies 2 ...Bk,t ck 1 , Bk,t 2...Bt 1, Bt (cid:26) (cid:20) (cid:21)(cid:27) Bt = El(cid:54)=k max Bk,t x max y H(Bk,t x , Bl,t y ) y y X∩Y X∪Y . where Bk,t is the set of nodes in the xth block of the kth x partition at time t, ck is the number of blocks in partition k, c is the median number of blocks across all K partitions of time t, and Ek(cid:54)=l[·] is the expectation taken over all pairs of different iterations. Here we use similarity function H(X, Y ) = This resolution step can also be extended to "smooth" the output by incorporating Bl,t−1 (as well as additional time segments in each direction) into the bounds of the expectation, emphasizing connections between clus- ters over time if an algorithm with memory is desired. This smoothing approach compares the possible partitions available for time step t to those from t − 1 and t + 1 in order to find blocks which are consistent across different time steps. and/or Bl,t+1 A slight overhead is incurred during the resolution of the ensemble of partitions. For a given time step, each partition k out of the K different partitions of the graph has ck blocks, and all pairs of blocks from different partitions must be compared for the intersection of their member nodes. If using a hashed implementation, each intersection operation costs Θ(min(nx, ny)), where nx and ny are the size of blocks x and y. For every pair of partitions, all pairs of block must be compared, and each are an average size of N ck . The cost of all comparisons is then or N cl , Θ ck N cl ckclK(K − 1)min = Θ (N · min(ck, cl)) . Assuming that the number of blocks returned in any par- tition is less than log2N, the computational complexity of the ensemble resolution step is equal to or less than the cost of clustering the graph once at any particular time segment on average. Each iteration of the graph clustering algorithm in the ensemble can be run in parallel, needing no communication across the iterations until the resolution step, making the cost of each time segment Θ(cid:0)N log2N(cid:1). In the memoryless, unsmoothed version of the framework the individual time segments are also independent and can be run in parallel. The smoothed version requires communica- tion between adjacent time segments after graph clustering but before the cluster resolution step, where the clouds of clusterings obtained by the K iterations are exchanged. 2.3. Visualization of Output Finally, after producing a representative clustering at each time segment with smoothing between adjacent seg- ments, we produce a visualization showing how the com- munity structure changes over time. An example of the (cid:18) N (cid:19)(cid:19) (cid:18) visualization obtained from a real world network can be seen in Figure 7. In this community graph, the nodes represent communities with the node diameter indicating community size, and the communities are connected in sequential time segments if they share members with a thicker edge in- dicating more overlap between the two sets of members. This visualization allows for a greater understanding of the evolution of community structure than merely reporting on the number of communities at each time segment. Through it, community evolution can be observed, including events such as communities merging, splitting, forming out of background noise, and dissolving into background noise. 3. Experiments To validate our method we performed a series of experi- ments on synthetic, semi-synthetic, and real-world datasets. We evaluated the clustering methods using both pairwise- precision and pairwise-recall, as well as number of blocks reported where applicable. The pairwise-precision measures the percentage of node-pairs which are correctly identified as being in the same cluster: (true positives)/(true positives + false positives). Pairwise-recall considers all pairs of nodes belonging to the same ground-truth cluster and measures the fraction of them which are in the same identified clus- ter: (true positives)/(true positives + false negatives). More information can be found in [14], [15]. 3.1. Synthetic Experiments The first experiment comparing the ensemble model to the baseline is performed on a synthetic dataset where edges emerge from a network stream over time and the clustering uses all edges observed up to the current time. The block model from which the edges are sampled is considered constant; observing more edges merely increases the sample size and considerations like model drift are ignored. The data consists of 500 node synthetic graphs: each of these has 8 ground truth blocks with a 1:5 external/internal edge ratio, and the number of edges observed was scaled from 1000 to 1900. The results from this experiment are shown in Figure 3. Red lines represent the baseline MCMC algorithm while blue lines are the ensemble algorithm; smoothing was not used as it is designed for dynamic models not static ones. The dashed lines show the standard errors over ten runs of the algorithms. The left-hand panel reports the number of clusters found by both algorithms. Here, the black line shows the true number of clusters. The middle panel and right-hand panels report the pairwise-precision and pairwise- recall, respectively, as more edges are sampled. With fewer than approximately 1400 edges sampled neither method has good performance due to the number of blocks being under estimated. However, as more edges are sampled, the ensemble method has strictly superior pairwise-precision. The pairwise-recall performance appears about the same for both the baseline and the ensemble. Figure 3: Performance on emerging edges experiment. Red lines represent the baseline algorithm and blue represent the ensemble approach, with dashed lines showing the standard error over 10 runs of the algorithms. The left-hand panel shows the number of blocks reported (compared to ground truth shown by the black line) as more edges are sampled. Pairwise-precision among nodes as more edges are sampled is shown in the middle panel. The right-hand panel shows the change in pairwise-recall among nodes as more edges are sampled. structure (until the cluster has fully split). As the clustering algorithm will misidentify nodes in the splitting cluster as being in the same community, the precision suffers. Likewise at the start of the merge event the number of communities is underestimated, placing members of the merging communities as separate which causes recall to drop. For number of blocks found, the ensemble method per- forms slightly better than the baseline, especially in more quickly identifying a split. The smoothed ensemble matches the ground truth cluster number almost exactly. In terms of the pairwise-precision, the ensemble algorithm has strictly superior precision to the basic algorithm while the smoothed algorithm is better still. The same relationship holds for recall but the difference is not statistically significant at all times. 3.2. Semi-Synthetic Experiment The third experiment is performed on a semi-synthetic dataset which was produced by inserting nodes and edges sampled from a block model into a real-world network, the "email-Eu-core-temporal" dataset from the SNAP database [16]. The synthetic blocks inserted consist of a 100-node block which splits into two 50-node blocks between time segments 2 and 4 and two 50-node blocks which merge into one 100-node block between time segments 6 and 8. This is similar to the dynamic structure in the previous experiment, except that now nodes from the synthetic blocks are ran- domly connected to nodes in the real-world email dataset, rather than nodes from other synthetic blocks. This more accurately measures the ability of the algorithms to identify dynamic community structure under conditions which more closely represent those found in real-world data. The synthetic data has 960 edges per time segment, 800 of those being between pairs of synthetic nodes and 160 going to nodes in the real-world data; the real-world data consists of the first 10 weeks of the dataset which has 986 nodes and 38330 edges, split fairly evenly across those weeks. Pairwise-precision and pairwise-recall are reported for the synthetic nodes only. Figure 4: A high level view of the evolving community dynamics from the second experiment on synthetic data. The blocks are shown in the "spy plot" of the adjacency matrices associated with the underlying block model. The second experiment is another synthetic dataset, but here the edges are drawn from a dynamic model where the block structure is changing over time. The model consists of a 100-node block which splits into two 50-node blocks between time segments 2 and 4; two 50-node blocks which merge into one 100-node block between time segments 6 and 8; five 50-node blocks which remain constant; and 50 nodes which connect to any node with equal probability, for a total of 500 nodes. 2000 edges were sampled for each time segment. A high level view of these dynamics is shown in Figure 4. The results of this experiment are shown in Figure 5 Here, the red and blue lines represent the baseline and ensemble algorithms, respectively, as in Figure 3. The purple lines represent the ensemble with smoothing. The gray regions represent times when the block structure is in flux; the ground truth is set to be that of the blocks at the end of the transition, hence the drop in performance at the start of each region when the behavior observed has yet to fully align with the new ground truth. The left-hand panel in Figure 5 reports the number of clusters found by the various algorithms used. The black line represents the number of clusters in the ground-truth clustering. The middle panel reports the pairwise-precision over time and the right-hand panel shows the pairwise-recall. At the start of the split event, the precision drops as the ground truth has more clusters than the immediate graph Figure 5: Performance on synthetic split and merge dynamic event. Red and blue lines represent the baseline algorithm and the ensemble, respectively, with dashed lines showing the standard error over 10 experiments, as in Figure 3. Purple lines, solid and dashed, show the same for the ensemble approach with smoothing between time segments. The grey regions represent time periods during which the underlying stochastic block model is changing. The left-hand panel shows the number of blocks identified in each time segment using the baseline and the ensemble as compared to the ground truth, shown by the black line. The middle panel shows the pairwise-precision for all three methods over time and the right-hand panel shows the pairwise-recall. Figure 6: Performance on semi-synthetic split and merge dynamic event. Red, blue, and purple lines represent the baseline MCMC clustering algorithm, the ensemble approach, and the ensemble approach with smoothing, respectively, with the dashed lines showing the standard error over 10 runs. The grey regions represent time periods during which the underlying stochastic block model on the synthetic nodes is evolving. The left-hand panel shows the number of clusters found by each algorithm over time, the middle panel shows the pairwise-precision for the synthetic nodes, and the right-hand panel shows the pairwise-recall for the synthetic nodes. The results of this experiment on semi-synthetic data are shown in Figure 6. As previously, the red, blue, and purple lines represent the baseline MCMC, ensemble, and ensemble with smoothing algorithms, respectively. The left- hand panel shows the number of clusters reported by all three algorithms. Clusters discovered in the real data are included in this count, which is the cause of the drop in cluster number in the final two large merges occurred in the real dataset at this time reducing the overall number of clusters. These merges are visible in the community structure shown in Figure 7. time segment: The grey bars highlight the time segments in which the clusters of synthetic nodes are undergoing a split or merge. Unlike the previous experiment, the ground-truth number of clusters is not reported due to the fact that the ground truth communities are not known a priori for the "email-Eu-core-temporal" dataset. The middle and right- hand panels show the pairwise-precision and pairwise-recall over time. In this experiment the pairwise-precision does not seem significantly different across the different algorithms, however the recall for the baseline algorithm is much worse than either the ensemble or the smoothed ensemble. 3.3. Real-World Experiment Finally, we ran the smoothed ensemble framework on the first 10 weeks of the EU email dataset. The dynamic community structure obtained is shown in Figure 7. The horizontal axis represents time (in weeks), with community detection taking place weekly. The points represent the communities found during each week, with the diameter of the points representing the size of the communities at a given time and the thickness of the lines between communities in different weeks representing the node set overlap between two communities. Many of the communities are constant, forming long horizontal lines with no splitting or merging behavior. Others undergo prominent splitting and/or merging behavior. There are also a number of communities which appear to "die" over the course of the ten weeks, with the community terminating without merging into another. Of particular note are the pair of large merge events occurring in the final time segment where multiple communities join to form large blocs. graphs so sparse as to be useless. Other possible improve- ments exist for the visualization of the dynamic network: for example, the variability in the node membership of the representative clusters could be visualized to represent the level of "confidence" the algorithm has in a particular cluster assignment. Acknowledgements This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. References [1] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang, "Complex networks: Structure and dynamics," Physics Reports, vol. 424, pp. 175–308, 2006. [2] U. Brandes and T. Erlebach, eds., Network Analysis: Methodologi- cal Foundations, vol. 3418 of Lecture Notes in Computer Science. Springer, 2005. [3] M. E. J. Newman, Networks: An Introduction. Cambridge University [4] [5] Press, 2010. S. Fortunato, "Community detection in graphs," Physics Reports, vol. 486, no. 3, pp. 75–174, 2010. F. D. Malliaros and M. Vazirgiannis, "Clustering and community detection in directed networks: A survey," Physics Reports, vol. 533, no. 4, pp. 95–142, 2013. Figure 7: Timeline of the clustering behavior of SNAP European University dataset over the first 10 weeks of data using the ensem- ble algorithm with smoothing. 4. Conclusions The ability to perform dynamic clustering on real-world networks and produce an interpretable visualization is a powerful tool for network analysis. The ensemble frame- work presented in this paper has the ability to produce high- level summations of the evolving community structure of dynamic networks and can be applied to large networks due to its efficient parallel implementation . We also presented a version of the ensemble with memory which allows for more stable representations of the community structure. We demonstrated using synthetic, semi-synthetic and real-world scenarios that the ensemble algorithm has superior precision and recall performance compared to the basic clustering algorithm, and we also presented the types of dynamic community representations that the ensemble algorithm can produce from real graphs. Future directions for this framework include incorporat- ing more information from previous time segments when performing clustering: for example, edges from graphs in previous segments could be included in the current graph us- ing an exponential decay scheme. This would allow a more narrow resolution of the time segments without making the [6] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, "Fast unfolding of communities in large networks," Journal of statistical mechanics: theory and experiment, vol. 10, p. P10008, 2008. [7] T. P. Peixoto, "Efficient monte carlo and greedy heuristic for the inference of stochastic block models," Physical Review E, vol. 89, no. 1, p. 012804, 2014. [8] T. P. Peixoto, "Parsimonious module inference in large networks," Physical Review Letters, vol. 110, no. 14, p. 148701, 2013. [9] T. P. Peixoto, "Entropy of stochastic blockmodel ensembles," Physical Review E, vol. 85, no. 5, p. 056122, 2012. [10] D. Greene, D. Doyle, and P. Cunningham, "Tracking the evolution of communities in dynamic social networks," in International Con- ference on Advances in Social Networks Analysis and Mining, 2010. [11] J. Sun, S. Papadimitriou, P. S. Yu, and C. Faloutsos, "Graphscope: parameter-free mining of large time-evolving graphs," in Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM, 2007. [12] Y. Hong, C. Fu, Q. Huang, Z. Fang, J. Zeng, and L. Han, "Com- munities evolution analysis based on events in dynamic complex network," in IEEE 14th Intl Conf on Dependable, Autonomic and Secure Computing, 14th Intl Conf on Pervasive Intelligence and Computing, 2nd Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology Congress, 2016. [13] S. Pandit, D. H. Chau, S. Wang, and C. Faloutsos, "Netprobe: a fast and scalable system for fraud detection in online auction networks," in Proceedings of the 16th international conference on World Wide Web, pp. 201–210, ACM, 2007. [14] W. M. Rand, "Objective criteria for the evaluation of clustering methods," Journal of the American Statistical Association, vol. 66, no. 336, pp. 846–850, 1971. [15] L. Hubert and P. Arabie, "Comparing partitions," Journal of classifi- cation, vol. 2, no. 1, p. 193218, 1985. [16] J. Leskovec and A. Krevl, "SNAP Datasets: Stanford large network dataset collection." "http://snap.stanford.edu/data", June 2014.
1712.01690	1	1712	2017-12-03T02:24:41	Size Matters: A Comparative Analysis of Community Detection Algorithms	[ "cs.SI", "physics.soc-ph" ]	Understanding community structure of social media is critical due to its broad applications such as friend recommendations, user modeling and content personalizations. Existing research uses structural metrics such as modularity and conductance and functional metrics such as ground truth to measure the qualify of the communities discovered by various community detection algorithms, while overlooking a natural and important dimension, community size. Recently, anthropologist Dunbar suggests that the size of stable community in social media should be limited to 150, referred to as Dunbar's number. In this study, we propose a systematic way of algorithm comparison by orthogonally integrating community size as a new dimension into existing structural metrics for consistently and holistically evaluating the community quality in social media context. we design a heuristic clique based algorithm which controls the size and overlap of communities with adjustable parameters and evaluate it along with five state-of-the-art community detection algorithms on both Twitter network and DBLP network. Specifically, we divide the discovered communities based on their size into four classes called close friend, casual friend, acquaintance, and just a face, and then calculate the coverage, modularity, triangle participation ratio, conductance, transitivity, and the internal density of communities in each class. We discover that communities in different classes exhibit diverse structural qualities and many existing community detection algorithms tend to output extremely large communities.	cs.SI	cs	Size Matters: A Comparative Analysis of Community Detection Algorithms Paul Wagenseller III Feng Wang School of Mathematical and Natural Sciences School of Mathematical and Natural Sciences Arizona State University [email protected] Arizona State University [email protected] 7 1 0 2 c e D 3 ] I S . s c [ 1 v 0 9 6 1 0 . 2 1 7 1 : v i X r a Abstract-Understanding community structure of social me- dia is critical due to its broad applications such as friend recommendations, user modeling and content personalizations. Existing research uses structural metrics such as modularity and conductance and functional metrics such as ground truth to measure the qualify of the communities discovered by various community detection algorithms, while overlooking a natural and important dimension, community size. Recently, anthropologist Dunbar suggests that the size of stable community in social media should be limited to 150, referred to as Dunbar's number. In this study, we propose a systematic way of algorithm comparison by orthogonally integrating community size as a new dimension into existing structural metrics for consistently and holistically evaluating the community quality in social media context. we design a heuristic clique based algorithm which controls the size and overlap of communities with adjustable parameters and evaluate it along with five state-of-the-art community de- tection algorithms on both Twitter network and DBLP network. Specifically, we divide the discovered communities based on their size into four classes called close friend, casual friend, acquaintance, and just a face, and then calculate the coverage, modularity, triangle participation ratio, conductance, transitivity, and the internal density of communities in each class. We discover that communities in different classes exhibit diverse structural qualities and many existing community detection algorithms tend to output extremely large communities. Index Terms-community detection, Dunbar's number, over- lapping community, Clique I. INTRODUCTION Community is a natural and fundamental element that exists in a wide variety of networked systems, such as biology networks, social networks, and neural networks. Identifying communities in these networks is a crucial step for gaining an in-depth understanding on network structure, dynamics and interactions. However, there is no unique and widely accepted goodness measure of community quality in literature. Informally, a good community is a densely-connected group of nodes that is sparsely connected to the rest of the network. Extensive research has been devoted to designing community detection algorithms to uncover communities with the goal to minimize or maximize structural metrics, such as modularity, triangle participation ratio, or conductance of the discovered communities. Many comparative research [4], [6], [7], [9], [12] has also been conducted to evaluate and compare state-of- the-art community detection algorithms from the perspective of different structural metrics. In parallel, [10] validates the communities discovered by different community detection algorithms with a functional metric called ground truth, which is the a priori knowledge of user's community membership. However, the majority of previous research in community detection algorithm design and evaluation has overlooked an important metric, community size, which has significant mean- ing in the context of social network. The primary objective of forming community in social network is for social interaction and influence, while it is hard for large communities to achieve such goal. Anthropologist Dunbar [1] suggests that the size of communities with strong ties in both traditional social networks and Internet-based social networks should be limited to 150 (called Dunbar's number) due to the cognitive constraint and time constraint of human being. Large communities of size over 150 contain weak connections among their members therefore are not stable, while small communities of size 2 or 3 cannot provide the strong sense of team or community. Therefore, we refer to communities of size 4 to 150 as desirable community in this paper and carry out extensive experiments to systematically evaluate and compare various community detection algorithms taking the size of identified communities into consideration. To make the comparison results reproducible, we adopt the following five well known community detection algo- rithms implemented in igraph package: Infomap [15], Mul- tilevel [16], FastGreedy [21], Eigenvector [19], and Label Propagation [18]. Additionally, we propose a clique based al- gorithm called Clique Augmentation Algorithm (CAA) which augments the cliques in the network into communities while using growing threshold and overlapping threshold to control the size of the community and the amount of overlap among communities. We then evaluate these six algorithms on two datasets with different characteristics, one is a Twitter follower graph and the other is a DBLP network [5]. Specifically, we first study the size distribution of the communities identified by each algorithm, then calculate the percentage of users that are assigned to desirable communities for each algorithm. Next, we divide the communities into four different size groups and compare the goodness of communities in each group with the following metrics: 1) extended modularity, 2) triangle participation ratio (TPR), 3) conductance, 4) internal density, and 5) transitivity. Our experimental results reveal that communities in differ- ent groups exhibit diverse structural qualities. We discover that Infomap and CAA outperform others in terms of both community size and coverage by consistently producing de- sirable communities covering a large portion of users in both Twitter and DBLP networks. Nevertheless, both modularity maximization algorithms including Multilevel, FastGreedy, and Eigenvector and node labeling algorithm Label Propa- gation tend to output extremely large communities of size over 30,000. Specifically, the largest community identified by Eigenvector holds 71.6% of all users in Twitter graph and 99.5% of all users in DBLP network respectively, therefore lacking practical significance. For modularity, Infomap outperforms others because unlike Multilevel, FastGreedy, and Eigenvector, whose modularity scores are mostly decided by the large communities of size [501+], Infomap achieves high score that is contributed by all four groups of it's communities. For conductance, modularity maximization algorithms perform better than other algorithm indicating their identified communities are better separated from the external compared to other algorithms. With regard to triangle participation ratio, CAA outperforms others reaching above 90% in DBLP network and close to 90% in Twitter network. All algorithms perform comparably in terms of internal connectivity metrics, transitivity and internal density, with CAA performs slightly better for desirable community sizes. Our contributions in this paper are twofold: 1) We introduce Dunbar's number into community detec- tion and evaluation research 2) We carry out extensive experiment to systematically evaluate and compare various community detection al- gorithms taking the size of identified communities into consideration. To the best of our knowledge, this is the first effort to investigate this overlooked but important metric in a systematic way. Even many online communities have hundreds and even thousands of members, most of those members are inactive, non-participatory in the group on a regular basis. Evidence of Dunbar's number exists ubiquitously, from the size of villages, to the size of modern hunter-gatherer societies, to the number of active members in online gaming communities and online forum, and to the number of active administrators in Wikipedia. Additionally, technical forum tends to break down when it reaches about 80 active contributors, requiring a forum split before continued growth could occur [3]. By scrutinizing the top 5000 highest quality communities provided in DBLP, a co-authorship network [5], we found that 4951 of them are of size 4 to 150. This means 99% of the highest quality ground-truth communities in DBLP network conform to Dunbar's number. Furthermore, The dataset claims 13,477 ground truth communities in total and 94% of them have size of 4 to 150. This is another strong evidence of Dunbar's number. In addition to the well-known number of 150, [1] also gives a series of numbers which correspond to the closeness of a person to people he interacts. According to Dunbar's model, each person can maintain about five people in their support group of closest friends, about 15 people in a sympathy group who are close enough to confide in, about 50 close friends, about 150 casual friends, and about 500 acquaintances. In all, it indicates that any given human can identify about 1,500 faces in total. We illustrate these social circles in Table I. TABLE I: Social Circle Size 4-5 6-15 16-50 51-150 151-500 501-1500 Closeness Support clique Sympathy group Close friend Casual friend Acquaintance Just a face The remainder of this paper is organized as follows. Section 2 describes the importance of size in social networks. Section 3 introduces the proposed clique augmentation algorithm. Performance comparison is given in Section 4. Section 5 summarizes related work in community detection algorithm design and evaluation, and Section 6 concludes this paper and outlines our future work. II. COMMUNITY SIZE MATTERS IN SOCIAL MEDIA Community size is an important factor in social media. On one hand, groups of size 1, 2, and 3 are too small to be called a community. On the other hand, large communities cannot facilitate communication or interaction therefore members in the community have limited influence upon each other. Dunbar's number of 150, which is the limit of stable online community size, is recently introduced to social media [1]. It is calculated as a direct function of relative neocortex size which decides the neocortical processing capacity that limits the number of individuals with whom a stable inter-personal relationship can be maintained, and this in turn limits group size. In this paper, seeking to gain a better understanding of how community goodness metrics change with relation to community size, we use these numbers as guideline to di- vide the identified communities into different size groups. Specifically, to conform to Dunbar's breakdown and to avoid overcrowded figures for clear presentation, we group the first three categories in Table I into one group. In this way, the six categories are turned into four classes of communities of size [4-50], [51-150], [151-500], and [501+] respectively. We call these four classes close friend, casual friend, acquaintance, and just a face communities. It is worth noting that we are not stating that communities of large size are not important. Large communities are not prac- tical in social media, but they have their own applications in other types of networks such as biological network and neural network. Furthermore, we are not reducing the significance of a community only to its size, but it is worth to give in-depth analysis of the quality of communities an algorithm produces by adding community size as an orthogonal dimension to the existing structural measurements. III. CLIQUE AUGMENTATION ALGORITHM In this section we propose a clique based community detection algorithm called Clique Augmentation Algorithm (CAA). CAA is built on the following two principles: 1) Users in a maximal clique belong to a stable community since a clique is densely connected internally; 2) A neighboring node that is highly connected to a clique should be part of the community since it keeps the triadic closure property among all nodes in the community. Given a social network topology, CAA algorithm discovers communities in the topology using the following steps: 1) Find all maximal cliques in the topology; 2) Filter the overlapping cliques. We sort the cliques based on their size then use an overlapping threshold to control the amount of overlap between two cliques. The overlapping threshold is defined as the percentage of overlapping nodes in the smaller clique. For example, given two cliques c1 and c2 where c1 is of size 10, and c2 is of size 5. Suppose the overlapping threshold is 0.7. If c2 only has 2 nodes overlapping with c1, we consider c1 and c2 as two independent cliques since 2 < 5 ∗ 0.7 which is 3.5. If c2 had 4 overlapping nodes with c1, we would discard c2 since 4 > 3.5; 3) Grow each clique into a community by adding new nodes in one by one. Growing threshold is utilized for controlling the growth of each community. The growing threshold is defined as the ratio of the number of incoming edges from the new node to other nodes in the community over the size of a community. For example, if a community has size of 10, and the growing threshold is set to 0.7, then for a neighboring node to be added into the community, it must have at least 7 edges coming into the community. The algorithm checks the neigh- boring nodes for each node within the current community. This process is repeated for the updated community until no more nodes can be added. The growing threshold allows us to zoom in or out of the graph around the clique. An analysis of the effect of overlapping and growing threshold is conducted in Section IV-C. The implementation detail of the algorithm is presented in Algorithm 1 The time complexity of CAA is dominated by the time com- plexity of maximal clique algorithm, which is NP-Complete. However, consider the moderate clique size in social network, maximal clique algorithm implemented in NetworkX package runs in polynomial time in practice. After the maximal cliques are obtained, the time to filter overlapping cliques is O(N 2) where N is the number of nodes. The time to augment clique into communities also takes O(N 2). It is worth noting that CAA takes a different approach than Clique Percolation Method (CPM) where two adjacent cliques are merged into a community structure. In CAA, instead of merging neighboring cliques, we simply grow the commu- nity structure by adding individual node to the community sequentially. CAA has a few nice features: 1) it is faster than CPM and manages to produce similar results. 2) CAA captures the natural growth process of a community in the sense that if a user befriends with many users in a densely connected Algorithm 1 Clique Augmentation Algorithm 1: INPUT: A graph G, overlapping threshold ω, growing threshold φ 2: OUTPUT: communityList in graph G 3: cliques = findMaximalCliques(G) 4: filteredCliqueList = filterOverlappingCliques(cliques, ω) 5: foreach clique c in filteredCliqueList: 6: 7: 8: 9: 10: 11: 12: growingThreshCount = (len(curCommunity) -1) * φ) neighborList = [] foreach neighbor w of v: if w not in curCommunity && w not in neighborList curCommunity = [] curCommunity.add(c.nodes()) foreach node v in curCommunity: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: add w to neighborList growingList = [] foreach node u in neighborList: incomingEdgeCount = 0 foreach neighbor w of u: if w is in curCommunity incomingEdgeCount++ if incomingEdgeCount ≥ growingThreshCount add node u to growingList curCommunity.add(growingList) communityList.add(curCommunity) community, the user will be most likely grow as part of the community. IV. EMPIRICAL ANALYSIS A. Description of Data Sets We carry out the comparative research on two datasets: 1) Twitter user follower topology collected over a 3-month period in summer 2013. It contains 318, 233 Twitter users with 3, 545, 258 directed edges. Because not every community detection algorithm we measure in this paper supports directed graph, we derive an undirected graph by removing all non- mutual edges and the isolated nodes. We call the undirected graph Twitter network in this paper. It contains 190, 520 nodes and 1, 001, 528 undirected edges. 2) DBLP dataset [5] which is a co-authorship network in computer science where two au- thors are connected if they publish at least one paper together. Each node represents an author and each edge indicates co- authorship of a paper. There are totally 317,080 nodes and 1,049,866 undirected edges in DBLP network. To better understand the characteristic of these two net- works, we plot the log-log graph for their node degree dis- tribution in Fig. 1. X-axis is the rank of a node, y-axis is the degree of the node, and a node with larger degree has lower rank. It is clear that node degrees in both networks follow power law distribution. Furthermore, these two networks are different in density and users in Twitter network have more neighbors than users in DBLP network in general. Name CAA Infomap Multilevel Eigenvector FastGreedy Label Propagation TABLE II: Summary of Algorithms Exponential, f ast in practice f or small cliques Time Complexity O(E) O(N log N ) O(N (E + N )) O(N log2(N )) O(E) Overlap Yes No No No No No Category Clique based Network coding and random walk Modularity maximization Modularity maximization Modularity maximization Node labeling implementation of community detection algorithms in open- access python package igraph and NetworkX respectively. More specifically, we run Infomap, Multilevel, FastGreedy, Eigenvector, Label Propagation, Edge Betweenness, Walk- Trap, and SpinGlass algorithms in igraph, Clique Percolation Method (CPM) [20] algorithm in NetworkX, and CAA algo- rithm on both Twitter network and DBLP network. In our experiment, we set the growing threshold of CAA to 0.7 and its overlapping threshold to 0. The communities grow from cliques of size ≥ 3. However, only Infomap, Multilevel, FastGreedy, Eigenvector, Label Propagation, and CAA can handle the scale of the data. Other algorithms hang due to their time complexity and the limitation of the hardware configuration of our experiment environment. Table II summa- rizes the six algorithms that finish running. These algorithms can be classified into four categories based on their design ideas as follows: clique based (CAA), network coding and random walk (Infomap), modularity maximization (Multilevel, FastGreedy, Eigenvector), and node labeling (Label Propaga- tion). In Table II, overlap indicates whether an algorithm can produce overlapping communities which is more desirable for social media since a user in social media usually belongs to multiple communities. Overlapping community detection is not as extensively studied as disjoint algorithms, thus we are unable to find open source implementation of overlapping algorithms to compare besides CPM in NetworkX. However, we are unable to run CPM on our datasets due to the scale of the data. TABLE III: Total number of detected communities and the size of the largest community in Twitter network Algorithm CAA Infomap Multilevel Eigenvector FastGreedy Label Propagation Number of Communities 12,312 18,537 7,409 5,834 9,350 7,648 Largest Community Size / Ratio 680 / 0.35% 13,126 / 6.89% 34,955 / 18.35% 136,403 / 71.6% 51,781 / 27.18% 149,022 / 78.22% TABLE IV: Total number of detected communities and the size of the largest community in DBLP network Algorithm CAA Infomap Multilevel Eigenvector FastGreedy Label Propagation Number Of Communities 28,213 16,999 565 2 3,206 21,156 Largest Community Size / Ratio 270 / 0.085% 587 / 0.185% 30,427 / 9.6% 315,569 / 99.5% 54,787 / 17.28% 84,363 / 26.6% 1) Community Size Distribution: Table III and IV summa- rizes the number of communities and the size of the largest Fig. 1: Node degree distribution of Twitter network and DBLP network Fig. 2: Desirable Community Coverage B. Comparison Of Community Detection Algorithms As we mentioned in Section II, although a lot of research effort has been devoted to the comparison of different com- munity detection algorithms and proposed different evaluation metrics, they ignore a very important factor, the size of the community. In this section, we first study the size distribution of the detected communities by different algorithms and calcu- late the percentage of users assigned to a desirable community of size between 4 to 150. Then we divide the community by their size into four groups of [4-50], [51-150], [151-500], and [501+] and compare the quality of communities in each group with the following criteria: extended modularity, triangle participation ratio (TPR), conductance, internal density, and transitivity. To make our results reproducible, we adopt the 0%10%20%30%40%50%60%70%80%90%100%CAAInfoMapFastGreedyMultiLevelEigenVectorLabelPropPercentageAlgorithmTwitterDBLPTABLE V: Community Size Distribution in Twitter network FastGreedy Label Propagation Size Range 1 - 3 4 - 50 51 - 150 151 - 500 501+ CAA 241 11,359 641 69 2 Infomap Multilevel 8,811 9,394 223 91 18 6,612 754 11 8 24 Eigenvector 5,468 364 0 0 2 TABLE VI: Community Size Distribution in DBLP network Infomap Multilevel FastGreedy Label Propagation Size Range 1 - 3 4 - 50 51 - 150 151 - 500 501+ CAA 578 27,176 454 5 0 757 15,475 751 14 2 0 415 5 18 127 Eigenvector 0 0 0 0 2 7,416 1,889 20 10 15 258 2,782 85 35 46 6,039 1,546 36 22 5 3,065 17,737 329 24 1 community revealed by each algorithm. We also include the percentage of users that are assigned to the largest commu- nity. As can be seen, modularity maximization algorithms (Multilevel, FastGreedy, Eigenvector) and Label Propagation all produce extremely large communities for both networks. For example, the largest community Eigenvector produces in Twitter network has size of 136, 403, that is, over 70% of all users are grouped into one large community. In DBLP network, Eigenvector algorithm produces largest community of size 315, 569, that is 99.5% of all users are grouped into one large community. There lack of strong connections among community users in such large communities and we can hardly put communities of such large size to practical use. Table V and VI provide more detailed breakdown of the number of communities in different size ranges. As can be seen, CAA and Infomap produce more communities in desir- able size of 4-150 than others. Another interesting observation is that Label Propagation algorithm performs significantly better on DBLP network than on Twitter network. Eigenvector performs poorly on both networks, especially on DBLP net- work since all it discovers are two large communities of size greater than 501. As we can see in Figure 3-7, Eigenvector does not have any score in range [4-50], [51-150] and [151- 500] for DBLP network since it does not generate communities in those size ranges. From the community size distribution point of view, we would recommend CAA and Infomap which can produce decent number of communities with desirable sizes. 2) Community Coverage: In this section, we define a new metric called desirable community coverage to measure the number of users assigned to desirable communities, that is, communities of size 4 to 150. Generally speaking, the higher the desirable community coverage, the better the al- gorithm. Note that even Infomap, Multilevel, Eigenvector, and FastGreedy all assign every single node in the graph to a community, they do not provide 100% coverage in our definition of desirable community coverage. Fig. 2 shows the performance of each algorithm with regard to desirable community coverage while taking the community size into account. It is clear that Infomap performs the best in terms of desirable community coverage since it assigns 62% of all users in the Twitter network and 98% of all users in DBLP network into desirable communities. CAA also performs well in both networks when compared to other methods. It is worth noting that in this experiment, CAA starts from non-overlapping cliques by setting the overlapping threshold to 0. In fact, CAA can achieve higher coverage if we allow overlapping cliques to start with. Label Propagation performs inconsistently in two networks and FastGreedy, Multilevel, and Eigenvector failed to assign the majority of users in the network to meaningful communities. 3) Extended Modularity: Extended modularity has been defined as a popular metric to measure the goodness of overlapping communities in [14]. We give the definition of this metric in Equation (1), where Ci is the ith community, Ov is the number of communities the vertex v belongs to, similarly Ow is the number of communities vertex w belongs to. A is the adjacency matrix, that is, Avw = 1 means there exists an edge between vertex v and vertex w, otherwise it is 0. kvkw 2m describes the expected number of edges between vertex v and vertex w. kv and kw are the node degree of v and w in the whole topology respectively. m is the total number of edges in the whole topology. The range of extended modularity is [−1, 1] and the higher the value, the better the community in terms of modularity. The intuition behind extended modularity is that communities should have more internal connectivity than random graph of the same degree sequence. Extended modularity is a variation to modularity, which is the most widely used metric to measure the goodness of disjointed community. If there is no overlap between the communities, is 1 and Equation (1) becomes the traditional modularity OvOw as defined in [23]. In general, overlapping communities have lower modularity than disjoint communities since overlapping communities have many connections outside the community. 1 (cid:88) (cid:88) (cid:104) EQ = 1 2m v∈Ci,w∈Ci i OvOw 1 Avw − kvkw 2m (1) (cid:105) As can be seen from Equation (1), each community con- tributes a value towards the modularity score. To measure to what degree communities of different sizes impact the modularity, we divide the sum in Equation (1) into four parts, that is, the sum over all communities of size [4-50], [51-150], [151-500], and [501+] and present the partial modularity score contributed by each community group in Figure 3. Infomap outperforms others because unlike Multilevel, FastGreedy, and Eigenvector, whose modularity scores are mostly decided by the large communities of size [501+], Infomap achieves high score that is contributed by all four groups of it's communities. its modularity score is very close to the highest In fact, score achieved by Multilevel. The observation that larger communities achieve higher modularity score is consistent with the resolution limit of modularity as indicated in [17]. This discovery also suggests we should maximize modularity while control the community size at the same time. It is also worth noting that in general overlapping communities have lower modularity than non-overlapping ones thus the low modularity of CAA is expected behavior. Fig. 3: Extended Modularity Fig. 4: Triangle Participation Ratio 4) Triangle Participation Ratio: TPR was proposed in [10] as a metric for community evaluation, where it is defined as the number of nodes in a community that form a triad, divided by the total number of nodes in the community. [10] found that the communities discovered by algorithms to optimize community modularity do not align with the ground truth communities, that is, the known membership in real life, while TPR is a good metric when looking for ground- truth communities. As indicated by Fig. 4, CAA achieves significantly higher TPR than others on both networks, with close to 0.9 TPR score on Twitter network and above 0.9 TPR score on DBLP network. This is expected since CAA starts from a clique with TPR as 1 and grows the clique with new nodes that are highly connected to the growing community. We can also see than all algorithms show satisfactory performance on DBLP network in terms of TPR. 5) Conductance: Conductance is a metric that takes into consideration both external and internal connections of a com- munity [12]. It is defined as the ratio of the number of edges on the boundary of the community over the sum of degrees of nodes in the community. Lower conductance score indicates better community. We calculate the average conductance of communities in different size groups in Fig. 5. We find that Multilevel achieves the lowest average conductance score and all three modularity maximization algorithms perform well. CAA performs poorly in conductance as expected since it produces overlapping communities. 6) Internal Density: Internal density [12] is defined in Equation (3), where mS is the number of edges in the subset S divided by the total possible edges between all nodes nS(nS−1)/2. Internal density is a measurement of the internal structure within the community. f (S) = mS nS(nS − 1)/2 (3) Fig. 6 gives average internal density for communities in each size group. We find that smaller community size typically indicates higher internal density value. This is not surprising as smaller community sizes tend to form cliques. All algorithms show similar performance and CAA consistently performs slightly better across all size groups on DBLP network. 7) Transitivity: Transitivity [24] measures ties between individuals and is defined in Equation (4). This value indicates the probability that two friends of a single person are also friends. A clique has transitivity of 1.0. Higher transitivity value is better since it indicates that the community is more 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropModularityAlgorithmTwitter[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropModularityAlgorithmDBLP[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropTriangle Participation RatioAlgorithmTwitter[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropTriangle Participation RatioAlgorithmDBLP[4-50][51-150][151-500][501+]Fig. 5: Conductance Fig. 6: Internal Density tightly connected to one another and has a higher probability that users within the community know each other. 3 × number of triangles number of triads (4) The result of the average Transitivity score is presented in Fig. 7. We notice that CAA performs the best in range [4-50] for both networks. All algorithms perform similarly in range [51-150]. In summary, the experimental results suggest that Infomap and CAA outperform others in terms of both community size and coverage by consistently producing desirable communi- ties covering a large portion of users in both Twitter and DBLP networks. Nevertheless, both modularity maximization algorithms including Multilevel, FastGreedy, and Eigenvector and node labeling algorithm Label Propagation tend to output extremely large communities of size over 30,000. In terms of modularity, InfoMap performs the best in general. The identified community of different sizes all contribute to the modularity score and its total modularity is comparable to modularity maximization algorithms. For conductance, mod- ularity maximization algorithms perform better than other algorithm indicating their identified communities are better separated from the external compared to other algorithms. With regard to triangle participation ratio, CAA outperforms others reaching above 90% in DBLP network and close to 90% in Twitter network. All algorithms perform similarly in terms of internal connectivity metrics, transitivity and inter- nal density, with CAA performs slightly better for desirable community sizes. C. Growing Threshold and Overlapping Threshold In this section, we measure the impact of growing threshold and overlapping threshold on community size and the number of communities in order to give suggestions on the parameter selection for CAA algorithm. First we look into the effect of growing threshold. We find all cliques of size 3 and larger in Twitter network and set overlapping threshold to 0 to find all non-overlapping cliques. We set the growing threshold to 0.5, 0.7, and 0.9 indicating a neighboring node can only be added to the community if it is connected to at least 50%, 70%, or 90% of nodes in the community. The result is plotted in Fig. 8a. x-axis is the community size range and y-axis is the number of communities whose size fall in the range. It can be seen that the cliques do not grow much with growing threshold 0.9 since most communities are of size between 3 and 9. With growing 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropConductanceAlgorithmTwitter[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropConductanceAlgorithmDBLP[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropInternal DensityAlgorithmTwitter[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropInternal DensityAlgorithmDBLP[4-50][51-150][151-500][501+]Fig. 7: Transitivity (a) The Effect of Growing Threshold (b) The Effect of Overlapping Threshold Fig. 8: CAA Parameter Analysis threshold set to 0.7, the number of communities with size in range [3-9] drops significantly, from around 11,000 to 5,000 and there are more communities in the range of [10-150]. Therefore, we recommend to set the growing threshold to 0.7. Another interesting observation is that there is no significant difference in the distribution of community size for growing threshold 0.5 and 0.7. Next we investigate the effect of overlapping threshold. We choose all cliques of size over 15 and increase the overlapping threshold from 0 to 1. Intuitively, by increasing the overlapping threshold, less cliques are filtered, therefore the number of communities increases. As can been seen in Fig. 8b, where x-axis is the overlapping threshold value, y-axis is the number of cliques, the number of cliques increases significantly for overlapping threshold ≥ 0.8. In general, we suggest to choose overlapping threshold less than 0.6 to avoid having heavily overlapping communities. V. RELATED WORK A lot of effort has been devoted to the area of community detection along with ways to determine the quality of the iden- tified communities. Existing community detection algorithms can be categorized into disjoint algorithms and overlapping algorithms, based on whether the identified communities have overlap or not. Infomap [15] stands out as the most popular and widely used disjoint algorithm. It is based on random walks on networks combined with coding theory with the intent of understanding how information flows within a network. Mul- tilevel [16] is a heuristic based algorithm based on modularity optimization. Multilevel first assigns every node to a separate community, then selects a node and checks the neighboring nodes attempting to group the neighboring node with the selected node into a community if the grouping results in an increase in modularity. Newman's Leading Eigenvector [19] works by moving the maximization process to the eigenspec- trum to maximize modularity by using a matrix known as the modularity matrix. Fast Greedy [21] is based upon modularity as well. It uses a greedy approach to optimize modularity. Label propagation [18] algorithm works by assigning a unique label to each node in the graph. The nodes are then listed in a random sequential order which the algorithm follows to diffuse the labels through the network. This causes nearby neighboring nodes to adopt the same label from the nodes nearest to them. This process causes community like sets of nodes to quickly 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropTransitivityAlgorithmTwitter[4-50][51-150][151-500][501+] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1CAAInfoMapMultilevelFastGreedyEigenVectorLabelPropTransitivityAlgorithmDBLP[4-50][51-150][151-500][501+] 0 2000 4000 6000 8000 10000 120000-1010-2020-3030-4040-5050-6060-7070-8080-9090-100100-110110-120120-130130-140140-150150+Community CountSize RangeGrowingThresh=0.5GrowingThresh=0.7GrowingThresh=0.9 0 5000 10000 15000 20000 25000 30000 35000 40000t=0.0t=0.05t=0.1t=0.15t=0.2t=0.25t=0.3t=0.35t=0.4t=0.45t=0.5t=0.55t=0.6t=0.65t=0.7t=0.75t=0.8t=0.85t=0.9t=0.95t=1.0Clique Count[2] http://www.socialsciencespace.com/2013/11/robin-dunbar-on-dunbar- numbers/ [3] http://www.newyorker.com/science/maria-konnikova/social-media- affect-math-dunbar-number-friendships [4] S. Fortunato, "Community Detection in Graphs", Physics Reports, 486.3-5, pp. 75-174, 2010. [5] http://snap.stanford.edu/data/com-DBLP.html [6] M. Wang, C. Wang, J. Xu Yu, and J. Zhang, "Community Detection in Social Networks: An In-depth Benchmarking Study with a Procedure- Oriented Framework," Proceedings of the 41st International Conference on Very Large Data Bases (VLDB), 2015. [7] Z. Yang, R. Algesheimer, and C. J. Tessone, "A Comparative Analysis of Community Detection Algorithms on Artificial Networks," Sc. Rep. 6, 30750, 2016. [8] Z. Ding, X. Zhang, D. Sun, and B. Luo, "Overlapping Community Detection based on Network Decomposition," Sci. Rep. 6, 24115, 2016. [9] S. Harenberg, G. Bello, L. Gjeltema, S. Ranshous, J. Harlalka, R. Seay, K. Padmanabhan, and N. Samatova, "Community Detection in Large- scale Networks: A Survey and Empirical Evaluation," WIREs Comput Stat 6.6, 2014. [10] J. Yang and J. Leskovec, "Defining and Evaluating Network Communi- ties Based on Ground-Truth," 2012 IEEE 12th International Conference on Data Mining, 2012. [11] J. Xie, S. Kelley, and B. K. Szymanski, "Overlapping Community Detection in Networks," CSUR ACM Comput. Surv. ACM Computing Surveys 45.4 (2013): 1-35. [12] J. Leskovec, K. J. Lang, and M. Mahoney, "Empirical Comparison of Algorithms for Network Community Detection," Proceedings of the 19th International Conference on World Wide Web (WWW), 2010. [13] A. Lazar, D. Abel, and T. Vicsek, "Modularity Measure of Networks with Overlapping Communities," Europhys. Lett. 90.1, 2010. [14] H. Shen, X. Cheng, K. Cai, and M. Hu, "Detect Overlapping and Hierarchical Community Structure in Networks," Physica A: Statistical Mechanics and Its Applications 388.8, 2009. [15] M. Rosvall and C. T. Bergstrom, "Maps of Random Walks on Complex Networks Reveal Community Structure," Proceedings of the National Academy of Sciences, 105.4., 2008. [16] VD. Blondel, JL. Guillaume, R. Lambiotte, and E. Lefebvre, "Fast Unfolding of Communities in Large Networks," J. Stat. Mech. Journal of Statistical Mechanics: Theory and Experiment, 2008. [17] S. Fortunato and M. Barthelemy, "Resolution Limit in Community Detection," Proceedings of the National Academy of Sciences 104.1, pp. 36-41, 2006. [18] UN. Raghavan, R. Albert, and S. Kumara, "Near Linear Time Algorithm to Detect Community Structures in Large-scale Networks," Phys. Rev. E, 76.3, 2007. [19] M. E. J. Newman, "Finding Community Structure in Networks Using the Eigenvectors of Matrices," Physical Review E Phys. Rev. E, 74.3 2006. [20] G. Palla, I. Dernyi, I. Farkas, and T. Vicsek, "Uncovering the Overlap- ping Community Structure of Complex Networks in Nature and Society," Nature, 435.7043, pp. 814-14, 2005. [21] M. E. J. Newman, "Fast Algorithm for Detecting Community Structure in Networks," Phys. Rev. E 69.6, 2004. [22] M. E. J. Newman, M. Girvan, "Finding and Evaluating Community Structure in Networks," Phys. Rev. E. 69, 026113, 2004. [23] A. Clauset, M. E. J. Newman, and C. Moore, "Finding Community Structure in Very Large Networks," Phys. Rev. E. 70, 06611, 2004. [24] M. Girvan and M. E. J. Newman. "Community Structure in Social and Biological Networks," Proceedings of the National Academy of Sciences, 99.12, pp. 7821-826, 2002. converge to a final label that uniquely identifies the group. In the category of overlapping algorithm, Clique Percolation Method [20] merges two cliques into a community if they overlap more than a threshold. Another relevant paper [11] discusses overlapping community detection algorithms along with various quality measurements. One comprehensive survey of recent advances [4] discusses a wide range of existing algorithms including traditional meth- ods, modularity based methods, spectral algorithms, dynamic algorithms, and more. Community detection often tries to optimize various metrics such as modularity as described by Girvan and Newman [22] or conductance. The work in [12] discusses many of the various objective functions currently in use and how they perform. Similarly, [10] points out that it is important for the community detection algorithm to extract functional communities based on ground truth, where func- tional ground-truth community is described as a community in which an overall theme exists. Another recent paper [7] conducts a comparative analysis of community detection al- gorithms on LFR benchmark in order to provide unbiased guidelines to choose algorithms for a given network. They propose a method to quantify the accuracy of the algorithm with mixing parameter. In addition, [6] provides a procedure- oriented framework for benchmarking and compare different algorithms. The authors also fit and implement 10 algorithms in their framework and compare and rate the algorithms. Our work is different from existing comparative research in that we propose a new angle of comparing the structural metrics for community detection algorithms by considering the size of the identified communities. Our work is the first to investigate previously overlooked community size in a systematic way. VI. CONCLUSION In this paper, we carry out a comparative analysis on community detection algorithms by introducing community size as a new orthogonal dimension to existing structural metrics. In order to make sense of communities of different sizes, we introduce the Dunbar's number in sociology into the community detection in social media. We find that both Infomap and CAA are able to discover communities of desir- able sizes. We also present and compare the performance of each evaluated algorithm in all evaluated structural metrics. Our future work includes exploring greedy approach to improve the performance of CAA in terms of conductance and modularity, conducting comparison with more overlapping community detection algorithms. Furthermore, we plan to explore new metrics such as whether and in what degree a community interest exists in the identified community for community quality evaluation and also make sense of the community. REFERENCES [1] R. I. M. Dunbar, "Do Online Social Media Cut through the Constraints That Limit the Size of Offline Social Networks?" R. Soc. Open Sci., 2016.
1604.03601	1	1604	2016-04-12T22:09:02	Community Detection with Node Attributes and its Generalization	[ "cs.SI", "physics.soc-ph", "stat.ML" ]	Community detection algorithms are fundamental tools to understand organizational principles in social networks. With the increasing power of social media platforms, when detecting communities there are two possi- ble sources of information one can use: the structure of social network and node attributes. However structure of social networks and node attributes are often interpreted separately in the research of community detection. When these two sources are interpreted simultaneously, one common as- sumption shared by previous studies is that nodes attributes are correlated with communities. In this paper, we present a model that is capable of combining topology information and nodes attributes information with- out assuming correlation. This new model can recover communities with higher accuracy even when node attributes and communities are uncorre- lated. We derive the detectability threshold for this model and use Belief Propagation (BP) to make inference. This algorithm is optimal in the sense that it can recover community all the way down to the threshold. This new model is also with the potential to handle edge content and dynamic settings.	cs.SI	cs	November 5, 2018 Community Detection with Node Attributes and Its Generalization Y. Li Department of Statistics Northwestern University, Evanston, USA Community detection algorithms are fundamental tools to understand organizational principles in social networks. With the increasing power of social media platforms, when detecting communities there are two possi- ble sources of information one can use: the structure of social network and node attributes. However structure of social networks and node attributes are often interpreted separately in the research of community detection. When these two sources are interpreted simultaneously, one common as- sumption shared by previous studies is that nodes attributes are correlated with communities. In this paper, we present a model that is capable of combining topology information and nodes attributes information with- out assuming correlation. This new model can recover communities with higher accuracy even when node attributes and communities are uncorre- lated. We derive the detectability threshold for this model and use Belief Propagation (BP) to make inference. This algorithm is optimal in the sense that it can recover community all the way down to the threshold. This new model is also with the potential to handle edge content and dynamic settings. 6 1 0 2 r p A 2 1 ] I S . s c [ 1 v 1 0 6 3 0 . 4 0 6 1 : v i X r a 1 Introduction Community detection is one of the critical issues when understanding social networks. In many real-world networks (e.g. Facebook, Twitter), in addition to topology of so- cial network, content information is available as well. Even though different sources of information about social networks can be collected via social media, node attributes and the structure of networks are often interpreted separately in the research of com- munity detection. Usually the primary attention of algorithms has only focused on the topology of the social networks while on the other hand, the decision of commu- nity assignments has been made solely based on node attributes. The partial use of data is tremendously inefficient. Sometimes, especially when the network is sparse, algorithms which are incapable of incorporating multiple data sources are often para- lyzed and unsuccessful in recovering community assignment. It is of great interests to study how to incorporate topology features and node attributes into one algorithm. Several papers address community detection with node attributes under the as- sumption that the observed node attributes are highly correlated with communi- ties. The two main approaches are: heuristic measure-based models and probabilis- tic inference-based models. The heuristic measure-based models combine topology structure and node attributes in a heuristic function. L. Akoglu et al. [1] proposed a parameter-free identification of cohesive subgroups (PICS) in attributed graphs by minimizing the total encoding costs.Y. Zhou et al. [2] proposed SA-Cluster based on structural and attribute similarities through a unified distance measure.The proba- bilistic inference-based approach usually assumes that the networks are generated by random processes and uses probabilistic generative models to combine both topology and attributes. J. Yang et al. [3] developed Communities from Edge Structure and Node Attributes (CESNA) for detecting overlapping networks communities with node attributes. In CESNA model, the links are generated by process of BigCLAM and node attributes can be estimated by separate logistic models. B.F. Cai et al. [4] pro- posed a popularity-productivity stochastic block model with a discriminative frame- work (PPSB-DC) as the probabilistic generative model. Y.Chen et al. [5] adopted Bayesian method and developed Bayesian nonparametric attribute (BNPA) model. A nonparametric method was introduced to determine the number of communities automatically. These probabilistic generative models can be further categorized based on two different ways of modeling the stochastic relationship between attributes X, communities F and graph G. CESNA and BNPA assume that communities generate both the network as well as attributes (Figure 1 (c) ) however PPSB-DC assumes that communities can be predicted based on attributes and then network are generated based on communities (Figure 1 (d) ). Even though many studies have shown that social ties are not made random but constrained by social position [6] [7], it is possible that the observed node attributes may not contribute much to social position so that they are uncorrelated with com- 1 munities. When communities and node attributes are not correlated, adding nodes attributes into the above models will not give more information about communities. In this paper we propose an approach that allows us to go beyond the similarity between communities and node attributes. One assumption we rely on is that node attributes will lead to heterogeneity in the degree of nodes (Figure 1 (e)). The idea of including heterogeneity in the degree in SBM was first introduced by Wang and Wong [8] and later revisited by Karrer [9]. By including this heterogeneity, our approach is able to solve the more challenging problem where node attributes and communities are uncorrelated. Our intuition is that the node attributes label not only the nodes but the edges as well. Due to heterogeneity in the degree, different types of edges carry different information of communities, therefore our approach should be able to recover the communities more accurately. Another important problem of interests is to understand to which extend the extra information of node attributes will improve the performance, especially when communities and node attributes are not correlated. Here we are focusing on the detectability threshold for our new model. E. Mossel et al. [10] have proven that there exists a phase transition in the detectability of communities for two equal size communities in stochastic block model. S. Heimlicher et al. [12] investigated the phase transition phenomena in more general context of labelled stochastic block model and generalized the detectability threshold. A. Ghasemian et al. [13] derived the detectability threshold in dynamic stochastic block model as a function of the rate of change and the strength of the communities. In this paper we derive the detectability thresholds for community structure in stochastic block model with node attributes and compare it with the original thresholds while no information of node attributes is available. 2 Model The stochastic block model (SBM) is a classic probabilistic generative model of com- munity structure in static networks [14] [15] [16]. Here, we develop a generative model by extending SBM to include heterogeneity due to node attributes in the degree of nodes. In our model, we first assign nodes with different nodes attributes to dif- ferent communities and then generate the topology of network based on both the community assignment and the node attributes (Figure 1 (e)). The graphical model in Figure 1 (e) can be seen as an extension of the graphic model in Figure 1 (d). The main reason for generalizing the graphic model in Figure 1 (d) instead of the graphic model in Figure 1 (c) is that the graphic model in Figure 1 (d) is a combination of graphical models in Figure 1 (a) and Figure 1 (b), which are corresponding graphical models for clustering problem and community detection in stochastic block model. Therefore we find the graphic model in Figure 1 (d) is a better candidate to combine 2 Figure 1: Ways of modeling the stochastic relationship between attributes X, com- munities F and graph G. Circles represent latent community assignment and squares represents observed variables. topology information and node attributes information. In our model, we also assume that all the node attributes are categorical variables. Finally we correct the degree of nodes based on node attributes, which leads to sub-communities structure (Figure 2). This assumption allows heterogeneity in communities and generalize the community in SBM. We formally describe the generative process of a graph G = {V, E, x1, x2, . . . , xm} under stochastic block model with node attributes, where x represents attributes, as follows. First, we construct an one-to-one map of node attributes from m-dimensional point {x1, x2, . . . , xm} to 1-dimensional point {Xr}, where m is the number of different types of observed attributes and r is from 1 to R. Then we assign each of the n nodes i ∈ V into R group according to node attributes and the number of nodes in each group is nr. Using a prior qk,r, we assign nr nodes in attributes category r into K communities. We then generate the (i, j)th element in adjacency matrix A ccording to a Bernoulli distribution with probability P{ki,ri},{kj ,rj}, where ki is the community assignment for node i, ri is the attributes category for node i and P{ki,ri},{kj ,rj} is the probability of forming an edge between a node from community ki with attributes Xri and a node from community kj with attributes Xrj .The full likelihood of graph under SBM with node attribute is: P (E, kX, P ) = ( qki,ri)( (1 − P{ki,ri},{kj ,rj})) (1) (cid:89) i (cid:89) i,j∈E (cid:89) i,j /∈E P{ki,ri},{kj ,rj} 3 Figure 2: heat map of block matrix, red squares represent two primary communities, green squares represent sub-communities in primary communities. Since P{ki,ri},{kj ,rj} = O( 1 n), sometimes its easier to work with the rescale matrix c{ki,ri},{kj ,rj}. When two nodes are from group K1, K2 with category of attributes a, b, the rescale matrix c{K1,a},{K2,b} = nP{K1,a},{K2,b}. For subsequent analysis, we will focus on the choice of uniform prior qk,r = 1 K since we are interested in the detectability threshold when attributes are not correlated with communities. We will also limit ourselves to an algorithmically difficult case of block model, where every group k has the same average degree conditional on the type of edge: (cid:88) (cid:88) k1 k2 (cid:88) k2 cab = nb K 2 P{k1,a},{k2,b} = nb K P{k1,a},{k2,b} for any k1. (2) If this is not the case, reconstruction can be achieved by labeling nodes based on their degrees. 3 Detectability threshold in SBM with node at- tributes The best-known rigorous detectability threshold in sparse SBM has been derived by E. Mossel et al. [10]. In the sparse partition model, where p = a n and a > b > 0 , the clustering problem is solvable in polynomial time if (a− b)2 > 2(a + b). However for K ≥ 3 it is still an open question to find a rigorous detectability threshold in SBM. The Kesten-Stigum (KS) threshold in statistical physics can be treated as a non-rigorous threshold for K ≥ 3 [17] [18]. Let G be generated by SBM(n, k, a, b) and n,q = b 4 define SN R = √ a−b k(a+(k−1)b) . If SN R > 1 then the clustering problem is solvable and the Kesten-Stigum (KS) threshold can be achieved in polynomial time. In the sparse regime, E = O(n), the graph generated by SBM is locally treelike in the sense that all most all nodes in the giant component have a neighborhood which is a tree up to distance O(log(n)). Therefore the threshold for reconstruction on tree can provide good insight into reconstruction on SBM. As mentioned before, one intuition is that node attributes label the edges, there- fore we consider a multi-type branching process of edges to generate the tree that ap- proximates the graph generated by SBM with node attributes. By defining a Markov chain on the infinite tree T = (V, E, X), we can derive the construction thresholds on SBM with node attributes. To construct the multi-type branching process, we first label the edge by the categories of node attributes at the two ends of the edge as L{Xa, Xb}, where Xa is the attributes for the node that is closer to the root, Xb is the attributes for the node at the far-end and a, b is from 1 to R. So there are R2 different types of edges. Map {Xa, Xb} to (a− 1)∗ R + b and relabel the L{Xa, Xb} type edge as L{(a− 1)∗ R + b}. Let R2 ∗ R2 dimensional matrix C be the matrix describing the expected number of children, where cij is the expected number of type L{i} edges induced by one type L{j} edge. Note that one L{Xa1, Xb1} type of edge will give birth to L{Xa2, Xb2}type R ] and z = j − [ j−1 of edges if and only if b1 = a2. Let x = [ i−1 R ], (cid:40) 0 cxy R ] + 1 and y = i − [ i−1 if x (cid:54)= z, if otherwise. (cid:80) 5 (3) When moving outward a type L{Xa, Xb} edge, the K ∗ K stochastic transition cij = matrix σ associate with the edge can be defined as: nb K P{k1,a},{k2,b} σk1k2 ab = (4) The largest eigenvalue for the K ∗ K stochastic transition matrix σ is 1 and let the second largest eigenvalue be λab. Define mij in the R2 ∗ R2 matrix M1 as cij ∗ λ2 . The robust reconstruction is possible when the absolute value of largest [ i−1 R ]+1,i−[ i−1 R ] eigenvalue for matrix M1 exceeds 1 [13][11]. cab . 4 Belief propagation To recover the community assignments in SBM with node attributes, we use Bayesian inference to learn the latent community: P (kE, X, P ) = P (k, EX, P ) g P (Eg, X, P ) , (5) The distribution is too complex to compute directly since(cid:80) where k is the estimated group assignment and g is the original group assignment. g P (Eg, X, P ) runs over exponential number of terms. In the regime E = O(n), the graph is locally treelike therefore belief propagation, which is exact to calculate the marginal probability of community assignment on a tree, can be applied to calculate Bayesian inference efficiently. Well show that BP is an optimal algorithm in the sense that it can reach the detectability thresholds for SBM with node attributes. bility, denoted as ψi→j the absence of node j. We can compute the messenger from i to j as: To write the belief propagation equation, we define conditional marginal proba- , which is the probability that node i belongs to group ki in ki ψi→j ki = 1 Z i→j qkiri cAil{kl,rl},{ki,ri}(1 − c1−Ail {kl,rl},{ki,ri} n )ψl→i ki ], (6) (cid:89) (cid:88) [ l∈∂i\j kl where Ail is the (i, l)th element in the adjacency matrix for the graph generated by SBM with node attributes, ∂i denotes all the nodes connected to i and Z i→j is a normalization constant ensuring ψi→j to be a probability distribution. The marginal ki probability ψi can be calculated as: ki ψi ki = 1 Z i qkiri cAil{kl,rl},{ki,ri}(1 − c1−Ail {kl,rl},{ki,ri} n )ψl→i ki ], (7) (cid:89) l∈∂i [ (cid:88) kl where Z i is a normalization constant ensuring ψi to be a probability distribution. In ki SBM with node attributes, we have interactions between all pairs of nodes, therefore we have n(n−1) messengers to update for one iteration. To reduce the computational complexity to O(n), we follow past work on BP for SBM citeDecelle. At the cost of making O( 1 n) approximation to the messenger, when there is no edge between i andj, the messenger from i to j can be calculated as: ψi ki→j = ψi ki . (8) Now only messengers sent on edges are needed to be calculated. By introducing an external field, the messenger from i to j when there is an edge between i and j can be approximated as: where the external field hkiri can be defined as: Its worth noting that ψi→j ki = qkiri is a fixed point in (9). 6 ψi→j ki = c{ki,ri},{kl,rl}ψl→i ki ], (cid:88) [ kl 1 (cid:89) Z i qkirie−hkiri (cid:88) (cid:88) l∈∂i hkiri = 1 n l kl c{ki,ri},{kl,rl}ψl kl . (9) (10) 5 Phase transition in BP and simulation In this section, we will study the stability of the fixed point under random perturba- tions. As discussed above, in the sparse regime, the graph generated by SBM with node attributes is locally treelike. Here consider such a tree with d levels. On the leave md the fixed point is perturbed as ψmd is i.i.d. kmd random variable. Then the influence of perturbation on leave md to the root m0 can be calculated as: , where md kmd = qkmd rmd + md kmd m0 = T dab ab md, (11) (cid:89) {ab} where dab is the number of type L{Xa, Xb} edges on the path from leave md to the root m0 and T ab is the transfer matrix for type L{Xa, Xb} edges, which, by following the calculation in [19], can be defined as: ab − 1). T k1k2 ab = qk1a(kσk1k2 As d → ∞, dab → ∞ as well,therefore m0 ≈(cid:81) (cid:89) ab md,where υab is the largest eigenvalue for T ab. Now let us consider the variance at root m0 induced by the random perturbation on all leaves at level d. Since the influence of each leaf is independent, the variance of the root can be written as: all{ab} υdab (cid:88) (12) υ2dab ab V ar(md). (13) V ar(m0) = all the path (r∼md) {ab} transition in BP, it's sufficient to calculate Zd =(cid:80) When the variances on leaves are amplified exponentially, the fixed point is un- stable and BP algorithm is able to recover the community assignment with high probability, otherwise the perturbation on leaves will vanish and the fixed point in stable under BP algorithm. From eq.(13), when md is i.i.d., to determine the phase . This calculation can be done by viewing this summation as a weight associated multi- type branching process. Consider thus a multi-type branching process with Possion distribution with mean cab if the parent-child edge in the corresponding multi-type branching process belongs to type L{Xa, Xb}. The variance amplified along the tree generated by the above multi-type branching process and the expected values of the variance at level d can be calculated as: allthepath(r∼md) all{ab} υ2dab (cid:81) ab E(Zdm0) = 1T M d 2 em0, (14) where the (i, j)th element of M2 is cijυ2 , em0 is an R2-dimensional unit vector with the rth element equal to 1 and r is the node attribute type of the root node m0. When the largest eigenvalue of M2 exceeds 1, the fixed point of BP is unstable and the community is detectable. Noting that λab = υab, therefore BP is an [ i−1 R ]+1,i−[ i−1 R ] 7 optimal algorithm in the sense that it can reach the detectability threshold in SBM with node attributes even when node attributes and communities are not correlated. Next, we compare the detectability thresholds for SBM with node attributes with the detectability threshold for the original SBM without information of node at- tributes. In the following discussion, we will limit ourselves to the case where nr = n R , qki,ri = 1 K and C{ki,ri},{kj ,rj} satisfy equation (15), a if ki = kjand ri = rj if ki = kjand ri (cid:54)= rj if otherwise, b c c{ki,ri},{kj ,rj} = (15) where a ≥ b ≥ c. For SBM with node attributes, the community is detectable if (a − c)2 a + (K − 1)c ξ1 = + (R − 1) (b − c)2 b + (K − 1)c > KR. (16) For SBM without information of node attributes, the community is detactable if (a + (R − 1)b − Rc)2 ξ2 = a + (R − 1)b + (K − 1)Rc (17) By simple calculation, it can be shown that ξ1 ≥ ξ2, therefore even in the situation where the observed node attributes are uncorrelated with communities, including node attributes into model will give us more infomation about communities. > KR. We conduct the following simulation to verify the claim of phase transition in BP. Considering for simplicity only two communities and two node attributes, we generate a series of graphs by SBM with node attributes for 4000 nodes and various choice of (a, b) when controlling average degree to be 5. We use η = a b to represnt different choices of (a, b) and = c b to represent the strength of communities. When = 0 the clusterings are maximally strong while at = 1 the clusterings are weak. The accuracy of reconstruction is measure by overlap matric introduced by [19]. In figure 3, we plot overlap metric against for different values of η and for each curve, we use a vertical dash line in the same color as the corresponding curve to indicate the detectability threshold. Figure 3 shows that BP can recover commu- nities that are positively correlated with true communities all the way down to the detectability thresholds for various choice of (, η). The algorithm has larger overlap metric with smaller . 6 Conclusion In this paper, we consider a model that uses information of nodes attributes in a different way such that this approach will provide more information of latent com- munities beyond the information carried by SBM even when node attributes are not 8 Figure 3: Overlap as a function of for various values of η. Dash lines mark the theoretical detectability thresholds for the choice of (, η). correlated with communities. We have derived a theoretical detectability threshold for SBM with node attributes, which coincides with phase transition in BP. We also conduct a numerical analysis of the phase transition in BP. While constricted to the two symmetric communities with two node attributes, this condition is sufficient to illustrate how the information of node attributes affects detectability even the node attributes are not correlated with communities. A nature extension will include edge contents and dynamic settings into the model. Our approach can be applied to this case by including different type of edges into the multi-branching process. On the theoretical front, it has been conjectured [10] that, for K ≥ 5, theres a regime that the clustering problem is solvable but not in polynomial time. Emmanuel Abbe and Colin Sandon [20] have developed a non- efficient algorithm that is shown to break down KS threshold at K = 5 in SBM. As a future work, well try to develop an algorithm that can break down the detectability threshold in our model for large numbers of groups. ACKNOWLEDGEMENTS I am grateful to Professor Wenxin Jiang and Professor Noshir Contractor for helpful discussion. 9 References [1] Akoglu, Leman, et al. "PICS: Parameter-free Identification of Cohesive Sub- groups in Large Attributed Graphs." SDM. 2012. [2] Zhou, Yang, Hong Cheng, and Jeffrey Xu Yu. "Graph clustering based on struc- tural/attribute similarities." Proceedings of the VLDB Endowment 2.1 (2009): 718-729. [3] Yang, Jaewon, Julian McAuley, and Jure Leskovec. "Community detection in networks with node attributes." Data mining (ICDM), 2013 ieee 13th interna- tional conference on. IEEE, 2013. [4] Chai, Bian-fang, et al. "Combining a popularity-productivity stochastic block model with a discriminative-content model for general structure detec- tion."Physical review E 88.1 (2013): 012807. [5] Chen, Yi, et al. "Network structure exploration in networks with node at- tributes." Physica A: Statistical Mechanics and its Applications (2016). [6] McPherson, Miller, Lynn Smith-Lovin, and James M. Cook. "Birds of a feather: Homophily in social networks." Annual review of sociology (2001): 415-444. [7] Centola, Damon, et al. "Homophily, cultural drift, and the co-evolution of cul- tural groups." Journal of Conflict Resolution 51.6 (2007): 905-929. [8] Wang, Yuchung J., and George Y. Wong. "Stochastic blockmodels for directed graphs." Journal of the American Statistical Association 82.397 (1987): 8-19. [9] Karrer, Brian, and Mark EJ Newman. "Stochastic blockmodels and community structure in networks." Physical Review E 83.1 (2011): 016107. [10] Mossel E, Neeman J, Sly A. Stochastic block models and reconstruction. arXiv preprint arXiv:1202.1499. 2012 Feb 7. [11] Janson, Svante, and Elchanan Mossel. "Robust reconstruction on trees is deter- mined by the second eigenvalue." Annals of probability (2004): 2630-2649. [12] Heimlicher, Simon, Marc Lelarge, and Laurent Massouli. "Community detection in the labelled stochastic block model." arXiv preprint arXiv:1209.2910 (2012). [13] Ghasemian, Amir, et al. "Detectability thresholds and optimal algo- for community structure in dynamic networks." arXiv preprint rithms arXiv:1506.06179(2015). 10 [14] Holland, Paul W., Kathryn Blackmond Laskey, and Samuel Leinhardt. "Stochas- tic blockmodels: First steps." Social networks 5.2 (1983): 109-137. [15] Faust, Katherine, and Stanley Wasserman. "Blockmodels: Interpretation and evaluation." Social networks 14.1 (1992): 5-61. [16] Snijders, Tom AB, and Krzysztof Nowicki. "Estimation and prediction for stochastic blockmodels for graphs with latent block structure." Journal of clas- sification 14.1 (1997): 75-100. [17] Kesten, Harry, and Bernt P. Stigum. "Additional limit theorems for indecompos- able multidimensional Galton-Watson processes." The Annals of Mathematical Statistics 37.6 (1966): 1463-1481. [18] Kesten, Harry, and Bernt P. Stigum. "Limit theorems for decomposable multi- dimensional Galton-Watson processes." Journal of Mathematical Analysis and Applications 17.2 (1967): 309-338. [19] Decelle, Aurelien, et al. "Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications." Physical Review E84.6 (2011): 066106. [20] Abbe, Emmanuel, and Colin Sandon. "Detection in the stochastic block model with multiple clusters: proof of the achievability conjectures, acyclic BP, and the information-computation gap." arXiv preprint arXiv:1512.09080(2015). 11
1909.06810	1	1909	2019-09-15T14:56:24	Centrality in dynamic competition networks	[ "cs.SI", "physics.soc-ph" ]	Competition networks are formed via adversarial interactions between actors. The Dynamic Competition Hypothesis predicts that influential actors in competition networks should have a large number of common out-neighbors with many other nodes. We empirically study this idea as a centrality score and find the measure predictive of importance in several real-world networks including food webs, conflict networks, and voting data from Survivor.	cs.SI	cs	Centrality in dynamic competition networks Anthony Bonato1, Nicole Eikmeier2, David F. Gleich3, and Rehan Malik1 1 Ryerson University, Toronto, Ontario, Canada 2 Grinnell College, Grinnell, IA, USA 3 Purdue University, West Lafayette, IN, USA Abstract. Competition networks are formed via adversarial interactions between actors. The Dynamic Competition Hypothesis predicts that influential actors in competition networks should have a large number of common out-neighbors with many other nodes. We empirically study this idea as a centrality score and find the measure predictive of importance in several real-world networks including food webs, conflict networks, and voting data from Survivor. 1 Introduction While social networks are often studied from the perspective of positive interactions such as friend- ship or followers, the impact of negative social interaction on their structure and evolution cannot be ignored. Structural balance theory posits positive and negative ties between actors in social networks, and assumes such signed networks will stabilize so that triples of actors are either all mutually friends or possess common adversaries; see [12], and [9] for a modern treatment. The pre- diction of the signs of edges in a social network was previously studied [15,18,21]. Further, negative interactions as a model for edges was studied in the context of negatively correlated stocks in mar- ket graphs [4], and in the spatial location of cities as a model to predict the rise of conflicts and violence [11]. Even in the highly cited Zachary Karate club network [22], the negative interaction between the administrator and instructor was the impetus for the split of the club participants into two communities. We propose that competition or negative interactions are critically important to the study of social networks and more broadly, real-world complex networks, and are often hidden drivers of link formation. In [6], we investigated properties inherent in social networks of competitors that evolve dynam- ically over time. Such networks are viewed as directed, where a directed edge from nodes u to v corresponds to some kind of negative social interaction. For example, a directed edge may represent a vote by one player for another in a social game such as the television program Survivor. Directed edges are added over discrete time-steps in what we call dynamic competition networks. Our main contribution in [6] was the presentation of a hypothesis, referred to as the Dynamic Competition Hypothesis, or (DCH), that served as a predictive tool to uncover alliances and leaders within dy- namic competition networks. We provided evidence for the hypothesis using U.S. voting record data from 35 seasons of Survivor. In the present paper, we focus on a particular implication of the DCH. Namely, the DCH pre- dicts that leaders and central actors in these networks should have large a large number of common out-neighbors with other nodes in the network. Consequently, this score should constitute a more accurate and interesting centrality score in competition networks where edges have a negative con- notation. We study this score in terms of its ranking of leaders in various kinds of networks ranging from additional international seasons of Survivor, to conflict networks, and to food webs. We organize the discussion in this paper as follows. In Section 2, we formally define dynamic competition networks, and review the DCH as stated in [6], with a focus on the common out-neighbor scores, called CON scores. In Section 3, we investigate using CON scores as centrality measures in three distinct sources: i) voting data from all international (that is, non-U.S.) seasons of Survivor, ii) from conflict networks arising from the countries of Afghanistan, India, and Pakistan, and iii) 2 Anthony Bonato, Nicole Eikmeier, David F. Gleich, and Rehan Malik in 14 food webs. We find that the CON scores predict influential actors in the networks with high precision. The final section interprets our results for real-world complex networks, and suggests further directions. We consider directed graphs with multiple directed edges throughout the paper. Additional background on graph theory and complex networks may be found in the book [5] or [7]. 2 The Dynamic Competition Hypothesis The Dynamic Competition Hypothesis (DCH) provides a quantitative framework for the structure of dynamic competition networks. We recall the statement of the DCH as first stated in [6]. Before we state the DCH, we present some terminology. A competition network G is one where nodes represent actors, and there is directed edge between nodes u and v in G if actor u is in competition with actor v. A dynamic competition network is a competition network where directed edges are added over discrete time-steps. For example, nodes may consist of individuals and edges correspond to conflicts between them; as another example, we may consider species in an ecological community, and directed edges correspond to predation. Observe that dynamic competition networks may have multiple edges if there were multiple conflicts; further, not all edges need be present. The central piece of the DCH we study here are the common out-neighbor scores. Without loss of generality, we assume that the node correspond to integers such that we can use the nodes to address an adjacency matrix as well. Consequently, let A be the adjacency matrix of given competition network. Entries in the matrix are 0 or positive integers for the number of competition interactions. For nodes u, v, and w, we say that w is a common out-neighbor of u and v if (u, w) and (v, w) are directed edges. Alternatively, AuvAvw ≥ 1. For a pair of distinct nodes u, v, we define CON(u, v) to be the number of common out-neighbors of u and v. Note that this common out-neighbor score k min(Auk, Avk), counts multiplicities based on the minimum number of interactions: CON(u, v) =(cid:80) which corresponds to a multiset intersection. For a fixed node u, define We call CON(u) the CON score of u. For a set of nodes S with at least two nodes, we define Observe that CON(S) is a non-negative integer. In the DCH, leaders are defined as members of a competition network with high standing in the network, and edges emanating from leaders may influence edge creation in other actors. In the context of conflict networks within a country, leaders may be actors who exert the strongest political influence within the country; note that these may not be the largest or most powerful actors. As another example, leaders in a food web would naturally have higher trophic levels (that is, higher position in a food chain). The DCH characterizes leaders as those nodes with high CON scores, low in-degree, high out-degree and high closeness. Recall that for a strongly connected digraph G and a node v, we define the closeness of u by  (cid:88) v∈V (G)\{u} −1 C(u) = d(u, v) where d(u, v) corresponds to the distance measured by one-way, directed paths from u to v. v∈V (G) (cid:88) (cid:88) u,v∈S CON(u) = CON(u, v). CON(S) = CON(u, v). Centrality in dynamic competition networks 3 In this paper, we focus on the implication that leaders in competition networks should have high CON scores, which suggests this is a natural centrality measure for these networks. The DCH also involves the notion of alliances, that does not factor into our present study. Alliances are defined as groups of agents who pool capital towards mutual goals. In the context of social game shows such as Survivor, alliances are groups of players who work together to vote off players outside the alliance. Members of an alliance are typically less likely to vote for each other, and this is the case in strong alliances. This is characterized in terms of near independent sets; see [6] for the formalism. In summary, the Dynamic Competition Hypothesis (or DCH ) asserts that dynamic competition networks satisfy the following four properties. 1. Alliances are near independent sets. 2. Strong alliances have low edge density. 3. Members of an alliance with high CON scores are more likely leaders. 4. Leaders exhibit high closeness, high CON scores, low in-degree, and high out-degree. Our focus in this work will be on the validation of the DCH with regards to detecting leaders; in particular, we will focus on items (3) and (4) of the DCH. Note that while we expect leaders to be in alliances (that is, have prominent local influence), leaders are determined via global metrics of the network. 3 Methods and Data 3.1 Survivor In [6], we studied the voting history of U.S. seasons of Survivor, which is a social game show where players compete by voting each other out. In Survivor, strangers called survivors are placed in a location and forced to provide shelter and food for themselves, with limited support from the outside world. Survivors are split into two or more tribes which cohabit and work together. Tribes compete for immunity and the losing tribe goes to tribal council where one of their members is voted off. At some point during the season, tribes merge and the remaining survivors compete for individual immunity. Survivors voted off may be part of the jury. When there are a small number of remaining survivors who are finalists (typically two or three), the jury votes in favor of one of them to become the Sole Survivor who receives a cash prize of one million dollars. Figure 1 represents a graphical depiction of the voting history of a season of U.S. Survivor. We extend the analysis of the 35 U.S. seasons in [6] to 82 international seasons of Survivor. Data used in our analysis was obtained from the Survivor wiki pages https://survivor.fandom.com/ wiki/Main Page. Several seasons (beyond the 82) were excluded for varying reasons. In some cases, a wiki page exists, but there was no voting data. In other cases, much of the voting information was missing, or the rules are significantly different than the traditional version of the game shows. Nevertheless, the number of seasons collected exceeds the number in [6]. In Table 1 we display some of the CON scores for a few example seasons. We distinguish which players are finalists, since the rules change in determining who is the last player eliminated. For example, instead of eliminating the last player via votes against players, in survivor many players may return for a final vote for who they would like to win. In Table 2, we detail relevant statistics on these networks. For each network, we consider whether the winner of the season had one of the top three or top five CON scores and list the percentage of such networks. For example from Table 2 we see that 81.7 percent of Survivor winners had one of the largest 5 CON scores. For comparison, we also compute PageRank (on the reversed-edge network, where we change the orientation of directed edge) and Jaccard similarity scores, which are both standard ranking scores. Jaccard similarity is a type of normalized CON score; see [10]. We find that the CON scores are a more accurate predictor for determining finalists of Survivor than both PageRank and Jaccard similarity. We observe that these results are consistent with 4 Anthony Bonato, Nicole Eikmeier, David F. Gleich, and Rehan Malik Fig. 1. The Survivor Heroes vs. Healers vs. Hustlers co-voting network. Nodes are scaled by closeness, and color-coded according to their original tribe. Thicker edges represent multiple votes. Australian Survivor (2002) Name CON ID OD C Robinson 2009 Name ID OD C CON Survivor South Africa Malaysia Name CON ID OD C Robert 5 10 0.714 44 0.652 37 Sciona 1 9 7 8 Joel 3 9 Katie 3 8 Sophie 9 6 Jane 8 5 Lance 8 8 Craig 8 7 Naomi 10 6 Caren Sylvan 3 5 Deborah 4 4 5 1 Jeff David 6 3 Tim 4 2 Lucinda 8 1 0.625 35 0.652 38 0.652 38 0.625 36 0.577 27 0.577 18 25 0.5 0.5 25 0.417 30 0.395 23 0.395 4 0.441 23 0.294 10 7 0.0 0.563 36 0 6 Ellenor 7 8 0.557 34 Jarmo 2 10 0.645 54 Anna 0.557 38 Nina 4 7 0.612 47 Erik Bl. 7 9 Lukas 4 7 0.557 31 0.612 46 6 8 Angela 30 0.51 Ranjit 5 5 24 0.51 Christian 3 4 Rafael 5 4 0.49 28 0.438 26 Erik Bi. 9 5 0.422 18 5 4 Erik R. 0.306 13 Mika 5 3 0.265 17 Josefine 0 2 Erika 7 2 0.306 15 12 0.35 Beatrice 6 2 Micha 12 1 0.299 7 Lorette Grant 4 9 5 8 Amanda 4 8 Mandla 0 8 Angie 6 9 Angela 4 6 Dyke 4 6 Hein 5 4 Irshaad 3 5 Lisa 11 4 Rijesh 4 3 Nichal 6 2 Elsie 8 2 Viwe 5 2 Nicola 5 1 Nomfundo 4 1 0.653 0.653 0.622 0.652 0.688 0.594 0.568 0.484 0.544 0.484 0.408 0.363 0.436 0.344 0.304 0.335 35 33 26 32 31 28 17 22 25 16 13 12 9 11 6 8 Table 1. Three international Survivor seasons. Players are listed by first name, in order from top to bottom with the winner at the top, and the first eliminated player at the bottom. For each player we list the in- degree, out-degree, closeness, and CON score. The horizontal line separates finalists from the rest of the group. the analysis in [6] for the U.S. seasons. As an added comparison, we list the probability of the winner appearing in a random set of three or five players; note that there is a range of percentages depending on how many players are in a given season. In the interest of space, we refer the reader to https://eikmeier.sites.grinnell.edu/uncategorized/competition-show-data/ where we house all data on these seasons. Centrality in dynamic competition networks 5 CON PageRank Survivor Top 3 Top 5 57.3 81.7 43.9 78.0 Jaccard Similarity 47.6 72.0 random set 11.1-27.3 18.5-45.5 Table 2. Statistics on international Survivor seasons. 3.2 Political Conflicts For our second competition network, we extracted data from The Armed Conflict Location and Event Data Project (or ACLED), which may be found at https://www.acleddata.com/data/. ACLED catalogs information about political conflict and protests across the world. In these conflict networks, nodes correspond to actors in a given region, and edges correspond to conflicts between the actors. Many types of metadata are recorded corresponding to each event. Our particular interest is in the actors involved in each event, and where the event took place. More information about this project can be found on the project website. An important note about the ACLED data is that we do not know which actor initiated a given event. Therefore, we do not consider the majority of edges (events) to be directed. The only events which we assume knowledge about directed-ness is when civilians are involved. We restricted our study to a set of events to a particular country; keeping the scale at the country level allows us to keep a larger set of actors. We selected three countries that have a large number of actors and events to analyze: Afghanistan, India, and Pakistan. We first consider the rankings for Pakistan with commentary. Fig. 2. A Slope Graph to compare the rankings via CON and PageRank. On the left, the top actors in the conflict network Pakistan via CON metrics, while on the right, the top actors in Pakistan via PageRank on the reverse network. Actors are labeled in black if the difference in rankings is less than or equal to three. Actors are labeled in red if the CON ranking is at least four places higher than the PageRank, and in green if the PageRank is at least 4 places higher than the CON ranking. Note that no line appears to connect the left and right side if the actor does not show up in the top 15 of the other ranking. Pakistan has faced terrorism activities since 2000, with many militant groups attacking civilians and Pakistan armed forces. TTP (Pakistan) is one of the largest radical extremist groups, which is Unidentified Armed Group (Pakistan)TTP: Tehreek-i-Taliban PakistanRioters (Pakistan)Police Forces of Pakistan (2013-2018)Military Forces of Pakistan (2013-2018)Police Forces of Pakistan (2008-2013)LeI: Lashkar-e-IslamMilitary Forces of Pakistan (2008-2013)Aman LashkarUnidentified Armed Group (Afghanistan)Pakistan RangersFrontier CorpsPolice Forces of Pakistan (2018-)Islamic State (Pakistan)BLF: Baloch Liberation FrontCONUnidentified Armed Group (Pakistan)TTP: Tehreek-i-Taliban PakistanPolice Forces of Pakistan (2013-2018)Military Forces of Pakistan (2013-2018)Police Forces of Pakistan (2008-2013)Military Forces of Pakistan (2008-2013)Rioters (Pakistan)Police Forces of Pakistan (2018-)ProtestersBLF: Baloch Liberation FrontJatoi Communal Militia (Pakistan)Chandio Communal Militia (Pakistan)Private Security Forces (Pakistan)Pakistan RangersMilitary Forces of United StatesPageRank6 Anthony Bonato, Nicole Eikmeier, David F. Gleich, and Rehan Malik an umbrella organization of many militant groups such as Lashkar-e-Islam, Islamic State (Pakistan), and Jamaat-ul-Ahrar. In Figure 2, we find that TTP has one of the highest CON scores. TTP has alliances with another terrorist organizations in Pakistan and neighboring countries, which lends to its prominence. In addition, due to the Afghan war, TTP has a strong influence and hold over many Islamic institutions in Pakistan. The Police Forces of Pakistan and Military Forces of Pakistan ensure national security, and they share information for achieving their goals. The Police Forces of Pakistan are an influential actor in the conflict network with another one of the highest CON scores. They perform their duties in all provinces of Pakistan with the help of their paramilitary forces such as Pakistan Rangers and Frontier Corps, and they maintain law and order, as well as border control. Military Forces of Pakistan (2013-2018) has one of the largest CON scores owing to their increased activities against terrorist groups in recent years. We also offer commentary on some of the lower ranked actors. The Baloch Liberation Front (or BLF) is an ethnic-separatist political front and militant organization that is currently fighting against the Pakistani government for an independent Balochi state. The BLF is the strongest and most influential militant group of Baluchistan, but there has been no confirmed coordination between the BLF and other Balochi and non-Balochi groups, and they operate independently of one another. This is a large reason that BLF have low CON and closeness scores. The Islamic State is a part of the militant Islamist group: Islamic State of Iraq and Levant (ISIS). The Islamic State was formed by some of the TTP leaders and is more successful in Afghanistan. This organization has had less success in Pakistan largely carrying out isolated, small scale attacks. The Police Forces of Pakistan actively participated with the support of paramilitary forces of Pakistan in 2008-2018 for war against terror. The Police Forces of Pakistan mostly work to maintain the daily law and order in their respective provinces. Likely for these reasons, they have lower CON scores than the years between 2008-2018. Fig. 3. Top actors in Afghanistan via CON score and PageRank. We note that the ranking of the top actors using the CON score (on the left in Figure 2) is not dissimilar to the one using PageRank on the reversed-edge network (on the right in Figure 2). To quantify the difference in the rankings we used Spearman's rank correlation coefficient. Note that we cannot use Pearson correlation because our data is not at all Gaussian. Recall that Spearman's correlation coefficient is defined as 1 − 6(cid:80)N i=1 d2 i N (N 2 − 1) , TalibanPolice Forces of Afghanistan (2004-)Unidentified Armed Group (Afghanistan)Military Forces of Afghanistan (2004-)Islamic State (Afghanistan)NATOMilitary Forces of Afghanistan (2004-) Special ForcesHQN: Haqqani NetworkAfghan Local PolicePrivate Security Forces (Afghanistan)Militia (Pro-Government)Military Forces of Pakistan (2008-)National Directorate of SecurityTaliban and/or Islamic State AfghanistanHezbi IslamiCONTalibanMilitary Forces of Afghanistan (2004-)Islamic State (Afghanistan)Police Forces of Afghanistan (2004-)NATOUnidentified Armed Group (Afghanistan)Military Forces of Afghanistan (2004-) Special ForcesNational Directorate of SecurityHQN: Haqqani NetworkTaliban and/or Islamic State AfghanistanAfghan and/or NATO ForcesAfghan Local PolicePrivate Security Forces (Afghanistan)Militia (Pro-Government)TTP: Tehreek-i-Taliban PakistanPageRankCentrality in dynamic competition networks 7 Fig. 4. Top actors in India via CON score and PageRank. where N is the total number of actors, and di is the difference in rankings between actor i. A value close to 1 means that the two rankings are very well positively correlated. The Spearman correlation for Pakistan is -0.341, which suggests that the rankings are not that similar. In fact, the negative value implies that as the CON ranking decreases, the PageRank score increases. There are 741 total actors we consider in the Pakistan data set, and the later rankings clearly vary greatly. We finish this section with the rankings for Afghanistan and India in Figures 3 and 4. The Spearman coefficients are 0.604 and -.267 respectively, indicating that the rankings provided by CON matches more similarly to PageRank in the Afghanistan dataset. While we do not provide in-depth commentary on these rankings, we find influential actors in both countries with the largest rankings against DCH metrics. 3.3 Food Webs As a third and final type of data that we analyzed against the DCH, we studied food web datasets from the Pajek website: http://vlado.fmf.uni-lj.si/pub/networks/data/bio/foodweb/foodweb.htm [3]. These are 14 food webs in total. In food webs, the nodes are species, and a weighted edge (u, v) exists with weight w if u inherits carbon from v (that is, u preys on v) [2]. We interpret this as a directed negative interaction from node u to node v. A noteworthy difference in these networks (vs. Survivor, say) is that the movement of energy is balanced, meaning the in-degree and out-degree for each species is equal. Rankings of selected food web datasets are in Figures 5 and 6; the rankings for all the datasets may be found at https://eikmeier.sites.grinnell.edu/uncategorized/competition-show-data/ along with the computed CON scores, closeness, and PageRank on the reversed-edge network. In studying the rankings of these 14 food webs, we see a difference between the CON rank- ings and PageRank. PageRank has been used to study the importance of species in regards to co-extinction [1,16,14], which we expect is likely reflected in the rankings we see here using vanilla PageRank. However, we find a substantially different ranking when using the CON scores; for exam- ple, see the placement of Heteroflagellates in Figure 5. The average Spearman correlation coefficient across these 14 datasets is 0.271, and the range is between 0.004 and 0.554. (Recall that a value close to 1 means very well correlated.) Therefore, we suggest that the CON scores are giving a different ranking, which is much closer to trophic levels of species. In particular, the CON scores reflect a natural hierarchical structure in ecosystems, and this is consistent with the DCH. Rioters (India)Police Forces of India (2014-)Unidentified Armed Group (India)Military Forces of India (2014-)Protestors (India)CPI (Maoist): Communist Party of India (Maoist)Police Forces (2014-) Rashtriya RiflesBJP: Bharatiya Janata PartyPolices Forces (2014-) Assam RiflesTMC: Trinamool Congress PartyINC: Indian National CongressJJMP: Jharkhand Jan Mukti ParishadNSCN-K: National Socialist Council of Nagaland-KhaplangPolice Forces (2014-) Central Reserve ForceUnidentified Armed Group (Pakistan)CONPolice Forces of India (2014-)Rioters (India)Military Forces of India (2014-)Unidentified Armed Group (India)CPI (Maoist): Communist Party of India (Maoist)Police Forces (2014-) Border Security ForceProtestors (India)Police Forces (2014-) Central Reserve ForcePolices Forces (2014-) Assam RiflesPolice Forces (2014-) Rashtriya RiflesNSCN-IM: National Socialist Council of Nagaland-Isak MuivahINC: Indian National CongressBJP: Bharatiya Janata PartyPolice Forces (2014-) Forest Range OfficerSAD: Shiromani Akali DalPageRank8 Anthony Bonato, Nicole Eikmeier, David F. Gleich, and Rehan Malik Fig. 5. The Chesapeake Bay Lower food web dataset. On the left, organisms are listed in decreasing order, with the largest CON score at the top. On the right, organisms are listed by decreasing PageRank score. 4 Conclusion and future directions We studied an implication of the Dynamic Competition Hypothesis (DCH) for competition networks across several different types of real-world networks. We found that the DCH prediction that high CON scores should correspond to leaders is supported in predicting winners in international seasons of Survivor, in predicting species with high trophic level species in food web, and for determining influential actors in conflict networks in Afghanistan, India, and Pakistan. Metrics such as CON scores outperformed PageRank as an indicator of influential actors in the competition networks we studied. While our results provide support for the DCH, more work needs to be done. We did not address items (1) and (2) of the DCH regarding alliances in our data sets, and that would be an important next step. Another direction is to consider an aggregate score, based on the CON score, closeness, and in- and out-degree, as a measure of detecting leaders in competition networks. An interesting direction would be to study more closely the dynamic aspects of competition networks, analyzing them over time to predict leaders. For example, we could analyze the co-voting network of Survivor of each episode of a season, and determine if temporal trends in network statistics predict finalists. An open question is whether CON score centrality is applicable to large-scale networks exhibiting adversarial interactions, such as in Epinions and Slashdot (which give rise to signed data sets with Bay AnchovyWeakfishHogchokerChrysaoraCtenophoresHerrings and ShadsBlue CrabWhite PerchAmerican EelCatfishMeroplanktonOystersSuspension Feeding BenthosBluefishCroakerStriped BassMeiofaunaMenhadenSpotMicrophytobenthosCiliatesMesozooplanktonDeposit Feeding BenthosRotifersSAVHeteroflagellatesCONHeteroflagellatesDeposit Feeding BenthosRotifersOystersMeroplanktonChrysaoraBlue CrabSuspension Feeding BenthosCiliatesWhite PerchCatfishHerrings and ShadsHogchokerCtenophoresMicrophytobenthosAmerican EelBay AnchovyWeakfishBluefishCroakerMeiofaunaMenhadenMesozooplanktonSAVSpotStriped BassPageRankCentrality in dynamic competition networks 9 Fig. 6. The CrystalC food web dataset. On the left, organisms are listed in decreasing order, with the largest CON score at the top. On the right, organisms are listed by decreasing PageRank score. tens of thousands of nodes, and available from [13]). Epinions was an on-line consumer review site, where users could trust or distrust each other. Slashdot is a social network that contains friend and foe links. A challenge with these data sets from the view of validating the DCH is that there is no inherently defined ranking, as there is in Survivor (via the order contestants were voted off), food webs (trophic level), and in conflict graphs (via political and strategic influence). 5 Acknowledgments The research for this paper was supported by grants from NSERC and Ryerson University. Gleich and Eikmeier acknowledge the support of NSF Awards IIS-1546488, CCF-1909528, the NSF Center for Science of Information STC, CCF-0939370, and the Sloan Foundation. References 1. S. Allesina, M. Pascual, Googling food webs: can an eigenvector measure species' importance for coex- tinctions? PLoS Computational Biology (2009) 59. 2. D. Baird, R.E. Ulanowicz, The seasonal dynamics of the Chesapeake Bay ecosystem, Ecological Mono- graphs (1989) 594 329 -- 364. 3. V. Batagelj, A. Mrvar, Pajek food web datasets, http://vlado.fmf.uni-lj.si/pub/networks/data/ 4. V. Boginski, S. Butenko, P.M. Pardalos, On structural properties of the market graph, In: Innovation in Financial and Economic Networks, Edward Elgar Publishers, pp. 29 -- 45. 5. A. Bonato, A Course on the Web Graph, American Mathematical Society Graduate Studies Series in Mathematics, Providence, Rhode Island, 2008. SilversideGoldspotted KillifishGulf FlounderLongnosed KillifishSheepshead KillifishNeedlefishPinfishSheepsheadSilver JennyGulf KillifishBay AnchovyMulletMoharraBlacktip SharkStingrayStriped AnchovyBenthic InvertebratesZooplanktonCONStingrayBlacktip SharkSilversideBenthic InvertebratesSilver JennyGoldspotted KillifishGulf FlounderBay AnchovyMoharraSheepsheadStriped AnchovyLongnosed KillifishZooplanktonMulletSheepshead KillifishNeedlefishGulf KillifishPinfishPageRank10 Anthony Bonato, Nicole Eikmeier, David F. Gleich, and Rehan Malik 6. A. Bonato, N. Eikmeier, D.F. Gleich, R. Malik, Dynamic competition networks: Detecting alliances and leaders. In: Proceedings of Algorithms and Models for the Web Graph (WAW'18), 2018. 7. A. Bonato, A. Tian, Complex networks and social networks, invited book chapter in: Social Networks, editor E. Kranakis, Springer, Mathematics in Industry series, 2011. 8. U. Brandes, T. Erlebach (Eds.) Network Analysis: Methodological Foundations, LNCS 3418, Springer, 2005. 9. D. Easley, J. Kleinberg, Networks, Crowds, and Markets Reasoning about a Highly Connected World, Cambridge University Press, 2010. 10. J.C. Gower, M.J. Warrens, Similarity, dissimilarity, and distance, measures of, Wiley StatsRef: Statistics Reference Online (2006). 11. W. Guo, X. Lu, G.M. Donate, S. Johnson, The spatial ecology of war and peace, Preprint 2019. 12. F. Heider, The Psychology of Interpersonal Relations, John Wiley & Sons, 1958. 13. J. Leskovec, The Stanford large network dataset collection, http://snap.stanford.edu/data/index.html 14. E. McDonald-Madden, R. Sabbadin, E.T. Game, P.W.J. Baxter, I. Chad`es, H.P. Possingham, Using food-web theory to conserve ecosystems, Nature Communications (2016) 7. 15. J. Leskovec, D. Huttenlocher, J. Kleinberg, Predicting positive and negative links in online social net- works, In: Proceedings of the 19th international conference on World wide web (WWW '10), 2010. 16. D.B. Stouffer, J. Bascompte, Compartmentalization increases food-web persistence, In: Proceedings of the National Academy of Sciences (2011) 1089 3648-3652. 17. Survivor Wiki http://survivor.wikia.com/wiki/Main Page. 18. J. Tang, S. Chang, C. Aggarwal, H. Liu, Negative link prediction in social media, In: Proceedings of the Eighth ACM International Conference on Web Search and Data Mining (WSDM '15), 2015. 19. D.J. Watts, S.H. Strogatz, Collective dynamics of 'small-world' networks, Nature 393 (1998) 440-442. 20. D.B. West, Introduction to Graph Theory, 2nd edition, Prentice Hall, 2001. 21. S-H. Yang, A.J. Smola, B. Long, H. Zha, Y. Chang. Friend or frenemy?: predicting signed ties in social networks. In: Proceedings of the 35th international ACM SIGIR conference on Research and development in information retrieval (SIGIR '12), 2012. 22. W.W. Zachary, An information flow model for conflict and fission in small groups, Journal of Anthro- pological Research 33 (1977) 452-473.
1802.08066	3	1802	2018-03-06T01:49:32	Reputation Systems for News on Twitter: A Large-Scale Study	[ "cs.SI" ]	Social networks offer a ready channel for fake and misleading news to spread and exert influence. This paper examines the performance of different reputation algorithms when applied to a large and statistically significant portion of the news that are spread via Twitter. Our main result is that simple algorithms based on the identity of the users spreading the news, as well as the words appearing in the titles and descriptions of the linked articles, are able to identify a large portion of fake or misleading news, while incurring only very low (<1%) false positive rates for mainstream websites. We believe that these algorithms can be used as the basis of practical, large-scale systems for indicating to consumers which news sites deserve careful scrutiny and skepticism.	cs.SI	cs	Reputation Systems for News on Twitter: A Large-Scale Study Luca de Alfaro1, Massimo Di Pierro2, Rakshit Agrawal3, Eugenio Tacchini4, Gabriele Ballarin5, Marco L. Della Vedova6, Stefano Moret7 1. Department of Computer Science, UC Santa Cruz, CA, USA. [email protected] 2. School of Computing, DePaul University, Chicago, IL, USA. [email protected] 3. Department of Computer Science, UC Santa Cruz, CA, USA. [email protected] 4. Universit`a Cattolica, Piacenza, Italy. [email protected] 5. Independent researcher. [email protected] 6. Universit`a Cattolica, Brescia, Italy. [email protected] 7. ´Ecole Polytechnique F´ed´erale de Lausanne, Switzerland. [email protected] (https://truthvalue.org) Jan 8, 2018 Abstract Social networks offer a ready channel for fake and misleading news to spread and exert influence. This paper examines the performance of different reputation algorithms when applied to a large and statistically significant portion of the news that are spread via Twitter. Our main result is that simple algorithms based on the identity of the users spreading the news, as well as the words appearing in the titles and descriptions of the linked articles, are able to identify a large portion of fake or misleading news, while incurring only very low (< 1%) false positive rates for mainstream websites. We believe that these algorithms can be used as the basis of practical, large-scale systems for indicating to consumers which news sites deserve careful scrutiny and skepticism. 1 Introduction Social networks such as Facebook and Twitter offer fake news the means to spread, and the res- onance chambers where users can consume and reshare them Allcott and Gentzkow (2017). This situation has prompted the research community to create the means to identify and label fake news as they spread across social networks; Shu et al. (2017) provides an excellent and up to date overview of such research. Previous experiments with fake news detection focused on curated sets that captured only a portion of what is shared. Our interest here lies in the question of how fake news detection systems perform when applied to the full variety of news being shared via Twitter. What performance can be obtained? How do the algorithms behave on news from well-known websites? How effective are the methods in unearthing previously unknown websites that spread fake or misleading news? To answer these questions, over a period of several months, we collected a large subset of URLs shared on Twitter pointing to news articles, including all the news posted on mainstream sites and agencies, all the news from selected scientific and peer reviewed publications, all the news from hundreds of sites that have appeared in published curated lists of sites with dubious reputation or low editorial controls, and all news being spread by a selected group of Twitter users we chose to follow. This initial set was then augmented with the result of queries aimed at dataset completion, finding additional users who shared the news, and additional news shared by users in our dataset. The result is a dataset consisting of 5.5 million news articles, 88 million tweets, and 9 million users. For each news article, we collected both the item title and description, and the identities of the users 1 who shared the item. Our experiments are carried out over temporal slices generally consisting of 1.4 million news and 20 million tweets. We experimented with three algorithms for identifying fake or misleading news. The first algorithm, from Tacchini et al. (2017), uses the users that spread a piece of news, and optionally the title and description words, as features, and builds a logistic-regression classifier. The second algorithm is a variant in which the user identities are first aggregated via topic modeling Blei et al. (2003). The third algorithm is a graph-based crowdsourcing algorithm from de Alfaro et al. (2015); Tacchini et al. (2017). All of these algorithms are well suited for large-scale implementation. Our ground truth consists in two independently developed lists of low quality news sites: the Opensources list Opensources (2017), and a list from Metacert1; each list comprises about 500 sites. Our main result consists in showing that the algorithms can be adapted, trained, and tuned to successfully discover a significant portion of the low-quality news in our test dataset (above 50% by default, or above 85% when limiting the comparison to fake news with more than 10 tweets, which constitute the most "viral" news), while incurring a false positive rate on mainstream news below 1%. In preliminary user trials, we observed this low false positive rate to be essential: mainstream reliable media such as the major newspapers produces a very large amount of news daily, and even a small fraction labeled erroneously as fake translates in a large number of incorrectly flagged news, eroding user confidence in the system. We provide extensive results on how the algorithms perform on individual news sites, both mainstream and lesser known. Furthermore, we show that these algorithms are able to generalize from known low-quality news sites to other sites that deserve to be scrutinized. In particular, we show that starting from one of our ground truths, we can recover a majority of the sites belonging to the other ground truth. Together, these results indicate that the algorithms can perform well when applied to the full set of news items shared via Twitter, an essential milestone for their wider adoption. We like to think of these algorithms as reputation systems for news, rather than proper fake news detection algorithms. These algorithms are too shallow semantically to be able to actually discern genuine from fake news; this is a task best carried out by humans, or by systems performing a much deeper analysis of the text, and requires taking into account a broader context. Our aim is more modestly to show that reputation systems with high predictive value and low false positive rate on mainstream news are achievable in practice. The system for fake news detection described in this paper has been implemented as part of the Truth Value Project, and it is available as a service at the URL https://truthvalue.org The Truth Value Project system continuously collects information from Twitter, and it allows uses to query the system for the reputation of any arbitrary news article, identified by its URLs. Users can also vote news articles as being reliable, or fake/misleading; the users' opinions are weighted (positively or negatively) depending on their reputation. Additionally, we offer two other ways to interact with our system: we provide a Twitter bot (hoaxbot) that can respond to enquiries from the users, and a browser bookmarklet which enables users to quickly access the scores of sites they are accessign with their browser. 2 Related Work Fake news is a phenomenon that has received wide attention of late, and motivated much research aimed at detecting fake news automatically; a comprehensive and recent survey is given in Shu et al. (2017). Many approaches are based on the content of the news. Some of these approaches 1http://www.metacert.com 2 find their roots in the much older work aimed at detecting spam; see for instance Mason (2002); Vukovi et al. (2009). Wang (2017) uses deep learning, LSTMs, and SVMs to classify about 12,000 sort text statements from Politifact; Riedel et al. (2017); Ahmed et al. (2017) used a text TF-IDF model followed by a neural net to classify a few thousands text items. The writing style has been used in Potthast et al. (2017) to classify 1,600 items from BuzzFeed. An extensive study was performed in Castillo et al. (2011), which collected tweets verging on 2,500 topics that were trending on Twitter between April and July 2010. The topics were classified on the basis of the features of the users and the tweets belonging to the topics, showing high classification accuracy for the topics; the best-performing classification algorithm was based on a decision-tree classifier. The social dynamics on Twitter and the temporal propagation of posts has been the subject of a very extensive study in Shao et al. (2017), with the aim of quantifying the effect of bot accounts, detecting them, and tracing how the information spreads. The study is similar to ours in extent, involving millions of news items and Twitter users. Temporal aspects of news spreading have been studied and correlated to the reliability of news in Ma et al. (2015), where about a million tweets are used to analyze a hundred news items, and in Shao et al. (2016), where large-scale measurements are provided. The present work is inspired by work on reputation systems, in which the specific identity of users is used as a feature. Some notable recent work in this direction is Tacchini et al. (2017), some of whose algorithms we study in this paper, and Ruchansky et al. (2017), where user identities are fed along with text into a deep-learning system. 3 Dataset 3.1 Dataset composition To characterize the performance of reputation systems for news spread via Twitter, it is essential to have a dataset that covers the full spectrum of news that are shared on Twitter, from mainstream sites, to sites whose news quality is less certain. Since our concern is news reputation, rather than individual tweet reputation, we identify news with URLs, and we restricted our attention to tweets that contain at least one URL. Ideally, we would like our dataset to include all URL-containing tweets and construct the full bipartite graph with URLs on one side, and Twitter users mentioning those URLs on the other side. However, acquiring this much data would be prohibitively expensive. We therefore settled on a sampling strategy that allowed us to collect a large number of news-related tweets and re-tweets, and build a significative sub-graph. We achieved this by gathering URL-containing tweets using the Twitter streaming API. We collected all Tweets from August 1, 2017 until December 8, 2017 that satisfied at least one of the following filters: • Mainstream news: all tweets containing URLs from a list of mainstream news sites, including ABC News, Breitbart, BuzzFeed, CBS News, Channel 7 News, CNN, Fox News, MSNBC, NBC News, The Huffington Post, The Economist, The Guardian, The Hill, The Onion, The New York Post, The New York Times, The Times, The US Herald, The Washington Post, USA Today, and US News; • News agencies: all tweets from The Associated Press and Reuters; • Scientific, and peer-reviewed news: all tweets from ArXiV, Nature, and Science Magazine; • Fact-checking sites: all tweets from Politifact and Snopes; 3 % Site % Site 4.319 wordpress.com 0.754 youtube.com 0.625 3.145 nytimes.com nypost.com 0.619 theguardian.com 2.904 thehill.com huffingtonpost.com 1.964 0.616 latimes.com 0.609 washingtonpost.com 1.944 breitbart.com 0.563 1.585 arxiv.org cbsnews.com usatoday.com 1.504 0.426 reuters.com reddit.com 1.458 indiatimes.com 0.388 foxnews.com dailycaller.com 0.367 1.262 blogspot.com 1.202 newsmax.com 0.336 Table 1: The 20 news sites with the most URLs in our dataset in the period from September 1 to November 30, 2017. • News from selected users: all tweets from a set of about 160 users, which included many of the most active tweeters in news and science; and • Low-quality news: all tweets that mentioned URLs from many of the sites mentioned as fake, bias, unreliable, clickbait, or consipiracy in the lists ?Opensources (2017). The original lists consisted of 1,000 sites; we excluded many that seemed to be no longer active, or that published news only sporadically. In addition, we ran specific queries aimed at completing the graph, by periodically querying (within the limits of the Twitter API) for URLs that were mentioned only once or twice, in the attempt to find more tweets referencing these URLs; and users from whom we had a limited number of tweets, in the attempt to discover further activity from such users. Overall, between the streaming and the graph-completion queries, we gathered about a million URL-containing tweets per day. Upon gathering each tweet, we parsed it, and we downloaded the HTML of the reference page. From the semantic tags (og:url, og:title, and og:description) in the HTML we extracted and stored the canonical URL, the title, and the description. We report in Table 1 the list of news sites with the most URLs in the dataset for the period from September 1 to November 30, 2017. Once the tweets and the URLs had been obtained, we processed the dataset and we associated with each URL the following information: • URL date: the date at which we first saw the URL. • Users: the list of Twitter usernames who shared the URL. • Title: the title of the news article, as given in the og:title social tag. • Description: the description of the news article, as given in the og:description tag. 3.2 Ground Truth All the machine learning algorithms that we utilize follow into the category of semi-supervised learning algorithms. In order to train them we need a set of example URLs that we can reliably categorize as good (reliable news) or bad (fake and misleading news). We refer to this set of categorized URLs as ground truth. 4 Ideally we would like to utilize a large, unbiased, and possibly crowdsourced, ground truth. Unfortunately, such a set is not currently available to the authors. Most sets of crowdsourced URLs are fairly small, generally numbering less than 1,000 URLs. It is difficult to build, on the basis of such small sets, algorithms that generalize well to the whole set of news being shared on Twitter. Furthermore, and perhaps even more importantly, if the testing set is small it does not yield useful information on the algorithms' performance because we cannot extrapolate the results to the full set of URLs collected on Twitter, which is the question that interests us here. As the next-best solution, we use as ground truth, two curated lists of news sites publishing (in the opinions of the curators) fake, misleading, unchecked, and biased news: the Metacert list, and the Opensources list. • Opensources list. The Opensources list consists of sites derived from a list originally posted by Melissa Zimdars Opensources (2017). We kept all sites in that list with type bias, fake, junksci, hate, conspiracy, or clickbait. This resulted in 581 sites. • Metacert list. The Metacert list consists of sites flagged as fake, misleading, or unreliable news sources using a crowdsourcing method by Metacert2, and kindly shared with the authors. The list consists of 500 sites. These two lists have partial overlap, sharing 331 sites. Using these two lists as ground truth entails two problems. First, the lists consist of sites, not of URLs. There is thus the risk that our algorithms simply learn the sites, rather than the characteristics of fake news. We have taken two steps to counter this. As mentioned above, we ensure that no information used in the machine learning directly mentions the site. In particular, we remove all mentions of the sites where the news appear (and their aliases) from the content we use as ML feature. Further, we will present extensive results on cross-learning, in which we use one set of sites (for instance, the Metacert list) to infer the reliability of URLs belonging to another set of sites (for instance, the Opensources list). The second problem is that the Metacert and Opensources lists are only partial lists of sites containing unreliable information. Even when we use the union of the two lists of sites, we certainly cannot assume that all URLs not belonging to these sites contain reliable news. In other words, the false positives of our machine learning - the URLs that are labeled as low reputation, but do not belong to a site in the list - may not be mistakes, in the sense that they may consist in part of URLs that, upon examination, reveal themselves to be fake or misleading news. For this reason, when reporting the performance of the ML algorithms, we will not blindly accept the classification given by the ground truth. Rather, we will report detailed results on which news sites are reported to contain low-reputation news, and in what proportion. As we will see, the ML algorithms we study in general have the useful ability to discover fake and unreliable news beyond those that belong to the ground truth. To give an idea of the breadth of news collected, we give in Figure 1 the correlation between the main news sites, and the sites appearing in the union of the Opensources and Metacert lists on one side, and the scientific site arxiv.org on the other. We compute the correlation between two sites i, j as Tij/(cid:112)TiTj where Tij counts the number of users that tweeted about both domains weighted by their number of tweets; Ti and Tj count the number of users that twitted about domain i and j respectively, also weighted by the number of tweets. 2https://metacert.com/ 5 Figure 1: Average correlation of most tweeted news sites with known fake news sites (Opensources + Metacert lists) vs correlation with science related (arxiv.org). The area of each circle is proportional to the number of URLs we have from each site. The data is from November 1 to November 7, 2017. 6 4 Methods We build reputation systems for news shared on Twitter following three different approaches. In the first approach, we consider each user who shared the news article, and each word that appears in the title and description of the article itself, as an individual feature. This gives rise to a large number of features, on which we train a simple logistic regression model Tacchini et al. (2017). In the second approach, we aggregate users and title/description words using topic analysis, and we then train neural network classifiers based on the resulting topics. In the third approach, we implement the "harmonic" crowdsourcing algorithm described in de Alfaro et al. (2015); Tacchini et al. (2017). 4.1 Logistic-regression classifier based on users and words 4.1.1 Architecture Our first and simplest model is from Tacchini et al. (2017): we use as list of features for a piece of news (a URL after canonicalization) the individual users who shared the URL by tweeting or retweeting it, alongside with the words that appear in the title and description of the news article itself. This gives rise to a large set of features, numbering in the hundred of thousands: each user and each word is a feature. On this large, and rather sparse feature set, we train a simple logistic regression model. Logistic regression models are well suited to coping with large and sparse feature sets. Intuitively, the coefficient associated with each user is positive if the user tends to share high-reputation news, and negative if the user tends to share low-reputation news. Similarly, the coefficient of a word is positive if the word tends to appear in the title or description of high-reputation news, and negative if the word tends to appear in the title or description of low-reputation news. We consider three kind of models: • LR-U, built on users only. • LR-UT, built on users and title/description text. • LR-T, built on title/description text only. The models are all implemented using the scikit-learn Python package. Since our ground truth is based on sites, rather than individual URLs, we removed from the title and descriptions all mentions of the news site to which the individual URLs belong, to avoid having a "cheat factor" in the experiments. 4.1.2 Training and testing datasets To evaluate the logistic-regression based classification methods, we split the dataset into training and testing datasets, as follows. • The full training dataset consists of all URLs that have been first seen from September 1st, 2017, to October 20, 2017. For each URL, we keep all the tweets whose timestamp is strictly before October 30, 2017. • The full testing dataset consists of all URLs that have been first seen from October 31, 2017, to November 26, 2017. For each URL, we keep all the tweets that mention it. 7 Dataset Training dataset Min-2 training dataset Testing dataset URLs 787,601 275,400 607,299 Tweets 14,587,984 14,075,783 7,967,170 Table 2: Sizes of training and testing datasets for logistic-regression based classifiers. Number of URLs Train Min-2 train Test Opensources Metacert Common 14,339 9,104 9,210 21,263 13,213 13,388 21,146 13,305 11,740 Total 787,601 275,400 607,299 Table 3: Number of URLs in the Opensources and Metacert lists that appear in our training and testing sets for logistic-regression classifiers, alongside the total number of URLs in the datasets. The construction of the full training and testing datasets ensures that the URLs they contain are disjoint, and that the testing dataset is built only from information that chronologically precedes all of the information in the testing dataset. The interval from October 20 to October 30, 2017, allows us to collect tweets mentioning the URLs seen first on October 20 for 9 additional days, until midnight of October 29 (all dates used in the experiment are in UTC). As the full training and testing datasets are quite large, we subsampled them by keeping only the tweets appearing on alternating days, starting from September 1, and October 31, respectively. Furthermore, we obtained a min-2 training dataset by considering only URLs that were retweeted at least twice; this weeded out from the training set many low-significance URLs, and let to higher classification precision for mainstream news, as we will see. The characteristics of the resulting datasets are reported in Table 2. The composition of the two ground truths is summarized in Table 3. As the number of positive and negative instances in the training set is highly unbalanced (in the training dataset, only 21,263 of 787,601 URLs are considered fake news), we train the logistic classifier giving class weights that are inversely proportional to class sizes, thus minimizing per-class errors. 4.2 Topic-based models A limitation of the above logistic-based classification model is that we can assign coefficients only to users and words that we have seen in our ground truth. In particular, if a user u(cid:48) has been recently sharing similar URLs as users u1 and u2, and we have extensive data on the quality of the URLs shared by u1 and u2 but not by u(cid:48), we still cannot attribute a coefficient (a positive or negative influence) to u(cid:48). Topic-based models extract the common behavioral similarities among users, summarizing them in a number of topics that each user is likely to share. We can then train our models on the basis of topics, rather than individual users. In the example above, the topic preference of u(cid:48) would be close to the one of u1 and u2, so if we knew how the topic preference of u1 and u2 and how those topics are associated to high or low reputation news, we could transfer this information to u(cid:48). 8 4.2.1 Architecture We create our topic models by following the classical Latent-Dirichelet Allocation models (LDA) for text Blei et al. (2003). We considered each URL as a document, with the users that shared that URL as the words, and we applied standard LDA-based topic-modeling to the resulting document corpus via Gensim Rehurek and Sojka (2010). We constructed the topic modeling on the joint dataset consisting of both training and testing data. Constructing topic models is a fully unsupervised operation which can be easily repeated periodically (e.g., once a day), so that in practice it is feasible to maintain an updated topic model encompassing the whole dataset. Constructing the topic modeling on the joint dataset enabled us to exploit behavior similarity between the users in the training dataset, and those appearing only in the testing dataset. Once the topic model was constructed, we used Gensim to map each URL, with its sharing users, to a topic vector. We first experimented with training a logistic classifier on topic vectors, but this worked rather poorly. The topics provide rich semantical information, and apparently it is necessary to consider the conjunctions, and disjunctions, of topics in order to arrive at a good classification of news URLs. Thus, after some experimentation, we settled on a neural-net classifier with one 100-neuron hidden layer with rectified liner activation, and one final neuron outputting the classification. Again, due to class imbalance between the true and fake classes, in the learning we use class weights that are inversely proportional to class size. 4.2.2 Training and testing datasets Due to computational limitations (topic decomposition is an expensive operation), we used some- what smaller datasets than the ones previously described. Precisely: • The raw training datasets consists of all URLs with at least two tweets that were first seen in 2017 on September 1, 5, 9, 13, 17, 21, 25, 29, and October 3, 7, 11, 15, 19. This dataset contained 144,137 URLs, and encompassed 7,395,756 tweets. • The raw testing dataset consisted, as in Section 4.1, in all URLs first seen on alternating days, starting on October 31, and ending in November 26, 2017. This dataset contained 607,299 URLs, encompassing 7,967,170 tweets. To further simplify the problem computationally, we noted that URLs that are shared only a few times, and users who share only occasionally, do not contribute much information to the topic factorization. Hence, we computed the topic factorization on the basis of URLs and users in the union of the raw training and testing datasets that appeared at least 5 times. Once the topic modeling was computed, we used our raw training set above to train the neural net on the basis of the URL topics. As testing set, again to simplify the problem computationally ,we considered a random subsampling of 20% of the raw testing set mentioned above, resulting in 121,460 URLs. The compositions of the training and testing sets is given in Table 4. 4.3 Crowdsourcing-based models The third method we applied is based on the crowdsourcing-inspired "harmonic boolean label crowdsourcing" model of de Alfaro et al. (2015); Tacchini et al. (2017). As for topic modeling, the crowdsourcing models are developed on the entire dataset; indeed, the system we use for Twitter data acquisition is able to run the harmonic model computation every few hours, ensuring that up-to-date results are always available. 9 Number of URLs Opensources Metacert Common 4,876 1,810 7,069 2,664 7,032 2,331 Total 144,137 121,460 Train Test Table 4: Number of URLs in the Opensources and Metacert lists that appear in our training and testing sets for topic modeling. The different proportion of URLs that belong to the Opensources and Metacert sets, compared to the total URLs, depends on the fact that the training set consists only of URLs that were shared at least twice. In boolean label crowdsourcing models, users provide True/False labels for a set of items, and the crowdsourcing algorithms then attribute to each item an estimate of truth or falsehood, and to each user an estimate of how much the user is likely to tell the truth or to lie Karger et al. (2011); Liu et al. (2012). As in Tacchini et al. (2017), we use a variant of these methods in which the truth or falsehood of a subset of items is taken as known; the algorithms then propagate the information to the remaining users and items. We use the harmonic algorithm of de Alfaro et al. (2015), both because it is efficient on very large-scale data, and because it has been shown to give good results in Tacchini et al. (2017). 4.3.1 The harmonic algorithm We represent our entire dataset as a bipartite graph (I ∪ U, T ), where I are the news items, U are the users, and T ⊆ I × U are the shares, so that a share (i, u) represents user u tweeting item i. We denote by ∂i = {u (i, u) ∈ L} and ∂u = {i (i, u) ∈ L} the 1-neighborhoods of an item i ∈ I and user u ∈ U , respectively. The harmonic algorithm maintains for each node v ∈ I ∪ U two non-negative parameters αv, βv defining a beta distribution: intuitively, for a user u, αu − 1 represents the number of times we have seen the user share a reliable news item, and βu − 1 represents the number of times we have seen the user share an unreliable item. Similarly, for a news item i, αi − 1 is the number of shares from reliable users, and βi − 1 the number of shares from unreliable users. For each node v, let qv = (αv − βv)/(αv + βv): for a news item i, positive values of qi indicate a likely reliable news, and negative values, a likely hoax. Let the training set consist of two subsets IF , IN ⊆ I of fake and non-fake news. The algorithm sets qi := −1 for all i ∈ IF , and qi := 1 for all i ∈ IN ; it sets qi = 0 for all other posts i ∈ I\(IF ∪IN ). The algorithm then proceeds by iteratively propagating the information from items to users, and from users to items. First, for each user u ∈ U , it lets: (cid:88){qi i ∈ ∂u, qi > 0} βu := c −(cid:88){qi i ∈ ∂u, qi < 0} . αu := c + The positive constant c = 0.02 acts as regularization. In the second part of the iteration, the algorithm updates the values for each item i ∈ I \ (IF ∪ IN ) (without affecting the ground truth) by: (cid:88){qu u ∈ ∂i, qu > 0} βi := c −(cid:88){qu u ∈ ∂i, qu < 0} . αi := c + 10 All URLs URLs with ≥2 shares URLs with ≥5 shares URLs with ≥10 shares Hoax recall: Nonhoax recall: Hoax precision: Hoax recall: Nonhoax recall: Hoax precision: Hoax recall: Nonhoax recall: Hoax precision: Hoax recall: Nonhoax recall: Hoax precision: Full Training Min-2 Training LR-U 57.64 97.40 33.30 72.12 95.62 41.44 82.57 95.47 50.24 87.14 95.64 54.28 LR-UT 61.14 97.75 38.15 71.69 96.66 49.19 81.27 96.66 57.43 85.43 96.63 60.16 LR-T 57.25 94.18 18.15 62.69 91.73 24.54 67.38 90.14 27.46 69.90 88.91 27.24 LR-U 46.84 98.76 46.14 63.94 97.55 46.14 79.03 97.00 52.83 85.43 96.77 61.10 LR-UT 53.25 98.51 44.61 66.76 97.84 57.07 78.94 97.49 63.56 84.80 97.33 65.37 LR-T 53.21 94.82 18.81 59.19 93.25 27.34 64.93 92.24 31.66 68.06 91.50 32.23 Topics 54.80 93.88 16.50 66.56 92.23 26.60 76.32 92.83 36.12 83.13 93.20 40.70 Harmonic 1x 4x 91.20 89.74 16.64 91.20 89.74 16.64 91.83 85.44 24.79 93.52 84.54 25.44 74.34 96.69 32.87 74.34 96.60 32.87 69.79 95.32 43.75 70.83 95.43 46.62 Table 5: Hoax and non-hoax recall, and hoax precision, expressed as percentages, for the news reputation systems compared in this paper. The methods are LR-U: logistic regression based on users; LR-UT: logistic regression based on users and text; LR-T: logistic regression based on text; Topics: topic analysis, and Harmonic: harmonic crowdsourcing algorithm. For logistic regression, we report results for two training sets: the full one, and the one consisting of URLs that have been shared at least twice. For the other methods, we only report results on the full training set. For Harmonic, we report the results with both 1x and 4x sampling of good URLs. The results are based on the Opensources ground truth. We experimented with various numbers of iterations, and we settled on using four iterations, which seem sufficient to spread information from the labeled nodes. Once the iterations are concluded, we classify a news item i as fake if qi < 0, and as reliable otherwise; thus, we take qi as the reputation of i. 4.3.2 Training and testing datasets The harmonic algorithm, as mentioned, is run on our entire graph of URL and tweets; the graph consists to date of 5.5 million URLs and 88 million tweets. We use as training set the approximately 2.5 million URLs collected before October 15, 2017, and as testing set, the approximately 2.5 million URLs collected after November 1st, 2017. In the training set, we take as negative ground truth all URLs that belong to the Opensources or Metacert ground truths; let n be the number of such URLs. As n (cid:28) 2, 500, 000, if we labeled positively all the remaining URLs in the training set, the training set would have a strong class imbalance. The harmonic algorithm does not come with a "control knob" comparable to class weight normalization for logistic regression. To obtain a balance between the number of URLs labeled positively and negatively for training, we sample a subset of URLs not in the Opensources and Metacert lists, and we will label only those positively. We consider two subsamplings: the 1x subsampling, in which we subsample n URLs, and the 4x subsampling, in which we subsample 4n URLs. 5 Results Table 5 summarizes the results for the various algorithms under consideration. For each algorithm, and each test-set, we report: • Hoax recall: the percentage of hoax news in the testing dataset that the algorithm correctly labels as hoaxes. • Nonhoax recall: the percentage of non-hoax news in the testing dataset that the algorithm correctly labels as non-hoaxes. 11 • Hoax precision: the percentage of news in the hoax-labeled news in the testing dataset that are indeed hoaxes. The hoax and nonhoax recalls, together, give a picture of the classification correctness for each of the two classes. Generally, we see that the best-performing algorithm is the simple logistic regression based on individual users and words, trained on the whole dataset. We note that the topic modeling algorithm is not competitive with the one based on simple logistic regression. What is more, the benefit of the algorithm based on simple features over topic modeling is maintained in spite of the fact that the topics are generated over the union of the testing and training sets, as explained in Section 4.2.1. We also see that using textual information helps mostly for URLs that are shared only once. Text information alone yields inferior results, chiefly due to the low precision of the resulting classifiers. However, these overall results tell only half of the story. For the true insight, we need to look at the outcome of the algorithms on individual news sites. 5.1 A look at individual news sites To gain a better understanding of the performance of the algorithms, it is useful to look at the classification results for individual news sites, reported in Table 6. The results indicate that if we train on the full dataset, the percentages of major news sites URLs that the algorithms label as hoaxes are rather high. For instance, about 2% of URL of the Washington Post are classified as hoaxes, and so are 1% of the URLs from Reuters and The New York Times. In our preliminary experiments, users were negatively impressed when the system labeled as possible hoaxes news that were obviously reliable and came from news sites with strong editorial controls. Precision of hoax detection on mainstream news sites was highly valued by our users, much more so than missing the occasional hoax that was shared only a few times (and was thus, likely, a fairly obscure news piece). Compounding the problem, these are among the largest and most shared news sites, so that even a small percentage of errors translates into a large number of URLs labeled as hoaxes. For this reason, we find the results obtained with the Min-2 training set to be preferable in practice. Further, we note that not all sites that have a high percentage of URLs labeled as hoaxes in Table 6 are known for their high standards of journalistic accuracy and editorial controls. Indeed, the reputation systems seem to be able to use the Opensources and Metacert list to discover other unreliable sites. This occurs because the reputation systems assign low reputation (low factors) to the users who share known unreliable news, and then follow those users to other news they have shared. As a consequence, the precision figures reported in Table 5 may be underestimates: many of the URLs labeled as hoaxes that belong to sites beyond those in the two ground-truth lists may indeed be hoaxes or misleading news. 5.2 Cross-ground-truth prediction To validate the ability of the reputation systems for fake news to discover additional unreliable news and sites, we conducted an experiment. The two ground truths at our disposal, which have been developed independently, share only 331 of 581 (Opensources) and 500 (Metacert) sites respectively. Thus, we can ask the question: starting from one ground truth, how good are our reputation systems at discovering the additional URLs and sites in the other ground truth? From Table 6, we see that sites with over 5% hoax detection are of uncertain quality, worthy of manual investigation by a human. Taking 5% as the threshold for fake or misleading, and starting from a ground truth, we measured: 12 nytimes.com theguardian.com huffingtonpost.com washingtonpost.com arxiv.org usatoday.com foxnews.com nypost.com thehill.com latimes.com breitbart.com cbsnews.com reuters.com dailycaller.com townhall.com truthfeednews.com hotair.com freedomdaily.com conservativedailypost.com lucianne.com redstate.com theblaze.com newsbusters.org zerohedge.com LR-UT LR-U Harmonic Full Train OS 0.88 1.16 2.61 1.92 0.08 1.09 1.91 3.42 3.39 0.21 5.10 1.53 1.16 25.00 31.41 21.31 9.40 88.89 93.20 15.65 39.23 43.06 18.22 28.18 MC 0.83 1.02 1.96 2.12 0.08 1.13 2.01 3.20 3.43 0.17 3.39 1.88 1.11 41.38 16.74 2.95 2.03 87.96 92.37 96.56 14.67 24.07 21.78 19.86 Min-2 Train OS 0.42 0.48 1.12 0.77 0.00 0.56 1.08 1.99 1.79 0.13 3.73 0.57 0.44 16.76 21.33 16.8 2.99 89.81 95.88 2.67 38.92 37.96 14.22 18.24 MC 0.35 0.40 0.65 0.80 0.00 0.42 0.91 1.71 1.60 0.13 1.79 0.44 0.44 32.07 14.13 2.33 0.96 86.11 95.46 96.56 9.57 14.81 21.78 6.47 Full Train OS 0.83 0.97 1.43 3.07 0.11 0.84 3.07 3.70 4.64 0.38 11.04 1.62 1.07 50.58 38.88 18.82 15.60 81.48 84.54 89.31 39.23 16.67 29.33 31.18 MC 0.73 0.93 1.40 2.98 0.11 0.63 2.77 3.36 3.55 0.42 9.37 1.22 0.93 56.59 23.49 1.87 10.15 78.70 81.03 96.56 18.18 12.50 31.56 23.33 Min-2 Train OS 0.19 0.29 0.53 0.38 0.05 0.17 1.42 1.12 1.91 0.17 6.32 0.52 0.62 39.63 28.71 15.71 10.58 79.63 79.38 77.48 26.48 5.09 20.89 16.40 MC 0.14 0.16 0.46 0.33 0.06 0.14 0.78 1.28 1.09 0.17 4.76 0.39 0.36 50.58 16.2 2.64 7.48 77.78 73.61 96.56 11.96 4.63 31.11 9.24 Entire, 1x OS 4.10 5.13 7.91 7.32 0.59 4.43 40.43 17.94 14.71 2.71 77.96 11.07 3.54 86.41 78.53 98.28 87.38 95.88 98.18 99.14 78.81 76.01 89.98 80.24 MC 3.95 4.81 7.24 7.46 0.45 4.23 37.34 17.01 14.09 2.75 82.20 10.46 3.26 85.99 76.85 97.58 86.22 96.30 98.18 98.45 71.72 73.99 89.74 67.62 Entire, 4x OS 1.37 1.75 2.72 2.45 0.19 1.48 8.05 4.90 3.96 0.83 20.75 3.92 1.29 29.07 31.31 81.92 11.40 79.42 92.66 26.42 23.60 20.61 17.53 45.88 MC 1.18 1.60 2.36 2.10 0.22 1.12 5.67 3.92 3.31 0.56 13.05 3.36 0.98 20.77 18.96 35.75 7.01 71.60 87.89 34.37 10.77 11.92 11.20 17.12 Table 6: Percentage of URLs that are classified as hoaxes for some news sites, including the top news websites of Table 1. Min-2 Train is the training set consisting of URLs that were shared at least twice; Full Train is the full training set. OS stands for Opensources ground truth; MC stands for Metacert ground truth. N. of Detection GT Mtd. Train LR-UT Min-2 OS OS LR-U Min-2 MC LR-UT Min-2 MC LR-U Min-2 LR-UT Full OS Full LR-U OS MC LR-UT Full MC LR-U Full URLs in Direct URL 13.9 32.7 9.5 9.1 21.7 42.6 15.0 18.3 diff 2530 2530 4178 4178 2530 2530 4178 4178 Site 75.0 66.7 56.7 56.7 91.7 75.0 76.7 76.7 Suspicious URL 93.2 89.1 70.0 68.2 95.5 93.3 88.9 83.9 Table 7: Cross dataset detection percentages of fake news and misleading URLs, starting from the ground truth (GT) Opensources (OS) and Metacert (MT). 13 • Direct URL detection: what percentage of URLs appearing only in the other ground truth do we label as hoaxes? • Site detection: how many sites of the other ground truth with at least 20 URLs in the dataset do we detect as fake or misleading? • Suspicious URL detection: what percentage of URLs appearing only in the other ground truth belong to suspicious sites? For site detection, the limitation of sites with at least 20 URLs removes both inactive sites, and sites on which we have so few URLs that computing statistics is unreliable. The results are given in Table 7. As we see, direct URL detection is low: this is in line with the results in Table 6, where we see that only part of URLs belonging to other fake or misleading sites are flagged as hoaxes. On the other hand, the detection of fake or misleading URLs is high, above 80% using full training sets, and above about 70% for the more conservative Min-2 training set. We observe that the Opensources ground truth is more predictive of the Metacert ground truth than the other way round. 6 Discussion We believe that the challenge of building a reputation system for online news sources consists in creating an algorithm that is able to flag a large portion of fake news, while only minimally affecting mainstream news coming from publications with high journalistic standards and editorial controls. We obtained the best results via logistic-regression approaches trained on URLs shared via Twitter that were shared more than once, considering users and title/description words as features. These models had false positive rates for mainstream news sites and they were generally below 1%, as can be seen from Table 6, while yielding an 85% recall for low-quality news URLs shared more than 10 times. Algorithms including only users as features worked almost as well as algorithms that included also title and description as features, except on URLs that were shared only once. Our experi- ments also indicate that algorithms that treat individual users as topics, such as the logistic and crowdsourcing ones, offer superior performance compared to those that consider common features of users as topics. We believe this likely happens because on Twitter some users acts as indicators for particular types of news, systematically (re)tweeting a large portion of them. Indeed, we have reasons to suspect that some of the users having large influence (weights) in the logistic-regression based methods may be bots and this is something we intend to explore in future work. We note that our results on topic modeling are preliminary: there may well be more sophisticated ways of applying topic modeling that lead to superior results. The "harmonic" crowdsourcing-inspired approach worked fairly well. Its main benefit seems to lie in its ability to spread information from a ground truth to a much wider set of news items and users. Another benefit is that, uniquely among the algorithms we considered, it supports incremental learning. In Table 6, we note that the harmonic method leads to the least differences in classification starting from our two different ground truths. The harmonic method however has two weaknesses. First, it is based in a very specific graph algorithm that cannot be extended with additional features in the same way as other machine-learning algorithms can. Second, tuning the harmonic method to deliver good results is a very time-consuming process. In most machine- learning algorithms, one can start with small training and testing sets, experiment with hyper- parameters such as class weights, and then run on larger datasets. The harmonic algorithm, on the other hand, relies heavily on the connectivity structure of the graph, and it is not obvious how to 14 properly sub-sample the graph to conduct small-scale experiments whose results can be extended to the complete graph. Our results also highlight an interesting ability of the reputation algorithms considered, namely, their ability to identify fake and misleading news sites that were not known to us in advance. Indeed, some of the news sites in Table 6 first came to the attention of the authors when looking for false positives. This suggests an additional use of reputation systemsm for news: as a tool that helps human moderators and fact checkers find more sites that deserve inspection. Acknowledgements. We thank Paul Walsh, CEO of Metacert, for making available to us the Metacert dataset of low- quality sites. References Hadeer Ahmed, Issa Traore, and Sherif Saad. 2017. Detection of Online Fake News Using N-Gram Analysis and Machine Learning Techniques. In Intelligent, Secure, and Dependable Systems in Distributed and Cloud Environments (Lecture Notes in Computer Science). Springer, Cham, 127–138. https://doi.org/10.1007/978-3-319-69155-8_9 Hunt Allcott and Matthew Gentzkow. 2017. Social Media and Fake News in the 2016 Election. https://doi.org/10.1257/ Journal of Economic Perspectives 31, 2 (May 2017), 211–236. jep.31.2.211 David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research 3, Jan (2003), 993–1022. Carlos Castillo, Marcelo Mendoza, and Barbara Poblete. 2011. Information credibility on twitter. In Proceedings of the 20th international conference on World wide web. ACM, 675–684. Luca de Alfaro, Vassilis Polychronopoulos, and Michael Shavlovsky. 2015. Reliable aggregation of boolean crowdsourced tasks. In Third AAAI Conference on Human Computation and Crowd- sourcing. http://www.aaai.org/ocs/index.php/HCOMP/HCOMP15/paper/view/11604 David R. Karger, Sewoong Oh, and Devavrat Shah. 2011. Iterative learning for reliable crowd- sourcing systems. In Advances in neural information processing systems. 1953–1961. http:// papers.nips.cc/paper/4396-iterative-learning-for-reliable-crowdsourcing-systems Qiang Liu, Jian Peng, and Alexander T. Ihler. 2012. Variational inference for crowdsourcing. In Advances in neural information processing systems. 692–700. http://papers.nips.cc/paper/ 4627-variational-inference-for-crowdsourcing Jing Ma, Wei Gao, Zhongyu Wei, Yueming Lu, and Kam-Fai Wong. 2015. Detect rumors using time series of social context information on microblogging websites. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management. ACM, 1751–1754. Justin Mason. 2002. Filtering spam with spamassassin. In HEANet Annual Conference. 103. Opensources. 2017. Curated lists of credible and non-credible online sources, available for public use. (2017). https://github.com/OpenSourcesGroup/opensources 15 Martin Potthast, Johannes Kiesel, Kevin Reinartz, Janek Bevendorff, and Benno Stein. 2017. A Stylometric Inquiry into Hyperpartisan and Fake News. arXiv preprint arXiv:1702.05638 (2017). Radim Rehurek and Petr Sojka. 2010. Software Framework for Topic Modelling with Large Corpora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks. ELRA, Valletta, Malta, 45–50. http://is.muni.cz/publication/884893/en. Benjamin Riedel, Isabelle Augenstein, Georgios P. Spithourakis, and Sebastian Riedel. 2017. A simple but tough-to-beat baseline for the Fake News Challenge stance detection task. arXiv:1707.03264 [cs] (July 2017). http://arxiv.org/abs/1707.03264 arXiv: 1707.03264. Natali Ruchansky, Sungyong Seo, and Yan Liu. 2017. CSI: A Hybrid Deep Model for Fake News Detection. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Man- agement. ACM, 797–806. Chengcheng Shao, Giovanni Luca Ciampaglia, Alessandro Flammini, and Filippo Menczer. 2016. Hoaxy: A platform for tracking online misinformation. In Proceedings of the 25th International Conference Companion on World Wide Web. International World Wide Web Conferences Steer- ing Committee, 745–750. Chengcheng Shao, Giovanni Luca Ciampaglia, Onur Varol, Alessandro Flammini, and Filippo Menczer. 2017. The spread of fake news by social bots. arXiv preprint arXiv:1707.07592 (2017). Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and Huan Liu. 2017. Fake news detection on social media: A data mining perspective. ACM SIGKDD Explorations Newsletter 19, 1 (2017), 22–36. Eugenio Tacchini, Gabriele Ballarin, Marco L Della Vedova, Stefano Moret, and Luca de Alfaro. 2017. Some Like it Hoax: Automated Fake News Detection in Social Networks. In Proceed- ings of the Second Workshop on Data Science for Social Good, Vol. 1960. CEUR-WS, Skopje, Macedonia. Marin Vukovi, Kreimir Pripui, and Hrvoje Belani. 2009. An Intelligent Automatic Hoax Detection System. In Knowledge-Based and Intelligent Information and Engineering Systems. Springer, Berlin, Heidelberg, 318–325. https://doi.org/10.1007/978-3-642-04595-0_39 William Yang Wang. 2017. "Liar, Liar Pants on Fire": A New Benchmark Dataset for Fake News Detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL 2017). ACL, Vancouver, BC, Canada. bibtex: wang liar 2017. 16
1703.03366	1	1703	2017-03-09T17:41:18	Knowledge Acquisition: A Complex Networks Approach	[ "cs.SI", "physics.soc-ph" ]	Complex networks have been found to provide a good representation of the structure of knowledge, as understood in terms of discoverable concepts and their relationships. In this context, the discovery process can be modeled as agents walking in a knowledge space. Recent studies proposed more realistic dynamics, including the possibility of agents being influenced by others with higher visibility or by their own memory. However, rather than dealing with these two concepts separately, as previously approached, in this study we propose a multi-agent random walk model for knowledge acquisition that incorporates both concepts. More specifically, we employed the true self avoiding walk alongside a new dynamics based on jumps, in which agents are attracted by the influence of others. That was achieved by using a L\'evy flight influenced by a field of attraction emanating from the agents. In order to evaluate our approach, we use a set of network models and two real networks, one generated from Wikipedia and another from the Web of Science. The results were analyzed globally and by regions. In the global analysis, we found that most of the dynamics parameters do not significantly affect the discovery dynamics. The local analysis revealed a substantial difference of performance depending on the network regions where the dynamics are occurring. In particular, the dynamics at the core of networks tend to be more effective. The choice of the dynamics parameters also had no significant impact to the acquisition performance for the considered knowledge networks, even at the local scale.	cs.SI	cs	Knowledge Acquisition: A Complex Networks Approach Henrique Ferraz de Arruda,1, ∗ Filipi Nascimento Silva,2, † Luciano da Fontoura Costa,2, ‡ and Diego Raphael Amancio1, § 1Institute of Mathematics and Computer Science at São Carlos, University of São Paulo, PO Box 668, 13560-970, São Carlos, SP, Brazil. 2São Carlos Institute of Physics, University of São Paulo, PO Box 369, 13560-970, São Carlos, SP, Brazil 7 1 0 2 r a M 9 ] I S . s c [ 1 v 6 6 3 3 0 . 3 0 7 1 : v i X r a Complex networks have been found to provide a good representation of the structure of knowledge, as understood in terms of discoverable concepts and their relationships. In this context, the discovery process can be modeled as agents walking in a knowledge space. Recent studies proposed more realistic dynamics, including the possibility of agents being influenced by others with higher visibility or by their own memory. However, rather than dealing with these two concepts separately, as previously approached, in this study we propose a multi-agent random walk model for knowledge acquisition that incorporates both concepts. More specifically, we employed the true self avoiding walk alongside a new dynamics based on jumps, in which agents are attracted by the influence of others. That was achieved by using a Lévy flight influenced by a field of attraction emanating from the agents. In order to evaluate our approach, we use a set of network models and two real networks, one generated from Wikipedia and another from the Web of Science. The results were analyzed globally and by regions. In the global analysis, we found that most of the dynamics parameters do not significantly affect the discovery dynamics. The local analysis revealed a substantial difference of performance depending on the network regions where the dynamics are occurring. In particular, the dynamics at the core of networks tend to be more effective. The choice of the dynamics parameters also had no significant impact to the acquisition performance for the considered knowledge networks, even at the local scale. I. INTRODUCTION Understanding how science works and evolves has be- come an important subject of study over the last few years. Science itself can be regarded as being a complex system, which can be approached through concepts from many disciplines such as physics, statistics, linguistics, and information science. The knowledge achieved by hu- manity can be understood as a subset of the knowledge in nature. Moreover, the knowledge is constantly evolv- ing and growing with new discoveries. Such evolution have been studied, for instance in [1, 2]. In particular, scientometry [3] emerged as a new research area inves- tigating how science evolves. Among such studies are those related to the modeling of the discovery process [4– 6], which are, in general, based on investigations of the structure and dynamics existing in a knowledge space. Such dynamics usually involves researchers acquiring in- formation while exchanging information. Furthermore, the discovery process must take into account how knowl- edge is organized [3, 7–9]. The study of how knowledge acquisition takes place allow us to identify the effects influencing the efficiency of such a dynamics and, consequently, to better under- stand how they can improve or constrain the process of learning. An example of such study was conducted by ∗Electronic address: [email protected] †Electronic address: [email protected] ‡Electronic address: [email protected] §Electronic address: [email protected] Silva et al [10], in which it was shown that discoveries in mathematics, in particular theorems, are more likely to be incorporated at the borders of the respective knowl- edge space. In order to understand the knowledge ac- quisition dynamics, some key elements should be con- sidered: (i) how researchers choose their own research topic; (ii) the way in which they spread the obtained results; (iii) how fast the information spreads; (iv) the visibility of specific researchers and (v) the organization of the knowledge itself. Some of these elements were previously considered in literature, such as a model of spreading ideas in a group of multiple agents [11]. Other studies addressed techniques to optimize the exploration of a knowledge space [12–14]. Another important as- pect of interest is knowing how the interactions among researchers impacts the discovery process. For instance, what would be the effects implied in case researchers were strongly influenced by their more famous colleagues. Many works tackled the characterization and modeling of science by using complex networks (e.g. [6, 10, 14–19]). Typically, each node represents a concept while the edges stand for the relationship between them. The dynamics of learning in such networks can be modeled as random walks, in which an agent randomly moves along the net- work through its edges [15, 17]. In this context, the nodes already visited by an agent represent the already learnt concepts. For example, in the methodology proposed by Costa [15], an agent performs a random walk along a multilayer network, which represents the hierarchy of the knowledge structure. Additionally, the movement among layers is only permitted if a certain milestone of learned concepts is reached. Other characteristics can be incor- porated in the random walk, such as memory, in which the agent tends to move to places already known in a net- work [17]. Apart from single agent dynamics, some pa- pers studied the collective discovery process, in which the interactions among multiple researchers are taken into consideration [6, 14, 16, 18]. For example, Forsman et [18] suggests that social and academic networks, as al well as their dynamics, are influenced by the relation- ships among university-level students. In perspective to the previous studies, we propose a model for knowledge acquisition that incorporates two main aspects typically approached separately: knowl- edge memory and multiple agents. The memory aspect is encompassed into the proposed dynamics by an spe- cific type of random walk, the true self-avoiding walk (TSAW), which was found to be one of the most efficient models to explore networks [12]. In the TSAW dynamics, the agent tends to avoid passing through already visited nodes. The dynamics is based on a set of a random walk- ers that simultaneously explore the network. Aside from these two aspects, we also incorporate into the dynamics the possibility of a researcher to change its research focus based on discoveries made by others. In particular, the agents have a probability of performing a jump, which can be understood as a long-range travel across the net- work [20]. In the analyzes made by Foster et al [1], the jump dynamics was found to be related to the risk of re- search. On the other hand, such kind of risk can produce innovation with high impact in science. In this paper, we perform knowledge dynamics for a set of network models and two real networks: a cita- tion [9] and a Wikipedia [8] network. In both networks, the nodes represent articles and the edges, the citations among them. In particular, for the Wikipedia network, an edge exists between two nodes if there is a hyperlink connecting their corresponding articles. Bearing in mind that our approach simulates collec- tive discovery, we measure the performance in terms of the fraction of the total number of explored nodes by the agents after a certain number of iterations. Through our approach, we seek to address some important questions regarding the collective discovery process, such as: if fre- quent changes of the area of study by researchers can positively contribute to the performance in the discov- ery process; investigating influence of researchers with very high visibility to the dynamics; how important is the overall organization of knowledge to the dynamics; and how the dynamics behaves in specific regions of a network (e.g. its borders). The current paper is organized as follows: Section II describes the basic concepts related to the particularities of the adopted random walk model. In Section III, we present, in detail, our model for knowledge acquisition as well as the description of the used dataset. Section IV describes the results obtained from the proposed method- ology. Finally, Section V provides the conclusions and suggestions for further work. II. TRUE SELF-AVOIDING RANDOM WALK 2 The problem of random walks was initially proposed by Karl Pearson, in 1905 [21], in order to study the dy- namics of mosquito swarms in forests. Pearson assumed a mosquito as an agent under a dynamics based on the walking of a drunkard. Such a process starts with the agent placed at a point O along a plane. Next, it moves straight ahead l yards, then, it turns through a random angle and moves another l yards in a straight line. This process is repeated n times. In a short letter to Na- ture, Pearson asked for help on solving the problem of devising the probability distribution P (d, n) of finding such an agent at a certain distance d from O after a high amount of n iterations. The solution was found by Lord Rayleigh [22], who showed that P (d, n) ≈ 2d n e−d2/n as n → ∞, a relationship that is statistically similar to the behavior of the diffusion dynamics. The established link between these two concepts (random walks and diffu- sion), paved the way to the development of the theoretical approaches to understand the characteristics of matter, such as the concept of Brownian motion [23], leading to the discovering of several macroscopic characteristics and the modeling of many real-world phenomena. More recently, some variants of the random walk dy- namics were introduced, which were employed for mod- eling of many systems in a wide range of disciplines, in- cluding the analysis of insect movements [24], modeling of neural activity [25], and other biological processes [26]. Subsequently, random walks were applied in conjunction with complex networks [27–29], which have been used to represent several real-world complex systems. Examples of such developments are: finding influential spreaders in rumor propagation [30], characterizing proteins [31], text mining [32] and knowledge representation [3, 33]. Among the proposed variations of random walk dy- namics on networks, there are those based on self- avoiding random walks (SAW) [34, 35]. In such a walking dynamics, each agent behaves identically as it would do in a traditional random walk; however, the same agent is not allowed to return to the same node. Because at each iteration the agent always visit a new node, this dy- namic is particularly suitable for optimally exploring a network with no a priori information of its global struc- ture. By definition, paths generated from SAWs are finite in length [36], thus a path restart mechanism is needed for exploring the whole network. A drawback associated with SAWs is the creation of discontinuous paths, as a consequence of the path restart mechanism. To over- come this pitfall, variations of the self-avoiding random walk have been proposed, including the true self-avoiding walk (TSAW) [12, 37–39]. In such dynamics, also known as myopic walk, the probability of visiting a neighbor is higher if it has not already been frequently accessed. Note that the TSAW resembles the SAW dynamics, with the difference of not being fully restrictive with regard to neighbors already accessed in the previous visits. In a typical TSAW dynamics, the next node v(t+1) to be taken by an agent must lie in its current neighbor- hood vi ∈ N (v(t)). The selection of v(t+1) is conditioned to a probability Pi that diminishes exponentially with the number of times vi have already been visited, f (vi). Thus, the probability Pi is proportional to wi = α−f (vi), (1) where α is a constant. Given wi, the probability Pi of visiting the next node vi can be computed as wi(cid:80) vj∈N (v(t)) Pi = , wj (2) Henceforth, α = 2. Since both TSAW and SAW dynamics must maintain a memory of the path already taken by the agent, the an- alytical study of such both processes becomes very com- plex. The TSAW dynamics; nonetheless, is more ana- lytically tractable than SAW because the former does not depend on a path restart mechanism [13]. Recently, Kim et al [12] showed that the TSAW dynamics is an efficient way to explore complex networks with agents having only the knowledge of the local structure of the network. III. KNOWLEDGE ACQUISITION MODELING The process of collective discovery can be modeled as a population of agents exploring the knowledge space under certain dynamics, which should also incorporate the interactions among the agents. In this context, the knowledge space can be represented by a complex net- work [3, 33] where potential discoveries are indicated by nodes, while relationships between such discoveries are represented as edges. Examples of such networks include citation or co-citation networks [7, 9, 40–42], Wikipedia networks [8, 43], and learning objects [44]. In this study we consider the organization of the knowl- edge as a complex network, and simulate the researchers as agents under a TSAW dynamics. An example of the proposed TSAW dynamics for an agent is illustrated in Figure 1, where three situations of an agent under such dynamics are shown. Such example illustrates an agent and its respective transitions in the network, where we show how the transition probabilities modifications along some iterations. In addition to the TSAW mechanism, we also incor- porate the concept of stochastic flight [45, 46] to the proposed dynamics. Such a modification is needed to account for a scientific environment of multiple inter- acting agents, which is usually done by reading each other papers, by attending conferences, or via collabo- rations [7, 47–50]. In particular, the flight dynamics is employed to simulate the events when a agent changes its object of study based on the discoveries made by an- other agent. In summary, the TSAW accounts for the 3 local exploration of the knowledge network by a agent, while the stochastic flight allows agents to reach nodes farther away through jumps to knowledge entities close to the recent discoveries. A variety of flight dynamics exists in the literature, in- cluding the Lévy, Cauchy and Rayleigh flights [45]. In our model, we use a flight mechanism with a probability γ. If this mechanism is activated, the agent does not per- form the TSAW dynamics at the current time step. As an alternative, the agent jumps to an arbitrary destination node in the network. The destination node is stochas- tically determined according to an influence field ema- nating from the other agents. In our model, each agent vj emits a field Ea(i) along the topological structure of the network to simulate knowledge dissemination. Such a field is defined in a way so that dissemination decays exponentially with the topological distance d(i, j) from an observation vertex vi to vj. Thus, for each agent, a, this field is computed as Ea(i) = ηi exp(−d(i, j)τ ), (3) (cid:88) j∈V with V being the set of nodes in the knowledge network and τ a constant of the dynamics. The set of values ηi constitutes another parameter of the dynamics and mod- els the fitness of each agent, which allows us to simulate different fields magnitudes emanating from each agent. In other words, by considering this parameter, it is pos- sible to emulate a case where the influence of the agents is not homogeneous. Here, the reduction of the influence field was modeled in terms of exponential decay instead of a power law, so as to constrain the spreading of the influence irrespectively to the network dimension [51–53]. In the proposed model, the field acts as a stochastic attractor on the agents, inducing them to make a jump. The parameters τ and ηi appearing in the definition of the field (equation 3) can be interpreted, respectively, as the dissemination decay and the individual fitness of each agent. The parameter τ defines the overall locality of the influence field. On the other hand, η allows distinct performance to be assigned to each agent. The adopted influence field is illustrated in Figure 2 (a). For the simulations, we adopted a binary distribution of η, in which η can assume only two possible values: ηcommon and ηinfluential. A parameter Dη defines the por- tion of agents having η = ηinfluential. In our model, the fields are combined via superposition, thus, the total field E(i) acting on a vertex vi can be calculated as (cid:88) a∈A E(i) = Ea(i). (4) where A is the set of agents. An example of this field can be seen in Figure 2 (b). Note that when the agent jumps it is attracted to the resultant field E and tends to go to a position near the other agents. Finally, the destination of the jump is chosen among all the nodes of the network 4 FIG. 1: Example of the TSAW dynamics on a 2D lattice along three distinct time steps: (a) first iteration, (b) second iteration and (c) 13th iteration. Node colors indicate the number of times f that a vertex was visited. The colors orange, blue, green, and red represent f = 0, f = 1, f = 2, and f = 3, respectively. For each situation the permitted transitions are shown as arrows between nodes alongside with the transition probability given by equation 2. Network database In order to characterize the performance of our model among different knowledge organizations, we put to- gether a small dataset of networks having many distinct properties. This selection starts with one of the simplest network models, a bidimensional lattice (LA), in which nodes are regularly distributed over a squared grid and are connected by proximity. In this work, each node in such a network is connected to its nearest four nodes, except those lying at the borders of the networks. In addition, in order to maintain the same degree in the whole network and to eliminate the boundary effect, we also incorporate a lattice to our dataset, with the bor- der nodes being toroidally connected, which is hence- forth called toroidal Lattice (TLA). In this case, all nodes of the network become indistinguible among themselves. We also incorporate some random network models to the dataset, namely the Watts-Strogatz (WS) [54], the Barabási-Albert model (BA) [55], the community model developed by Lancichinetti et al (CN ) [9] and the Wax- man model (WAX ) [56]. social network [57], The WS model reproduces the small-world phe- nomenon commonly found in many real-world networks, e.g. food web [58], brain net- works [59, 60], among others [61]. In such networks, most of the nodes can be reached from the other nodes in a small number of steps [62]. Another important characteristic present in many real- world networks is their scale-free nature. The BA model reproduces this characteristic through two mech- anisms: preferential attachment and network growth. More specifically, new nodes are progressively added to the network and are more likely to connect with nodes al- ready presenting many connections, resulting in a power law node degree distribution [62]. We also incorporated a geographic model to the FIG. 2: Example of fields emitted by agents. Item (a) repre- sents the field emitted by a single agent, Ea. Note that for a given agent, the further the node, weaker its respective field will be. Item (b) shows an example of a resultant field, E, which is a superposition of the fields generated by agents A and B. Item (b) also illustrates agent C being attracted by the resulting influence field. In this example ηC = 0 so agent C has no influence over the others. . according to a probability Π(i): E(i)(cid:80) j∈V E(j) Π(i) = . (5) In summary, our model starts with a set of A agents randomly displaced over a network representing the knowledge structure. For each iteration, the agents walk along the network according to the two proposed dynam- ics, TSAW or flight, depending on the jump probability γ. We should emphasize that the TSAW dynamics in our model is not dependent on the interaction between agents. Such influence takes place only during the deci- sion process of a jump. 2/72/72/71/7f = 11/41/41/41/44/313/3112/3112/31f = 3f = 2f = 2f = 2f = 2f = 4f = 2(a)(b)(c)A11/21/21/41/41/41/41/41/81/81/161/81/81/41/21/2A1.1250.750.750.3750.750.750.3750.281250.1410.18751/160.3750.281250.56250.5625B1.125C(a)(b)dataset. In this case, we choose the traditional Waxman model [56], in which nodes are randomly displaced along a two dimensional space and are connected according to a probability that decays exponentially with the distance between each pair of nodes. In addition to the traditional complex network mod- els, we also considered the use of networks presenting community structure, which can be found in many real- world networks representing knowledge, such as citation networks [9] and Wikipedia [8]. In order to do so, we gen- erated benchmark networks of Lancichinetti et al. [63], as we call Community Networks (CN). This model generates scale-free networks with a given number of communities. In addition, there are other parameters in this model, such as the mixing parameter, µ, which defines how often nodes from a community connect with nodes from other communities. Furthermore, there are parameters to con- trol the average degree of the network, kmin and kmax which are minimum and maximum degree, respectively, and the minimum and maximum community size, smin and smax, respectively, where the condition smin > kmin and smax > kmax should be enforced. Further informa- tion about this model can be found in [63]. We also included in our dataset two real-world net- works representing knowledge: a Wikipedia [8] (WIKI ) network and a citation network obtained from the Web of Science [9] (WOS). The WIKI network was obtained from a subset of the Wikipedia incorporating only arti- cles from two main categories: Biology and Mathematics. In such a network, each node represents an article and the edges indicate a reference between two articles. The WOS network was obtained from the set of articles and citations resulting from a query on the Web of Science encompassing only the complex networks field. The employed models were configured to have similar number of nodes as in the selected real networks. In addition, we considered all networks as unweighted and undirected. The dataset of networks is summarized in table I. Additionally, we also obtained a visualization for each network, which is shown in Figure 3. Performance evaluation In order to characterize the performance of the pro- posed dynamics for distinct parameters configurations, we employed a measurement to quantify the learning per- formance of the agents. The dynamics is evaluated ac- cording to their collective discovery performance, more specifically the total number of explored nodes, εT , which is defined as ta(cid:88) εT (ta) = ε(t), (6) t=1 where ta is the actual time (i.e. the current iteration) and ε(t) is the number of newly explored nodes at t, not taking into account those explored on iterations before t. 5 A full understanding of the presented dynamics can be accomplished by studying how the learning performance is affected by distinct topological characteristics, such as by how central a node is in a network, as the dynam- ics may vary substantially depending on the region of the network they are being analyzed. Complex networks can display very distinct local characteristics that may affect the dynamics arising from their structure. For in- stance, it is known that a random walk dynamics can be employed to find the community structure of a given network [64]. For this reason, we also considered the eval- uation of the collective discovery for distinct regions of the networks. In this case, each region corresponds to sets of nodes sharing a similar property, which can be quantified by a topological measurement. In this work, we use the accessibility [65] measurement to characterize distinct regions of the networks. Such measurement is known to detect borders and influential nodes in a high variety of real-world networks [30, 66–68]. In our study we estimate the accessibility as in [68]. The accessibility of a vertex i is defined as −(cid:88)  , A(h) i = exp p(h) ij log p(h) ij (7) j where p(h) is the probability of reaching a vertex j having ij departed from i, after h steps. In this study, we consid- ered h = 3 to avoid the limited size effects as the diame- ter of some networks in our database can reach very low values, such as 6 or even 5. IV. RESULTS AND DISCUSSION We compared the learning performance of our dynam- ics among different networks and for many distinct sets of parameters. Moreover, by considering the structure of each network, we discuss the performance of the pro- posed dynamics in different network regions. Henceforth, the analyses take into consideration 300 realizations for each combination of parameters. As a result, the per- formance of the dynamics is measured in terms of the average total number of explored nodes ((cid:104)εT(cid:105)) calculated over all these realizations. Network dynamics evaluation We first analyzed the lattice network because of its simplicity. We found that the knowledge acquisition per- formance (cid:104)εT(cid:105) in LA and TLA did not vary given sev- eral combinations of parameters (Dη, τ, and γ). For this reason, we only present the results obtained for the LA network. First, we observed that the variations of Dη parameter were weakly reflected in (cid:104)εT(cid:105), as shown in Figure 4 (a). We can infer that the variation of the number of agents with high influence, represented by the TABLE I: Description of the analyzed networks. The number of nodes (N) and the average network degree (k) are shown. 6 Network Description k Parameters Refs. 3.96 4.00 4.00 4.00 p = 0.001 p = 0.005 N – 10k – 10k 10k [54] [54] 10k ≈10k 6.02 α = 1, β = 0.015, and L = 1 [56] 10k [55] 6.00 [63] 5.63 10k ≈12k 7.29 [8] ≈11k 17.08 2 communities, µ = 0.2 – – [9] Bidimensional lattice LA TLA Bidimensional toroidal lattice WS-1 Watts-Strogatz model WS-2 Watts-Strogatz model WAX Waxman model BA CN WIKI Subset of wikipedia encompassing articles from Barabasi-Albert model Fortunato model Biology and Mathematics WOS Citation network obtained for the query "Com- plex Networks" on the Web of Science FIG. 3: Visualization of all selected networks, with colors representing the accessibility of each node (A(h=3) i ). vector Dη, does not change the dynamics performance. The evaluated performance of the other parameters, τ (the locality of the field) and γ (the jump probability), are shown in Figure 4 (b) and Figure 5 (a), respectively. Even though all the curves present high standard devi- ation, the dynamics performance, (cid:104)εT(cid:105), was greatly af- fected by varying τ or γ. We also observe that the per- formance decreases when any of the these two parameters increases. Regarding the parameters Dη and τ, the re- sults obtained for the spatial models, WS-1, WS-2 and WAX are markedly similar to those obtained by the anal- ysis of the LA model. (i) WS-1(j) WS-2(f) WAX(a) Lattice(c) CN(b) BA(g) WIKI(h) WOS(g) WIKI(h) WOS7 ever, that the effects of changing γ is much lower than for the spatial networks. The curves obtained for the BA network also present this same behavior, as shown in Figure 7(a). In addition, the WIKI network presented a general performance substantially lower than the other models even when compared with the CN network, which reproduces many of the WIKI network characteristics, such as the number of nodes, node degree distribution and its community structure. Regarding the other parameters, we found no signifi- cant influence of Dη and τ to the dynamics performance for the considered knowledge networks. In these cases, the curves are very similar to those obtained for the BA network, which are shown in Figures 7(b) and (c). Dynamics evaluation in network regions We analyzed above the networks in terms of the av- erage performance of (cid:104)εT(cid:105) by measuring the dynamics performance globally. Now, we consider different net- work regions. For this purpose, we understand a net- work region as being defined by a set of nodes presenting a certain topological characteristic (such as being at the borders of the network, or presenting similar accessibil- ity). To measure the performance, we consider the value of (cid:104)εT(cid:105) after 1000 iterations of the dynamics. This num- ber of iterations was chosen because, if we consider a small number of iterations, the exploration becomes lo- cal and, consequently, similar for all considered networks. On the other hand, for high values, the regions are totally explored resulting in similar performance. Considering our database, we start the analysis with the LA model. The toroidal configuration is not taken into account because it is impossible to define particular regions in this network. For the LA model, we calculated the Chebyshev distance [69], known as maximum value distance, between the position of each node and the ge- ographical center of the network, instead of accessibility measurement. In order to do so, we used the average position of all network nodes as the geographical center. The obtained results for the LA simulation can be seen in Figure 8, in which the performances of different Dη values were not shown because, in this case, there are no significant variations among (cid:104)εT(cid:105) performance. The mea- sured (cid:104)εT(cid:105) in different network regions was not found to present significant variations when the jump probability, γ, is changed (shown in Figure 8(a)). A similar effect was observed when a range of Dη values was tested. Further- more, as in the first analysis, varying the τ parameter provides a better performance for the regions near the geographical center of the network, but when τ ≤ 0.01 there were no significant variations of (cid:104)εT(cid:105) (shown in Fig- ure 8(b)). A more extensive analysis was done by considering the accessibility measurement, α(3) to characterize the net- works regions. According to this analysis, the variations of Dη vector were distinctly reflected in the dynamics (a) Dη. (b) τ. FIG. 4: Representations of the dynamics executed in LA, where the items (a) and (b) represent the results of the pa- rameter changes of Dη and τ, respectively. By focusing the analysis on the γ parameter, we found similarities among the curves of (cid:104)εT(cid:105) obtained for the net- works LA, WS-1, and WS-2, as shown in Figure 5(a-c). In these cases, we also observed that the standard devi- ations of (cid:104)εT(cid:105) are similar among all configurations. In addition, increasing γ was found to improve the dynam- ics performance significantly for these networks. Despite these similarities, the impact of increasing γ is diminished as the probability of rewiring increases for the WS model (recall that LA also corresponds to a WS model with zero rewiring probability). This indicates that (cid:104)εT(cid:105) tends to be more sensitive to the γ parameter when the network is more spatial or organized. However, the results ob- tained for the WAX, shown in Figure 6, indicate that this behavior is not a direct consequence of the spatiality of networks. The WAX network, a spatial structure, pre- sented particularly low performance variation with the increase of γ. As in the previous cases, it is better to explore the network without using jumps. Furthermore, we also observed that varying τ produced similar results. We also analyzed how the performance can be affected by varying the jump probability (γ) for the WOS, WIKI and CN networks. In contrast to the spatial networks discussed above, these networks are inherently more re- lated to knowledge structures. The performance curves obtained for these networks are shown in Figure 5(d-f). Differently from the results obtained for spatial networks, increasing γ improves the performance. We note, how- 5%10%15%0200400600800100012001400Time0.00.20.40.60.81.020%(cid:682)εT (cid:683)00.00010.0010.010.10.20.40.810200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)8 (a) LA (b) WS-1 (c) WS-2 (d) WOS (e) WIKI (f) CN FIG. 5: Average performances of the proposed dynamics ((cid:104)εT(cid:105)) and corresponding standard deviations obtained by varying the jump probability (γ) considering six networks: (a)LA; (b)WS-1, (c)WS-2, (d)WOS, (e)WIKI and (f)CN. The parameter γ is indicated by colors according to the legend. as shown in Figures 9(c-f). From these results, we con- clude that the number of visible agents did not signifi- cantly impact on the network performance. The factor that most influenced the performance was the centrality of the region being explored, because normally the net- work center is easier to be explored. An exception to this trend is the BA network (f), which did not present sig- nificant performance variations when considering distinct network regions. For the WAX model, we found that no significant dif- ferences of (cid:104)εT(cid:105) when Dη is changed for regions at the borders of the network. However, more expressive differ- ences of (cid:104)εT(cid:105) were observed in the central region. The standard deviations were also higher, as shown in Fig- ure 9(b). The simulation for the WS networks showed that the percentage of influent agents, Dη, is directly re- lated to (cid:104)εT(cid:105). These results are shown in Figure 9(c). The results obtained for the analyses involving jump probability γ variation are shown in Figure 10. For all curves, the dynamics performance (cid:104)εT(cid:105) was found to be higher at the central regions of the networks (i.e. those with nodes having high accessibility). Except for the WOS(d) and BA(f) networks, the curves monotonically increase with the average accessibility of the regions. Similarly to the results obtained by varying η, the per- formance for the BA network does not vary substantially along the network regions nor by changing γ. However, FIG. 6: Performance, (cid:104)εT(cid:105), of our dynamics, varying the val- ues of γ adopted in the WAX networks. for each network region. Since the results obtained for the parameter τ are almost the same as those obtained for Dη, the following discussion regarding this parameter also holds for Dη. The performance obtained for the considered networks by varying Dη are shown in Figure 9. In most cases, variations of Dη did not imply in performance variation for any of the network regions. However, when compar- ing among regions in the same network, the most central regions tended to have higher values of (cid:104)εT(cid:105). Such an effect was observed in all the considered networks repre- senting knowledge structure (CN, WOS, WIKI and BA), 0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)0200400600800100012001400Time0.00.20.40.60.81.00.0.0010.010.10.5(cid:682)εT (cid:683)0.0.0010.010.10.50200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)9 (a) γ (b) Dη (c) τ FIG. 7: The performance, (cid:104)εT(cid:105), obtained from the simulation of our dynamics on the BA network while varying the parameters: (a)γ, (b)Dη and (c)τ. We note that the influence of the variations of Dη and τ were not significative. On the other hand, the network exploration is more efficient when the jump probability γ is high. is that, the standard deviations obtained for the knowl- edge networks are substantially lower compared to those obtained for the spatial networks. This indicates that the peculiar behavior observed for the WOS networks is not caused by statistical fluctuations but may be the conse- quence of a more complex topological trait. In general, for the considered networks representing knowledge structure (WOS, WIKI, CN and BA), in con- trast with the spatial networks (LA, WS-2 and WAX), the dynamics do not change much by varying its param- eters. In the average, the performance increases only by a small factor, for jump probability and changes even less for the other parameters. However, for the WOS network, depending on the network region, the improve- ment of performance by varying this parameter can be substantially higher than the other networks. V. CONCLUSION The problem of how humans represent and acquire knowledge has received growing attention along the last years as a consequence of its potential for understand- ing and improving the speed of research and learning (e.g. [1, 2]). These problems are considerably com- plex because they involve several aspects that can in- fluence the efficiency of how knowledge is achieved, ex- changed and disseminated. Such aspects include, for in- stance, the complexity of the knowledge structure itself, the visibility of researchers, the available memory, the strategies adopted for deciding the next possible sub- jects of research, amongst many others. One particularly promising approach to understanding knowledge acqui- sition consists in representing the knowledge as a com- plex networks, and researchers as agents that move along such networks. A number of approaches adopting such a framework have been proposed in the literature [4–6] to study the learning process, i.e. the discovery of new concepts by researchers. In this study, we propose a sys- tematic approach focusing on two important elements in- (a) γ (b) τ FIG. 8: The performance (cid:104)εT(cid:105) obtained from the dynamics simulation for each network region. Such regions were defined according to the distance between each node and the central position of LA. Two parameters, γ (a) and τ (b), were varied. it is interesting to note that, for regions presenting very low accessibility in the BA network, the performance has an opposite trend to the global behavior, as it decreases with γ. In most cases, the influence of varying γ to the perfor- mance tends to decrease with the average accessibility of the considered region. However, in the case of the WOS network, for a certain range of accessibility values around 1700, the performance have a significantly drop of perfor- mance. This is interesting because other models did not present such behavior. Another interesting observation 0.0.0010.010.10.50200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)5%10%15%20%0200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)00.00010.0010.010.10.20.40.810200400600800100012001400Time0.00.20.40.60.81.0(cid:682)εT (cid:683)152025303540450.00.20.40.60.81.00.0.0010.010.10.5Avg. distance (cid:682)εT (cid:683)(cid:682)εT (cid:683)00.00010.0010.010.10.20.40.81152025303540450.00.20.40.60.81.0Avg. distance10 (a) WS-2 (b) WAX (c) CN (d) WOS (e) WIKI (f) BA FIG. 9: The performance (cid:104)εT(cid:105) along the network regions for distinct values of Dη. The regions were defined in terms of the accessibility measurement with h = 3. For this analysis, the considered networks are: (a)WS-2; (b)WAX; (c)CN, (d)WOS, (e)WIKI, (f)BA. (a) WS-2 (b) WAX (c) CN (d) WOS (e) WIKI (f) BA FIG. 10: The performance (cid:104)εT(cid:105) for the WS-2 (a), WAX (b), CN (c), WOS (d), WIKI (e) and BA (f) networks, measured from the dynamics simulation for each network region, where the regions were computed according to the α(3) measurement. In this test, we vary the values of the parameter γ. 0200400600800100012000.20.30.40.50.60.70.80.9Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)020040060080010001200020.30.40.50.60.70.80.9Avg. accessibility -(cid:682)α3(cid:683) .(cid:682)εT (cid:683)0.20.30.40.50.60.70.80.9Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)02004006008001000120002.0200400600800100012000.20.30.40.50.60.70.80.902004006008001000120002.Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)0200400600800100012000.20.30.40.50.60.70.80.902004006008001000120002.Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)(cid:682)εT (cid:683)0200400600800100012000.20.30.40.50.60.70.80.95%10%15%20%Avg. accessibility -(cid:682)α3(cid:683) 212223242526272829Avg. accessibility -(cid:682)α3(cid:683) 0.30.40.50.60.70.80.9(cid:682)εT (cid:683)4060801001201401600.30.40.50.60.70.80.9Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)01002003004005006007008000.30.40.50.60.70.80.9Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)(cid:682)εT (cid:683)Avg. accessibility -(cid:682)α3(cid:683) (cid:682)εT (cid:683)Avg. accessibility -(cid:682)α3(cid:683) 20040060080010000.0.0010.010.10.5Avg. accessibility -(cid:682)α3(cid:683)(cid:682)εT (cid:683)0.000000.34.5.6.7.8.9fluencing knowledge acquisition, namely the memory and visibility of agents, in order to better understand the col- lective discovery process. The memory aspect was inves- tigated in terms of true self-avoiding dynamics occurring in a knowledge space. Additionally, interactions among researchers were modeled by means of a flight dynamics biased towards the most visible researchers. We execute our dynamics in many networks, including two real networks (WIKI and WOS) and a set of net- work models with distinctive topological properties. We observed that the performance of knowledge acquisition can be distinctly optimized for different characteristics of the networks. For example, by increasing the jump prob- ability (γ), the speed of learning also rises significantly in the case of BA networks. This could be a consequence of the fact that, in such networks, the agents need to pass through hubs to access other nodes up to the point that it starts to avoid the hubs completely. This results in such an agent becoming trapped by a set of nodes be- cause the majority of the shortest paths go across hubs. In a BA network, a jumping agent can move without crossing the hubs, thus having more possibilities to ac- cess new nodes. On the other hand, such a configuration was found to be less effective in WAX networks, since when an agent jumps towards other nodes, it will tend to navigate through already learned concepts. We also investigated the performance by considering distinct network regions, such as the borders and the center of networks. We found that, in most cases, col- lective discovery occurs faster at the core of the net- work and becomes slower at the borders. Regarding the three parameters controlling the dynamics, in contrast with the results obtained for the global analysis, some dynamics configurations can indeed change the learning performance depending on the properties of regions being 11 explored. However, for the networks representing knowl- edge structure, the average gain in performance is low irrespectively to the dynamics parameters. An exception to this trend is the WOS network, which presents a set of regions where the performance can vary substantially under influence of the jump probability. In this case, it is possible to enhance knowledge acquisition by a sub- stantial factor. These results indicate that, for a typical knowledge network, the heuristics governing the way re- searchers seek for new knowledge does not substantially affect the global performance of the collective discovery, but still can have influence depending on the properties of the region of knowledge under investigation. The current study adopted a random walk approach incorporating jumps in a multi-agent dynamics. In fu- ture works, we intend to probe the effects of considering other network topologies, including directed structures, on the proposed collective learning model. Additionally, some dynamic characteristics can also be investigated, such as other visibility mechanisms and agents with dif- ferent features (e.g. memory, speed, etc.). Furthermore, our study can be extended to incorporate two layers: one representing the knowledge organization, and the other a network of interactions among researchers. Acknowledgements The authors acknowledge financial from Foundation Capes-Brazil, 2016/19069-9, 2015/08003-4, (FAPESP) (grant no. 2014/20830-0 and 2011/50761-2), CNPq-Brazil (grant no. 307333/2013-2) and NAP-PRP-USP. Paulo Research support São [1] J. G. Foster, A. Rzhetsky, and J. A. Evans, American Sociological Review 80, 875 (2015). [2] M. Cokol, I. Iossifov, C. Weinreb, and A. Rzhetsky, Na- ture biotechnology 23, 1243 (2005). [3] L. Leydesdorff and I. Rafols, Journal of the American Society for Information Science and Technology 60, 348 (2009). [4] I. T. Koponen and M. Nousiainen, Journal of Complex Networks 2, 187 (2014). [10] F. N. Silva, B. A. Travencolo, M. P. Viana, and L. da Fon- toura Costa, Journal of Physics A: Mathematical and Theoretical 43, 325202 (2010). [11] J. Burridge and M. Gnacik, Physical Review E 94, [12] Y. Kim, S. Park, and S.-H. Yook, Physical Review E 94, 062319 (2016). 042309 (2016). [13] V. M. López Millán, V. Cholvi, L. López, and A. Fernán- dez Anta, Networks 60, 71 (2012). [5] L. Pareschi and G. Toscani, Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 372 (2014). [6] J. Batista and L. d. F. Costa, Physical Review E 82, [14] A. Rzhetsky, J. G. Foster, I. T. Foster, and J. A. Evans, Proceedings of the National Academy of Sciences 112, 14569 (2015). [15] L. da Fontoura Costa, Physical Review E 74, 026103 016103 (2010). [7] T. Martin, B. Ball, B. Karrer, and M. E. J. Newman, Phys. Rev. E 88, 012814 (2013). [8] F. N. Silva, M. P. Viana, B. A. N. Travençolo, and L. d. F. Costa, Journal of informetrics 5, 431 (2011). [9] F. N. Silva, D. R. Amancio, M. Bardosova, L. F. Costa, and O. N. Oliveira Jr., Journal of Informetrics 10, 487 (2016). (2006). [16] J. C. González-Avella, V. M. Eguíluz, M. Marsili, F. Vega-Redondo, and M. San Miguel, PLoS One 6, e20207 (2011). [17] D. Boyer and C. Solis-Salas, Physical review letters 112, 240601 (2014). [18] J. Forsman, R. Moll, and C. Linder, Physical Review Special Topics-Physics Education Research 10, 020122 [19] J. Guan, Y. Yan, and J. J. Zhang, Journal of Informetrics 173 (Macmillan, 1983). (2014). 11, 407 (2017). [20] E. Estrada, J.-C. Delvenne, N. Hatano, J. L. Mateos, R. Metzler, A. P. Riascos, and M. T. Schaub, arXiv preprint arXiv:1612.08631 (2016). [21] K. Pearson, Nature 72, 342 (1905). [22] L. Rayleigh, Nature 72, 318 (1905). [23] I. Horváth, B. Tóth, and B. Vető, Probability Theory and Related Fields 153, 691 (2012). [24] P. Kareiva and N. Shigesada, Oecologia 56, 234 (1983). [25] G. L. Gerstein and B. Mandelbrot, Biophysical journal 4, 41 (1964). of the Royal Society Interface 5, 813 (2008). [27] L. d. F. Costa, O. N. Oliveira Jr, G. Travieso, F. A. Rodrigues, P. R. Villas Boas, L. Antiqueira, M. P. Viana, and L. E. Correa Rocha, Advances in Physics 60, 329 (2011). [28] L. da Fontoura Costa and G. Travieso, Physical Review [29] J. D. Noh and H. Rieger, Physical review letters 92, E 75, 016102 (2007). 118701 (2004). [30] G. F. de Arruda, A. L. Barbieri, P. M. Rodríguez, F. A. Rodrigues, Y. Moreno, and L. d. F. Costa, Physical Re- view E 90, 032812 (2014). [31] F. A. Rodrigues and L. da Fontoura Costa, Molecular BioSystems 6, 234 (2009). [32] R. Mihalcea, in International Conference on Intelli- gent Text Processing and Computational Linguistics (Springer, 2006), pp. 249–262. [33] F. N. Silva, F. A. Rodrigues, O. N. Oliveira Jr., and L. d. F. Costa, Journal of Informetrics 7, 469 (2013). [34] C. P. Herrero, Physical Review E 71, 016103 (2005). [35] C. P. Herrero and M. Saboyá, Physical Review E 68, 026106 (2003). [36] I. Tishby, O. Biham, and E. Katzav, Journal of Physics A: Mathematical and Theoretical 49, 285002 (2016). 301 (1984). [38] A. Stella, S. de Queiroz, P. Duxbury, and R. Stinch- combe, Journal of Physics A: Mathematical and General 17, 1903 (1984). [39] D. J. Amit, G. Parisi, and L. Peliti, Physical Review B 27, 1635 (1983). 391, 3533 (2012). 12 [45] B. B. Mandelbrot, The fractal geometry of nature, vol. [46] D. Volchenkov, Communications in Nonlinear Science and Numerical Simulation 16, 21 (2011). [47] T. ZHOU, B.-H. WANG, Y.-D. JIN, D.-R. HE, P.-P. ZHANG, Y. HE, B.-B. SU, K. CHEN, Z.-Z. ZHANG, and J.-G. LIU, International Journal of Modern Physics C 18, 297 (2007). [48] M. P. Viana, D. R. Amancio, and L. F. Costa, Journal of Informetrics 7, 371 (2013). [49] D. R. Amancio, O. Oliveira Jr., and L. da F. Costa, EPL (Europhysics Letters) 99, 48002 (2012). [50] M. E. J. Newman, Proceedings of the National Academy [51] J. D. Jackson, Electrodynamics (Wiley Online Library, 1975). [52] F. N. Silva and L. d. F. Costa, arXiv preprint [53] D. Wei, B. Wei, Y. Hu, H. Zhang, and Y. Deng, Physics arXiv:1209.2476 (2012). Letters A 378, 1091 (2014). [54] D. J. Watts and S. H. Strogatz, nature 393, 440 (1998). [55] A.-L. Barabási and R. Albert, science 286, 509 (1999). [56] B. M. Waxman, IEEE Journal on Selected Areas in Com- munications 6, 1617 (1988). [57] S. Wasserman and K. Faust, Social network analysis: Methods and applications, vol. 8 (Cambridge university press, 1994). [58] J. A. Dunne, R. J. Williams, and N. D. Martinez, Pro- ceedings of the National Academy of Sciences 99, 12917 (2002). [59] O. Sporns and J. D. Zwi, Neuroinformatics 2, 145 (2004). [60] D. S. Bassett and E. Bullmore, The neuroscientist 12, 512 (2006). [61] L. A. N. Amaral, A. Scala, M. Barthelemy, and H. E. Stanley, Proceedings of the national academy of sciences 97, 11149 (2000). [62] L. d. F. Costa, R. F.A., T. G., and B. P.R.V., Advances in Physics 56, 167 (2007), ISSN 0001-8732. review E 78, 046110 (2008). [64] M. Rosvall and C. T. Bergstrom, Proceedings of the Na- tional Academy of Sciences 105, 1118 (2008). [65] F. A. Rodrigues and L. da Fontoura Costa, Journal of Neuroscience Methods 183, 57 (2009). [66] B. A. Travençolo, M. P. Viana, and L. da Fontoura Costa, [67] D. R. Amancio, Journal of Statistical Mechanics: Theory [26] E. A. Codling, M. J. Plank, and S. Benhamou, Journal of Sciences 98, 404 (2001). [37] P. Lam, Zeitschrift für Physik B Condensed Matter 57, [63] A. Lancichinetti, S. Fortunato, and F. Radicchi, Physical [40] F.-X. Ren, H.-W. Shen, and X.-Q. Cheng, Physica A New Journal of Physics 11, 063019 (2009). [41] D. R. Amancio, O. N. Oliveira Jr., and L. F. Costa, Jour- and Experiment 2015, P03005 (2015). nal of Informetrics 6, 427 (2012). [68] B. A. N. Travençolo and L. d. F. Costa, Physics Letters [42] D. R. Amancio, M. G. V. Nunes, O. N. Oliveira Jr., and A 373, 89 (2008). L. F. Costa, Scientometrics 91, 827 (2012). [43] M. Ibrahim, C. M. Danforth, and P. S. Dodds, Journal of Computational Science 19, 21 (2017). [44] P. Korytkowski and K. Sikora, in International Con- ference on Database and Expert Systems Applications (Springer, 2007), pp. 710–719. [69] F. van der Heijden, R. P. W. Duin, D. de Ridder, and D. M. J. Tax, Classification, parameter estimation and state estimation - an engineering approach using Matlab (John Wiley & Sons, Chichester, 2004).
1205.5509	1	1205	2012-05-24T17:33:18	Four Degrees of Separation, Really	[ "cs.SI", "physics.soc-ph" ]	We recently measured the average distance of users in the Facebook graph, spurring comments in the scientific community as well as in the general press ("Four Degrees of Separation"). A number of interesting criticisms have been made about the meaningfulness, methods and consequences of the experiment we performed. In this paper we want to discuss some methodological aspects that we deem important to underline in the form of answers to the questions we have read in newspapers, magazines, blogs, or heard from colleagues. We indulge in some reflections on the actual meaning of "average distance" and make a number of side observations showing that, yes, 3.74 "degrees of separation" are really few.	cs.SI	cs	Four Degrees of Separation, Really Paolo Boldi Sebastiano Vigna Dipartimento di Informatica Università degli Studi di Milano Italy 2 1 0 2 y a M 4 2 ] I S . s c [ 1 v 9 0 5 5 . 5 0 2 1 : v i X r a Abstract—We recently measured the average distance of users in the Facebook graph, spurring comments in the scientific community as well as in the general press [1]. A number of interesting criticisms have been made about the meaningfulness, methods and consequences of the experiment we performed. In this paper we want to discuss some methodological aspects that we deem important to underline in the form of answers to the questions we have read in newspapers, magazines, blogs, or heard from colleagues. We indulge in some reflections on the actual meaning of “average distance” and make a number of side observations showing that, yes, 3.74 “degrees of separation” are really few. FOUR DEGREES OF SEPARATION In 2011, together with Marco Rosa, we developed a new tool for studying the distance distribution of very large (unweighted) graphs, called Hyper- ANF [2]: this algorithm built on powerful graph compression techniques [3] and on the idea of diffusive computation pioneered in [4]. The new tool made it possible to accurately study the dis- tance distribution of graphs orders of magnitude larger than it was previously possible. The work on HyperANF was presented at the 20th World- Wide Web Conference, in Hyderabad (India), and Lars Backstrom happened to listen to the talk; he was intrigued by the possibility of experimenting our software on the Facebook graph and suggested a collaboration. Experiments were performed in the summer of 2011, resulting in the first world-scale social- network graph-distance computations, using the en- tire Facebook network of active users (721 million users, 69 billion friendship links). The average dis- tance (i.e., shortest-path length) observed was 4.74, corresponding to 3.74 intermediaries (or “degrees Partially supported by a Yahoo! faculty grant and by by the EU- FET grant NADINE (GA 288956). of separation”, in Milgram’s parlance). These and other findings were finally presented in [1] and made public by Facebook through its technical blog on November 19, 2011. Immediately after the announcement, the news appeared in the general press, starting from the New York Times [5]1 and soon spreading worldwide in newspapers, blogs and forums. A number of interesting criticisms have been made about the meaningfulness, methods and con- sequences of the experiment we performed. In this paper we want to discuss some methodological aspects that we deem important. We shall consider such issues in an answer-to-question style, with the double aim of replying to doubts and attacks and of stimulating new discussions and further interest. I. NOT ALL PAIRS ARE CONNECTED: HOW CAN THE AVERAGE DISTANCE BE EVEN FINITE? If by “average distance” we mean “average of the distances between all pairs”, of course Facebook has an infinite average distance, as we know that there is a very large connected component containing almost all (99.9%) nodes, but there are also some (few) unreachable pairs. This is an interesting comment, as it shows an actual black hole in all the literature: people study- ing social problems (starting with the 50s, at least) had in mind very small groups, possibly groups that would fit one room (actually, in some cases, just sitting around a table). Or small communities. The very idea of “unreachable” was not part of the picture. In the famous paper by Travers and Milgram [6], the vast majority of postcards did not 1Incidentally, with an off-by-one error, as 4.74 is the average distance, whereas the average number of degrees of separation is 3.74 (see [1]). We discuss here in some detail reach the target2. Nonetheless, the “six degrees of separation” idea came from the average distance (5.4 to 6.7, depending on the group) obtained in the experiment, computed just on reachable pairs.3 two possible mathematical solutions to this problem—not only because they are interesting, but because we want to urge researchers to take the problem into considera- tion more seriously, and to remark to those objecting to the use of reachable pairs that old results would be really stated differently if unreachable pairs were correctly taken into account. An obvious patch is to quote the average distance between reachable pairs, sided by the percentage of reachable pairs, which should be considered as a sort of confidence on the measure. If the percentace of reachable pairs is low, the average distance is telling us little. On a completely disconnected graph, the average distance is 0, but with “confidence” 1/n. On a perfect match,4 the average distance is 1/2, but the “confidence” is 2/n (in both cases, almost zero for large graphs). Seen in this perspective, Milgram’s experiment proposes an average distance of 6.2 but provides an incredibly low level of confidence—just 22%,5 whereas in our case we can claim confidence 99.9% for our value (4.74). The problem is that we like to compare results, and comparing two pairs of numbers can be difficult, if not impossible (see, e.g., the plethora of methods used to combine somehow precision and recall in information retrieval). A solution that does not show the latter drawback is to consider harmonic means when working with distances. We recall that the harmonic mean is the is reciprocal of the mean of the reciprocals. It 2It should be noted, as an aside, that in Milgram’s experiment the interrupted chains do not actually imply unreachability, a point that will be better discussed later. 3Indeed, the authors of one of the first studies of the web as a whole [7] noted the same problem, and proposed the name aver- age connected distance. We refrain, however, from using the word “connected” as it somehow implies a bidirectional (or, if you prefer, undirected) connection. The notion of average distance between all pairs is useless in a graph in which not all pairs are reachable, as it is necessarily infinite, so no confusion can arise. 4A perfect match is an undirected 1-regular graph, that is, a set of disconnected edges. 5Travers and Milgram’s paper [6] reports 29%, as this is the percentage of chains that started and completed with respect to those that started. Some of the chains did not start at all, and we are considering them as incomplete, which explains the slightly slower value we are reporting. 2 always smaller than the arithmetic mean, as it tends to give less relevance to large outliers and more relevance to small values, and it is used in a number of contexts6. The important feature of the harmonic mean is that if we stipulate that 1/∞ = 0, it can take in ∞ as a perfectly valid distance. Its effect is that of making the mean larger in a hyperbolic fashion. This is why Marchiori and Latora [9] proposed to consider the harmonic mean of all distances between distinct nodes7, which we call harmonic diameter following Fogaras [10] (rather than “average distance between reachable pairs”), as a measure of tightness of a network. For instance, a disconnected graph has av- erage distance zero, but infinite harmonic diameter; and a perfect match has average distance 1/2, but harmonic diameter n − 1. What happens if we switch from the average dis- tance to the harmonic diameter? On highly discon- nected network, with many missing paths, we get a larger number. On the LAW web site8 you can find the basic statistics of several web-graph snapshots, and the harmonic diameter is always significantly larger than the average distance between reachable pairs. In the case of Facebook, the harmonic diameter is 4.59—even smaller than the average distance. The situation, however, is quite different if we make the same computation with Milgrams’ experiment and assume that incomplete chains correspond to unreachable pairs: overall, the harmonic mean is 18.29, almost four times larger than the average distance. If we restrict to the Nebraska random group (i.e., we avoid geographical or cultural clues), the harmonic mean is more than five times larger. By this measure, the improvement described in [1] is even more impressive. The problem with the harmonic diameter is that even if it is a clearly and sensibly defined mathe- matical feature, it deprives us from the “degree of separation” metaphore. The fact that in 2007 the harmonic diameter of it was more than 15 000 does not mean, of course, that you need to pass through 6Incidentally, the HyperLogLog counters [8] used by Hyper- ANF [2], the algorithm with which the average distance of Face- book was computed, use the harmonic mean to perform stochastic averaging. 7The fact that we do not consider the distances d(x, x) is essential, as otherwise the harmonic mean becomes zero. 8http://law.dsi.unimi.it/ HARMONIC DIAMETER OF THE GRAPHS FROM [1]. TABLE I 2007 2008 2009 2010 2011 current it 15083.99 (±298.82) 23.66 (±0.75) 4.74 (±0.11) 3.92 (±0.13) 3.76 (±0.11) 3.68 (±0.10) se 51.07 (±1.50) 4.37 (±0.15) 4.37 (±0.11) 3.90 (±0.16) 3.93 (±0.16) 3.69 (±0.20) itse 3760.77 (±161.28) 6.44 (±0.21) 4.71 (±0.11) 4.24 (±0.18) 4.29 (±0.18) 3.90 (±0.13) us 4.16 (±0.14) 4.61 (±0.16) 4.67 (±0.16) 4.68 (±0.15) 4.23 (±0.13) 4.45 (±0.11) fb 6.33 (±0.26) 5.74 (±0.24) 5.07 (±0.21) 5.03 (±0.21) 4.70 (±0.30) 4.59 (±0.13) 3 TABLE II THE HARMONIC MEAN AND THE MEAN OF ALL DISTANCES (INCLUDING ∞ FOR BROKEN CHAINS) FOR THE GROUPS DETAILED IN TRAVERS AND MILGRAM’S PAPER [6]. NOTE THE SIGNIFICANTLY LOWER VALUE OF THE HARMONIC MEAN FOR THE BOSTON GROUP. Group Harmonic mean Median distance Nebraska random Nebraska stockholders All Nebraska Boston random All 26.68 19.37 22.40 12.63 18.29 ∞ ∞ ∞ ∞ ∞ 15 000 friendship links! Another possibility for taking into account infinite distances is to use the median of all distances as a measure of closeness. That is, we list in increasing order the n2 values of d(x, y), and we take that of index ⌊n2/2⌋ (numbering from zero). This number is significantly larger than the average distance if several pairs are unreachable because the ∞ values at the end of the list “push” the median to the right. Again, on the LAW web site you can see that in several web graphs the median of all distances is significantly larger than the average distance, as it takes into account the existence of unreachable pairs. It is a good idea to complement the median with the fraction of pairs within its value: in any case, we know that at least 50% of the pairs (of all pairs, not just the reachable ones) are within its value, which gives us a concrete handle. The median of all distances for Facebook is 5 (and 92% of all pairs is within this distance). So, again, “four degrees of separation”. Obviously, for Milgram in all cases the median is ∞. So, using this measure we progressed really a lot. With the collaboration of Jure Leskovec we were able to compute similar measures for Horvitz and Leskovec’s Messenger experiment [11]: the average distance, 6.618, has confidence 71.3%; the harmonic diameter is 8.935, whereas the median distance is 7, covering 78.7% of all pairs.9 Note that these figures are due to the presence of isolated nodes, that is, nodes that did not participate in any communication in the observed month: if the graph is reduced to non isolated nodes, essentially all values collapse. II. THE SAMPLE IS BIASED, AND ANYWAY IT JUST REPRESENTS 10% OF HUMANITY! As a first consideration, we invite the reader to observe that there is no such things as a “uniform” or “unbiased” sample of a graph. One can, of course, sample the nodes or the arcs of a graph, and consider the induced subgraph, but there is no guarantee that the induced subgraph preserves the properties of interest of the whole graph—much more sophisticated strategies are necessary, and in any case, it must be proved beforehand that the selected strategy creates an induced subgraph that is sufficiently similar to the whole graph (whatever notion of “similar” we want to take into account). In any case, let us take a step back and look for a moment at the conditions of Milgram’s experiment: • number of pairs examined: 296; • sample of the population: 100 United States cit- izens living in Boston, 96 random United States citizens living in Nebraska, 100 stockholders living in Nebraska; • completed chains: ≈ 22%; • definition of link: instructions to send the letter only to a “first-name acquaintance”. Our case: • number of pairs examined: 250 millions of billions; • sample of the population: 721 million people spread in several continents; • completed chains: ≈ 99.8%; 9We cannot report statistical metadata such as the standard error, because we were provided with already-aggregated breadth-first sam- ples only. • definition of link: sharing a friendship link on Facebook. We realize, obviously, that Facebook is not a random sample, and that being on Facebook im- plies already sharing a mindset, or certain areas of interest. We are also aware of the digital divide problem (that introduces a strong geopolitical and economical bias) and that there are links on Face- book between people that never met each other in person (e.g., gamers). On the other hand, a random sample of 96 people from Nebraska is not a random sample of the world population, either. And, again, we will never know if some letters in the experiment actually passed through, say, two pen pals who never met in person. What a lot of people did not realize is that, essentially, the only thing we know about how people were involved in Milgram’s experiment is that the sender judged that it had a “first-name acquaintance” with the receiver. The link between sender and receiver might have been in some cases even weaker than sharing a friendship link of Face- book. There is, moreover, another important factor to take into account: since there will be many first- name acquaintances who are not on Facebook (and hence not Facebook friends) some short paths will be missing. These two phenomena will likely, at least in part, balance each other; so, although we do not have (and cannot obtain) a precise proof of this fact, we do not think we are losing or gaining much in considering the notion of Facebook friend as a surrogate of first-name friendship. All in all, we see a definite progress in stating that the world is small. Thanks to Facebook, which is the largest ever-created database of human rela- tionships, we have been able to make Milgram’s experiment (or at least the part of it that has to do with measuring shortest paths) much more concrete and objectively measurable. Nonetheless, let us take another step back and consider, for a moment, the genius of a man who approached a mind-boggling (even for us, now) problem on a worldwide scale armed with three hun- dred postcards and an incredibly clever experiment. Obtaining a result almost unbelievably close to what we obtained using a number of pairs that is fifteen orders of magnitude larger. One is tempted to draw a comparison with Galileo’s celebrated mental ex- periment in the Dialogo sopra i due massimi sistemi 4 del mondo [12]: you do not need an expensive lab to test the principle of relativity—you just need a ship, some butterflies and some fish. Of course, once you do it, an expensive lab to check it thoroughly is definitely not a bad idea. III. YOU MEASURED THE AVERAGE DISTANCE, BUT DEGREES OF SEPARATION ARE ALGORITHMIC Just after we disseminated our paper, we learned that an experiment was trying to settle the “degree of separation” problem, which was “still unresolved” using Facebook.10 We were, of course, quite sur- prised. While we certainly did not “resolve” any- thing, it was difficult to imagine an experiment at present time with a larger sample or significantly more precise measurements. The point is the distinction between “routing” and “distance”. Milgram’s postcard were routed locally (each sender did not know whether the recipient was the best choice to get to the destination, i.e., if it lay on a shortest path to the destination). Apparently, the question is still unresolved because by studying Facebook we have only computed the “topological”, not the “algorithmic” degrees of separation. that We believe, however, this is a red her- ring. Reading carefully Travers and Milgram’s pa- pers [13], [6], it is clear that the very purpose of the authors was to estimate the number of inter- mediaries: the postcards were just a tool, and the details of the paths they followed were studied only as an artifact of the measurement process. In the words of Milgram, the problem was defined by “given two individuals selected randomly from the population, what is the probability that the minimum number of intermediaries required to link them is 0, 1, 2, . . . , k?”. Said otherwise, Milgram was interested in estimating the distance distribution of the acquaintance graph. The interest in efficient routing lies more in the eye of the beholder (e.g., the computer scientist) than in Milgram’s: if he had at his disposal an actual large database of friendship links and algorithms like the ones we used, he would have dispensed with the postcards altogether. Thus, the fact that we measured actual shortest paths between individuals, instead of the paths of a greedy routing, is a definite progress. Routing is an interesting computer-science 10http://smallworld.sandbox.yahoo.com/. (and sociological) problem, but it had little or no in- terest for Milgram—actually, the main interest in the routing process was understanding the convergence of paths. From the paper: The theoretical machinery needed to deal with social networks is still in its infancy. The empirical technique of this research has two major contribution to make to the development of that theory. First it sets an upper bound on the minimum number of intermediaries required to link widely sep- arated Americans. Since subjects cannot always foresee the most efficient path to a targer, our trace procedure must inevitably produce chains longer than those gen- erated by an accurate theoretical model which takes full account of all paths em- anating from an individual. That said, the results obtained in Milgram’s experiment are even more stunning because the average routing distance they computed (with the provisos about uncompleted chains discussed above) is so close to the average shortest-path length. The latter observation seems to suggest that human beings are extremely good at routing, so good that they almost route messages along the shortest possible path. However, taking uncompleted paths into consideration gives a slightly different twist to this remark: it seems that when someone felt confident enough to continue the experiment, (s)he did so almost in the best possible way; but more often than not, the experiment was stopped probably because the message arrived at an individual that did not know how to route it further efficiently. Apart for the attempts to measure the routing distance in real-world social graphs, there is an ever increasing focus on developing a theory of distributed efficient routing on small worlds, starting from Kleinberg’s intriguing notion of navigabil- ity [14], [15]; this is however outside of the scope of our paper. IV. JUST ADD A FEW LINKS HERE AND THERE AND WE’LL ALL BE AT ONE DEGREE OF SEPARATION Another, closely related, question is: “We have seen that the degree of separation has constantly decreased since 2008, reaching its current value. What can we expect for the future?” 5 To answer the above comment/question, notice that the average distance is kPk/r, X k>0 where Pk is the number of pairs at distance exactly k and r is the number of reachable pairs, which is n2 if and only if the graph is strongly connected. Of course, if we have bounds Bk ≥ Pk for some 1 ≤ k ≤ ℓ, it is immediate to see that, if Pℓ−1 k=1 Bk ≤ r then kPk ≥ X k>0 ℓ−1 X k=1 kBk + ℓ(cid:16)r − X k>0 Bk(cid:17). (1) Now, depending on how much you want to consider a graph similar to the Facebook graph described in [1], there are many ways to generate some Bk’s. a) First bound (depending on n, m and D).: There are instrinsic bounds on the number of short paths you can generate when the number of neigh- bours of a node is limited. The simplest observation is that (letting D be the maximum degree and m be the number of arcs in the graph, i.e., twice the number of edges) you cannot have more than m pairs at distance one, mD pairs at distance 2, and so on; more precisely, we can set Bk = mDk−1, getting (from (1)) the lower bound kPk ≥ m + 2mD + 3(r − m − mD) X k>0 provided that m+mD ≤ r; in the case of Facebook (D = 5000, n ≈ 721 × 106, r = 5 × 1017, m ≈ 69 × 109) the inequality m + mD ≤ r is satisfied and the lower bound obtained is ≈ 2.999. In other words, no graphs with the same number of nodes, arcs and maximum outdegree of the graph we considered can have an average distance smaller than 2.999. b) Second bound (depending on the degree sequence).: To improve over the previous trivial bound, we can use the actual degree distribution.11 This is a bit like answering to the question: what if some omniscent being “rewired” Facebook in an optimised way to reduce the average distance as much as possible, but leaving each user with its current number of friends? Let us first notice that P2 can be bounded by Px d(x)2, which, being the sum of entries of the square of the adjacency matrix, is 11The degree distribution is publicly available as part of the dataset associated with [1]. an upper bound for the number of pairs at distance 2. Providing a good bound for P3 is slightly more difficult: Theorem 1 Let d0 ≥ d1 ≥ . . . dn−1 be the degree sequence of the graph, s = Pn−1 i and define, for every t, i=0 d2 δ(t) = dt−1 X i=0 di. Then P3 (the number of pairs of nodes at distance exactly 3) can be bounded by P3 ≤ ℓ X k=0 dkδ(k) + dℓ+1(cid:16)s − ℓ X k=0 δ(k)(cid:17) ℓ is the greatest integer such that where Pℓ k=0 δ(k) < s. in at Proof: We can bound P3 from above by tuples (ui, vi, wi, zi) counting the number p of corresponding to paths of length 3. Let V = {v0, . . . , vk−1} be the set of nodes appearing as second component least one such tuple, sorted by non-increasing node degree; clearly p ≤ d(v0)π(v0) + · · · + d(vk−1)π(vk−1) where d(x) is as usual the degree of x and π(x) is the number of paths of length 2 starting from x: this is because ev- ery single path of length 3 of the form (−, vi, −, −) is obtained by choosing a neighbor of vi and a path of length 2 leaving from vi. Observe that π(v0)+· · ·+π(vk−1) cannot be larger than s (because the latter is an upper bound to the number of paths of length 2 in the graph). Now, of course, for every t = 0, . . . , k − 1, d(vt) ≤ dt, so p ≤ d0π(v0) + · · · + dk−1π(vk−1); it is convenient to think of the latter as a summation of a list L of length s ≥ π(v0) + · · · + π(vk−1), where d0 occurs π(v0) times, d1 occurs π(v1) times etc., and at the end of the list 0 occurs enough times to reach the desired length. Now π(vt) can be bounded from above by the number of paths of length 2 leaving from a node of degree dt. But the latter can be obtained by choosing at the first step the dt nodes with largest degree, and summing up their degree; that is, π(vt) ≤ δ(t). So we can safely substitute the above list L with another list L′ of the same length where d0 is repeated δ(0) ≥ π(v0) times, d1 is repeated δ(1) ≥ π(v1) times etc. The resulting list L′ dominates L elementwise, hence the thesis. 6 i=0 d2 Plugging B1 = m, B2 = Pn−1 i and B3 as in Theorem 1, and using the actual degree sequence of Facebook, we obtain ≈ 3.6. Thus, Facebook is essentially just one step (distance or degree doesn’t matter) away from the best possible, given that every individual keeps the current number of friends. V. IT’S JUST BECAUSE OF THE NODES WITH VERY HIGH DEGREE THAT WE OBSERVE SUCH A LOW VALUE Since the first studies on the structure of complex graphs [16], and in particular of social networks, the degree distributions have been a central topic on which many authors focused, concluding that both in- and out-degrees exhibit a heavy-tailed distribu- tion: this fact implies that there are many nodes whose degree largely exceeds the average. It is a widely assumed tenet that those nodes, sometimes referred to as hubs, represent a sort of “social glue” that keeps the whole network structure together and that shortcut friendship paths. In the case of social networks, such as Twitter or Facebook, hubs are superstars like Lady Gaga or Barack Obama, whose account often do not even correspond to real persons. But, is this the case? In our analysis of the Facebook graph we excluded pages (the accounts that people may “like”), and standard accounts have a hardwired limit of 5 000 friends. Nonetheless, we cannot rule out the possibility that there are some fake celebrity accounts remaining in the graph we studied. The general question we are asking can be re- stated as follows: take a social network and start re- moving the nodes of largest degrees; how much does the distribution of distances change? in particular: how does the average distance change (presumably: increase)? We considered this question in a previ- ous paper [17] (see also [18]), where we actually studied the more general problem of which removal strategies are more disruptive under the viewpoint of distance distributions. We report an anticipation of a subset of the results of [18], as they suggest that high-degree node removal is not going to cause drastic changes in the structure of the network. We show results for a TABLE III TABLE IV CHANGE IN AVERAGE DISTANCE OF WEB AND SOCIAL GRAPHS CHANGE IN HARMONIC DIAMETER OF WEB AND SOCIAL GRAPHS AFTER REMOVING THE LARGEST (IN-)DEGREE NODES. THE REMOVAL PROCESS IS STOPPED WHEN THE NUMBER OF ARCS AFTER REMOVING THE LARGEST (IN-)DEGREE NODES. THE REMOVAL PROCESS IS STOPPED WHEN THE NUMBER OF ARCS REMOVED REACHES THE 10% AND 30%. REMOVED REACHES THE 10% AND 30%. Graph .in Hollywood LiveJournal Orkut original 15.34 3.92 5.99 4.21 10% 30% 16.11 (+5.0%) 4.02 (+2.5%) 6.15 (+2.7%) 4.43 (+5.2%) 18.98 (+23.7%) (+7.9%) 4.23 6.55 (+9.3%) 4.67 (+10.9%) Graph .in Hollywood LiveJournal Orkut original 32.26 4.08 7.36 4.06 10% 30% 47.03 (+45.8%) (+1.0%) (+5.2%) (+6.7%) 4.12 7.74 4.33 87.68 (+171.8%) (+7.8%) (+17.8%) (+13.6%) 4.40 8.67 4.61 7 small12 snapshot of the Indian web (.in), for the Hollywood co-starship graph, for a snapshot of the LiveJournal network kindly provided by the authors of [19], and a snapshot of the Orkut network kindly provided by the authors of [20].13 The results we obtained are the following. Re- moving largest-degree nodes does affect the average distance on web graphs: after the removal of 30% of the arcs14 the average distance gets increased of about 24%. Nonetheless, the same removal strategy seems to have a weaker impact on genuine social networks: under the same condition, the increase in average distance ranges between 8% and 11% (see Table III). Nonetheless, we are actually missing a very im- portant point: in the social networks we studied, removing 30% of the arcs actually does not change the percentage of reachable pairs, whereas in web graphs the percentage (which is already lower) is reduced by a half. As we discussed in Section I, the average distance turns out again to be a very rough and unrealiable measure when the number of unreachable pairs is large. Thus, in Table IV we show what happens to the harmonic diameter. The results show that the increase for social networks is very modest (less than 20% after the removal of as many as the 30% of the arcs), whereas for web graphs the harmonic 12Similar results have been obtained with a lesser degree of precision on a snapshot of a 100 million pages in [17]; computations are underway to obtain high-precision data similar to what we report here about the smaller snapshot, and the results will be included in the final version of this paper. 13All these datasets are at http://law.dsi.unimi.it/. The identifiers of the datasets are in-2004, hollywood-2011, ljournal-2008 and orkut-2007. available public and 14We emphasize that we remove nodes (in decreasing order of their in-degree) and all incident edges, but count how many arcs are removed, because it is the number of deleted arcs that determines the expected loss in connectivity. We invite the reader to consult [17] for more details. diameter almost triplicates! This confirms again that the harmonic diameter is more reliable value to be associated to the “tightness” or “connectedness” of a network. We remark that LiveJournal and Orkut are people- to-people friendship networks as Facebook (note, however, that LiveJournal is directed). We believe that the resistance to high-degree removal is actually a common phenomenon in such networks, which prompts us to conjecture that similar node-removal prodedures will not change Facebook average dis- tance or harmonic diameter significantly, albeit we have no empirical data to support our hypothesis at this point. Actually, a more general conclusion obtained in the cited paper [17] is that social networks seem very robust to node removal, and we could not find any node order that determined radical changes in the distance distribution. This observation leaves an intriguing question still open to debate: if hubs are not the inherent cause behind short distances, then what is the real reason of this phenomenon? VI. ARE YOU SAYING THAT FACEBOOK REDUCED THE AVERAGE DISTANCE BETWEEN PEOPLE? Some of the comments in the general press took the outcomes of our experiments as an evidence that online social networks (such as Facebook) reduced the average distance between people; of course, this was not the purpose (neither the content) of the experiment and in any case there is no direct way to know if this is true or not, because our measurements are performed on Facebook. We can see, however, that the distance between Facebook users constantly decreased over time: it used to be 5.28 in 2008, 5.06 in 2010 and 4.74 in our most recent dataset. Whether this decrease is due to Facebook, or whether it simply Facebook reflecting 8 [10] D. Fogaras, “Where to start browsing the web?” in Innovative Internet Community Systems, Third International Workshop, IICS 2003, ser. Lecture Notes in Computer Science, vol. 2877. Springer, 2003, pp. 65–79. [11] J. Leskovec and E. Horvitz, “Planetary-scale views on a large instant-messaging network,” in Proceeding of the 17th inter- national conference on World Wide Web. ACM, 2008, pp. 915–924. [12] G. Galilei, Dialogo sopra i due massimi sistemi del mondo. Landini, 1632. [13] S. Milgram, “The small world problem,” Psychology Today, vol. 2, no. 1, pp. 60–67, 1967. [14] J. M. Kleinberg, “Navigation in a small world,” Nature, vol. 406, no. 6798, pp. 845–845, 2000. [15] ——, “The small-world phenomenon: an algorithm perspec- tive,” in Proceedings of the 32nd ACM symposium on theory of computing. ACM, 2000, pp. 163–170. [16] A.-L. Barabási, R. Albert, H. Jeong, and G. Bianconi, “Power- law distribution of the World Wide Web,” Science, vol. 287, p. 2115a, 2000. [17] P. Boldi, M. Rosa, and S. Vigna, “Robustness of social net- works: Comparative results based on distance distributions,” in Social Informatics, Third International Conference, SocInfo 2011, ser. Lecture Notes in Computer Science, vol. 6894. Springer, 2011, pp. 8–21. [18] ——, “Removal strategies for web and social graphs,” 2012, submitted for publication. [19] F. Chierichetti, R. Kumar, S. Lattanzi, M. Mitzenmacher, A. Panconesi, and P. Raghavan, “On compressing social net- works,” in KDD ’09: Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. New York, NY, USA: ACM, 2009, pp. 219–228. [20] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee, “Measurement and Analysis of Online Social Networks,” in Proceedings of the 5th ACM/Usenix Internet Measurement Conference (IMC’07), San Diego, CA, October 2007. [21] M. Granovetter, “The Strength of Weak Ties,” The American Journal of Sociology, vol. 78, no. 6, pp. 1360–1380, 1973. better and better the situation in the “real world” is hard to say. In the former case, as someone suggested, we would be observing a reduction in path lengths due probably to the presence of weak ties [21] that hardly correspond to a real friendship relation and would probably not even show up in a non–electronically-mediated environment. Understanding how online social networks are changing our way of interacting, communicating and thinking is absolutely beyond the scope of our paper, whose aim was much humbler and certainly not as far-reaching. We believe, however, that giving a concrete and realistic explanation of what is going on requires a co-ordinated effort and calls for an interdisciplinary endeavor, putting together sociology, psychology, computer science and math- ematics. This is, we think, one of the most important challenges for people working in these disciplines, with yet unknown consequences of philosophical, social and even economical value. REFERENCES [1] L. Backstrom, P. Boldi, M. Rosa, J. Ugander, and S. Vigna, “Four degrees of separation,” Arxiv preprint arXiv:1111.4570, 2012, accepted at ACM Web Science 2012. [2] P. Boldi, M. Rosa, and S. Vigna, “HyperANF: Approximating the neighbourhood function of very large graphs on a budget,” in Proceedings of the 20th international conference on World Wide Web, S. Srinivasan, K. Ramamritham, A. Kumar, M. P. Ravindra, E. Bertino, and R. Kumar, Eds. ACM, 2011, pp. 625–634. [3] P. Boldi and S. Vigna, “The WebGraph framework I: Compres- sion techniques,” in Proc. of the Thirteenth International World Wide Web Conference (WWW 2004). Manhattan, USA: ACM Press, 2004, pp. 595–601. [4] C. R. Palmer, P. B. Gibbons, and C. Faloutsos, “Anf: a fast and scalable tool for data mining in massive graphs,” in KDD ’02: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. New York, NY, USA: ACM, 2002, pp. 81–90. [5] J. Markoff and S. Sengupta, “Separating you and me? 4.74 degrees,” The New York Times, no. 325, p. B1, 21 November 2011. [6] J. Travers and S. Milgram, “An experimental study of the small world problem,” Sociometry, vol. 32, no. 4, pp. 425–443, 1969. [7] A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. Wiener, “Graph structure in the Web: experiments and models,” Computer Networks, vol. 33, no. 1–6, pp. 309–320, 2000. [8] P. Flajolet, É. Fusy, O. Gandouet, and F. Meunier, “Hyper- LogLog: the analysis of a near-optimal cardinality estimation algorithm,” in Proceedings of the 13th conference on analysis of algorithm (AofA 07), 2007, pp. 127–146. [9] M. Marchiori and V. Latora, “Harmony in the small-world,” Physica A: Statistical Mechanics and its Applications, vol. 285, no. 3-4, pp. 539 – 546, 2000.
1606.05730	1	1606	2016-06-18T08:41:16	A Comparison of Methods for Cascade Prediction	[ "cs.SI", "physics.soc-ph" ]	Information cascades exist in a wide variety of platforms on Internet. A very important real-world problem is to identify which information cascades can go viral. A system addressing this problem can be used in a variety of applications including public health, marketing and counter-terrorism. As a cascade can be considered as compound of the social network and the time series. However, in related literature where methods for solving the cascade prediction problem were proposed, the experimental settings were often limited to only a single metric for a specific problem formulation. Moreover, little attention was paid to the run time of those methods. In this paper, we first formulate the cascade prediction problem as both classification and regression. Then we compare three categories of cascade prediction methods: centrality based, feature based and point process based. We carry out the comparison through evaluation of the methods by both accuracy metrics and run time. The results show that feature based methods can outperform others in terms of prediction accuracy but suffer from heavy overhead especially for large datasets. While point process based methods can also run into issue of long run time when the model can not well adapt to the data. This paper seeks to address issues in order to allow developers of systems for social network analysis to select the most appropriate method for predicting viral information cascades.	cs.SI	cs	A Comparison of Methods for Cascade Prediction Ruocheng Guo, Paulo Shakarian Arizona State University Tempe, AZ Email: {rguosni,shak}@asu.edu 6 1 0 2 n u J 8 1 ] I S . s c [ 1 v 0 3 7 5 0 . 6 0 6 1 : v i X r a Abstract- Information cascades exist in a wide variety of platforms on Internet. A very important real-world problem is to identify which information cascades can "go viral". A system addressing this problem can be used in a variety of applications including public health, marketing and counter-terrorism. As a cascade can be considered as compound of the social network and the time series. However, in related literature where methods for solving the cascade prediction problem were proposed, the experimental settings were often limited to only a single metric for a specific problem formulation. Moreover, little attention was paid to the run time of those methods. In this paper, we first formulate the cascade prediction problem as both classification and regression. Then we compare three categories of cascade prediction methods: centrality based, feature based and point process based. We carry out the comparison through evaluation of the methods by both accuracy metrics and run time. The results show that feature based methods can outperform others in terms of prediction accuracy but suffer from heavy overhead especially for large datasets. While point process based methods can also run into issue of long run time when the model can not well adapt to the data. This paper seeks to address issues in order to allow developers of systems for social network analysis to select the most appropriate method for predicting viral information cascades. I. INTRODUCTION Identifying when a piece of information goes "viral" in so- cial media is an important problem in social network analysis. This is often referred to as "cascade prediction". Recently, the cascade prediction problem attracted considerable atten- tion from researchers from communities of machine learning, data mining and statistics. Researchers attempted to predict the final size of information cascades based on approaches inspired by knowledge in various areas. Pei et al. [1] measured influence of the root node by k-shell number and related heuristics. Weng et al. [2] and Guo et al. [3] uitilized features describing both structural and temporal properties of early- stage cascades. The work described in [4] and [5] modelled cascades by one-dimensional point process. However, in this line of research, the experimental settings varied from paper to paper. Furthermore, as the cascade prediction problem can be treated as either classification or regression, most of previous work only dealt with one or the other and using just a single evaluation metric.With deployment of a counter- extremism messaging system (i.e. an enhanced version of [6]) as one of the primary goals in our group, cascade prediction can play a crucial role in detection of early-stage extremism message that is potential to go viral on social network sites. Other applications include the spread of information following a disaster, promotion of health behaviors and applications to marketing. Therefore, it is important to understand how well the existing methods stemming from different research area could perform in near real-world experimental settings. An ideal cascade prediction method for counter-extremism messaging system should provide acceptable accuracy with ability to make near real-time prediction. In this paper, we compare performance of a variety of cascade prediction methods originating from different research areas as both classification and regression problems with multiple evaluation metrics. We also measure the run time of the tasks required by the methods to complete cascade prediction – another key deployment concern not explored in most research. In this paper, the main contribution can be summarized as: • We compare cascade prediction methods in three cate- gories: centrality based, feature based and point process based, therefore providing comparison between methods orginating from different research areas. • The cascade prediction problem is considered from both the aspect of regression and classification. we also con- duct a comprehensive comparison between methods by various evaluation metrics. • We also compare the run time of tasks needed for the cascade prediction methods are also measured in a task by task style. The rest of this paper is organized as the follows: In Section II, definitions relevant to the methods considered in this paper are introduced along with a formal problem statement of cascade prediction. Section III summarizes the mechanism of the three categories of cascade prediction methods. Section IV and V presents the setup of experiments and performance of each method in terms of both accuracy and run time. Section VI reviews related work. At last, Section VII concludes the paper and discusses the main issues of these methods. II. TECHNICAL PRELIMINARIES In this section, related concepts for the three categories of methods are defined. Then we formulate the cascade prediction problem as regression and classification respectively. A. Definitions Network and Cascade: The social network is a directed graph G = (V, E) where each node v ∈ V represents a user and each edge eij = (vi, vj) denotes that user vi is followed by user vj. Identified by the original message or the corresponding hashtag, a cascade is a time-variant subgraph of the social network d(t) = (V (t), E(t)). Each node v ∈ V (t) denotes a user reposted the original message of cascade d(t) (for the Aminer dataset [7]) or a user posted the hashtag defining cascade d(t) (for the Twitter dataset [2]) within time t. The time variable t denotes number of time units since the microblog including the original message or the hashtag. For each node v ∈ V (t) we record their adoption time of cascade d(t) as tv. For v ∈ V (t), tv ≤ t while for v (cid:54)∈ V (t) we define tv = ∞. Thus we can get an ascendingly sorted vector tv(t) including all tv ≤ t for each cascade, which plays an important role in both feature based methods and point process based methods for cascade prediction. The kth element of tv(t) can be denoted as tv(t)[k]. For convenience, we use tend to denote the time when the last adoption of a cascade happened. Besides the cascade d(t) = (V (t), E(t)), the neighborhood of V (t) also can provide information about the potential of the cascade. Here we define the out-neighborhood reachable by any node in V (t) in step i as ith surface Fi(t). To show how 'fresh' the cascade is for a node v ∈ Fi(t), we define a function f∆t : v → ∆t that maps such a node to the number of time units since v become a member of first surface to current time t. As time makes a big difference in social influence and diffusion, we divide the first surface F1(t) into two sets of nodes depends on f∆t(v) for all v ∈ F1(t). With a selected threshold tλ. The first set named as frontiers includes all nodes v ∈ F1(t) such that f∆t(v) ≤ tλ and the other set non- adopters consists the other nodes v ∈ F1(t) with f∆t(v) > tλ. In this paper, x denotes absolute value of scaler x and x denotes cardinality of set x. Communities: We can treat a community partition of a social network as a function fC : V → C which maps a set of nodes V to a set of communities C. With this function, given a cascade d(t) = (V (t), E(t)), it enables us to describe the distribution of nodes over communities by features such as fC(V ), the number of communities among set V . Point Process: Each adoption in a cascade can be represented as an event from the aspect of point process as in [4]. Thus, for cascade prediction, we can use tv(t − ∆t) to describe the history of a point process strictly before t. The core of a point process is the conditional density function λ(t). Conditioned on tv(t− ∆t), the conditional density is the limit of expected number of adoptions would happen in time interval [t, t + ∆t] by taking ∆t → 0+: λ(t) = lim ∆t→0+ E{V (t + ∆t) − V (t)} (1) Given the density function λ(t) and target prediction time t(cid:48), the predicted cascade size can be computed by: (cid:90) t(cid:48) Regression Problem: Given a early stage cascade d(t) = (V (t), E(t)) and the corresponding node attribute vector tv(t) with constraint V (t) = n, the target is to predict the final size of the cascade V (tend). Classification Problem: A threshold T hres is selected to label each cascade. For a given cascade if its V (tend) ≥ T hres, we define it as a viral sample labeled as 1, otherwise, we label it as non-viral labeled as 0. Then the problem is to classify a given early-stage cascade d(t) to the viral class or the non-viral class. III. METHODS In this section we introduce several recently published methods for solving the cascade prediction problem. Diffusion process in social network includes information of time series, network structure, sometimes with microblog content and node attributes, therefore, methods originated from knowledge in various research area like social network analysis, random point process and non-linear programming can be applied. The methods can be categorized into: centrality based methods, feature based methods and point process based methods. A. Centrality Based Methods Previous work [1] discovered that the k-shell value of a node is highly correlated to the average cascade size it initiates. In this paper, we also consider eigenvector centrality, out- degree and Pagerank of the root node of cascades to deal with the cascade prediction problem. We refer to centrality based approaches as method C in this paper. B. Feature Based Methods In this paper, we consider two recently proposed methods [3] and [2] and call them method A and method B respectively for convenience. The features computed by the two methods can be categorized into network features, community based features and temporal features. Both of the feature based methods require to take advantage of community detection algorithms. Given the social network, community detection algorithms such as [8] and [9] can be applied to it and assign each node to one or multiple communities. Based on the communities detected, features can be computed to numerically describe how the nodes that par- ticipate in a cascade are distributed over communities. Thus, we can quantitatively measure structural diversity from [10] or influence locality from [7] as features. Network Features: In method B proposed by [2], the authors consider several types of network features: • Neighborhood size, including first surface (F1(Vt)) and V (t(cid:48)) = V (t) + λ(τ )dτ (2) second surface (F2(Vt)). B. Problem Statement t In this paper, we focus on comparison of different methods which can solve the cascade prediction problem. This prob- lem can be formulated as either a regression problem or a classification problem: • Path length, consisting average step distance and coef- ficient of variation of it, and diameter of the cascade. Step distance is the length of shortest path between two consecutive adopters vi and vi+1. Where coefficient of variation is defined as the ratio of the standard deviation to the mean. Community Based Features: In both [3] and [2], community features are extracted and contribute to the predictive methods. • Community features for adopters, including the number of communities (fC(V (t))), entropy and gini entropy. • Community features for frontiers and non-adopters, in- cluding the number of communities (fC(F1(t))), en- tropy and gini entropy. • The number of shared communties between any two groups of adopters, frontiers and non-adopters. Temporal Features: In [3], the authors computed average of tv(t) while average step time and its corresponding coefficient of variation are calculated in [2] as two features. C. Point Process Based Methods To discover patterns in the temporal dynamics of cascades, authors of both [5] and [4] both consider a cascade as an instance of one-dimensional point process in time space. They proposed novel density functions to characterize time series of cascades. The two methods are quite similar, in terms of the formulation of conditional density function λ(t). In both cases, λ(t) consists of an element modeling the popularity of the cascade and another describing the probablity distribution of an adoption behavior over time. The Reinforced Poisson Process (RPP) Method: In [5], the authors consider the density function for a cascade d as a product of three elements: λd(t) = αdfd(t; θd)V (t) (3) For cascade d, αd denotes the intrinsic attractiveness, fd(t; θd) is defined as the relaxation function which models how likely an adoption would happen at time t without considering αd and V (t). For each cascade d, parameters αd and θd are learned by maximization of the likelihood of tv(t). Thus, the predicted cascade size at time t(cid:48) > t can be computed by: V (t(cid:48)) = V (t) + αdfd(τ ; θd)V (τ ) dτ (4) The SEISMIC Method: In [4], authors model the density function as a modified Hawkes Process made up of three ele- ments: infectiousness pt, node degree ni and human reaction time distribution φ(s): niφ(t − tvi) λ(t) = pt (5) Where tvi ∈ tv(t) is the time when each adoption happens. Similar to αd in the Reinforced Poisson Process model, pt is computed by maximization of the likelihood function: i=1 (cid:90) t(cid:48) t V (t)(cid:88) V (t)−1(cid:89) i=0 (cid:40) TABLE I: Dataset Statistics Property Directed Nodes Edges Number of communities Modularity Average Out-degree Average Eigenvector Cen- trality Average K-shell Average Pagerank Cascades (≥ 50 nodes) Twitter Dataset Weibo Dataset undirected 595,460 7,170,209 24,513 0.7865 47.94 0.001783 24.6032 1.6794e−6 14,607 directed 1,787,443 216,511,564 2,802 0.5581 231.3381 0.0186 52.3064 5.596e−7 99,257 straint (cid:82) ∞ As φ(s) is a probability distribution function, with the con- 0 φ(s)ds = 1 and power-law decay factor θ esti- mated by training data, c can be computed. With the density function λ(t), the predicted cascade size can be computed by equation (2). IV. EXPERIMENTAL SETUP For comprehensiveness, we evaluate the performance of each method by treating cascade prediction problem as both regression and classification problem. We only consider cas- cades that end up with at least 50 adopters. Thus we can treat first 50 nodes of each cascade as its early stage. In this section, an introduction of the datasets is followed by descriptions of setup of the classification and regression experiments. All the experiments are carried out on an Intel(R) Xeon(R) CPU E5-2620 @ 2.40 GHz machine with 256GB RAM running Windows 7. All the methods are implemented in Python 2.7. A. Dataset Description The statistics of the two datasets used in this paper for evaluation of the cascade prediction methods are shown in Table I. Twitter Dataset: Twitter1 is the most well-known microblog platform throughout the world. The dataset was used in [2]. This dataset includes a friendship network with undirected edges, cascades identified by hashtags and corresponding mentions and retweets. Weibo Dataset: Sina Weibo2 is the largest Chinese microblog social network. The dataset was used in [7]. It consists of a directed followership network and retweet cascades. B. Regression For the regression problem, the m × 1 ground truth vector y is made up of final size of each cascade (V (tend)), where m is the number of cascade. Each regression model is able to output a m × 1 vector y. Thus each element yi ∈ y is the predicted size of the ith cascade. For point process models, with different prediction time, the predicted results can change. Thus, for each early-stage cascade, we set t as −(cid:82) tvi+1 tvi pt = arg max pt λ(tvi) exp λ(τ )dτ (6) While the human reaction time distribution φ(s) is formulated as a piece-wise function consists of a constant piece and a power-law piece with parameter c and θ: φ(s) = s ≤ s0 s > s0 )−(1+θ) c c( s s0 (7) 1https://twitter.com 2https://weibo.com the time when we observed the 50th adoption and prediction time as {2, 4, 6, 8, 10} × tv(t)[50]. To evaluate a method for the regression problem, the difference between its prediction results y and the ground truth y can be described by various error functions. In addition, ytop10% denotes the set of top 10% cascades in prediction result while ytop10% is the set top of 10% cascades of ground truth. In this paper we choose following metrics to compare the prediction made by different methods, as they are widely used in related literatures such as [11], [5], [12] and [4]: • APE (average percentage error): 1 m • RMSE (root mean square error): yi−yi i=1 yi (cid:80)m i=1( yi − yi)2 m (cid:114)(cid:80)m (cid:114)(cid:80)m i=1(log yi − log yi)2 (cid:12)(cid:12)ytop10% ∩ ytop10% m (cid:12)(cid:12) • RMLSE (root mean logrithm square error): • Top 10% coverage: 10 m C. Classification For classification, we apply three predetermined thresholds (50th, 75th and 90th percentiles) to final size of cascades to assign each of them a class label, which provides the m × 1 ground truth vector L = {l0, ..., lm−1} one for each threshold. The cascades with size larger than threshold are labelled as viral class with li = 1. Table II shows the thresholds and counts of samples for both classes. Then the methods for solving the classification problem can output predicted label vector L. Comparing L with L results in standard metrics: precision, recall and F1 score. To examine the effectivess of the methods, we focus on reporting the metrics on the minority class (viral) as it is more difficult to do good predictions for it than the other. Specially, for point process based mothods, as they are capa- ble to predict the final cascade size without being trained with class labels (once parameters are determined and prediction times are selected), we carry out the evaluation on them in this way: prediction results (by setting different prediction times) are treated as features for each sample. As the time when each cascade stop growing is not easy to determine. D. Run time We also take the run time of tasks into account for the cas- cade prediction methods. To understand how computationally expensive the methods are in terms of run time, it is neces- sary to analyze the procedure of them. For centrality based methods, the prediction can be divided into three steps: com- putation of centrality, training and prediction. Similarly, for feature based methods, computation of features, training and prediction are required to be done. In addition, preprocessing like community detection, computation of shortest path length are needed, which can be computationally expensive. While point process based methods require little preprocessing. For each cascade, parameters are computed independently through TABLE II: Thresholds for Classification Percentile Threshold Viral samples Non-viral samples 50% 75% 90% 50% 75% 90% Twitter Dataset 7,303 3,652 1,461 Weibo Dataset 49,628 24,814 9,925 106 226 587 152 325 688 7,304 10,955 13,146 49,629 74,443 89,332 MLE of the observed time vector tv(t) and properties of the adopters V (t). Then prediction is made by integral of density functions. Thus, we consider the following processes one by one and then combine them to estimate the run time of a certain method. Proprecessing: There are three types of proprecessing con- sidered: loading the graph, computation of centralities and community detection. Computation of Features: For feature based methods, we measure the run time of computation of the features , which takes the product of preprocessing as input. Training and Prediction: For centrality and feature based methods, the run time of training and prediction is measured for ten-folds. For point process based methods, we measure the run time of parameter estimation and prediction for the whole batch of data. V. EXPERIMENTAL RESULTS In this section we show the experimental results including both accuracy of cascade prediction and the run time for each method. For convenience, we call method of [3], [2] and the centrality based method as method A, B and C respectively. For method A, B and C, 10-fold cross-validation is applied. For results where we compare these three methods, we report only the best-performing centrality measure amongst out- degree, Pagerank, Shell number and eigenvector centrality as the method C for each dataset. As shown in Fig. 1, eigenvector centrality outperforms others in the classification task when the two classes are imbalanced. Thus we take eigenvector centrality as the method C. The results for regression is not shown here for limited space as the difference between results produced by different centralities is trivial. For the Reinforced Poisson Process (RPP) method [5], as the parameter estimation task for each cascade is independent of others, the cross- validation is skipped and predictions are made by parameters learned from first 50 nodes of each cascade. For the SEISMIC method [4], we also skip the 10-fold cross-validation. We set the cutoff time s0 = 30000(s) for the Twitter dataset and s0 = 300(s) for the Weibo dataset then fit the parameters (θ, c) for the human reaction time distribution function φ(s) with all samples in the dataset. While in the original paper [4], the authors set θ and c just by 15 tweets they manually picked. The power-law fitting is done as per [13], which (a) Twitter Dataset: 50th percentile (b) Twitter Dataset: 75th percentile (a) Twitter Dataset: APE (b) Twitter Dataset: RMSE (c) Twitter Dataset: 90th percentile (d) Weibo Dataset: 50th percentile (c) Twitter Dataset: RMSLE (d) Twitter Dataset: Top 10% Cover- age (e) Weibo Dataset: 75th percentile Fig. 1: Classification results of centrality based methods: error bar stands for one standard deviation. (f) Weibo Dataset: 90th percentile returns (θ, c) = (0.440, 1.018e−5) and (0.282, 7.332e−4) for the Twitter dataset and Weibo dataset respectively. A. Regression For centrality based methods, we apply linear regression with least squared error. We carry out the training and predic- tion with random forest regressor, SVR and linear regression model provided by [14] for feature based methods. We only show the results produced by SVR as it outperformes others. For the point process based mothods, we only report the best result among prediction time out of {2, 4, 6, 8, 10}×tv(t)[50]. For the Twitter dataset, Fig. 2a, 2b, 2c and 2d show the experimental results for the regression problem. Feature based methods and SEISMIC outperform RPP and method C w.r.t. APE. Concerning RMSE, method A shows more predictive power than others. As to RMSLE, feature based methods result in less error than the other two categories. From the aspect of Top 10% coverage, RPP, method A are more likely to track the trending cascades than others. Fig. 2e, 2f, 2g and 2h show the regression result for the Weibo dataset, Regarding APE, SEISMIC, method A and B have comparable performance and outperform others. In terms of RMSE, method A, B are measured to be more predictive than the rest. Feature based methods also make predictions with least RMSLE. For top 10% coverage, RPP is more likely to detect popular cascades than others. An interesting observation is that the prediction accuracy measured by different error metrics can be contrary to each other. For example, in Fig. 2a, compared to SEISMIC, pre- diction made by method C results in more error measured by (e) Weibo Dataset: APE (f) Weibo Dataset: RMSE (g) Weibo Dataset: RMSLE (h) Weibo Dataset: Top 10% Coverage Fig. 2: Regression results APE, however, comparable error w.r.t. RMSE and less error regarding RMSLE (See Fig. 2b and 2c). This implies that it is better for researchers to show more than one type of error for evaluation of regression results. B. Classification We show the precision, recall and F1 score for the viral class with all the three percentile thresholds. For each dataset, we choose the 50th, 75th and 90th percentile of the final size of all cascades as the thresholds for assigning the cascades into viral or non-viral class. The number of samples in each class is shown in Table II. Thus we can evaluate the cascade prediction methods with balanced and imbalanced classes. For each method, we only show the best result among those produced by different classifiers or various training methods. As a result, for feature based methods, random forest outperforms others. While for point process based methods we treat cascade size predicted by setting prediction time as {2, 4, 6, 8, 10}×tv(t)[50] as features. Here we show the results produced by classifiers trained by these features. Fig. 3a, 3b and 3c show the classification results for the Twitter dataset. With all three thresholds, feature based methods A and B outperform others. In addition, they also PrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvectorPrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvectorPrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvectorPrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvectorPrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvectorPrecisionRecallF1 Score0.00.20.40.60.81.0Out-degreePagerankShell NumberEigenvector0.0%50.0%100.0%150.0%200.0%250.0%300.0%A(SVR)B(SVR)C(DTR)SEISMICRPP050010001500200025003000A(SVR)B(SVR)C(DTR)SEISMICRPP0.00.51.01.52.0A(SVR)B(SVR)C(DTR)SEISMICRPP0.000.050.100.150.200.250.30A(SVR)B(SVR)C(DTR)SEISMICRPP0.0%50.0%100.0%150.0%200.0%A(SVR)B(SVR)C(DTR)SEISMICRPP0500100015002000A(SVR)B(SVR)C(DTR)SEISMICRPP0.00.51.01.52.0A(SVR)B(SVR)C(DTR)SEISMICRPP0.000.050.100.150.200.250.300.35A(SVR)B(SVR)C(DTR)SEISMICRPP(a) Twitter Dataset: 50th percentile (b) Twitter Dataset: 75th percentile (c) Twitter Dataset: 90th percentile (d) Weibo Dataset: 50th percentile (e) Weibo Dataset: 75th percentile Fig. 3: Classification results: error bar stands for one standard deviation. (f) Weibo Dataset: 90th percentile show more robustness than others to imbalance of two classes in dataset. In terms of point process based methods, SEISMIC outperforms RPP especially when the two class are imbal- anced. RPP suffers from relatively large standard deviation, as the Newton's method is not always able to achieve conver- gence. Thus the parameters learned through the MLE approach can vary as a result from random initialization. Method C (eigenvector centrality) shows little predictive power with any of the three thresholds for the Twitter dataset, even if it outperforms other centrality based methods. For the Weibo dataset, as shown in Fig. 3d, 3e and 3f, feature based methods outperform others again with all three thresholds. Regarding point process based methods, contrary to the results for Twitter dataset, RPP achieves better F1 score than SEISMIC when threshold value becomes large. Method C (eigenvector centrality) performs comparably to RPP. C. Run time In this subsection, we show the run time of tasks for the cascade prediction methods considered in this paper. On one hand, preprocessing, computation of centralities and features suffer from high overhead as immense amount of data needs to be loaded. The run time of these tasks are listed in Table III. On the other hand, training and prediction tasks barely have the overhead issue. Preprocessing: We carry out the community detection task by the java implementation of Louvain algorithm [15] with 10 random start and 10 iterations for each start. For computation of centralities, we load edgelist of the social networks as a graph object in igraph-python [16]. As shown in Table III, community detection, computation of Pagerank and loading graph are the tasks suffer the most when the size of dataset increases. Community detection, computation of Pagerank and loading graph for the Weibo dataset take 80.32, 66.855 and 19.80 times the run time of those for the Twitter dataset respectively. Computation of Features: As shown in Table III, for the feature computation task, it takes method B 12.37 and 8 times the run time method of A for the Twitter Dataset and the Weibo Dataset respectively. To explain this observation, an analysis of what computation is carried out in each iteration for method A and B. For method A, computation of the features can be done without loading the graph (a heavy overhead). Moreover, for each cascade, method B also requires expensive computation of shortest path length for each pair of nodes in cascade subgraphs and size of 2-hop neighborhood. Training and Prediction: The run time of training and pre- diction is not directly related to the size of the social network. On one hand, it is correlated to the number of cascades for training and prediction. On the other hand, it is decided by the complexity of the method: for example, number of parameters to be learned, the complexity for learning each paramter and the comsumption to work out the prediction. Here we only measure the run time for solving the classification problem. We run each method with single process, overhead run time such as graph loading is ignored. For feature based methods the training and prediction time are also correlated to the number of features. For centrality based methods, we only show the run time for k-shell (method C) as all methods in this category are trained and tested with one feature: the centrality measure of the root node. Compared to RPP, SEISMIC is a deterministic method with closed form solution. The run time for each sample can be distributed with little variance. For the RPP method, as the log-likelihood function is non-convex, it is not guaranteed that global maximum can be reached in limited number of iterations. Therefore, the run time for a sample running out of the maximum number of iterations can be thousands times that of another, which reaches the convergence condition in the first iteration. As the log-likelihood function of RPP is twice-differentiable, Newton's method can be applied. In our experiments, with the maximum number of iterations setted as 100, the convergence is more likely to be achieved by Newton's method than gradient descent. Thus we only show the run time of RPP with Newton's method. For the Weibo dataset, as shown in Fig. 4b, the run time Fig. 4 shows the run time for each method to complete training and prediction tasks for all cascades in the two datasets. For feature based methods, it shows the run time needed for random forest (RF), SVM and logistic regression (LR). For method C, it shows that of decision tree (DT), SVM and logistic regression (LR). Concerning the Twitter dataset (See Fig. 4a), taking ad- vantage of decent implementation of classifiers, feature based methods comparable run time to point process based methods w.r.t. the training and prediction task with random forest and SVM (rbf kernel). PrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPPrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPPrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPPrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPPrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPPrecisionRecallF1 Score0.00.20.40.60.81.0A(RF)B(RF)C(DT)SEISMICRPPTABLE III: Run time: Preprocessing & Feature Computation Type Task Total time (s) Time sample (s) per Twitter Dataset 275 60.033 0.016 2.757 20.444 26.298 267.144 3252.7562 Louvain Loading Graph Degree K-shell Eigenvector Pagerank A B Weibo Dataset Louvain Loading Graph Degree K-shell Eigenvector Pagerank A B 22087 1188.486 0.045 139.128 391.140 1758.164 11181.453 87651.213 – – – – – – 0.018 0.2227 – – – – – – 0.110 0.883 Preprocessing Feature Computation Preprocessing Feature Computation (a) Twitter Dataset (b) Weibo Dataset Fig. 4: Run Time of Trainig and Prediction feature based methods comsume is comparable to SEISMIC with random forest. But the SVM with rbf kernel suffers from the order-of-magnitude increase of the number of training and testing samples. Thus leads to the observation that the run time becomes approximately 10 times that of random forest. Comparing Fig. 4b with Fig. 4a, the run time of RPP method increases the most. This means it is much more difficult for the Newton's method to converge for samples in the Weibo datasets. There are two possible reasons to explain this: 1). the uniform distribution used in random initialization can not produce good initial values that are closed to local optimal points; 2). the choice of log-norm distribution as function fd(t; θd) can not provide fairly good description of cascades in this dataset. VI. RELATED WORK Influence Maximization Since the proposal of Influence max- imization problem by Kempe et al. [17], related work emerged, focusing on estimation of influence for a selected set of nodes that can be measured by expected number of infectees under a certain influence model, such as [18] and [19]. Recently, a scalable randomized algorithm designed by Du et al. [20] estimates influence initiated by selected source nodes and thus select seed set with maximum expected influence. Cascade Prediction Although in [1], k-shell and heuristics of k-shell were shown to be effective indicator of long-term influence of nodes, in [21], experimental results showed that the shell number of the root node is not effectively predictive in the cascade by cascade scenario. Feature based methods from Jenders et al. [22] Chen et al. [23] were designed to solve the cascade prediction problem formulated as binary classification on balanced dataset, however, these methods are more or less dependent on content features from specific social media sites. Ma et al. [24] focused on applying content features to classify hashtag cascades by how much their size increases. Regarding to point process based methods, model designed with the intuition of mutual exciting nature of social influence, Zhou et al [25] applied multi-dimensional Hawkes process to rank cascades (memes) by their popularity. Recently, the model introducted by Yu et al. [11] combined feature engineering and human reaction time distribution function widely used in point process based methods to aggregate adoptions in subcascades for cascade prediction. Besides feature based methods and point process based methods studied in this paper, knowledge from related research fields could also be applied to cascade prediction. Goyal et al. [26] proposed the credit distribution model to learn pair-wise influence based on IC model proposed by Kempe et al. [17]. Cui et al. [27] proposed a feature selection approach for binary classification of cascades. Wang et al. [28] proposed a model to decouple the influence measured in a pair-wise way into two latent vectors representing influence and susceptibility of a node. This work differs from all the past efforts in that it is the most thorough comparison of methods general enough to be applied to different datasets without relying on features specific to a certain social media site. VII. DISCUSSION AND CONCLUSION In this paper, we evaluate three categories of recently proposed methods with both the classification and regression formulaton of cascade prediction. Feature based methods generally provide better prediction accuracy for the cascade prediction problem, no matter it is considered as classification or regression. However, they suffer from heavy overhead such as community detection and computation of features. Random point process based methods enable us to achieve the prediction with little preprocessing but are shown to be less accurate than feature based methods. The run time of methods in this category can also suffer from the situation when the data can not be well modelled by the proposed density function λ(t). In regression experiments, we find the inconsitancy between evaluation with different error metrics. A method that performs well w.r.t. one metric could result in large error measured by another. A predictive method should be able to perform fairly well measured by various error metrics. How to deal with changes in the social network and progress of cascades to update features is the biggest issue that both centrality based and feature based methods encounter. The 10-1100101102103Run time (seconds)A(RF)A(LR)A(SVM)B(RF)B(LR)B(SVM)C(DT)C(LR)C(SVM)SEISMICRPP100101102103104105106Run time (seconds)A(RF)A(LR)A(SVM)B(RF)B(LR)B(SVM)C(DT)C(LR)C(SVM)SEISMICRPPheavy overhead introduced by preprocessing and computation of features limits these methods from near real-time prediction. Point process based methods require little preprocessing and the training and prediction process are parallelable as they consider each cascade is indenpendent of others. This advantage in terms of run time over feature based methods can also be amplified as the size of the social network and the number of cascades. Moreover, point process based methods encounter little cold start problem. These two characteristics of point process based methods make them more suitable for real- time cascade prediction task. But how to secure the accuracy of prediction is the biggest issue for them. The point process based models are faced with two more problems: sensitivity to scale of time unit and requirement of prediction time as an input variable. In real-world application, given a early stage cascade, estimation of when it will stop progressing is a non- trivial problem. On balance, this paper explored various methods in the academic literature of predicting viral information cascades in a more comprehensive manner. Our aim is to provide im- portant insights into which methods based on graph topology or temporal dynamics performed best - as these results can generalize to a variety of application domains. In our ongoing work on developing a deplyable system for identifying viral extremist messages, this represents an important consideration. Our next step is to consider microblog content as well - which tends to be more domain specific. ACKNOWLEDGMENTS Some of the authors are supported through the AFOSR Young Investigator Program (YIP) grant FA9550-15-1-0159, ARO grant W911NF-15-1-0282, the DoD Minerva program and the EU RISE program. REFERENCES [1] S. Pei, L. Muchnik, J. S. Andrade Jr, Z. Zheng, and H. A. Makse, "Searching for superspreaders of information in real-world social me- dia," Scientific reports, vol. 4, 2014. [2] L. Weng, F. Menczer, and Y.-Y. Ahn, "Predicting successful memes using network and community structure," in Eighth International AAAI Conference on Weblogs and Social Media, 2014. [3] R. Guo, E. Shaabani, A. Bhatnagar, and P. Shakarian, "Toward order-of- magnitude cascade prediction," in Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining 2015. ACM, 2015, pp. 1610–1613. [4] Q. Zhao, M. A. Erdogdu, H. Y. He, A. Rajaraman, and J. Leskovec, "Seismic: A self-exciting point process model for predicting tweet popularity," in Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2015, pp. 1513–1522. [5] H. Shen, D. Wang, C. Song, and A.-L. Barab´asi, "Modeling and predicting popularity dynamics via reinforced poisson processes," in Twenty-Eighth AAAI Conference on Artificial Intelligence, 2014. [6] N. Kim, S. Gokalp, H. Davulcu, and M. Woodward, "Lookingglass: A visual intelligence platform for tracking online social movements," in Advances in Social Networks Analysis and Mining (ASONAM), 2013 IEEE/ACM International Conference on. IEEE, 2013, pp. 1020–1027. [7] J. Zhang, B. Liu, J. Tang, T. Chen, and J. Li, "Social influence locality for modeling retweeting behaviors." in IJCAI, vol. 13, 2013, pp. 2761– 2767. [8] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, "Fast unfolding of communities in large networks," Journal of statistical mechanics: theory and experiment, vol. 2008, no. 10, p. P10008, 2008. [9] M. Rosvall and C. T. Bergstrom, "Maps of random walks on complex the National networks reveal community structure," Proceedings of Academy of Sciences, vol. 105, no. 4, pp. 1118–1123, 2008. [10] J. Ugander, L. Backstrom, C. Marlow, and J. Kleinberg, "Structural diversity in social contagion," Proceedings of the National Academy of Sciences, vol. 109, no. 16, pp. 5962–5966, 2012. [11] L. Yu, P. Cui, F. Wang, C. Song, and S. Yang, "From micro to macro: Uncovering and predicting information cascading process with behavioral dynamics," arXiv preprint arXiv:1505.07193, 2015. [12] S. Gao, J. Ma, and Z. Chen, "Modeling and predicting retweeting dynamics on microblogging platforms," in Proceedings of the Eighth ACM International Conference on Web Search and Data Mining. ACM, 2015, pp. 107–116. [13] J. Alstott, E. Bullmore, and D. Plenz, "powerlaw: a python package for analysis of heavy-tailed distributions," PloS one, vol. 9, no. 1, p. e85777, 2014. [14] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander- plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch- esnay, "Scikit-learn: Machine learning in Python," Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [15] L. Waltman and N. J. van Eck, "A smart local moving algorithm for large-scale modularity-based community detection," The European Physical Journal B, vol. 86, no. 11, pp. 1–14, 2013. [16] G. Csardi and T. Nepusz, "The igraph software package for complex network research," InterJournal, vol. Complex Systems, p. 1695, 2006. [Online]. Available: http://igraph.org [17] D. Kempe, J. Kleinberg, and ´E. Tardos, "Maximizing the spread of influence through a social network," in Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2003, pp. 137–146. [18] W. Chen, Y. Wang, and S. Yang, "Efficient influence maximization in social networks," in Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2009, pp. 199–208. [19] A. Goyal, W. Lu, and L. V. Lakshmanan, "Celf++: optimizing the greedy algorithm for influence maximization in social networks," in Proceedings of the 20th international conference companion on World wide web. ACM, 2011, pp. 47–48. [20] N. Du, L. Song, M. Gomez-Rodriguez, and H. Zha, "Scalable influence estimation in continuous-time diffusion networks," in Advances in neural information processing systems, 2013, pp. 3147–3155. [21] P. Shakarian, A. Bhatnagar, A. Aleali, E. Shaabani, and R. Guo, Diffusion in Social Networks. Springer, 2015. [22] M. Jenders, G. Kasneci, and F. Naumann, "Analyzing and predicting the 22nd international conference International World Wide Web tweets," in Proceedings of viral on World Wide Web companion. Conferences Steering Committee, 2013, pp. 657–664. [23] J. Cheng, L. Adamic, P. A. Dow, J. M. Kleinberg, and J. Leskovec, "Can cascades be predicted?" in Proceedings of the 23rd international conference on World wide web. ACM, 2014, pp. 925–936. [24] Z. Ma, A. Sun, and G. Cong, "On predicting the popularity of newly emerging hashtags in twitter," Journal of the American Society for Information Science and Technology, vol. 64, no. 7, pp. 1399–1410, 2013. [25] K. Zhou, H. Zha, and L. Song, "Learning social infectivity in sparse low- rank networks using multi-dimensional hawkes processes," in Proceed- ings of the Sixteenth International Conference on Artificial Intelligence and Statistics, 2013, pp. 641–649. [26] A. Goyal, F. Bonchi, and L. V. Lakshmanan, "A data-based approach to social influence maximization," Proceedings of the VLDB Endowment, vol. 5, no. 1, pp. 73–84, 2011. [27] P. Cui, S. Jin, L. Yu, F. Wang, W. Zhu, and S. Yang, "Cascading outbreak prediction in networks: a data-driven approach," in Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2013, pp. 901–909. [28] Y. Wang, H. Shen, S. Liu, and X. Cheng, "Learning user-specific latent influence and susceptibility from information cascades," in Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
1809.09747	3	1809	2018-12-18T20:04:58	Analyzing CDR/IPDR data to find People Network from Encrypted Messaging Services	[ "cs.SI", "cs.CR" ]	Criminals are increasingly using mobile based communication applications, like WhatsApp, that have end-to-end encryption to connect to each other. This makes traditional analysis of call graphs, or traffic analysis, virtually impossible and so is a hindrance for law enforcement personnel. Old methods of traffic analysis have been rendered useless and criminals, including arms dealers and terrorists, are able to engage in criminal activity undetected by police. At present, law enforcement must use extensive manual effort to parse data provided by cell companies to extract information. We have built a system that analyses cellular GPRS metadata and builds a profile and finds potential call connections explicitly which are implicit in the dataset. This paper describes our approach and system in detail and includes results of our evaluation using an anonymized dataset from Delhi Police. Our system permits call graph analysis to be done, and significantly reduces the time needed from the data analysis process.	cs.SI	cs	In proceedings of I-CAPCHA Workshop, at IEEE International Conference on Collaboration and Internet Computing (CIC 2018) Analyzing CDR/IPDR data to find People Network from Encrypted Messaging Services Adya V. Joshi Howard High School Ellicott City, MD, USA [email protected] Madan Oberoi Special Commissior of Police Special Cell and Technology Cell Delhi Police, Delhi, India Indian Institute of Technology Delhi Ranjan Bose New Delhi, India Abstract -- Criminals are increasingly using mobile based communication applications, like WhatsApp, that have end-to- end encryption to connect to each other. This makes traditional analysis of call graphs, or traffic analysis, virtually impossible and so is a hindrance for law enforcement personnel. Old methods of traffic analysis have been rendered useless and criminals, including arms dealers and terrorists, are able to engage in criminal activity undetected by police. At present, law enforcement must use extensive manual effort to parse data provided by cell companies to extract information. We have built a system that analyses cellular GPRS metadata and builds a profile and finds potential call connections explicitly which are implicit in the dataset. This paper describes our approach and system in detail and includes results of our evaluation using an anonymized dataset from Delhi Police. Our system permits call graph analysis to be done, and significantly reduces the time needed from the data analysis process. Keywords -- mobile traffic analysis, call graph, GPRS CDR data, profile I. INTRODUCTION In the past few years, people all over the world have rapidly adopted Internet based messaging applications, like WhatsApp [5], Facebook Messenger[4], Rakuten Viber [5], WeChat [7] etc., because of the cost savings they provide. Today over a billion people are using these messaging service to connect within and outside their social network. According to Statista [11], WhatsApp [5] is the most popular mobile messaging app in the world. Facebook Messenger [4] ranked second, followed by Chinese services QQ Mobile and WeChat [7]. Other popular messaging apps include Skype, Viber, LINE, BBM, Telegram and Kakaotalk. Facebook owns the top 2 mobile messaging apps -- WhatsApp and Facebook Messenger. that makes to intercept However, it is being increasingly observed that criminals are also adopting these applications since they often provide end-to-end encryption law it difficult for enforcement agencies them [1]. Traditional methods of mobile traffic analysis have been rendered useless due to the encryption and criminals are able to engage in criminal activity undetected by police. Specifically, given a potential criminal suspect, the law enforcement agencies have no way of identifying their network by analyzing the call graph generated from their mobile phones, since no phone calls are being made. WhatsApp, or other encrypted messaging services, are used instead. Governments in various countries have expressed concerns that encrypted messaging services can be a threat to national security as it can be used by terrorists and other criminals to harm the society [2][10]. Some countries are even changing laws requiring messaging service providers to provide law enforcement agencies with access to encrypted messages [8][9]. Hence, there is an urgent need to develop new techniques to analyze mobile phone logs to be able to connect potential actors in criminal activities who might be using encrypted messaging services for their criminal activities. Call Detail Records (CDR) and Internet Protocol Detail Records (IPDR) help track details of a telecommunication call or message generated by a phone device. These logs contain metadata that describe details of a specific call, like calling phone number, destination port, start date/time, end date/time etc. Law enforcement in Delhi has been analyzing CDR and IPDR logs to track potential criminals. This includes GPRS CDRs. While this specifically refers to 2G data related calls, we use the term loosely to describe all CDRs that cover "data" calls that provide the IP connectivity to the handset. However, the current process requires lot of manual effort and is a long and tedious process. Police must go through folders of data provided by cell companies to find one file of metadata. After this they must go through many lines of data, and extract information by hand. This information is typically limited to strings that indicate what website a suspect visited, or strings that might have been part of files locally opened etc. It does not include any attempt at constructing call graphs. We have developed a system that analyzes mobile Call Data Record (CDR) GPRS and IPDR datasets to generate a profile of the mobile user. It also correlates the mobile users with other users of the encrypted messaging service to make explicit the implicit call graphs in this data. Our key insight is the following. When two suspects are talking to each other on WhatsApp, analyzing their GPRS CDRs will reveal that both are connected to WhatsApp. The reverse is of course not true. Just because two people are connected to WhatsApp at the same time does not mean that they are talking to one another. However, the more two people are on WhatsApp at the same time, the higher the probability that they are talking to one another. This means that by correlating two GPRS CDRs, we can see how often two suspects are using the same service at Figure 1: Persona Generation Process the same time, and potentially inferring that they are connected to each other. This mobile dataset was obtained by anonymizing data of criminals collected by the special cell unit of Delhi Police [3]. No personally identifiable information (PII) data was included in the logs that were studied for this project. Hence, no prior information of the criminal's profile was available for our system design. Our system significantly reduces the time from the data gathering and analysis process. In this paper, we describe our approach and the system in detail. Section II describes the related work. Section III describes the system design approach and the results of our evaluation. Section IV describes planned future work that will better this software.. II. RELATED WORK A. CDR Logs Analysis logs to mining CDR Teng and Chou [12] have proposed a graphical analytics approach to determine potential communities. Zhou et. al. [13] have analyzed CDRs to deduce social attribute from the calling behavior. However, this approach will not work on encrypted messaging services as the connection between the two users cannot be determined from the call logs. The only information the logs provide us is the port number of the application to which the mobile user connected. B. WhatsApp Messaging Application Anglano [17] in his analysis used many types of artifacts generated on devices by WhatsApp in correlation to find pieces of information like contact lists, chronology of messages, deleted contacts, and deleted messages. This information could not be gathered by analyzing these artifacts on their own. This research was limited to Android devices. Researchers have also reviewed encrypted messaging applications, like WhatsApp to study their impact on human behavior. Montag et. al. [15] conducted demographic analysis of WhatsApp usage. They observed that WhatsApp is about 20% of a typical daily smartphone usage and females, younger people, and those with extraversion tend to use it more than others. The study was conducted by collecting data directly through the subjects' smartphone and analyzing said data. Church et. al. [16] have compared benefits of WhatsApp with traditional SMS technology for mobile communication. These studies have however been limited to mobile data and not CDR logs. C. Mobile Data Analysis for Networks Researchers have analyzed and correlated mobile data logs to determine social networks. Eagle et. al. [14] demonstrated that friendships can be predicted based on the observational mobile data alone, where friend dyads demonstrate distinctive temporal and spatial patterns in their physical proximity and calling patterns. III. SYSTEM DESIGN log correlating Our system consists of three main components -- as first step it generates the persona or profile of the mobile phone user. The process of this component is illustrated in Figure 1 and described in section III. A. The second system component generates a logs and determining if they were connecting via encrypted messaging services. Figure 2 illustrates this process that is detailed in section III B. The third component focuses on WhatsApp [5] application usage by the mobile user and generates daily and weekly trend graphs from the IPDR logs. We have developed this system using Python 3.7 programming language. two mobile user The dataset available to us included the information listed in Table 1. We used all the fields for our analysis. Fields DESTPORT and MSISDN were the most important fields for our analysis along with the instance date and time fields. A. Persona Generation The first component of our system generates the persona or profile of the mobile user. This is derived from the GPRS metadata files. Figure 1 illustrates the various steps of the persona generation process. We also detail these steps below: Step 1: The program is formatted to accept CSV files, so all input files must be in CSV format. The first part of the program opens the file using a with loop. Figure 2: Process to build Correlation File TABLE I. CDR METADATA FIELDS Metadata field Description PRIVATEIP PRIVATEPORT PUBLICIP PUBLICPORT DESTIP DESTPORT MSISDN IMSI IP Address of the Mobile Device Port of the Mobile Device Public IP Address presented by the Mobile Device Public Port presented by the Mobile Device IP Address of the record Destination Port of the record destination Mobile Station International Subscriber Directory Number - number used to identify a mobile phone number internationally International Mobile Subscriber Identity number can identify the user of a mobile network Record Start date and Time START_DATE, START_TIME END_DATE, END_TIME Record End Date and Time IMEI International Mobile Equipment Identity Number that can uniquely identify the mobile phone ID of the Cell Tower Amount of data Upload CELL_ID UPLINK_VOLUME DOWNLINK_VOLUME Amount of data Downloaded TOTAL_VOLUME I_RATTYPE Total volume of the Data Identifies whether it is 2G Data or 3G Data Step 2 : After opening the file, program cross references the destination ports. A vital part of this project is the destination port. The destination port is needed to figure out what exactly the suspect is doing on the Wi-Fi. One major obstacle that presented itself was linking destination ports to a destination. Step 2A: In the traditional method, the law enforcement analyst manually extracted all possible destination ports from the data. This was a time consuming process as over nine thousand points of data were analyzed, and destination ports were repeated many times. After this tedious process, one had to link each unique port to a destination. This proved to be the most challenging part, as there is no central database in the public domain with all of the ports and their associations. The process required us to manually browse through multiple websites, some which were helpful and some which were not. Sites that frequently came up were question and answer websites, such as Quora, and Yahoo Answers. For the different applications, we had to manually identify where each port lead to. Step 2B: Using the identified ports, we created lists of the major applications that were frequently accessed. These included WhatsApp, iTunes/QuickTime, Microsoft/Games, and Web access via HTTP and HTTPS. Table II lists all the ports and their corresponding applications that were identified by us. Step 3: The next part identifies the precise destination. Our program runs a reverse DNS search to find the site that was accessed. This is accomplished using the destination IP. If the search is successful, the system returns the web address accessed by the person. Otherwise, the destination IP is returned. Step 4: The next step was to link each port to a destination. Earlier in the process, the ports had already been cross referenced, this step linked each port which existed in the program's dictionary to a destination (see Table II). Step 5: The last step outputs all of the information to a text file in a format that is readable by the user. This output file is used to generate the User profile graph from the data processed. We illustrate such a profile chart in section IV (see Figure 4). B. Correlating Callers The second component of our system determines the correlation between two mobile users by reviewing their CDR logs and determining the time when the two were potentially accessing the same messaging application. Figure 2 illustrates the process of correlating two mobile callers. The steps followed by the system to identify the connected mobile users are detailed below: Figure 2: Process to build Correlation File Step 1: To find connections between any two mobile users, we begin with the GPRS metadata files of the users. The files are converted into CSV format and then inputted into the system program. Step 2: The next step is to establish if two events in the GPRS file have the same destination port. All the instances in the two files with the same destination port are identified. TABLE II. Step 3: The next part involves determining that the time difference is within the specified threshold. We calculate the difference between the start times of events in the two files that have the same destination port. If the time difference is within the specified threshold, then the program will mark the two as a connection. The threshold time range can be changed at the analyst request. Step 4: The next part is to find out on what application the two people were on at the same time. This is done by linking the destination port to a destination, this is done by referencing the dictionary of ports that was established earlier in table II. Step 5: Once these details are gathered, the system appends them to a file. Figure 3 illustrates this output file generated by comparing some CDR logs of two mobile callers. We provide details about how often connections were made and what fractions of connections were in common. The analyst can provide a threshold above which this degree of common connectivity can be assumed to be a connection. There are clear tradeoffs between false positives and false negatives here, and the choice is a function of the tradecraft of the analyst or the policies of the organization. C. Caller's WhatsApp Usage Trends The third component of our system generates the trends of WhatsApp usage by a mobile user by analyzing their IPDR logs. This component looks at one or more files to generate usage trends only for WhatsApp application. With one input file of IPDR data, the system creates two output files of results. One file, a text file, lists all the individual connections. The second, a csv file, has a break down for the day based on three- hour intervals (i.e. 12am -- 3 am). When a directory is inputted into this component, the only result is a csv file with a breakdown of usage trends within three-hour intervals. is individual record Like the other components of the system, in this component too an identified as a WhatsApp connection using the destination port. If the destination port in a record is a valid WhatsApp port number (see table II), then it is analyzed further. To provide trends for a three-hour interval, the start time of the record is compared with the time intervals and placed within the correct one. The time intervals are logged in their own CSV file which can be further analyzed if PORT AND APPLICATION LIST Description / Protocol Destination Application WhatsApp WhatsApp WhatsApp WhatsApp WhatsApp Skype Skype Skype Web Connection Web Connection Web Connection IMAP iTunes McAfee E-Policy Orchestrator Black Ice Cap TCP UDP UDP+TCP Application Vendor Facebook Inc Facebook Inc Facebook Inc Facebook Inc Facebook Inc Microsoft Inc Microsoft Inc Microsoft Inc Apple Inc. Secure Sockets Layer Apple iTunes music Apple Inc.'s storage area network (SAN) or clustered file system for macOS. Xsan Apple Inc. Microsoft Inc. Microsoft Various Online Games Destination Port 5223 5228 4244 5222 5242 443 3478-3481 49152-65535 80, 8080 443 8081 993, 143 8024, 8027, 8013, 8017, 8003, 7275, 8025, 8009 58128, 51637, 61076 40020, 40017, 40023, 40019, 40001, 40004, 40034, 40031, 40029, 40005, 40026, 40008, 40032 Figure 3: Correlation Result file IV. SYSTEM EVALUATION For evaluating our system design, we used the anonymized dataset provided by Delhi Police. This data consisted of CDR and IPDR logs of potential criminals being tracked by the law enforcement agencies. The results of the persona generation component of our system is illustrated in Table III and Figure 4. It gives an analyst a quick sense of what the suspect is spending time on, what their interests might be etc. TABLE III. PERSONA GENERATED FOR MOBILE USER Application WhatsApp iTunes/QuickTime Microsoft or was playing games 38 Xsan File System Email Web via HTTPS Web via HTTP 3 199 4983 721 Frequency Usage percent 1997 18 25.09 0.22 0.47 0.03 2.5 62.6 9.05 The results of the User correlation generation component of our system is illustrated in figure 3 and figure 5. The figure shows how much time correlation takes in a worst-case situation, in this case, with all destination ports matching. Even Figure 4: User Profile generated by the system Figure 5: User Connections results generated by the system needed (see figures 8 and 9). In the appropriate case, the rest of the details are logged in a separate text file (see figure 7). figure 9 for WhatsApp trends for a week by Time of the day. These graph were found to be very relevant by the law enforcement agencies to determine the time when a suspect is active on the messaging application and help build his profile. with 200 data points, it takes about a second. Using the chart editor on Google sheets, we found the equation for the trendline of the data in Figure 6. It is a linear equation. However, since we've naively implemented correlation, we expected the big Oh scaling to be n2. To validate the accuracy of our work, we manually correlated the data, and then checked those results against the results generated by the system. This was done under multiple different scenarios. The base case was anonymized data from a real suspect's GPRS dump. This was used to create another synthetic GPRS dump which had no common usage with the real suspect dump. This created the first test, and as expected, the system found no overlap. We then created a series of test cases with varying degrees of constructed "overlap". The overlap was also manually verified. This was then compared against the output of our program, and it was found to confirm in each case Figure 8: WhatsApp connections trend for a week based on day Figure 9: WhatsApp Usage for a week based on Time of Day To the best of our knowledge, there are no similar programs that are available publicly. Such programs might exist internally with law enforcement or intelligence communities, but they are not discussed in the public domain. V. CONCLUSION AND FUTURE WORK In this paper, we have described the system we developed to generate a profile of a mobile user based on their GPRS CDR metadata. We also developed a technique to determine potential connection between two mobile users based on simultaneous user access of an encrypted messaging service or other similar application ports. This is far less time-consuming than the manual methods in use currently. This system is very useful to the law enforcement agencies who are faced with the challenge of monitoring messaging applications that have end- to-end encryption. Note that this is a proof of concept that demonstrates the power of our approach. Due to legalities associated with data access, we can't directly test our system Figure 6: Performance Trends of the correlation Figure 7: WhatsApp connections identified from the IPDR log The results of the WhatsApp trends component of our system is illustrated in figures 7,8 and 9. The system output identifies that a connection is being made to WhatsApp based on the port number . (Illustrated in figure 7). Figure 8 illustrates the graphical output for WhatsApp usage for a week by day and on multiple real data sets. However, our code has recently been turned over to Delhi Police, who plan to use it in their work. As part of our on-going work, we are also developing a user-friendly graphical interface for our system. In the future, we plan to enhance this system to allow analysis of other data sets such as Internet Protocol Data Records, IPDRs. We also plan to make some features such as the reverse DNS search optional in an effort to shorten the time taken by the user persona builder. ACKNOWLEDGEMENT This work was done while the first author visited Delhi and used the facilities of IIT Delhi and Delhi Police. The first author would like to acknowledge help and guidance provided by Mr. Pradip Kushwaha DANIPS (DCP Special/Technology Cell), Mr. Deepak Vats and Ms. Chanda Sahijwani of Delhi Police, and Mr. Miftah Siddiqui of IIT Delhi. REFERENCES [1] WhatsApp encryption: Online criminal activity no longer tapped, http://gulfnews.com/news/uae/crime/whatsapp-encryption-online- criminal-activity-no-longer-tapped-1.1707921, last retrieved 2/5/18 [2] WhatsApp must be accessible to authorities, says Amber Rudd, https://www.theguardian.com/technology/2017/mar/26/intelligence- services-access-whatsapp-amber-rudd-westminster-attack-encrypted- messaging, last retrieved 2/5/18 [3] Delhi Police , http://www.delhipolice.nic.in/ [4] Facebook Messenger, https://www.messenger.com/ [5] WhatsApp, https://www.whatsapp.com/ [6] Rakuten Viber , https://www.viber.com/ [7] WeChat, https://www.wechat.com/en/ [8] Australian PM's statement on Defense and National Security, June 2017, https://www.pm.gov.au/media/national-security-statement [9] How the Turnbull government plans to access encrypted messages, http://www.smh.com.au/federal-politics/political- David news/how-the-turnbull-government-plans-to-access-encrypted- messages-20170609-gwoge0.html, retrieved 2/7/18 Wroe, [10] Deputy Attorney General Rod J. Rosenstein Delivers Remarks on Encryption at the United States Naval Academy, October 2017, https://www.justice.gov/opa/speech/deputy-attorney-general-rod-j- rosenstein-delivers-remarks-encryption-united-states-naval, retrieved 2/7/2018 last [11] Mobile messenger apps - Statistics & Facts, Statista.com, https://www.statista.com/topics/1523/mobile-messenger-apps/ [12] Wei-Guang Teng and Ming-Chia Chou, Mining communities of acquainted mobile users on call detail records. In Proceedings of the 2007 ACM symposium on Applied computing (SAC '07). ACM, New York, 957-958. DOI=http://dx.doi.org/10.1145/1244002.1244212 USA, NY, [13] C. Zhou, Z. Xu, and B. Huang, "Activity recognition from call detail record: Relation between mobile behavior pattern and social attribute using hierarchical conditional random fields." IEEE, 2010, pp. 605 -- 611. [14] Nathan Eagle, Alex (Sandy) Pentland, David Lazer, 'Inferring friendship network structure by using mobile phone data', Proceedings of the National Academy of Sciences Aug 2009, pnas.0900282106; DOI: 10.1073/pnas.0900282106 [15] Montag, C., Błaszkiewicz, K., Sariyska, R., Lachmann, B., Andone, I., Trendafilov, B., … Markowetz, A. (2015). Smartphone usage in the 21st century: who is active on WhatsApp? BMC Research Notes, 8, 331. http://doi.org/10.1186/s13104-015-1280-z [16] Karen Church and Rodrigo de Oliveira, What's up with whatsapp?: comparing mobile instant messaging behaviors with traditional SMS, In Proceedings of the 15th international conference on Human-computer interaction with mobile devices and services (MobileHCI '13). ACM, New DOI: https://doi.org/10.1145/2493190.2493225 352-361. York, USA, NY, [17] Cosimo Anglano. Forensic analysis of WhatsApp Messenger on Android smartphones. Digit. Investig. 11, 3 (February 2015), 201-213. DOI=http://dx.doi.org/10.1016/j.diin.2014.04.003
1904.08114	1	1904	2019-04-17T07:54:06	Variational principle for scale-free network motifs	[ "cs.SI", "physics.soc-ph" ]	For scale-free networks with degrees following a power law with an exponent $\tau\in(2,3)$, the structures of motifs (small subgraphs) are not yet well understood. We introduce a method designed to identify the dominant structure of any given motif as the solution of an optimization problem. The unique optimizer describes the degrees of the vertices that together span the most likely motif, resulting in explicit asymptotic formulas for the motif count and its fluctuations. We then classify all motifs into two categories: motifs with small and large fluctuations.	cs.SI	cs	Variational principle for scale-free network motifs Clara Stegehuis1,*, Remco van der Hofstad1, and Johan S.H. van Leeuwaarden1 1Department of Mathematics and Computer Science , Eindhoven University of Technology April 18, 2019 Abstract For scale-free networks with degrees following a power law with an exponent τ ∈ (2, 3), the structures of motifs (small subgraphs) are not yet well understood. We introduce a method designed to identify the dominant structure of any given motif as the solution of an optimization problem. The unique optimizer describes the degrees of the vertices that together span the most likely motif, resulting in explicit asymptotic formulas for the motif count and its fluctuations. We then classify all motifs into two categories: motifs with small and large fluctuations. Introduction Many real-world networks, like communication networks, social networks and biological networks, were found to be scale free with power-law degree distributions and infinite variance in the large- network limit [1, 2, 3, 4, 5]. The heavy tail of the power law comes with hubs, vertices of extremely high degree. Hubs create ultra-small distances, ultra-fast information spreading and resilience against random attacks, while the average node degree is small. Attested by real-world data and explained by mathematical models, the consequences of power-law connectivity patterns now belong to the foundations of network science. Scale-free networks can be studied using random network models that connect vertices through edges forming power-law degree distributions. Connectivity patterns beyond edges are commonly described in terms of small subgraphs called motifs (or graphlets). There is increasing interest in algorithmic methods to count motifs [6, 7, 8, 9, 10], or the relation between motifs and network functions, like the spread of epidemics [11, 12, 13, 14, 15, 16]. Motifs can describe the tendency for clustering and other forms of network organization [17, 18, 19]. Much existing work focuses on the classification of complex networks in terms of motif counts or frequencies. The occurrence of specific motifs such as wedges, triangles and cliques have been proven important for understanding real-world networks. Indeed, motif counts might vary con- siderably across different networks [20, 21, 22] and any given network has a set of statistically significant motifs. Statistical relevance can be expressed by comparing real-world networks to mathematically tractable models. A popular statistic takes the motif count, subtracts the ex- pected number of motifs in the mathematical model, and divides by the standard deviation in the mathematical model [22, 23, 24]. Such a standardized test statistic predicts whether a motif is overrepresented in comparison to some mathematical model. This comparison, however, filters out the effect of the degree sequence, the network size and possibly other features that are needed to understand the structure and frequency of motifs. With the goal to explain the occurrence of motifs beyond counting, we develop a method to identify, for any given motif, the composition that dominates the motif count as the solution of an optimization problem. The unique optimizer describes the degrees of the vertices that together span the most likely motif, as well as predicts the leading asymptotic order for the motif count in 1 the large-network limit. Our method can potentially be applied to various random network models, but is developed first for the hidden-variable model [25, 26, 27, 28, 29, 30, 31], a random network model that generates graphs with power-law degrees. Given n vertices, the hidden-variable model associates to each node a hidden variable h drawn from the probability density ρ(h) = Ch−τ (1) for some constant C and h ≥ hmin. Next, conditionally on all the hidden variables, each pair of vertices is joined independently with probability p(h, h(cid:48)) = min(hh(cid:48)/(µn), 1). (2) with h and h(cid:48) the hidden variables associated with the two vertices, and µ the mean of the hidden variables. For any given motif, we now seek for its most likely structure. The probability P (H) of creating motif H on k uniformly chosen vertices can be written as P (H) = P (H on h1, . . . , hk) P (h1, . . . , hk) dh, (3) (cid:90) h where the integral is over all possible hidden-variable sequences on k vertices, with h = (h1, . . . , hk) and P (h1, . . . , hk) the density that a randomly chosen set of k hidden variables is proportional to h1, . . . , hk. The degree of a node is asymptotically Poisson distributed with its hidden variable as mean [32], so (3) can be interpreted as a sum over all possible degree sequences. Therefore, our optimization method then needs to settle the following trade-off, inherently present in power- law networks: On the one hand, large-degree vertices contribute substantially to the number of motifs, because they are highly connected, and therefore participate in many motifs. On the other hand, large-degree vertices are by definition rare. This should be contrasted with lower-degree vertices that occur more frequently, but take part in fewer connections and hence fewer motifs. Therefore, our method give rise to a certain 'variational principle', because it finds the selection of vertices with specific degrees that together 'optimize' this trade-off and hence maximize the expected number of such motifs. We leverage the optimization method in two ways. First, we derive sharp expressions for the motif counts in the large-network limit in terms of the network size and the power-law exponent. Second, we use the method to identify the fluctuations of motif counts. We present two versions of the method that we call free and typical variation. Free variation corresponds to computing the average number of motifs over many samples of the random network model. Typical variation corresponds to the number of motifs in one single instance of the random graph model. Remarkably, for τ ∈ (2, 3) these can be rather different. After that, we apply the method to study motif count fluctuations. Finally, we provide a case study where we investigate the presence of motifs in some real-world network data. Results Free variation We first show that only hidden-variable sequences h with hidden variables of specific orders give the largest contribution to (3). Write the hidden variables as hi ∝ nαi for some αi ≥ 0 for all i. Then, using (2), the probability that motif H exists on vertices with hidden variables h = (nα1, . . . , nαk ) satisfies (cid:89) P (H on h) ∝ (vi,vj )∈EH :αi+αj <1 nαi+αj−1. (4) The hidden variables are an i.i.d. sample from a power-law distribution, so that the probability n(1−τ )(cid:80) that k uniformly chosen hidden variables satisfy (h1, . . . , hk) ∝ (nα1, . . . , nαk ) is of the order i αi (see Supplementary Material 2). Taking the product of this with (4) shows that the 2 3 2 (3−2τ ) (a) n (b) n4−τ (c) n6−2τ (d) n6−2τ log(n) (e) n6−2τ log(n) (f) n4−τ (g) n4−τ (h) n4−τ log(n) n √n 1 non-unique Figure 1: Scaling of the expected number of motifs on 3 and 4 vertices in n, where the vertex color indicates the dominant vertex degree. Vertices where the optimizer is not unique are gray. maximum contribution to the summand in (3) is obtained for those αi ≥ 0 that maximize the exponent (5) αi + (1 − τ ) (i,j)∈EH :αi+αj <1 (αi + αj − 1) , (cid:88) i (cid:88) which is a piecewise-linear function in α. In Supplementary Material 2, we show that the maximizer of this optimization problem satisfies αi ∈ {0, 1 2 , 1} for all i. Thus, the maximal value of (5) is attained by partitioning the vertices of H into the sets S1, S2, S3 such that vertices in S1 have αi = 0, vertices in S2 have αi = 1 and vertices in S3 have αi = 1 2 . Then, the edges with αi +αj < 1 are edges inside S1 and edges between S1 and S3. If we denote the number of edges inside S1 by ES1 and the number of edges between S1 and S3 by ES1,S3, then maximizing (5) is equivalent to maximizing (cid:20) (cid:21) Bf (H) = maxP S1 − S2 − 2ES1 + ES1,S3 τ − 1 (6) (7) over all partitions P of the vertices of H into S1, S2, S3. This gives the following theorem (a more elaborate version is proven in Supplementary Material 2): Theorem 1 (Expected motif count). Let H be a motif on k vertices such that the solution to (6) is unique. As n → ∞, the expected number of motifs H grows as and is thus fully determined by the partition P E [N (H)] = nkP (H) ∝ n 3−τ 2 k+ τ−1 2 Bf (H), ∗ that optimizes (6). Theorem 1 implies that the expected number of motifs is dominated by motifs on vertices with hidden variables (and thus degrees) of specific orders of magnitude: constant degrees, degrees ∗ proportional to √n or degrees proportional to n. Figure 1 and Figure 2 show the partitions P that dominate the expected number of motifs on three, four and five vertices. Typical variation The largest degrees (hubs) in typical samples of the hidden-variable model scale as n1/(τ−1) with high probability. The expected number of motifs, however, may be dominated by network samples where the largest degree is proportional to n (see Theorem 1). These samples contain many motifs because of the high degrees, and therefore contribute significantly to the expectation. Nevertheless, the probability of observing such a network tends to zero as n grows large. We therefore now adapt the variational principle with the goal to characterize the typical motif structure and hence the typical number of motifs. We again assume degrees to be proportional to nαi, but now limit to degree sequences where the maximal degree is of order n1/(τ−1), the natural cutoff in view of the typical hub degrees. 3 5 2 (3−τ ) (a) n 5 2 (3−τ ) (b) n 5 2 (3−τ ) (c) n 5 2 (3−τ ) (d) n 13 2 −2τ (e) n 5 2 (3−τ ) (f) n 5 2 (3−τ ) (g) n (h) n7−2τ (i) on τ depends 13 2 −2τ (j) n (k) n7−2τ 5 2 (3−τ ) (l) n (m) n7−2τ (n) depends on τ 5 2 (3−τ ) (o) n (p) n5−τ (q) n7−2τ (r) n7−2τ (s) n 11−3τ 2 log(n) (t) n5−τ (u) n6−τ Figure 2: Scaling of the expected number of motifs on 5 vertices in n, where the vertex color indicates the dominant vertex degree, as in Fig. 1. (a) n6−2τ (b) n6−2τ log(n) (c) n6−2τ log(n) (d) n7−2τ− 1 τ−1 (e) n 3 τ−1 (f) n4−τ log(n) 1 n(τ−2)/(τ−1) √n n1/(τ−1) non-unique Figure 3: Typical scaling of the number of motifs on three or four vertices in n. The vertex color indicates the dominant vertex degree. 1 τ−1}, and obtain an optimization problem similar to (6). The dominant typical motif structure is then obtained by maximizing (5), with the additional constraint that αi ≤ 1 τ−1 . In Supplementary Material 3 we show that the possible optimizers are one of four values αi ∈ {0, τ−2 τ−1 , 1 2 , This shows that the typical degree of a motif is of constant order or proportional to n1/(τ−1), √n or n(τ−2)/(τ−1). Figure 3 and Supplementary Figure 1 show the most likely motifs on three, four and five vertices. Observe that the dominant structures and the number of motifs of Fig. 1 and 3 may differ. For example, the scaling of the expected number of claws (Fig. 1b) and the typical number of claws (Fig. 3b) is different. This is caused by the left upper vertex that has degree proportional to n in the free dominant structure, whereas its typical degree is proportional to n1/(τ−1). Only when the solution to (6) does not involve hub vertices, the two scalings coincide. Hub vertices in the dominant structure give a major contribution to the motif count. While typical hub degrees scale as n1/(τ−1), expected hub degrees may be much larger, causing the number of such motifs with hubs to scale faster in the free variation setting than in the typical variation setting. This indicates that the average and median motif count can differ dramatically. Graphlets It is also possible to only count the number of times H appears as an induced subgraph, also called graphlet counting. This means that an edge that is not present in graphlet H, should also be absent in the network motif. In Supplementary Material 4 we classify the expected and typical number of graphlets with a similar variational principle as for motifs. Supplementary 4 Figure 2 shows the typical behavior of graphlets on 4 vertices. This figure also shows that graphlet counting is more detailed than motif counting. For example, counting all square motifs is equivalent to counting all graphlets that contain the square as an induced subgraph: the square, the diamond and K4. Indeed, we obtain that the number of square motifs scales as n6−2τ log(n) by adding the number of square, diamond and K4 graphlets from Supplementary Figure 2. This shows that most square motifs are actually the diamond graphlets of Supplementary Figure 2b. Thus, graphlet counting gives more detailed information than motif counting. Fluctuations Self-averaging network properties have relative fluctuations that tend to zero as the network size n tends to infinity. Several physical quantities in for example Ising models, fluid models and proper- ties of the galaxy display non-self-averaging behavior [33, 34, 35, 36, 37, 38, 39]. We consider motif counts N (H) and call N (H) self-averaging when Var (N (H)) /E [N (H)]2 → 0 as n → ∞. Essential understanding of N (H) can then be obtained by taking a large network sample, since the sample- to-sample fluctuations vanish in the large-network limit. In contrast, if Var (N (H)) /E [N (H)]2 approaches a constant or tends to infinity as n → ∞, the motif count is called non-self-averaging, in which case N (H) shows (too) strong sample-to-sample fluctuations that cannot be mitigated by taking more network samples. Our variational principle facilitates a systematic study of such fluctuations, and leads to a classification into self-averaging and non-self-averaging for all motifs H. It turns out that whether N (H) is self-averaging or not depends on the power-law exponent τ and the dominant structure of H. We also show that non-self-averaging behavior of motif counts may not have the intuitive explanation described above. In some cases, motif counts in two instances are similar with high probability, but rare network samples behave differently, causing the motif count to be non-self- averaging. Thus, the classification of network motifs into self-averaging and non-self-averaging motifs does not give a complete picture of the motif count fluctuations. We therefore further divide the non-self-averaging motifs into two classes based on the type of fluctuations in the motif counts. For a given motif H, let H1, . . . , Hm denote all possible motifs that can be constructed by merging two copies of H at one or more vertices. We can then write the variance of the motif count as (see [40, 41, 42, 37] and the Methods section) Var (N (H)) = C1E [N (H1)] + ··· + CmE [N (Hm)] + E [N (H)]2 O(n−1). (8) for constants C1, . . . , Cm. Using (8), we can determine for any motif H whether it is self-averaging or not. First, we find all motifs that are created by merging two copies of H. For the triangle motif for example, these motifs are the bow-tie, where two triangles are merged at one single vertex, the diamond of Fig. 3b, and the triangle itself. We find the order of magnitude of the expected number of these motifs using Theorem 1 to obtain the variance of N (H). We divide by E [N (H)]2, also obtained by Theorem 1, and check whether this fraction is diverging or not. Table 1 shows for which values of τ ∈ (2, 3) the motifs on 3, 4 and 5 vertices are self-averaging. For example, the triangle turns out to be self-averaging only for τ ∈ (2, 5/2). Here is a general observation that underlines the importance of the dominant motif structure: Theorem 2. All self-averaging motifs for any τ ∈ (2, 3) have dominant free variation structures that consist of vertices with hidden variables Θ(√n) only. We prove this theorem in Supplementary Material 5. Note that the condition on the dominant motif structure is a necessary condition for being self-averaging, but it is not a sufficient one, as the triangle example shows. Table 1 shows the values of τ for which all connected motifs on 3, 4 and 5 vertices are self-averaging. Combining the classification of the motifs into self-averaging and non-self-averaging with the classification based on the value of Bf (H) from (6) as well as the difference between the expected and typical number of motifs yields a classification into the following three types of motifs: 5 self-averaging for subfigs of Fig. 1 subfigs of Fig. 2 (2,3) (2,5/2) (2,7/3) - c a b,d,e,f,g,h a,b,c,d f,g,l,o i e,h,j,k,m,n,p,q,r,s,t,u Table 1: The values of τ ∈ (2, 3) where the motifs of Figs. 1 and 2 are self-averaging. non-concentrated small variance concentrated, non-self-averaging (a) τ = 2.2, n = 105, triangles (b) τ = 2.8, n = 105, triangles (c) τ = 2.2, n = 105, wedges (d) τ = 2.2, n = 5 · 106, (e) τ = 2.8, n = 5 · 106, (f) τ = 2.2, n = 5 · 106, wedges triangles triangles Figure 4: Density approximation of the normalized triangle and wedge counts for various values of τ and n, obtained over 104 network samples. Type I: Motifs with small variance. Bf (H) = 0 and Var (N (H)) /E [N (H)]2 → 0. These motifs only contain vertices of degrees Θ(√n). The number of such rescaled motifs converges to a constant [43]. Furthermore, the variance of the number of motifs is small compared to the second moment, so that the fluctuations of these types of motifs are quite small and vanish in the large network limit. The triangle for τ < 5/2 is an example of such a motif, shown in Figs 4b and 4e. Type II: Concentrated, non-self-averaging motifs. Bf (H) = 0 and Var (N (H)) /E [N (H)]2 (cid:57) 0. These motifs also only contain vertices of degrees √n. Again, the rescaled number of such motifs converges to a constant in probability [43]. Thus, most network samples contain a similar amount of motifs as n grows large, even though these motifs are non-self-averaging. Still, in rare network samples the number of motifs significantly deviates from its typical number, causing the variance of the number of motifs to be large. Figs 4a and 4d illustrate this for triangle counts for τ ≥ 5/2. The fluctuations are larger than for the concentrated motifs, but most of the samples have motif counts close to the expected value. Type III: Non-concentrated motifs. Bf (H) > 0. These motifs contain hub vertices. The expected and typical number of such motifs therefore scale differently in n. By Theorem 2, these motifs are non-self-averaging. The rescaled number of such motifs may not converge to a constant, so that two network samples contain significantly different motif counts. Figs 4c and 4f show that the fluctuations of these motifs are indeed of a different nature, since most network samples have motif counts that are far from the expected value. 6 10−11001010123N(4)/E[N(4)]10−110010100.20.40.60.81N(4)/E[N(4)]10−11001010123N(∨)/E[N(∨)]10−11001010123N(4)/E[N(4)]10−110010100.20.40.60.81N(4)/E[N(4)]10−11001010123N(∨)/E[N(∨)]Figure 5: Number of graphlets on four vertices in five data sets on logarithmic scale: log(N (H))/ log(n). The ordering of the six different graphlets is from the most occurring in the hidden-variable model to the least. The values of the graphlet counts are presented in Sup- plementary Table 1. Data n m τ Gowalla Oregon Enron PGP Hep 196591 11174 36692 10680 9877 950327 23409 183831 24316 25998 2.65 2.08 1.97 2.24 3.50 Table 2: Statistics of the five data sets, where n is the number of vertices, m the number of edges, and τ the power-law exponent fitted by the procedure of [5]. We now investigate motifs in five real-world networks with heavy-tailed degree distributions: the Gowalla social network [44], the Oregon autonomous systems network [44], the Enron email network [45, 44], the PGP web of trust [46] and the High Energy Physics collaboration network (HEP) [44]. Table 2 provides detailed statistics of these data sets. Because the number of motifs can be obtained from the number of graphlets, we focus on graphlet counts. Figure 5 shows the graphlet counts on a logarithmic scale. The order of the graphlets is from the most occurring graphlet (the claw), to the least occurring graphlet (the square and K4) in the hidden-variable model, see Supplementary Figure 2. In three networks the motif ordering follows that of the hidden-variable model, while in two networks the ordering is different. In the HEP collaboration network, for example, K4 occurs more frequently than the square. While this is not predicted by the hidden-variable model, it naturally arises due to the frequently occurring collaboration between four authors, which creates K4 instead of the square. It would be interesting to see if this deviation from the ordering of the hidden-variable model can be linked to the specific nature of the data set in other examples. Supplementary Figure 2 enumerates all possible vertex types in graphlets on 4 vertices. In the hidden-variable model, vertex types t7 and t9 have typical degrees proportional to n1/(τ−1) in the hidden-variable model, vertex types t1 and t4 typically have degrees proportional to √n, vertex type t6 typically has degree proportional to n(τ−2)/(τ−1) and vertex types t5, t8, t10, t11 typically have constant degree. vertex types t2, t3 and t11 do not have a unique optimizer. The degrees of these vertex types are pair-constrained (see the proof of Lemma 3). Figure 6 shows the typical degree of all 11 vertex types in the five real-world data sets. Vertices with typical degree 1 in the hidden-variable model have the lowest degree in the five data sets. Vertices that have typical degree n1/(τ−1) in the hidden-variable model also have the highest degree among all vertex types in these five real-world data sets. Thus, typical degrees of vertices in a graphlet roughly follow the same ordering as in the hidden-variable model in these data sets. 7 Figure 6: Average degree of the vertex types displayed in Supplementary Figure 2 in 5 data sets on logarithmic scale. The curly brackets indicate the typical degree exponent of the vertex type in the hidden-variable model. The High Energy Physics collaboration network does not have a large distinction between the degrees of the different vertex types. This may be related to the fact that this network has less heavy-tailed degrees than the other networks (see Table 2). Discussion By developing a variational principle for the dominant degree composition of motifs in the hidden- variable model, we have identified the asymptotic growth of motif counts and their fluctuations for all motifs. This allowed us to determine for which values of the degree exponent τ ∈ (2, 3) the number of motifs is self-averaging. We further divide the non-self-averaging motifs into two classes with substantially different concentration properties. Hub vertices in dominant motif structures cause wild degree fluctuations and non-self-averaging behavior, so that large differences between the average motif count and the motif count in one sample of the random network model arise. Non-self-averaging motifs without a hub vertex show milder fluctuations. We expect that the variational principle can be extended to different random graph models, such as the hyperbolic random graph, the preferential attachment model and random intersection graphs. For example, for the hyperbolic random graph, the dominant structure of complete graphs is known to be √n degrees [47] like in the hidden-variable model, but the dominant structures of other motifs are yet unknown. In this paper, we presented a case study for motifs on 4 vertices in five scale-free network data sets. It would be interesting to perform larger network data experiments to investigate whether motifs in real-world network data also have typical vertex degrees, and to what extent these vertex degrees are similar to the ones of the hidden-variable model. Similarly, investigating the typical behavior of motifs on more than 4 vertices in real-world data compared to the hidden-variable model is another topic for future research. It would also be interesting to develop statistical tests for the presence of motifs in real-world data using the results from this paper. For example, one could compare the ordering of all motifs for size k from the most frequent occurring to the least frequent occurring motif, and compare this to the ordering in a hidden-variable with the same degree-exponent. This could shed some light on which motifs in a given data set appear more often than expected. Acknowledgements. This work is supported by NWO TOP grant 613.001.451 and by the NWO Gravitation Networks grant 024.002.003. The work of RvdH is further supported by the NWO VICI grant 639.033.806. The work of JvL is further supported by an NWO TOP-GO grant and by an ERC Starting Grant. 8 Methods Fluctuations Triangle fluctuations. We first illustrate how we can apply the variational principle to obtain the variance of the number of subgraphs by computing the variance of the number of triangles in the hidden-variable model. Let (cid:52) denote the number of triangles, and let (cid:52)i,j,k denote the event that vertices i, j and k form a triangle. Then, we can write the number of triangles as 1(cid:52)i,j,k , (9) (cid:88)(cid:48) i,j,k∈[n] 6 (cid:52) = 1 (cid:88)(cid:48) (cid:88)(cid:48) where (cid:80)(cid:48) can be written as Var ((cid:52)) = denotes the sum over distinct indices. Thus, the variance of the number of triangles i,j,k∈[n] s,t,u∈[n] P ((cid:52)i,j,k,(cid:52)s,t,u) − P ((cid:52)i,j,k) P ((cid:52)s,t,u) . (10) When i, j, k and s, t, u do not overlap, the hidden variables of i, j, k and s, t, u are independent, so that the event that i, j and k form a triangle and the event that s, t and u form a triangle are independent. Thus, when i, j, k, s, t, u are all distinct, P ((cid:52)i,j,k,(cid:52)s,t,u) = P ((cid:52)i,j,k) P ((cid:52)s,t,u), so that the contribution from 6 distinct indices to (10) is zero. On the other hand, when i = u for example, the first term in (10) denotes the probability that a bow tie (see Fig. 2i) is present with i as middle vertex. Furthermore, since the degrees are i.i.d. and the edge statuses are independent as well, P ((cid:52)i,j,k) is the same for any i (cid:54)= j (cid:54)= k, so that (cid:1) = E [(cid:52)] 6(cid:0)n P ((cid:52)i,j,k) = E [(cid:52)] 6n3 (1 + o(1)). (11) 3 This results in Var ((cid:52)) = 9E [# bow-ties] − 9n−1E [(cid:52)]2 + 18E [# diamonds] − 18n−2E [(cid:52)]2 + 6E [(cid:52)] − 6n−3E [(cid:52)]2 = 9E [# bow-ties] + 18E [# diamonds] + 6E [(cid:52)] + E [(cid:52)]2 O(n−1), (12) where the diamond motif is as in Fig. 3b. The combinatorial factors 9,18 and 6 arise because there are 9 ways to construct a bow tie (18 for a diamond, and 6 for a triangle) by letting two triangles overlap. The diamond motif does not satisfy the assumption in Theorem 1 that the optimal solution to (6) is unique. However, we can show the following result: Lemma 3. E [number of diamonds] = Θ(cid:0)n6−2τ(cid:1) log(n). Proof. Let i and j be the vertices at the diagonal of the diamond, and k and s the corner vertices. Then (5) is optimized for αi = β, αj = 1 − β, αk = β and αs = 1 − β for all values of β ∈ [1/2, 1] (see Supplementary Information 6). All these optimizers together give the major contribution to the number of diamonds. Thus, we need to find the number of sets of four vertices, satisfying hihj = Θ(n), hi > hj, hk = Θ(hi), hs = Θ(hj). (13) i j h1−τ Given hi and hj, the number of sets of two vertices k, s with hk = Θ(hi) and hs = Θ(hj) is given by n2h1−τ = Θ(n3−τ ), where we used that hihj = Θ(n). The number of sets of vertices i, j such that hihj = Θ(n) can be found using that the product of two independent power-law random variables is again distributed as a power law, with an additional logarithmic term [48, Eq. (2.16)] (where in our setting equality holds in [48, Eq. (2.16)], since we assume a pure power-law distribution). Thus, the number of sets of vertices with hihj = Θ(n) scales as n2n1−τ log(n). Then, the expected number of sets of four vertices satisfying all constraints on the degrees scales as n6−2τ log(n). By (4), the probability that a diamond exists on degrees satisfying (13) is asymptotically constant, so that the expected number of diamonds also scales as n6−2τ log(n). 9 Theorem 1 gives for the number of bow ties that (using [49] to find the optimal partition) and for the number of triangles (again using [49]) that E [(cid:52)] = Θ(n3(3−τ )/2). Combining this with (12) results in To investigate whether the triangle motif is self-averaging, we need to compare the variance to the second moment of the number of triangles, which results in (cid:40) (cid:40) (cid:40) E [# bow ties] = 2 (3−τ )) Θ(n 5 Θ(n4−τ ) τ < 7 3 , τ ≥ 7 3 , Var ((cid:52)) = 2 (3−τ )) Θ(n 5 Θ(n4−τ ) τ < 7 3 , τ ≥ 7 3 . Var ((cid:52)) E [(cid:52)]2 = 2 (τ−3)), Θ(n 1 Θ(n2τ−5), τ < 7 3 , τ ≥ 7 3 . (cid:40) (14) (15) (16) (17) Therefore, lim n→∞ Var ((cid:52)) E [(cid:52)]2 = τ < 5 0 2 , ∞ τ > 5 2 . For τ = 5/2, the limit in (17) is of constant order of magnitude. Thus, the number of triangles is self-averaging as long as τ < 5 2 the number of triangles is not self-averaging. General motif fluctuations. We now compute the variance of general motifs, similar to the triangle example. Let i = (i1, . . . , ik) be such that ip (cid:54)= iq when p (cid:54)= q. We can then write the variance as Var (N (H)) = P (Hi, Hj present) − P (Hi present) P (Hj present) . (18) (cid:17) 2 . When τ ≥ 5 (cid:16) (cid:88) (cid:88) i∈[n]k j∈[n]k The sum splits into several cases, depending on the overlap of i and j. The term where i and j do not overlap equals zero, since edges between vertices that do not overlap are independent. Now suppose i and j overlap at it1, . . . itr and js1 , . . . , jsr for some r > 0. Then P (Hi, Hj present) is equal to the probability that motif H is present on vertices i1, . . . , ik, j1, . . . , jk \ js1, . . . , jsr , where H denotes the motif that is constructed by merging two copies of H at it1 with js1 , at it2 with js2 and so on. Thus, this term can be written as (cid:88)(cid:48) (cid:17) as the connection probabilities, P (Hi present) = E [N (H)] /(cid:0)n (cid:16) Ht1,...,t2k−r present (cid:88)(cid:48) t1,...,t2k−r where = E P N ( H) (cid:104) (cid:105) (cid:1). Thus, (cid:88)(cid:48) P (Ht1,...,tk present) P(cid:0)Htk−r,...,t2k−r present(cid:1) = n−rE [N (H)]2 O(1). k denotes a sum over distinct indices. Furthermore, since the degrees are i.i.d. as well , (19) (20) t1,...,t2k−r Let H1, . . . , Hl denote all motifs that can be constructed by merging two copies of H at at least one vertex. We can then write the variance of the motif count as (see [40, 41, 42, 37]) Var (N (H)) = C1E [N (H1)] + ··· + ClE [N (Hl)] + E [N (H)]2 O(n−1). (21) where Ci is a combinatorial constant that denotes the number of distinct ways to merge two copies of H into Hi. These constants satisfy [40] Ci = (k − s)!. (22) s l(cid:88) (cid:18)k (cid:19)2 k−1(cid:88) i=1 s=0 10 References [1] R´eka Albert, Hawoong Jeong, and Albert-L´aszl´o Barab´asi. Internet: Diameter of the world- wide web. Nature, 401(6749):130 -- 131, 1999. [2] Michalis Faloutsos, Petros Faloutsos, and Christos Faloutsos. On power-law relationships of the internet topology. In ACM SIGCOMM Computer Communication Review, volume 29, pages 251 -- 262. ACM, 1999. [3] Hawoong Jeong, B´alint Tombor, R´eka Albert, Zoltan N Oltvai, and A-L Barab´asi. The large-scale organization of metabolic networks. Nature, 407(6804):651 -- 654, 2000. [4] Alexei V´azquez, Romualdo Pastor-Satorras, and Alessandro Vespignani. Large-scale topolog- ical and dynamical properties of the internet. Phys. Rev. E, 65:066130, 2002. [5] Aaron Clauset, Cosma Rohilla Shalizi, and M. E. J. Newman. Power-law distributions in empirical data. SIAM Rev., 51(4):661 -- 703, 2009. [6] N. Kashtan, S. Itzkovitz, R. Milo, and U. Alon. Efficient sampling algorithm for estimating subgraph concentrations and detecting network motifs. Bioinformatics, 20(11):1746 -- 1758, 2004. [7] Joshua A. Grochow and Manolis Kellis. Network motif discovery using subgraph enumeration and symmetry-breaking. In In RECOMB, pages 92 -- 106, 2007. [8] Saeed Omidi, Falk Schreiber, and Ali Masoudi-Nejad. MODA: An efficient algorithm for network motif discovery in biological networks. Genes & Genetic Systems, 84(5):385 -- 395, 2009. [9] F. Schreiber and H. Schwobbermeyer. MAVisto: a tool for the exploration of network motifs. Bioinformatics, 21(17):3572 -- 3574, 2005. [10] Sebastian Wernicke and Florian Rasche. Fanmod: a tool for fast network motif detection. Bioinformatics, 22(9):1152 -- 1153, 2006. [11] Thomas House, Geoffrey Davies, Leon Danon, and Matt J. Keeling. A motif-based approach to network epidemics. Bull. Math. Biol., 71(7):1693 -- 1706, 2009. [12] Pierre-Andr´e Noel, Antoine Allard, Laurent H´ebert-Dufresne, Vincent Marceau, and Louis J. Dub´e. Spreading dynamics on complex networks: a general stochastic approach. J. Math. Biol., 69(6-7):1627 -- 1660, 2013. [13] June Zhang and Jose M.F. Moura. Subgraph density and epidemics over networks. In 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2014. [14] Martin Ritchie, Luc Berthouze, and Istvan Z. Kiss. Beyond clustering: mean-field dynamics on networks with arbitrary subgraph composition. J. Math. Biol., 72(1-2):255 -- 281, 2015. [15] Martin Ritchie, Luc Berthouze, and Istvan Z. Kiss. Generation and analysis of networks with a prescribed degree sequence and subgraph family: higher-order structure matters. J. Complex Networks, page cnw011, 2016. [16] Clara Stegehuis, Remco van der Hofstad, and Johan S. H. van Leeuwaarden. Epidemic spreading on complex networks with community structures. Sci. Rep., 6:29748, 2016. [17] Erzs´ebet Ravasz and Albert-L´aszl´o Barab´asi. Hierarchical organization in complex networks. Phys. Rev. E, 67:026112, 2003. 11 [18] Austin R Benson, David F Gleich, and Jure Leskovec. Higher-order organization of complex networks. Science, 353(6295):163 -- 166, 2016. [19] Charalampos E. Tsourakakis, Jakub Pachocki, and Michael Mitzenmacher. Scalable motif- aware graph clustering. In Proceedings of the 26th International Conference on World Wide Web, WWW '17, pages 1451 -- 1460, 2017. [20] Ron Milo, Shai Shen-Orr, Shalev Itzkovitz, Nadav Kashtan, Dmitri Chklovskii, and Uri Alon. Network motifs: Simple building blocks of complex networks. Science, 298(5594):824 -- 827, 2002. [21] Stephan Wuchty, Zolt´an N Oltvai, and Albert-L´aszl´o Barab´asi. Evolutionary conservation of motif constituents in the yeast protein interaction network. Nat. Genet., 35(2):176 -- 179, 2003. [22] Ron Milo, Shalev Itzkovitz, Nadav Kashtan, Reuven Levitt, Shai Shen-Orr, Inbal Ayzenshtat, Michal Sheffer, and Uri Alon. Superfamilies of evolved and designed networks. Science, 303(5663):1538 -- 1542, 2004. [23] Jukka-Pekka Onnela, Jari Saramaki, J´anos Kert´esz, and Kimmo Kaski. Intensity and coher- ence of motifs in weighted complex networks. Phys. Rev. E, 71:065103, 2005. [24] C. Gao and J. Lafferty. Testing network structure using relations between small subgraph probabilities. arXiv:1704.06742, 2017. [25] K.-I. Goh, B. Kahng, and D. Kim. Universal behavior of load distribution in scale-free networks. Physical Review Letters, 87(27), dec 2001. [26] Mari´an Bogun´a and Romualdo Pastor-Satorras. Class of correlated random networks with hidden variables. Phys. Rev. E, 68:036112, 2003. [27] Juyong Park and M. E. J. Newman. Statistical mechanics of networks. Phys. Rev. E, 70:066117, 2004. [28] Ilkka Norros and Hannu Reittu. On a conditionally poissonian graph process. Adv. Appl. Probab., 38(01):59 -- 75, 2006. [29] F. Chung and L. Lu. The average distances in random graphs with given expected degrees. Proc. Natl. Acad. Sci. USA, 99(25):15879 -- 15882, 2002. [30] Bo Soderberg. General formalism for inhomogeneous random graphs. Physical Review E, 66(6), dec 2002. [31] G. Caldarelli, A. Capocci, P. De Los Rios, and M. A. Munoz. Scale-free networks from varying vertex intrinsic fitness. Physical Review Letters, 89(25), dec 2002. [32] Clara Stegehuis, Remco van der Hofstad, Johan S. H. van Leeuwaarden, and A. J. E. M Janssen. Clustering spectrum of scale-free networks. Phys. Rev. E, 96(4):042309, 2017. [33] Shai Wiseman and Eytan Domany. Lack of self-averaging in critical disordered systems. Phys. Rev. E, 52(4):3469 -- 3484, 1995. [34] Shai Wiseman and Eytan Domany. Finite-size scaling and lack of self-averaging in critical disordered systems. Phys. Rev. Lett., 81(1):22 -- 25, 1998. [35] Amnon Aharony and A. Brooks Harris. Absence of self-averaging and universal fluctuations in random systems near critical points. Phys. Rev. Lett., 77(18):3700 -- 3703, 1996. [36] Moupriya Das and Jason R. Green. Self-averaging fluctuations in the chaoticity of simple fluids. Phys. Rev. Lett., 119(11), 2017. 12 [37] Massimo Ostilli. Fluctuation analysis in complex networks modeled by hidden-variable mod- els: Necessity of a large cutoff in hidden-variable models. Phys. Rev. E, 89:022807, 2014. [38] L. A. Pastur and M. V. Shcherbina. Absence of self-averaging of the order parameter in the sherrington-kirkpatrick model. J. Statist. Phys., 62(1-2):1 -- 19, 1991. [39] F. Sylos Labini, N. L. Vasilyev, L. Pietronero, and Y. V. Baryshev. Absence of self- averaging and of homogeneity in the large-scale galaxy distribution. EPL (Europhysics Let- ters), 86(4):49001, 2009. [40] Ove Frank. Moment properties of subgraph counts in stochastic graphs. Ann. New York Acad. Sci., 319(1):207 -- 218, 1979. [41] Franck Picard, J-J Daudin, Michel Koskas, Sophie Schbath, and Stephane Robin. Assessing the exceptionality of network motifs. J. Comput. Biol., 15(1):1 -- 20, 2008. [42] Catherine Matias, Sophie Schbath, Etienne Birmel´e, Jean-Jacques Daudin, and St´ephane Robin. Network motifs: mean and variance for the count. Revstat, 4(1):31 -- 51, 2006. [43] Remco van der Hofstad, Johan S. H. van Leeuwaarden, and Clara Stegehuis. Optimal sub- graph structures in scale-free configuration models. arXiv:1709.03466, 2017. [44] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data, 2014. Date of access: 14/03/2017. [45] Bryan Klimt and Yiming Yang. Introducing the Enron Corpus. In CEAS, 2004. [46] Mari´an Bogun´a, Romualdo Pastor-Satorras, Albert D´ıaz-Guilera, and Alex Arenas. Models of social networks based on social distance attachment. Phys. Rev. E, 70(5), 2004. [47] Tobias Friedrich and Anton Krohmer. Cliques in hyperbolic random graphs. In INFOCOM proceedings 2015, pages 1544 -- 1552. IEEE, 2015. [48] Remco van der Hofstad and Nelly Litvak. Degree-degree dependencies in random graphs with heavy-tailed degrees. Internet Mathematics, 10(3-4):287 -- 334, 2014. [49] Clara Stegehuis. Dominant motif structures, code for solving optimization problem (6), Jan- uary 2019. 13 Supplementary Note 1 Additional motif structures Figure 1 presents the typical dominant motif structures for all connected motifs on 5 vertices. 5 2 (3−τ ) (a) n 5 2 (3−τ ) (b) n 5 2 (3−τ ) (c) n 5 2 (3−τ ) (d) n 13 2 −2τ (e) n 5 2 (3−τ ) (f) n 5 2 (3−τ ) (g) n (h) n9−3τ (i) on τ depends 13 2 −2τ (j) n (k) n6−2τ− 1 τ−1 5 2 (3−τ ) (l) n (m) n9−3τ (n) depends on τ 5 2 (3−τ ) (o) n (p) n7−2τ (q) n3−τ + 2 τ−1 (r) n6−2τ + 1 τ−1 11−3τ 2 (s) n (t) n4−τ + 1 τ−1 (u) n 4 τ−1 1 n(τ−2)/(τ−1) √n n1/(τ−1) non-unique Supplementary Figure 1: Typical magnitude of the number of motifs on 5 vertices. The vertex color indicates the dominant vertex degree. Supplementary Note 2 Proof of Theorem 1 We now investigate the relation between the expected number of motifs and the optimization problem Bf (H) = maxP S1 − S2 − 2ES1 + ES1,S3 τ − 1 . (23) Let N (H, α, ε) denote the number of times motif H occurs on vertices with degrees [ε, 1/ε](nαi)i∈[k] and let α∗ be defined as (cid:21) (cid:20) 0 1 1 2 α∗ i = if i ∈ S1, if i ∈ S2, if i ∈ S3. Then, the following theorem provides a more detailed version of Theorem 1: Theorem 4 (General motifs, expectation). Let H be a motif on k vertices such that the solution to (23) is unique. Then, for any α (cid:54)= α∗ and 0 < ε < 1, E [N (H, α, ε)] E [N (H, α∗, ε)] → 0. Furthermore, E [N (H, α∗, ε)] n 3−τ 2 k+ τ−2 2 B(H) = f (ε)Θ(1) for some function f (ε) not depending on n. 14 (24) (25) (26) We now prove Theorem 4. In the hidden-variable model, the connection probability of vertices with hidden variables hi and hj equals p(hi, hj) = min(hihj/(µn), 1). (27) Assume that hi = Θ(nαi) for some αi ≥ 0 for all i. Then, the probability that motif H exists on vertices (vi)i∈[k] with hidden variables h = (nαi )i∈[k] satisfies (cid:1) = Θ P(cid:0)H present on (v1,··· , vk) h = (nα1)i∈[k] (cid:90) 1/ε(µn)α P (hi ∈ [ε, 1/ε](µn)α) = ε(µn)α  (cid:89) (vi,vj )∈EH :αi+αj <1  . nαi+αj−1 cx−τ dx = K(ε)(µn)α(1−τ ) The hidden variables are an i.i.d. sample from a power-law distribution. Therefore, for some constant K(ε) not depending on n. The expected number of vertices with degrees in [ε, 1/ε](µn)α scales as Θ(n(1−τ )α+1). Then, the number of sets of vertices with degrees in [ε, 1/ε](nαi)i∈[k] scales as Combining (28) and (30) yields that the contribution from vertices with degrees nα = (nαi)i∈[k] to the expected number of motifs, E [N (H, α), ε] is (28) (29) (30) (31) (32) Θ . i αi (cid:17) (cid:16) nk+(1−τ )(cid:80) nk+(1−τ )(cid:80) i αi (cid:89) (cid:88) (i,j)∈EH :αi+αj <1 (cid:88) i  . E [N (H, α, ε)] = Θ nα1+αj−1 (i,j)∈EH :αi+αj <1 The maximum contribution is obtained for αi that maximize max(1 − τ ) αi + αi + αj − 1 for αi ≥ 0. The following lemma shows that this optimization problem attains its maximum for highly specific values of α: Lemma 5 (Maximum contribution to expected number of motifs). Let H be a connected graph on k vertices. If the solution to (32) is unique, then the optimal solution satisfies αi ∈ {0, 1 2 , 1} for all i. If it is not unique, then there exist at least 2 optimal solutions with αi ∈ {0, 1 2 , 1} for all i. Proof. Defining βi = αi − 1 2 yields for (32) 1 − τ 2 k + (1 − τ ) (cid:88) i max (cid:88) βi + βi + βj, (33) (i,j)∈EH :βi+βj <0 2 . Then, we have to prove that βi ∈ {− 1 over all possible values of βi ≥ − 1 2} for all i in the optimal solution. Note that (33) is a piecewise linear function in β. Therefore, if (33) has a unique maximum, it must be attained at the boundary for βi or at a border of one of the linear sections. Thus, any unique optimal value of βi satisfies βi = − 1 2 or βi + βj = 0 for some j. Note that this implies that βi ≤ 1 2 in (33), since it does not influence the optimal β values. Rewriting (33) without the constant factor yields 2 . We ignore the constant factor of (1 − τ ) k 2 , 0, 1 (cid:88) i max βi (1 − τ + #{ edges to j with βj < −βi}) . (34) 15 The proof of the lemma then consists of two steps: Step 1. Show that any unique solution does not have vertices i with βi ∈ (0, 1 2 ). Step 2. Show that any optimal solution that is not unique can be transformed into two different optimal solutions with βi ∈ {− 1 2} for all i. 2 , 0, 1 Let Step 1. We show that when the solution to (34) is unique, it is never optimal to have β ∈ (0, 1 2 ). (35) β = min i:βi>0βi . 2 , then we are finished, so assume that β < 1 If β = 1 2 . Then, there exist N β− vertices with their β value equal to − β, and N β+ vertices with value β, where N β+ + N β− ≥ 1. Furthermore, let E β− denote the number of edges from vertices with value − β to other vertices j such that βj < β, and E β+ the number of edges from vertices with value β to other vertices j such that βj < − β. Then, the contribution from these vertices to (34) is (cid:16) β (cid:16) (cid:17) (1 − τ ) N β+ − N β− + E β+ − E β− . (36) (cid:17) (cid:1) Because the optimizer of (34) is unique, the term inside the brackets cannot equal zero. Then, increasing β if the term inside brackets is larger than zero, or decreasing it if it is smaller than zero would improve the optimal contribution, which is not possible. Thus, if the optimal solution is unique, we must have β = 1 2} for all i. Step 2. If the solution to (34) is not unique, then by the same argument that leads to (36), 2 . This shows that any unique solution satisfies βi ∈ {− 1 (cid:16) (cid:17) there exist β1, . . . , βs > 0 for some s ≥ 1 such that 2 , 0, 1 (cid:16) (cid:17) βj (1 − τ ) N β+ j − N β − j + E β+ j − E β − j = 0 ∀j ∈ [s]. (37) Here we use the same notation as in (36). All other values of β must either be 0, 1 2 by the argument in Step 3. Thus, setting all βj to zero does not change the value of the solution, and setting all βj to 1 2 also does not change the value of the solution. Thus, if the solution to (34) is not unique, at least 2 solutions exist with βi ∈ {− 1 Completion of the proof of Theorem 4. We first rewrite (32) using Lemma 5. By Lemma 5, the maximal value of (32) is attained by partitioning VH into the sets S1, S2, S3 such that vertices in S1 have αi = 0, vertices in S2 have αi = 1 and vertices in S3 have αi = 1 2 . Then, the edges with αi + αj < 1 are edges inside S1 and edges between S1 and S3. If we denote the number of edges inside S1 by ES1 and the number of edges between S1 and S3 by ES1,S3, then we can rewrite (32) as 2} for all i. 2 or − 1 2 , 0, 1 2 S3 − ES1 − 1 2 ES1,S3 (38) maxP (1 − τ )(cid:0) over all partitions P of the vertices of H into S1, S2, S3. Using that S3 = k − S1 − S2 yields maxP 1 − τ 2 k + τ − 1 2 S1 − S2 − 2ES1 + ES1,S3 τ − 1 , (39) Since k is fixed and τ − 1 > 0, maximizing (32) is equivalent to maximizing S2 + 1 (cid:18) (cid:20) Bf (H) = maxP S1 − S2 − 2ES1 + ES1,S3 τ − 1 . (40) Furthermore, by Lemma 5, the optimal value of (40) is unique if and only if the solution to (32) is unique. 16 (cid:19) (cid:21) Let α∗ be the unique optimizer of (32). Then, by (31), for any α (cid:54)= α∗ = Θ(cid:0)n−η(cid:1) E [N (H, α, ε)] E [N (H, α∗, ε)] (41) for some η > 0. Combining this with (40) proves the first part of the theorem. By (31), the contribution of the maximum is then given by 1−τ 2 (k+Bf (H)) = n E [N (H, α∗, ε)] = nkn 3−τ 2 k+ τ−1 2 Bf (H), (42) which proves the second part of the theorem. Supplementary Note 3 Typical motif counts We now present equivalent versions of Lemma 5 and Theorem 4 for typical motif counts: Lemma 6 (Maximum contribution to typical number of motifs). Let H be a connected graph on k vertices. If the solution to (32) with the extra constraint αi ∈ [0, 1/(τ − 1)] for all i is unique, then the optimal solution satisfies αi ∈ {0, (τ − 2)/(τ − 1), 1 2 , 1/(τ − 1)} for all i. Furthermore, αi = 0 if and only if the degree of vertex i in H equals 1. If the solution is not unique, then there exist at least 2 optimal solutions with αi ∈ {0, (τ − 2)/(τ − 1), 1 2 , 1/(τ − 1)} for all i. Thus, the optimization problem can be translated into the optimal partition P of all vertices of H with degree at least 2 into three sets, S1, S2 and S3. The vertices in S1 correspond to vertices that have degree proportional to n(τ−2)/(τ−1) in the graph, vertices in S2 correspond to the maximal degree vertices with degrees proportional to n1/(τ−1), and S3 corresponds to the vertices of √n degrees. Let ESi,1 denote the number of edges between vertices in Si and vertices with degree 1 in H. Then, rewriting the optimal solution of (32) results in (cid:18) τ − 2 τ − 1 S1 + ES2,1 − 1 2 ES3,1, maxP (1 − τ ) τ − 2 τ − 1 − (cid:19) τ − 3 τ − 1 1 τ − 1 S2 + 1 2 S3 ES1 + τ − 3 2(τ − 1) ES1,S3 − ES1,1 τ − 1 (43) (cid:16) over all partitions P of the vertices of H with degree at least 2 into S1, S2, S3. Using that S3 = k − S1 − S2 − k1, ES3,1 = k1 − ES1,1 − ES2,1, where k1 denotes the number of degree 1 vertices in H, and extracting a factor (3 − τ )/2 results in maxP 1 − τ 2 k + (3 − τ ) S1 − S2 + τ − 1 Since k and k1 are fixed and 3 − τ > 0, we need to maximize 2 τ − 2 3 − τ k1 − 2ES1 + ES1,S3 − ES1,1 − ES2,1 τ − 1 (cid:17) , (44) Bt(H) = maxP S1 − S2 − 2ES1 + ES1,S3 + ES1,1 − ES2,1 . (45) τ − 1 Theorem 7 (General motifs, typical). Let H be a motif on k vertices such that the solution to (45) is unique and let k1 denote the number of vertices of degree 1 in H. With high probability, the number of motifs H in a hidden-variable model grows asymptotically as N (H) ∝ n 2 (k+Bt(H))+ τ−1 3−τ ∗ that optimizes (45). 2 k1, (46) and is thus fully determined by the partition P Proof. The proof of Theorem 7 as well as Lemma 6 follows the same lines as the proof of [43, The- orem 2.1 and Lemma 3.2] for the erased configuration model instead of the hidden-variable model. That proof relies on the fact that the connection probability between vertices with degrees Di and Dj in the erased configuration model has the same order of magnitude as min(DiDj/(µn), 1), which is the same as the connection probability in the hidden-variable model. Thus, the proof of [43, Theorem 2.1] also holds for the hidden-variable model. 17 Supplementary Note 4 Graphlets We now focus on graphlet counting. Thus, we now only count the number of times the subgraph H appears as an induced subgraph of a hidden-variable model. In particular, edges that are not present in H are also required not to be present in the graph. Because the probability that two edges between vertices of high degree are present equals one (see (27)), this puts a constraint on the number of vertices that typically have high degree. When we again let the degrees in the graphlet scale as nαi, we see that the probability that an edge (i, j) is present equals one as soon as α1 + αj > 1. Thus, for the expected number of graphlets, the optimization problem corresponding to (32) becomes (cid:88) max(1 − τ ) αi + (i,j)∈EH :αi+αj <1 αi + αj − 1 (47) where EH denotes the edge set of H. Again, this optimization problem is maximized for αi ∈ 0, 1 2 , 1, so that similarly, the optimization problem corresponding to (40) including the extra con- straint then becomes s.t. αi + αj ≤ 1 ∀(i, j) /∈ EH , (cid:88) i (cid:20) (cid:21) , Bg,f (H) = maxP S1 − S2 − 2ES1 + ES1,S3 τ − 1 s.t. (u, v) ∈ EH ∀u ∈ S2, v ∈ S2 ∪ S3. (48) This optimization problem again finds the most likely degrees of vertices that together form the graphlet H. Vertices in S1 have degrees proportional to a constant, vertices in S2 have degrees proportional to n and vertices in S3 have degrees proportional to √n. Using this optimization problem, we are also able to find the scaling of the number of graphlets: Theorem 8 (General graphlets, expected). Let H be a graphlet on k vertices such that the solution to (48) is unique. The expected number of graphlets Ng(H) in a hidden-variable model grows asymptotically as E [Ng(H)] ∝ n and is thus fully determined by the partition P 3−τ 2 (k+Bg,f (H)), ∗ that optimizes (48). (49) For the typical number of graphlets, we obtain a similar result, by adding the constraint τ−1 to (47). This results in the optimization problem (like its equivalent version for motifs αi ≤ 1 in (45)) Bg,t(H) = maxP S1 − S2 − τ − 1 s.t. (u, v) ∈ EH ∀u ∈ S2, v ∈ S2 ∪ S3. 2ES1 + ES1,S3 + ES1,1 − ES2,1 , (50) (cid:104) (cid:105) The corresponding theorem to Theorem 7 for graphlets then becomes: Theorem 9 (General motifs, typical). Let H be a graphlet on k vertices such that the solution to (50) is unique and let k1 denote the number of vertices of degree 1 in H. With high probability, the number of graphlets Ng(H) in a hidden-variable model grows asymptotically as and is thus fully determined by the partition P ∗ that optimizes (50). Ng(H) ∝ n 3−τ 2 (k+Bg,t(H))+ τ−1 2 k1, (51) Figure 3 shows the typical configuration of all graphlets on 5 vertices, similar to Fig. 1 for all motifs on 5 vertices. Note that the typical configuration of a motif and a graphlet may be substantially different (for example in Fig. 2(b) and its counterpart for motifs). 18 t1 t3 t2 t4 t4 t7 t6 t9 t8 t11 t1 t1 t1 t2 t3 (a) K4 n6−2τ (b) diamond n6−2τ log(n) t4 t4 (c) square n6−2τ 1 n(τ−2)/(τ−1) √n t11 t10 t5 t6 t8 t8 t10 (d) paw n7−2τ− 1 τ−1 n1/(τ−1) (e) claw 3 τ−1 n non-unique (f) path n4−τ log(n) Supplementary Figure 2: Scaling of the typical number of graphlets on 4 vertices in n. The vertex color indicates the typical vertex degree. The vertex labels indicate the vertex types used in Figure 6 5 2 (3−τ ) (a) n 5 2 (3−τ ) (b) n 5 2 (3−τ ) (c) n 5 2 (3−τ ) (d) n 13 2 −2τ (e) n 5 2 (3−τ ) (f) n 5 2 (3−τ ) (g) n (h) n9−3τ (i) on τ depends 13 2 −2τ (j) n (k) n6−2τ− 1 τ−1 5 2 (3−τ ) (l) n (m) n 5 2 (3−τ ) (n) depends on τ 5 2 (3−τ ) (o) n (p) n7−2τ (q) n3−τ + 2 τ−1 13 2 −2τ (r) n 11−3τ 2 (s) n (t) n4−τ + 1 τ−1 (u) n 4 τ−1 Supplementary Figure 3: Typical number of graphlets for all connected graphs on 5 vertices (constants ignored). The vertex color indicates the dominant vertex degree as in Figure 1. Supplementary Note 5 Proof of Theorem 2 Let H be a motif on k vertices such that the dominant motif structure of Theorem 1 contains non-√n vertices. A similar analysis as the triangle example shows that the only motif on 2 vertices is non-self-averaging. Thus, we may assume that k ≥ 3. Let the expected number of motifs of type H as predicted in Theorem 1 be denoted by nf (H), and the optimal contribution from vertices 1, . . . , k in H by (nαi)i∈[k]. By Theorem 1, E [N (H)] ∝ E(cid:2)# vertices having weights (nαi)i∈[k] (cid:3) P(cid:0)H present on weights (nαi )i∈[k] (cid:1) (52) Suppose there is a vertex v1 in H with an optimal contribution of weight n vertices, that is α1 = 1. Then, we study the contribution to the variance in (21) from the motif H which is the motif on 2k−1 vertices where two copies of motif H are merged at their vertex v1 (see Fig. 4 for an example). We now investigate the expected number of H motifs. In particular, we study the contribution to the expected number of H motifs on vertices with weights of the order of magnitude (nβi)i∈[2k−1] with βi = αt(i), 19 (53) v1 H H Supplementary Figure 4: Construction of H when H is a path of length 3. H is constructed by merging two copies of H at vertex v1 with most likely degree n. where t(i) is the vertex in H corresponding to vertex i in H. Figure 4 gives an example of this contribution when H is a path with 3 vertices. Because of the construction of H, (cid:16) H present on weights (nβi)i∈[2k−1] (cid:17) P = P(cid:0)H present on weights (nαi )i∈[k] (cid:1)2 . (54) Furthermore, the only difference between the vertices in H and two separate versions of H, is that H contains one less vertex of weight proportional to n. The expected number of vertices with weight proportional to n is given by n2−τ Therefore, E(cid:2)# vertices on weights (nβi)i∈[2k−1] (cid:3) E[N ( H)] ≥ P(cid:0)H present on weights (nαi)i∈[k] ∝ (cid:3)2 E(cid:2)# vertices on weights (nαi)i∈[k] (cid:1)2 E(cid:2)# vertices on weights (nαi)i∈[k] n2−τ . n2−τ Thus, ∝ E [N (H)]2 nτ−2. Combining this with (21) of the main text results in Var (N (H)) E [N (H)]2 ≥ E[N ( H)] E [N (H)]2 ≥ nτ−2, (cid:3)2 (55) (56) (57) (59) (60) which diverges, because τ ∈ (2, 3). Thus, if the optimal contribution to H satisfies S2 (cid:54)= ∅, H cannot be self-averaging for τ ∈ (2, 3). Now we study the case where H has optimal contribution with S1 (cid:54)= ∅, but no vertices of weights proportional to n so that S2 = ∅. Let v ∈ S1. The contribution from v to (23) is 1 − 2dv,S1 + dv,S3 τ − 1 , (58) where dv,Si denotes the number of edges from v to vertices in set Si. Moving v to S3 would change the contribution to Because we know that v in S1 is the optimal contribution, − dv,S1 τ − 1 . − dv,S1 τ − 1 < 1 − 2dv,S1 + dv,S3 τ − 1 , or dv,S1 + dv,S3 ≤ τ − 1, so that dv,S1 + dv,S3 ∈ {0, 1}. Thus, every vertex in S1 has at most 1 edge to other vertices in S1 or vertices in S3. Since we have assumed that S2 = ∅, and k > 2, this means that all vertices in S1 have degree 1 inside the motif, and are connected to a vertex in S3. W.l.o.g. assume that v1 is a vertex such that v1 ∈ S3 and v1 has at least one connection to a 20 v1 H H Supplementary Figure 5: Example of the construction of H. H is constructed by merging two copies of H at vertex v1: a vertex with most likely degree √n that is connected to a vertex of degree 1. vertex in S1. As in the previous proof, we consider H constructed by merging two copies of H at v1, as illustrated in Figure 5. Define (αi)i∈[k] as the maximal contribution to H. We define (cid:40) βi = αt(i) 1 t(i) (cid:54)= 1, t(i) = 1. (61) (cid:105) (cid:104) N ( H) That is, we study the contribution where all vertices in H except v1 have the same weight as in their counterpart in H. The weight of v1 is proportional to n, whereas the counterpart of v1 in H had weight proportional to √n, as in the illustration in Fig. 5. We again study the contribution to from vertices of weights (nβi)i∈[2k−1]. We now compare the probability that H exists on E vertices of weights (nβi)i∈[2k−1] to the probability that two copies of H exist on weights (nαi)i∈[k]. The difference between these two probabilities is that vertex v1 in H has weight n instead of √n in H. In H, v1 is connected to at least one vertex of weight proportional to 1. The probability of this connection to be present is proportional to n−1/2. In H, v1 has weight n, so that the probability that the corresponding connections occur in H is proportional to 1. The connection probabilities on vertices not connected to v1 do not change, so that P(cid:0)H present on weights (nαi)i∈[k] (cid:1)2 . (62) (cid:16) H present on weights (nβi)i∈[2k−1] (cid:17) P ≥ Since the difference between the vertices of two copies of H and H is that we remove two vertices of weight √n and add one vertex of weight proportional to n, we obtain E(cid:2)# vertices on weights (nαi)i∈[k] (cid:3)2 n2−τ , (63) where we have used that the number of vertices with degree proportional to √n is n(3−τ )/2 by (30) and the number of vertices with degree proportional to n scales as n2−τ . This results in (cid:3) E(cid:2)# vertices on weights (nβi)i∈[2k−1] E[N ( H)] ≥ P(cid:0)H present on degrees (nαi)i∈[k] ∝ ∝ E [N (H)]2 (cid:1)2 ≥ E(cid:2)# vertices on degrees (nαi)i∈[k] (cid:3)2 so that by (21) which does not converge to zero, so that the motif is not self-averaging. Var (N (H)) E [N (H)]2 ≥ E[N ( H)] E [N (H)]2 ∝ 1, n−1 n3−τ (64) (65) Graphlet fluctuations The fluctuations of the number of graphlets can be studied similarly as the motif fluctuations. Again, (21) holds, but now it only includes graphlets H1, . . . , Hl that can be constructed by merging two copies of H at one or more vertices, with the additional constraint that after merging 21 the copies of H, both copies still form induced subgraphs of H. As an example, consider 2-path fluctuations. The subgraphs that can be constructed from merging two 2-paths are the subgraphs in Fig. 3s, 3t, 3u, 2c, 2d, 2e, 2f, Fig. 3a and 3b. However, the merged subgraph of Fig. 2d and Fig. 2b does not contain two induced copies of graphlets, and therefore these subgraphs are excluded from equation (21). Other than that, the procedure to determine for any graphlet whether it is self-averaging or not is the same as the procedure for motifs, using Theorem 8 to find the expected order of magnitude of the number of merged graphlets. Supplementary Note 6 Optimal diamond motif We now study the optimizers of (6) for the diamond motif (Fig. 3b) for all τ ∈ (2, 3). Let i and j be the vertices at the diagonal of the diamond, and k and s the corner vertices. By [49], (6) is maximized by S3 = {i, j, k, s} or S1 = {k, s} and S2 = {i, j} and by Lemma 5 these are two optimizers of (6), yielding a contribution of 2(1 − τ ). By the argument in Lemma 5, for any vertex u, the optimizer of (6) is either given by αu = 0 or constrained by αu + αv = 1 for at least one edge (u, v) ∈ EH . The above two optimizers show that for all vertices in the diamond motif, there is at least one such constrained edge. Therefore, αi + αj + αk + αs = 2 must hold. Then, to find all optimizers, we solve (1 − τ )(αi + αj + αk + αs) + 1{αi+αj <1} (αi + αj − 1) + 1{αi+αk<1} (αi + αk − 1) + 1{αi+αs<1} (αi + αs − 1) + 1{αj +αs<1} (αj + αs − 1) + 1{αj +αk<1} (αj + αk − 1) = 2(1 − τ ). Using that αi + αj + αk + αs = 2 this reduces to αi + αj + αk + αs = 2, αi + αj ≥ 1, αi + αk ≥ 1, αi + αs ≥ 1, αj + αs ≥ 1, αj + αk ≥ 1. (66) (67) The constraints αi + αk ≥ 1, αj + αs ≥ 1 together with αi + αj + αk + αs = 2 yield αi + αk = 1 and αj +αs = 1. Similarly, αi +αs = 1 and αk +αj = 1, resulting in αi = αj =: β and αk = αs = 1−β. Furthermore, the constraint αi + αj ≥ 1 yields β ∈ [1/2, 1]. Supplementary Note 7 Graphlet counts Gowalla Oregon 3.13·1014 7.03·107 2.01·107 9.61·105 7.46·104 3.05·104 claw 5.96·1014 1.52·1014 path paw 3.11·1014 8.60·107 4.24·107 6.09·106 diamond square K4 Enron 4.48·1014 1.37·1014 3.76·108 2.25·107 6.76·106 2.34·106 PGP 4.04·106 2.72·106 1.96·106 2.74·105 2.16·104 2.39·105 HEP 1.23·106 2.12·106 5.31·105 3.55·104 6.84·103 6.56·104 Supplementary Table 1: Graphlet counts in the five data sets. 22
1507.04024	2	1507	2015-07-16T15:42:00	Analyzing the activity of a person in a chat by combining network analysis and fuzzy logic	[ "cs.SI" ]	Chat-log data that contains information about sender and receiver of the statements sent around in the chat can be readily turned into a directed temporal multi-network representation. In the resulting network, the activity of a chat member can, for example, be operationalized as his degree (number of distinct interaction partners) or his strength (total number of interactions). However, the data itself contains more information that is not readily representable in the network, e.g., the total number of words used by a member or the reaction time to what the members said. As degree and strength, these values can be seen as a way to operationalize the idea of activity of a chat-log member. This paper deals with the question of how the overall activity of a member can be assessed, given multiple and probably opposing criteria by using a fuzzy operator. We then present a new way of visualizing the results and show how to apply it to the network representation of chat-log data. Finally, we discuss how this approach can be used to deal with other conflicting situations, like the different rankings produced by different centrality indices.	cs.SI	cs	Analyzing the activity of a person in a chat by combining network analysis and fuzzy logic Sude Tavassoli Katharina Anna Zweig Graph Theory and Network Analysis Group Graph Theory and Network Analysis Group Department of Computer Science Technical University of Kaiserslautern Gottlieb-Daimler-Str. 48-668 67663 Kaiserslautern, Germany Telephone: +49 631 205 3340 Email: [email protected] Department of Computer Science Technical University of Kaiserslautern Gottlieb-Daimler-Str. 48-670 67663 Kaiserslautern, Germany Telephone: +49 631 205 3346 Email: [email protected] Abstract—Chat-log data that contains information about sender and receiver of the statements sent around in the chat can be readily turned into a directed temporal multi-network representation. In the resulting network, the activity of a chat member can, for example, be operationalized as his degree (number of distinct interaction partners) or his strength (total number of interactions). However, the data itself contains more information that is not readily representable in the network, e.g., the total number of words used by a member or the reaction time to what the members said. As degree and strength, these values can be seen as a way to operationalize the idea of activity of a chat-log member. This paper deals with the question of how the overall activity of a member can be assessed, given multiple and probably opposing criteria by using a fuzzy operator. We then present a new way of visualizing the results and show how to apply it to the network representation of chat-log data. Finally, we discuss how this approach can be used to deal with other conflicting situations, like the different rankings produced by different centrality indices. I. INTRODUCTION In network analysis, there are often multiple ways of mea- suring a structural property: centrality indices are well known for focusing on very different aspects of the prominence of a node’s position in a graph, weighted networks, temporal networks, or multi-networks further add to this amazing set of measures [1, 3, 4, 5]. For example, for weighted networks, one can now either measure the degree, the number of interaction partners, or the strength of a node, i.e., the sum of the weight of incident edges. Since both measures might result in opposing rankings of the nodes, Opsahl et al. argued that there might be situations in which it is necessary to scale between the degree and the strength of a node. They proposed a new formula in which a parameter is used to achieve this scaling idea [7]: degα(i) = deg(i) ×(cid:18) s(i) deg(i)(cid:19)α (1) where deg(i) is the degree of node i, s(i) is its strength, and α is the scaling parameter: if it is set to 0, the centrality of i is determined by deg(i) alone. If α is set to 1, it is determined by s(i) alone. However, Opsahl’s measure cannot be generalized to more than two opposing rankings. Therefore, in this paper, we use a well-known aggregation operator from fuzzy logic, to deal with all situations in network analysis, where some features of nodes in a network can be measured by multiple methods, whose result is conflicting in assigning a rank to the nodes. The concrete data used in this study is a network deduced from chat-logs of an online group psychotherapy [8]. This data contains time stamps for each statement submitted and the sentences themselves which cannot easily be repre- sented and analyzed in a graph; we thus propose to turn this data into features of the nodes and aggregate their values with those of the network analytic measure to enable an automatic, explorative analysis. The following section motivates why it can be necessary to take into account more than only the degree and the strength as features measuring the activity of a person in the concrete example of a chat-log data. II. ANALYZING THE ACTIVITY OF A PERSON USING A FUZZY AGGREGATION OPERATOR Recent studies show that the majority of mental disorders can be partly recovered in online group psychotherapy ses- sions [6]. Thus, constructing a social network deduced from large amounts of digitally archived interactions allows for a much more extensive exploration [8]. Due to the enormous scale of available data of more than 2, 000 sessions of at least 60 minutes, new, explorative measures have to be developed to find interesting patterns that indicate productive sessions and helpful therapists. Since the data also contains more information than just the fact that an interaction happened (i.e., a sentence from user A directed to user B), namely the number of words or the exact timing of when the sen- tence was submitted, measures should be able to use this information as well. The classic way to deal with temporal interaction data is to aggregate it; in the case of chat-logs of online group therapy, a natural aggregation unit is a single session. A network representation can be chosen such that each participant is represented by a node and an interaction is represented by an edge. Of course, in a chat, there are in general multiple interactions between the same persons. This can be represented by multi-edges or by edges weighted out-degree). • a2(i) : the total number of statements the member i sent (out-strength). • a3(i) : the total number of words the member i submitted. • a4(i) : the overall reaction time. Note that the specific characteristic of the features is not our main point. In the experimental result, we show that they are just chosen to present some concrete examples on which the benefit of using a fuzzy operator is demonstrated, as described in the following. A. Ordered Weighted Averaging (OWA) The idea of fuzzy operators is to take a set of values and to aggregate them into a single value. The aggregation operators are generally classified as either conjunctive or disjunctive, depending on whether they combine the values by a logical and or an or operator respectively. A number of studies show that these operators are very practical in Multi-criteria decision making where a solution for a problem needs to satisfy at least one, all, some, or most of the criteria [9, 2, 12]. These linguistic terms help to represent what we expect to get from the corresponding aggregation [12]. For some of the problems, it does not matter which of the criteria are actually satisfied as long as enough of them are satisfied. We will assume, that the degree to which some criteria j is satisfied can be expressed by a real positive number between 0 and 1, and that 1 means full satisfaction and 0 means no satisfaction. If we require all of the criteria to be fully satisfied, this can be stated as that the least satisfaction of any criterion must be 1. If we require that at least one criterion must be fully satisfied, this can be stated as that the largest satisfaction of any criterion must be 1. Thus, the idea of ordered weighted aggregation operator was to order the degree of satisfaction of all criteria and multiply it with a weight vector. Let A be the set of satisfaction values for all criteria, let A′ be descending version of it, and W = [w1, w2, ..., wm] (where m = A) be a weight vector with Pi wi = 1. Then the weight vector [1, 0, 0, ...0] returns the maximal satisfaction value. If the weight vector is [0, 0, ..., 1], the minimal satisfaction value is returned. The maximum operator is equated with the crisp existential quan- tifier (“there exists one criterion that is satisfied”) and with the OR operator; analogously, the minimum operator is equated with the crisp universal quantifier (“all criteria are satisfied”) and with the AN D operator [9]. Therefore, the operator OW A is defined as a mapping function I m → I (where I = [0, 1]) as follows: OW A(a1, a2, · · · , am) = m Xj=1 wjbj (2) where bj is the jth largest value of am. Of course, now the general idea of ordering the values and weighting can be taken further: for example, setting the weight vector to [1/m, 1/m, ..., 1/m] returns the regular average of the satis- faction. The orness of a weight vector like this is 0.5, the orness of the [1, 0, 0, ..., 0] vector is 1 and the orness of the [0, 0, 0, ..., 1] vector is 0 (andness = 1 − orness as defined Fig. 1: a) a temporal graph whose edges are time-stamped. b) the aggregated form of the graph (a) after ignoring the time sequence of the edges. c) the weighted graph of the graph (b) after summing up the number of edges between a pair of nodes. with the number of interactions (s. Fig. 1). Note that both kinds of representations should be directed as A might direct more sentence to B then vice versa. The weight could also be more fine-grained by summing up, e.g., the number of words contained in all interactions directed from A to B. Based on such a representation, the above mentioned scaling between degree and strength, i.e., the total weight, can be measured using Eq. 1. However, it seems to be reasonable to also explore measurable features of a single person in the chat, i.e., values associated to the node that might influence our concept of centrality in chat groups. In general, the responsibility of the therapist generally demands higher activity compared to the other members [11]. For example, the therapist might need to spend more time in communicating with each other or address more of the other members than a normal member. Sometimes, however, the therapist should talk less than other members and just manage the flow of the chat: because of the informal and basically unstructured nature of a chat, some members strive for dominating it and others isolate themselves. This might lead the therapist to have many interactions with the same persons (low degree, high strength) to either activate or moderate them. Without question, in most cases a reasonable measure should identify the therapist(s) as the most central person. In this paper, we aim at showing how additional features of a node can be analyzed together with network analytic based values by using a fuzzy aggregation operator, to explore the range of possible rankings that scale between the structural position of a node and its features. The two important features used in Eq. 1 are also included in the set of the features associated to the nodes in our method. One (very simple) way to operationalize the activity of a member is to count the number of words a user submits as one feature. Another way is to sum up the reaction time of the other members to his or her statements. We operationalize this idea by measuring the total reaction time, i.e., the total time until the next member starts sending a statement after i submitted a sentence (disregarding the 10 first and last sentences of a session which only contain hello and farewell statements). For illustrating the usage and usefulness of a fuzzy logic operator we thus assign to each node the following four features: • a1(i) : the number of communication partners (in- plus in [10]). Using a weighting function, we can express the con- cept of a wide range of linguistic quantifiers available between two absolute quantifiers of "there exists" and " for all" [12]. In addition, any relative quantifier such as must, few, or many, can be represented using a fuzzy subset where the quantifier specifies to which degree a given proportion of criteria satisfies the concept of the chosen linguistic term [10, 12]. For example, a weight vector W = [w1, w2, ..., wm] can be defined using a function Q(r) = rβ with a parameter β ≥ 0 (an increasing monotone quantifier) in the following way: wi = Q(cid:18) i m(cid:19) − Q(cid:18) i − 1 m (cid:19) (3) It can be easily seen that this vector sums to 1, since all terms of the form Q(cid:0) i m(cid:1) cancel each other for 1 ≤ i ≤ m − 1, and Q(m/m) = Q(1) = 1β = 1 and Q(0/m) = 0. The orness of such a vector is: orness(Qβ) = Z 1 0 rβdr = 1 β + 1 Thus, if β > 1 then orness < 0.5 and if β < 1 then orness > 0.5 [10]. The value of the β varies the property of compensation among (a1, a2, ...am) from full compensation (high-orness) to no compensation (high-andness). This means that a higher degree of satisfaction of one of the features can compensate for a lower degree of satisfaction of another features. Note that it is crucial to make sure that there is a weak order relation on the set of features for the nodes. Otherwise, the features place in the same positions after sorting in Eq. 2 and get the same importance corresponding to the weight vector; this might lead a feature to have a high importance in all aggregations and cause that it matters which of the features is satisfied for measuring the activity of a person. III. EXPERIMENTAL RESULT To show the explorative power of the OWA operator on the different activity features, we used a network including 52 nodes and 29590 multi-edges labeled by time stamps from one of the group-chat sessions. In order to put the features on the same scale, the features are normalized by their respective maximum and minimum values into [0, 1]. We computed the outcome for different values of the orness-parameter β and ranked the nodes by it. Remember that values close to 0 indicate highest orness, β = 1 reproduces the average of the four activity features and β = ∞ indicates an orness of 0 (an andness of 1). Fig. 2 shows the result: the most obvious and expected result is that the therapist is ranked highest for all values of β, i.e., independent of whether the operator favors the highest activity value or the lowest or any mixture of it, the therapist is the most active. The figure also shows that most participants’ rank is quite stable, independent of the orness of the OWA operator. However, three nodes show a very strong difference in their ranking on the extreme scales of β: the first one is P 7. He communicates with only 9 people out of the 52 members of the chat-log. If he was only ranked by the normalized degree, this would be a medium value (rank 28 out of 52) as listed in Table. I. However, his activity with those 9 persons, measured in the number of words typed, is quite high as his activity with respect to the number of statements and the total reaction time. In all of these other three values, he is ranked among the first quarter of ranks. Since he is never on the top rank of any of the measures, a high-orness makes him a medium active person. Decreasing the orness and increasing the andness, however, shows that he has a very stable medium to high rank in all of the criteria. P 17’s case is even more interesting: he has more distinct communication partners, namely 10. This gives him a higher rank in the high-orness (the low-andness) than P 7. However, all his other activity features are extremely close to 0, such that the higher the andness of the operator, the lower his ranking. Looking into the session, it turns out that P 17 is not actually a patient, but a psychologist who visited the chat-log and was a mere spectator. The last case, P 28, has a higher rank than P 7 using only the degree of the nodes for ranking as listed in Table. I. The same situation he has in the high-orness (the low-andness) as depicted in Fig. 2 but once the degree of compensation is decreased by increasing the value of β, his rank is gradually dropped from 33 to 19. Going through the chat-log, it shows that he has communicated with more distinct members than P 7 using very short, say, single word statements consecutively. Thus, considering a high-andness on the set of his activity features decreases his rank very fast as he needs to have most of them or all of them satisfied. We did not know this in advance and randomly pick this session out of the 2, 000 available sessions. IV. CONCLUSION AND FUTURE WORK Network analysis often has to deal with a multitude of information known about a node: in this paper, we concen- trated on some situations in which a node was assigned both structural features classically known from network analysis (degree and strength) together with data not directly present in the network (number of words, total reaction time). While the four features all focus on one possible operationalization of activity, they also create opposing ranks. The exploratory analysis with the OWA operator from fuzzy logic helps to understand how strong the ranks oppose, and thus give more insight then, e.g., a pairwise correlation of the values. Our examples identified some cases whose ranks got extremely changed by applying more features to the analysis of their activity. Looking into the chat-log data set, we found the main reasons which could not be recognized in the ranking using only two features the degree or the strength in such analysis. It is immediately clear that the proposed method can also help to deal with other conflicting information around nodes’ position in the network: (1) Consider as features the degree of a node in various time intervals. E.g., each session could be cut into intervals of 10 minutes, and the degree is measured for each of these intervals. Different values of β then identify whether someone is active for only some interval or for almost all intervals. (2) Consider as features the classic centrality val- ues, e.g., degree, closeness, and betweenness. Here, different 50 40 30 s k n a R 20 10 0 0 X: 0.5 Y: 33 X: 0.5 Y: 32 X: 0.5 Y: 20 0.5 1 1.5 2 X: 4.5 Y: 39 X: 4.5 Y: 19 X: 4.5 Y: 4 P7 P28 P17 3 3.5 4 4.5 5 2.5 0 < b < 5 Fig. 2: This figure shows how the ranks of the nodes from the highest 52 to the lowest 1 change over the different values of β. The possible values of β vary in the interval [0, 5] in this example and result in the andness [0, 0.8333] which indicates how the features can compensate each other’s value from full compensation (high-orness) to no compensation (high-andness). Three chosen samples are marked who have very strong differences in their ranks in the different values of β. These nodes are highlighted in Table. I to see their ranks using only the normalized degree. β values scale between the node who is most active in one of them and the one who has the highest value in all of them. TABLE I: Top 20 nodes are listed from the highest rank 52 to the lowest 20 with respect to the result of a1(i) which is the normalized degree of node i. Three nodes those which have some particular changes in their ranks in Fig. 2 are highlighted here. a2(i) a3(i) a4(i) 1 0.423 0.276 0.265 0.217 0.229 0.044 0.116 0.104 0.161 0.033 0.061 0.01 0.092 0.183 0.031 0.182 0.043 0.091 0.102 0.033 0.043 0.024 0.034 0.127 0.003 0.063 1 0.196 0.224 0.168 0.145 0.204 0.064 0.042 0.136 0.145 0.02 0.048 0.008 0.073 0.152 0.044 0.167 0.027 0.066 0.139 0.056 0.037 0.014 0.03 0.056 0.003 0.04 1 0.407 0.257 0.249 0.2 0.205 0.037 0.109 0.105 0.151 0.032 0.059 0.011 0.083 0.173 0.031 0.168 0.043 0.09 0.096 0.028 0.039 0.02 0.035 0.104 0.003 0.053 REFERENCES [1] A. Barrat, M. Barthelemy, R. Pastor-Satorras, and A. Vespignani. The architecture of complex weighted networks. PNAS, 101(11), 2004. [2] R.A. Bellman and L.A. Zadeh. Decision-making in a fuzzy environment. Management Sciences, Ser. B(17): 141–164, 1970. [3] S. Borgatti. Centrality and network flow. Social Net- works, 27, 55-71, 2005. [4] P. Holme and J. Saramäki. Temporal networks. Physics Reports, 519(3):97–125, 2012. [5] Dirk Koschützki, Katharina A. Lehmann, Leon Peeters, Stefan Richter, Dagmar Tenfelde-Podehl, and Oliver Zlo- towski. Network Analysis - Methodological Foundations, chapter Centrality Indices, pages 16–60. Springer Verlag, 2005. [6] M. Moessner, M. Schiltenwolf, and E. Neubauer. Internet-based aftercare for patients with back pain - a pilot study. Telemedicine and e-Health, 18:413–419, 2012. [7] T. Opsahl, F. Agneessens, and J. Skvoretz. Node cen- trality in weighted networks: Generalizing degree and shortest paths. Social Networks, 32(3):245–251, 2010. [8] S. Tavassoli, M. Moessner, and K.A. Zweig. Constructing social networks from semi-structered chat-log data. In the IEEE/ACM International Conference on Advances Rank 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 i Therapist P6 P1 P36 P37 P11 P22 P3 P30 P39 P27 P24 P28 P41 P5 P10 P18 P23 P25 P4 P16 P14 P20 P15 P7 P33 P51 P17 P42 P8 P31 P35 P43 a1(i) 1 0.588 0.51 0.49 0.451 0.431 0.431 0.392 0.392 0.373 0.333 0.314 0.314 0.314 0.294 0.294 0.294 0.294 0.294 0.275 0.255 0.235 0.235 0.216 0.196 0.196 0.196 0.176 0.176 0.157 0.157 0.157 0.157 0 0.027 0.022 0.003 0.002 0.004 0 0.02 0.024 0.004 0.003 0.006 0 0.026 0.02 0.002 0.003 0.004 in Social Networks Analysis and Mining (ASONAM), Beijing, China, August 2014. [9] R. R. Yager. On ordered weighted averaging aggregation operators in multi-criteria decision making. IEEE Trans- actions on Systems, Man and Cybernetics, 18:183–190, 1988. [10] R. R. Yager. Quantifier guided aggregation using owa operators. Technical Report MII-1504, Mashine Intelli- gence Institute, Iona College, New York, 1994. [11] I. D. Yalom and M. Lesczc. The Theory and Practice of Group Psychotherapy. Basic Books, 5th edition, 2005. [12] L. A. Zadeh. A computational approach to fuzzy quan- tifiers in natural languages. Computers & Mathematics with applications, 9(1):149–184, 1983.
1205.7009	1	1205	2012-05-31T14:36:44	Oriented and Degree-generated Block Models: Generating and Inferring Communities with Inhomogeneous Degree Distributions	[ "cs.SI", "cond-mat.stat-mech", "physics.soc-ph", "stat.ML" ]	The stochastic block model is a powerful tool for inferring community structure from network topology. However, it predicts a Poisson degree distribution within each community, while most real-world networks have a heavy-tailed degree distribution. The degree-corrected block model can accommodate arbitrary degree distributions within communities. But since it takes the vertex degrees as parameters rather than generating them, it cannot use them to help it classify the vertices, and its natural generalization to directed graphs cannot even use the orientations of the edges. In this paper, we present variants of the block model with the best of both worlds: they can use vertex degrees and edge orientations in the classification process, while tolerating heavy-tailed degree distributions within communities. We show that for some networks, including synthetic networks and networks of word adjacencies in English text, these new block models achieve a higher accuracy than either standard or degree-corrected block models.	cs.SI	cs	Oriented and Degree-generated Block Models: Generating and Inferring Communities with Inhomogeneous Degree Distributions Yaojia Zhu1(cid:63), Xiaoran Yan1, and Cristopher Moore2,1 1 University of New Mexico Albuquerque NM 87131 USA 2 Santa Fe Institute 1399 Hyde Park Road, Santa Fe NM 87501, USA Abstract. The stochastic block model is a powerful tool for inferring community structure from network topology. However, it predicts a Pois- son degree distribution within each community, while most real-world networks have a heavy-tailed degree distribution. The degree-corrected block model can accommodate arbitrary degree distributions within com- munities. But since it takes the vertex degrees as parameters rather than generating them, it cannot use them to help it classify the vertices, and its natural generalization to directed graphs cannot even use the ori- entations of the edges. In this paper, we present variants of the block model with the best of both worlds: they can use vertex degrees and edge orientations in the classification process, while tolerating heavy- tailed degree distributions within communities. We show that for some networks, including synthetic networks and networks of word adjacencies in English text, these new block models achieve a higher accuracy than either standard or degree-corrected block models. Keywords: complex networks, community detection, generative model, stochastic block model, degree distribution 1 Introduction In many real-world networks, vertices can be divided into communities or mod- ules based on their connections. Social networks can be forged by interactions in daily activities like karate training [24]. The blogosphere contains groups of linked blogs with similar political views [1]. Words can be tagged as different parts of speech based on their adjacencies in large texts [17]. Communities range from assortative clumps, where vertices preferentially attach to others of the same type, to functional communities of vertices that connect to the rest of the network in similar ways, such as groups of predators in a food web that feed on similar prey [4, 15]. Understanding this variety of community structures, and their relationships to functional roles of vertices and edges, is crucial to understanding network data. (cid:63) This work was supported by the McDonnell Foundation. 2 1 0 2 y a M 1 3 ] I S . s c [ 1 v 9 0 0 7 . 5 0 2 1 : v i X r a 2 Yaojia Zhu, Xiaoran Yan, Cristopher Moore The stochastic block model (SBM) [10, 12, 22, 2] is a popular and highly flexible generative model for community detection. It partitions the vertices into communities or blocks, where vertices belonging to the same block are stochas- tically equivalent [23] in the sense that the probabilities of a connection with all other vertices are the same for all vertices in the same block. This definition of community is quite flexible, letting block models capture many types of commu- nity structure, including assortative, disassortative, and satellite communities and mixtures of them [18, 19, 16, 15, 9, 8]. The SBM assumes that each edge is generated independently conditioned on the block memberships. Each entry Auv of the adjacency matrix is then Bernoulli-distributed, where the probability that Auv = 1 depends solely on the block memberships gu, gv of its endpoints. Since every pair of vertices in a given pair of blocks are connected with the same probability, for large n the degree distribution within each block is Poisson. As a consequence, vertices with very different degrees are unlikely to be in the same block. This leads to problems when modeling real networks, which often have heavy-tailed degree distributions within each community. For instance, both liberal and conservative political blogs range from high-degree “leaders” to low-degree “followers” [1]. Recently, Karrer and Newman [13] developed the degree-corrected (DC) block model for undirected networks. They add a parameter for each vertex, which controls its expected degree. By setting these parameters equal to the observed degrees, the DC can accommodate arbitrary degree distributions within commu- nities. This removes the model’s tendency to separate high-degree and low-degree vertices into different communities. Similar models were considered by Mørup and Hansen [16] and Reichardt, Alamino, and Saad [21]. On the other hand, the degree-corrected model cannot use the vertex degrees to help it classify the vertices, precisely because it takes the degrees as param- eters rather than as data that need to be explained. For this reason, DC may actually fail to recognize communities that differ significantly in their degree distributions. Thus we have two extremes: the SBM separates vertices by degree even when it shouldn’t, and DC fails to do so even when it should. For directed graphs, the natural generalization of DC, the directed degree- corrected (DDC) block model, has two parameters for each vertex: the expected in-degree and out-degree. But this model cannot even take advantage of edge orientations. For instance, in English adjectives usually precede nouns but rarely vice versa. Thus the ratio of each vertex’s in- and out-degree is strongly indicative of its block membership, and leveraging this part of the data is very helpful in the classification process. In this paper, we propose two new types of block model, which combine the strengths of the degree-corrected and uncorrected block models. The oriented degree-corrected (ODC) block model is able to utilize the edge orientations for community detection by only correcting the total degrees instead of the in- and out-degrees separately. We show that for networks with strongly asymmetric behavior between communities, including synthetic networks and networks of Degree Correction in Stochastic Block Models 3 word adjacencies in English text, ODC achieves a higher accuracy than either the original stochastic block model or the degree-corrected block model. We also propose the degree-generated (DG) block model, which treats the expected degree of each vertex as generated from prior distributions in each community, such as power laws whose exponents and cutoffs vary from one com- munity to another. By including the probability of these degrees in the likeli- hood of a given block assignment, the model captures the interaction between the degree distribution and the community structure. DG automatically strikes a balance between allowing vertices of different degrees to coexist in the same community on the one hand, and using vertex degrees to separate vertices into communities on the other. Our experiments show that DG works especially well in networks where com- munities have highly inhomogeneous degree distributions, but where the degree distributions differ enough between communities so that we can use vertex de- grees to help us classify the vertices. Both the standard and degree-corrected block models classify nodes solely on the basis of the relative density of connec- tions between communities, with different notions of “density.” DG block models let us leverage degree information as well. In some cases, DG has a further advan- tage in faster convergence as it reshapes the landscape of the parameter space, providing the inference algorithm a shortcut to the correct community structure. These new variants of the block model give us the best of both worlds. They can tolerate heavy-tailed degree distributions within communities, but can also use degrees and edge orientations to help classify the vertices. In addition to their performance on these networks, our models illustrate a valuable point about generative models and statistical inference: when inferring the structure of a network, you can only use the information that you try to generate. 2 The models In this section, we review the degree-corrected block model of [13], and present our variations on it, namely oriented and degree-generated block models. 2.1 Background: degree-corrected block models Throughout, we use N and M to denote the number of vertices and edges, and K to denote the number of blocks. The problem of determining the number of blocks is a subtle model selection problem, which we do not address here. In the original stochastic block model, the entries Auv of the adjacency matrix are independent and Bernoulli-distributed, with P (Auv = 1) = pgu,gv . Here gu is the block to which u belongs, where p is a K × K matrix. Karrer and Newman [13] consider random multigraphs where the Auv are independent and Poisson-distributed, Auv ∼ Poi(θuθvωgu,gv ) . 4 Yaojia Zhu, Xiaoran Yan, Cristopher Moore Here ω replaces p, and θu is an overall propensity for u to connect to other ver- tices. Note that since the Auv are independent, the degrees du will vary some- what around their expectations; however, the resulting model is much simpler to analyze than one that controls the degree of each vertex exactly. Ignoring self-loops, the likelihood with which this degree-corrected (DC) block model generates an undirected multigraph G is then P (G θ, ω, g) = (θuθvωgugv )Auv Auv! exp (−θuθvωgugv ) . (1) u<v divide ω by C 2, we can impose a normalization constraint(cid:80) each block r, where κr =(cid:80) To remove the obvious symmetry where we multiply the θ’s by a constant C and u:gu=r θu = κr for u:gu=r du is the total degree of the vertices in block r. Under these constraints, the maximum likelihood estimates (MLEs) for the θ parameters are θu = du. For each pair of blocks r, s, the MLE for ωrs is then (cid:89) where mrs is the number of edges connecting block r to block s (and edges within blocks are counted twice). Substituting these MLEs for θ and ω then gives the log-likelihood log P (G g) = 1 2 mrs log mrs κrκs . (2) ωrs = mrs κrκs , K(cid:88) r,s=1 2.2 Directed and oriented degree-corrected models The natural extension of the degree-corrected model to directed networks, which we call the directed degree-corrected block model (DDC), has two parameters u , θin θout u for each vertex. The number of directed edges from u to v is again Poisson-distributed, Auv ∼ Poi(θout u:gu=r θout u θin v ωgu,gv ) . and(cid:80) u:gu=r θin r for each block r, where κout u denote the total out- and in-degree of block r. As before, let mrs denote the number of directed edges from block r to block s. Then the likelihood is u:gu=r dout u = κout u:gu=r din u = κin and κin u r We impose the constraints(cid:80) (cid:0)θout r =(cid:80) (cid:89) (cid:81) P (G θ, ω, g) = uv u θin v ωgugv Auv! exp(−θout u(θout u )dout u (θin u )din = rs ωmrs rs uv Auv! v ωgugv ) u θin exp(−κout r κin s ωrs) , (3) r =(cid:80) (cid:1)Auv u (cid:81) (cid:81) Degree Correction in Stochastic Block Models 5 Ignoring constants, we get the log-likelihood as follows (cid:88) (cid:88) u log P (G θ, ω, g) = + u log θin u ) (dout u log θout u + din (mrs log ωrs − κout r κin s ωrs) . The MLEs for the parameters (see Appendix A) are rs θout u = dout u , θin u = din u , ωrs = mrs κout r κin s . Substituting these MLEs gives log P (G g) = K(cid:88) r,s=1 mrs log mrs r κin κout s . (4) (5) (6) In the DDC, the in- and out-degrees of each vertex are completely specified by the θ parameters, at least in expectation. Thus the DDC lets vertices with arbitrary in- and out-degrees to fit comfortably together in the same block. On the other hand, since the degrees are given as parameters, rather than as data that the model must generate and explain, the DDC cannot use them to infer community structure. Indeed, it cannot even take advantage of the orientations of the edges, and as we will see below it performs quite poorly on networks with strongly asymmetric community structure. To deal with this, we present a partially degree-corrected block model capa- ble of taking advantage of edge orientations, which we call the oriented degree- corrected (ODC) block model. Following the maxim that we can only use the information that we try to generate, we correct only for the total degrees of the vertices, and generate the edges’ orientations. Let ¯G denote the undirected version of a directed graph G, i.e., the multi- graph resulting from erasing the arrows for each edge. Its adjacency matrix is ¯Auv = Auv + Avu, so (for instance) ¯G has two edges between u and v if G had one pointing in each direction. The ODC can be thought of as generating ¯G according to the undirected degree-corrected model, and then choosing the orientation of each edge according to another matrix ρrs, where an edge (u, v) is oriented from u to v with probability ρgu,gv . Thus the total log-likelihood is log P (G θ, ω, ρ, g) = log P ( ¯G θ, ω, g) + log P (G ¯G, ρ, g) . (7) Writing ¯mrs = mrs + msr and κr = κin undirected model to their MLEs as in Section 2.1, giving r + κout r , we can set θu and ωrs for the log P ( ¯G g) = 1 2 ¯mrs log ¯mrs κrκs . (8) K(cid:88) r,s=1 6 Yaojia Zhu, Xiaoran Yan, Cristopher Moore The orientation term is log P (G ¯G, ρ, g) = (cid:88) rs mrs log ρrs = 1 2 (cid:88) (mrs log ρrs + msr log ρsr) , rs For each r, s we have ρrs + ρsr = 1, and the MLEs for ρ are ρrs = mrs/ ¯mrs . (9) (10) As (9) is maximized when the ρrs are near 0 or 1, the edge orientation term prefers highly directed inter-block connections. Since ρrr = 1/2 for any r, it also prefers disassortative mixing, with as few connections as possible within blocks. Substituting the MLEs for ρ and combining (8) with (9), the total log-likelihood is K(cid:88) r,s=1 log P (G g) = mrs log mrs κrκs . (11) We can also view the ODC as a special case of the DDC, where we add the constraint θin for all vertex u (see Appendix B). Moreover, if we set θu = 1 for all u, we obtain the original block model, or rather its Poisson multigraph version where each Auv is Poisson-distributed with mean ωgu,gv . Thus u = θout u SBM ≤ ODC ≤ DDC , where A ≤ B means that model A is a special case of model B, or that B is an elaboration of A. We will see below that since it is forced to explain edge orientations, the ODC performs better on some networks than either the simple SBM or the DDC. 2.3 Degree-generated block models Another way to utilize vertex degrees for community detection is to require the model to generate them, but according to some distribution derived from domain knowledge or an overall measurement of the network’s degree distribution. For instance, many real-world networks have a power-law degree distribution, but with parameters (such as the exponent, minimum degree, or leading constant) that vary from community to community. In that case, the degree of a vertex gives us a clue as to its block membership. Our degree-generated (DG) block models allow heavy-tailed degree distributions, unlike the simple block model, while taking advantage of vertex degrees to help it classify the vertices, unlike the degree-corrected model of Karrer and Newman. To maintain the tractability of the model, we do not generate the degrees directly. Instead, we generate the θ parameters of one of the degree-corrected block models discussed above, and use them to generate a random multigraph. Specifically, each θu is generated independently according to some distribution whose parameters ψ depend on the block gu to which u belongs. Thus the DG (cid:90) (cid:89) Degree Correction in Stochastic Block Models 7 model is a hierarchical model, which extends the previous degree-corrected block models by adding a degree generation stage on top, treating the θs as generated by the block assignment g and the parameters ψ rather than as parameters. We can apply this approach to the undirected, directed, or oriented versions of the degree-corrected model; at the risk of drowning the reader in acronyms, we denote these DG-DC, DG-DDC, and DG-ODC. In each case, the total log- likelihood of a graph G is log P (G ψ, ω, g) = log dθ P (G θ, ω, g) P (θ ψ, g) , where P (θ ψ, g) = P (θu ψgu ) . u For the directed models, we use θu as a shorthand here for θin As in many hierarchical models, computing this integral appears to be diffi- cult, except when P (θ ψ) has the form of a conjugate prior such as the Gamma distribution (see Appendix C). We approximate it by assuming that it is domi- nated by the most-likely value of θ, u and θout u . log P (G ψ, ω, g) ≈ log P (G θ, ω, g) + log P (θ ψ, g) . However, even determining θ is challenging where P (θ ψ) is, say, a power law with a minimum-degree cutoff. Thus we make a further approximation, setting θ just by maximizing the block model term log P (G θ, ω, g) as we did before, using (5) or the analogous equations for the DC or ODC. In essence, these approximations treat P (θ ψ, g) as a penalty term, imposing a prior probability on the degree distribution of each community with hyperparameters ψ. This leaves the door open for community structures that might not be as good a fit to the edges, but compensate with a much better fit to the degrees. We can either treat the degree-generating parameters ψ as fixed (say, if they are predicted by a theoretical model of network growth [3, 5, 14]) or infer them by finding the ψ that maximizes P (θ ψ). For instance, suppose the θu in block gu = r are distributed as a continuous power law with a lower cutoff θmin,r. Specifically, let the parameters in each block r be ψr = (αr, βr, θmin,r), and  P (θu ψr) = βr 0 (1−βr)(α−1) θmin,r (cid:16) θu θmin,r (cid:17)−αr θu = 0 0 < θu < θmin,r θu ≥ θmin,r . r r = (αin r , βin r , θin In the directed case, we have ψin min,r). Allowing βout to be nonzero, for instance, lets us directly include nodes with no outgoing neighbors; we find this useful in some networks. Alternately, we can choose (θin u ) from some joint distribution, allowing in- and out-degrees to be correlated in various ways. min,r) and ψout r = (αout u , θout , βout , θout r r 8 Yaojia Zhu, Xiaoran Yan, Cristopher Moore We fix θmin,r = 1. Given the degrees and the block assignment, the MLEs for αr and βr are as follows. Let Yr = {u : gu = r and θu (cid:54)= 0}, and let yr = Yr. Then the most-likely exponent of the power law is [6] (cid:44)(cid:88) u∈Yr αr = 1 + yr ln θi . (12) The MLE for βr is simply the fraction of vertices in block r with degree zero. 3 Experiments on synthetic networks In this section, we describe experiments on various synthetic networks. First, we generated undirected networks according to the DG-DC model, with two blocks or communities of equal size N/2. In order to confound the block model as much as possible, we deliberately designed these networks so that the two blocks have the same average degree. The degree distribution in block 1 is a power law with exponent α = 1.7, with an upper bound of 1850, so that the average degree is 20. The degree distribution in block 2 it is Poisson, also with mean 20. As described in Appendix D, the upper bound on the power law is larger than any degree actually appearing in the network; it really just changes the normalizing constant of the power law, and the MLE for α can still be calculated using (12). We assume the algorithm knows that one block has a power law degree distribution and the other is Poisson, but we force it to infer the parameters of these distributions. As in [13], we use a parameter λ to interpolate linearly between a fully random network with no community structure and a “planted” one where the communities are completely separated. Thus ωrs = λωplanted rs + (1 − λ)ωrandom where ωrandom rs = κrκs 2M , ωplanted = rs (cid:18) κ1 0 0 κ2 (cid:19) . We inferred the community structure with various models. We ran the Kernighan- Lin (KL) heuristic first to find a local optimum [13], and then ran the heat-bath MCMC algorithm with fixed number of iterations to further refine it if ever possi- ble. We initialized each run with a random block assignment; to test the stability of the models, we also tried initializing them with the correct block assignment. Since isolated vertices don’t participate in the community structure, giving us little or no basis on which we can classify them, we remove them and focus on the giant component. For λ = 1, where the community structure is purely the “planted” one, we kept two giant components, one in each community. We measured accuracy by the normalized mutual information (NMI) [7] be- tween the most-likely block assignment found by the model and the correct assignment. To make this more concrete, if there are two blocks of equal size Degree Correction in Stochastic Block Models 9 and 95% of the vertices in each block are labeled correctly, the NMI is 0.714. If 90% in each group are labeled correctly, the NMI is 0.531. For groups of unequal size, the NMI is a better measure of accuracy than the fraction of vertices labeled correctly, since one can make this fraction fairly large simply by assigning every vertex to the larger group. As shown in Fig. 1, DG-DC works very well even for small λ. This is because it can classify most of the vertices simply based on their degrees; if du is far from 20, for instance, then u is probably in block 1. As λ increases, it uses the connections between communities as well, giving near-perfect accuracy for λ ≥ 0.6. It does equally well whether its initial assignment is correct or random. The DC model, in contrast, is unable to use the vertex degrees, and has accuracy near zero (i.e., not much better than a random block assignment) for λ ≤ 0.2. Like the SBM [8, 9], it may have a phase transition at a critical value of λ below which the community structure is undetectable. Initializing it with the correct assignment helps somewhat at these values of λ, but even then it settles on an assignment far from the correct one. The original stochastic block model (SBM), which doesn’t correct the de- grees, separates vertices with high degrees from vertices with low degrees. Thus it cannot find the correct group structure even for large λ. Our synthetic tests are designed to have a broad degree distribution in block 1, and thus make SBM fail. Note that if the degree distribution in block 1 is a power-law with a larger exponent α, then most of the degrees will be much lower than 20, in which case SBM works reasonably well. Fig. 1. Tests on synthetic networks generated by the DG-DC model. Each point is based on 30 randomly generated networks with N = 2400. For each network and each model, we choose the best result from 10 independent runs, initialized either with random assignments (the suffix R) or the true block assignment (the suffix T ). Each run consisted of the KL-heuristic followed by 106 MCMC steps. Our degree-generated (DG) block model performs much better on these networks than the degree-corrected (DC) model. The non-degree-corrected (SBM) model doesn’t work at all. Next, we generated directed networks according to the DG-DDC model. We again have two blocks of equal size, with degree distributions similar to the 00.20.40.60.8100.20.40.60.81λNormalized Mutual Information DG−DC−TDG−DC−RDC−TDC−RSBM−TSBM−R10 Yaojia Zhu, Xiaoran Yan, Cristopher Moore undirected networks tested above. In block 1, both out- and in-degrees are power- law distributed with α = 1.7, with an upper bound 1850 so that the expected degree is 20. In block 2, both out- and in-degrees are Poisson-distributed with mean 20. To test our oriented and directed models, we interpolate between a random network ωrandom = κrκs/4M and a planted network with completely asymmetric connections between the blocks, rs (cid:18) (κ1 − ω12)/2 0 ωplanted = (cid:19) ω12 (κ2 − ω12)/2 , (13) where ω12 ≤ min(κ1, κ2). We choose ω12 = 1 2 min(κ1, κ2). As Fig. 2 shows, DG-ODC and DG-DDC have very similar performance at the extremes where λ = 0 and 1. However, DG-ODC works better than DG-DDC for other λ, and both of them achieve much better accuracy than the ODC or DDC models. As in Fig. 1, the degree-generated models can achieve a high accuracy based simply on the vertex degrees, and as λ grows they leverage this information further to achieve near-perfect accuracy for λ ≥ 0.8. Among the non-degree-corrected models, ODC performs significantly better than DDC for λ ≥ 0.4. Edges are more likely to point from block 1 to block 2 than vice versa, and ODC can take advantage of this information while DDC cannot. As we will see in the next section, ODC performs well on some real-world networks for precisely this reason. Fig. 2. Tests on synthetic directed networks with N = 2400. Left, DG-ODC and DG-DDC; right, ODC and DDC. The degree-generated models again perform very well even for small λ, since they can use in- and out-degrees to classify the vertices. ODC performs significantly better than DDC for λ ≥ 0.4, since it can use the edge orientations to distinguish the two blocks. The number of networks, runs, and MCMC steps per run are as in Fig. 1. 4 Experiments on real networks In this section, we describe experiments on three word adjacency networks in which vertices are separated into two blocks: adjectives and nouns. The first 00.20.40.60.810.60.70.80.91λNormalized Mutual Information DG−ODC−TDG−ODC−RDG−DDC−TDG−DDC−R00.20.40.60.8100.20.40.60.81λNormalized Mutual Information ODC−TODC−RDDC−TDDC−RSBM−TSBM−RDegree Correction in Stochastic Block Models 11 network consists of common words in Dickens’ novel David Copperfield [20]. The other two are formed by adjectives and nouns in the Brown corpus, which is a tagged corpus of present-day edited American English across various categories, including news, novels, documents, and many others [11]. We build two different networks from the Brown corpus. The smaller one contains words in the News category (45 archives) that appeared at least 10 times; the larger one contains all the adjectives and nouns in the giant component of the entire corpus. We considered both the simple version of these networks where Auv = 1 if u and v ever occur together in that order, and the multigraph version where Auv ≥ 0 is the number of times they occur together. The sizes, block sizes, and number of edges of these networks are shown in Table 1. In “News” and “Brown”, the block sizes are quite different, with more nouns than adjectives. As discussed above, the NMI is a better measure of accuracy than the fraction of vertices labeled correctly, since we could make the latter fairly large by labeling everything a noun. In each network, both blocks have heavy-tailed in- and out-degree distribu- tions. The connections between them are disassortative and highly asymmetric: since in English adjectives precede nouns more often than they follow them, and more often than adjectives precede adjectives or nouns precede nouns, ω12 is roughly 10 times larger than ω21, and ω12 is larger than either ω11 or ω22. The ω for each network corresponding to the correct block assignment (according to the stochastic block model) is shown in Table 2. Table 1. Basic statistics of the three word adjacency networks. S and M denote the simple and multigraph versions respectively. Network #words #adjective #noun #edges (S) #edges (M) David News Brown 112 376 23258 57 91 6235 569 1389 55 285 17023 66734 1494 2411 88930 Table 2. The matrices ωrs = mrs/(nrns) for the most-likely block assignment accord- ing to the stochastic block model. David(S) David(M) News(S) News(M) Brown(S) Brown(M) 0.039 0.118 0.080 0.358 0.010 0.015 0.012 0.028 9.1e-05 3.4e-04 1.1e-04 4.4e-04 0.018 0.006 0.025 0.011 0.002 0.010 0.003 0.019 2.0e-05 8.8e-05 2.4e-05 1.2e-04 12 Yaojia Zhu, Xiaoran Yan, Cristopher Moore 4.1 Performance of oriented and degree-corrected models Table 3 compares the performance of non-degree-generated block models, in- cluding SBM, DC, ODC, and DDC. When applying DC, we ignore the edge orientations, and treat the graph or multigraph as undirected (note that the re- sulting network may contain multi-edges even though the directed one doesn’t). In our experiments, we started with a random initial block assignment, ran the Kernighan-Lin (KL) heuristic to find a local optimum [13], and then ran the heat-bath MCMC algorithm. We also tested a naive heuristic (NH) which simply labels a vertex v as an adjective if dout v > dout . If dout v , NH labels v randomly with equal probabilities. v = din v > din v , and a noun if din v Table 3. For each model and each network, we pick the block assignment with highest likelihood and compute its NMI with the correct block assignment. Each run consisted of the KL-heuristic, starting with a random block assignment, followed by 106 MCMC steps. The results for “David” and “News” are based on 100 independent runs; for “Brown”, 50 runs are executed. The best NMI for each network is shown in bold. David(S) David(M) News(S) News(M) Brown(S) Brown(M) SBM DC ODC DDC NH .423 .566 .462 .128 .395 .051 .568 .470 8e-04 .449 .006 .084 .084 .084 .215 .018 .083 .029 .091 .233 .001 .020 .311 .016 .309 7e-04 .015 .318 .012 .314 For “David”, DC and ODC work fairly well, and both are better than the naive NH. Moreover, the mistakes they make are instructive. There are three adjectives with out-degree zero: “full”, “glad”, and “alone”. ODC mislabels these since it expects edges to point away from adjectives, while DC labels them correctly by using the fact that (undirected) edges are disassortative, crossing from one block to the other. The standard SBM works well on “David(S)” but fails on “David(M)” be- cause the degrees in the multigraph are more skewed than those in the simple one. Finally, DDC performs the worst; by correcting for in- and out-degrees sep- arately, it loses any information that the edge orientations could provide, and even fails to notice the disassortative structure that DC uses. Thus full degree- correction in the directed case can make things worse, even when the degrees in each community are broadly distributed. For “Brown”, all these models fail except ODC, although it does only slightly better than the naive NH. For “News”, all these models fail, even ODC. Despite the degree correction, the most-likely block assignment is highly assortative, with high-degree vertices connecting to each other. However, we found that in most runs on “News”, ODC used the edge orientations successfully to find the a block assignment close to the correct one; it found the assortative structure Degree Correction in Stochastic Block Models 13 only occasionally. This suggests that, even though the “wrong” structure has a higher likelihood, we can do much better if we know what kind of community structure to look for; in this case, disassortative and directed. To test this hypothesis, we tried giving the models a hint about the commu- nity structure by using NH to determine the initial block assignment. We then performed the KL heuristic and the MCMC algorithm as before. As Table 4 shows, this hint improves ODC’s performance on “News” significantly; it is able to take the initial naive classification, based solely on degrees, and refine it using the network’s structure. Note that this more accurate assignment actually has lower likelihood than the one found in Table 3 using a random initial condition— so NH helps the model stay in a more accurate, but less likely, local optimum. Starting with NH improves DC’s performance on “Brown” somewhat, but DC still ends up with an assignment less accurate than the naive one. Table 4. Results using the naive NH assignment as the initial condition, again followed by 106 MCMC steps. This hint now lets ODC outperform the other models on “News”. David(S) David(M) News(S) News(M) Brown(S) Brown(M) SBM DC ODC DDC NH .423 .566 .462 .015 .395 .051 .568 .470 .060 .449 .006 .084 .247 .084 .215 .021 .015 .270 .005 .233 .001 .160 .311 .005 .309 7e-04 .155 .318 .070 .314 4.2 Performance of degree-generated models In this section, we measure the performance of degree-generated models on the Brown network, and compare them to their non-degree-generated counterparts. As Fig. 3 shows, the in- and out-degree distributions in each block have heavy tails close to a power-law. Moreover, the out-degrees of the adjectives have a heavier tail than those of the nouns, and vice versa for the in-degrees. This is exactly the kind of difference in the degree distributions between communities that our DG block models are designed to take advantage of. Setting θmin = 1, we can estimate the parameters α and β for these dis- tributions as discussed in Section 2.3. We show the most likely values of these parameters, given the correct assignment, in Table 5. As Table 6 shows, degree generation improves DC and DDC significantly, letting them find a good assignment as opposed to one with NMI near zero. For ODC, the performance improvement is slight, making DG-ODC the best model overall, but there is another effect. We compare performance starting with the KL heuristic to performance using MCMC alone. We see that degree generation gives ODC almost as much benefit as the KL heuristic does. In other words, it 14 Yaojia Zhu, Xiaoran Yan, Cristopher Moore Fig. 3. Degree distributions in the Brown network. Table 5. MLEs for the degree generation parameters in the Brown network, given the correct assignment. block αin Brown(S) βin αout βout αin Brown(M) βin αout βout adjective 2.329 2.629 0.161 0.527 2.136 2.326 0.161 0.527 noun 2.721 2.248 0.716 0.021 2.576 2.134 0.716 0.021 10010110210310−410−310−210−1100out−degreeCCDFBrown Corpus (S) nounadjective10010110210310−510−410−310−210−1100in−degreeCCDFBrown Corpus (S) nounadjective10010110210310410−410−310−210−1100out−degreeCCDFBrown Corpus (M) nounadjective10010110210310−510−410−310−210−1100in−degreeCCDFBrown Corpus (M) nounadjectiveDegree Correction in Stochastic Block Models 15 speeds up the MCMC optimization process, letting ODC find a good assignment without the initial help of the (computationally expensive) KL heuristic. Table 6. Performance of degree-generated vs. non-degree generated models. KL in- dicates that we applied the KL heuristic and then 106 MCMC steps, as opposed to MCMC alone. DG indicates degree generation. Each number gives the NMI for the most-likely assignment found in 50 independent runs. The best model is DG-ODC. Moreover, degree generation helps ODC converge, providing much of the benefit of the KL heuristic while avoiding its long running time (see bold numbers). Brown(S) Brown(M) DC ODC DDC DC ODC DDC .010 – – KL – .020 – DG .267 KL DG .271 .188 .311 .302 .312 .008 .016 .213 .225 .007 .015 .278 .284 .203 .318 .310 .320 .011 .012 .149 .195 5 Conclusions Degree correction in stochastic block models provides a powerful approach to dealing with networks with inhomogeneous degree distributions. However, in a sense it denies information to the inference process, since a generative model can only help us learn from the data that it has to generate. We have introduced two new kinds of block models that allow for broad or heavy-tailed degree distributions, while using the degrees to help us detect com- munities. The oriented degree-corrected model (ODC) performs partial degree correction, taking the total degrees as parameters but generating edge orienta- tions. The degree-generated (DG) models don’t take the degrees as parameters, but assumes that they are generated according to some prior in each community. Unlike the directed degree-corrected (DDC) block model, which takes both in- and out-degrees as parameters, ODC is able to capture and account for certain correlations between the in- and out-degrees. Simply put, for ODC, two vertices are unlikely to be in the same community if one has high in-degree and low out-degree while another has high out-degree and low in-degree. If the network is highly directed or asymmetric, the edge orientations can help ODC find community structures that DDC fails to perceive. Our DG models use degree-corrected block models as a subroutine, but im- pose a penalty term based on the prior likelihood of the degree distribution in each community. They can take the (hyper)parameters of these priors as given, or infer them “on the fly.” DG models achieve high accuracy even when the den- sity of connections between communities is close to uniform, as we illustrated in synthetic networks for small λ. Augmenting block models, such as the ODC, 16 Yaojia Zhu, Xiaoran Yan, Cristopher Moore with degree generation also appears to speed up their convergence in some cases, helping simple algorithms like MCMC handle large networks without the benefit of expensive preprocessing steps like the KL heuristic. On the other hand, the effectiveness of DG depends heavily on knowing the correct form of the degree distribution in each community. Without some prior ground truth about the block assignment, or domain-specific knowledge, finding an appropriate family of degree distributions may be difficult for some networks. With all these variants of the block model, ranging from the “classic” version to degree-corrected and degree-generated variants, we now have a wide variety of tools for inferring structure in network data. Each model will perform better on some networks and worse on others. A better understanding of the strengths and weaknesses of each one—which kinds of structure they can see, and what kinds of structure they are blind to—will help us select the right algorithm each time we meet a new network. Acknowledgments. We are grateful to Terran Lane, Ben Edwards, Aaron Clauset, and Mark Newman for helpful conversations, and to the McDonnell Foundation for their support. Bibliography [1] L.A. Adamic and N. Glance. The political blogosphere and the 2004 US election: divided they blog. Proceedings of the 3rd international workshop on Link discovery, pages 36–43, 2005. [2] E.M. Airoldi, D.M. Blei, S.E. Fienberg, and E.P. Xing. Mixed membership stochastic blockmodels. The Journal of Machine Learning Research, 9:1981– 2014, 2008. [3] R´eka Albert and Albert-L´aszl´o Barab´asi. Statistical mechanics of complex networks. Rev. Mod. Phys., 74:47–97, Jan 2002. [4] Stefano Allesina and Mercedes Pascual. Food web models: a plea for groups. Ecology letters, 12(7):652–662, July 2009. [5] H. Bauke, C. Moore, J.-B. Rouquier, and D. Sherrington. Topological phase transition in a network model with preferential attachment and node re- moval. European Physical Journal B, 83:519–524, 2011. [6] A. Clauset, C.R. Shalizi, and M.E.J. Newman. Power-law distributions in empirical data. SIAM Review, 51:661–703, 2009. [7] Leon Danon, Albert D´ıaz-Guilera, Jordi Duch, and Alex Arenas. Compar- ing community structure identification. Journal of Statistical Mechanics: Theory and Experiment, 2005(09):P09008, September 2005. [8] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zde- borov´a. Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications. Physical Review E, 84(6), De- cember 2011. [9] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zde- Inference and Phase Transitions in the Detection of Modules in borov´a. Sparse Networks. Physical Review Letters, 107(6), August 2011. [10] S.E. Fienberg and S. Wasserman. Categorical data analysis of single socio- metric relations. sociological Methodology, pages 156–192, 1981. [11] W.N. Francis and H. Kucera. Brown Corpus Manual. Technical report, Department of Linguistics, Brown University, Providence, Rhode Island, US, 1979. [12] P.W. Holland, K.B. Laskey, and S. Leinhardt. Stochastic blockmodels: First steps. Social Networks, 5(2):109–137, 1983. [13] B. Karrer and M. Newman. Stochastic blockmodels and community struc- ture in networks. Physical Review E, 83(1), 2011. [14] Cristopher Moore, Gourab Ghoshal, and M. E. J. Newman. Exact solutions for models of evolving networks with addition and deletion of nodes. Phys. Rev. E, 74:036121, 2006. [15] Cristopher Moore, Xiaoran Yan, Yaojia Zhu, Jean-Baptiste Rouquier, and Terran Lane. Active learning for node classification in assortative and dis- assortative networks. In Proc. 17th KDD, pages 841–849, 2011. [16] M. Mørup and L.K. Hansen. Learning latent structure in complex networks. NIPS Workshop on Analyzing Networks and Learning with Graphs, 2009. 18 Yaojia Zhu, Xiaoran Yan, Cristopher Moore [17] M. Newman and E.A. Leicht. Mixture models and exploratory analysis in networks. Proceedings of the National Academy of Sciences, 104(23):9564– 9569, 2007. [18] M.E. Newman. Assortative mixing in networks. Physical Review Letters, 89(20):208701, 2002. [19] M.E. Newman. Mixing patterns in networks. Physical Review E, 67(2):026126, 2003. [20] M.E. Newman. Finding community structure in networks using the eigen- vectors of matrices. Physical Review E, 74(3):036104, 2006. [21] Jorg Reichardt, Roberto Alamino, and David Saad. The interplay between microscopic and mesoscopic structures in complex networks. PloS one, 6(8):e21282, January 2011. [22] T.A. Snijders and K. Nowicki. Estimation and prediction for stochastic blockmodels for graphs with latent block structure. Journal of Classifica- tion, 14(1):75–100, 1997. [23] S. Wasserman and C. Anderson. Stochastic a posteriori blockmodels: Con- struction and assessment. Social Networks, 9(1):1–36, 1987. [24] W. W. Zachary. An information flow model for conflict and fission in small groups. Journal of Anthropological Research, 33(4):452–473, 1977. A Maximum Likelihood Estimators for the directed degree-corrected (DDC) block model We maximize the log-likelihood function (4), (cid:88) (cid:88) u rs log P (G θ, ω, g) = + (cid:88) u log θin u ) (dout u log θout u + din (mrs log ωrs − κout r κin s ωrs) , (cid:88) (14) (15) where we have imposed the constraints on the θ parameters θout u = κout r and u = κin θin r . u:gu=r u:gu=r For each block r, we associate Lagrange multipliers λout r with these con- straints. For each vertex u, taking the partial derivative of the log-likelihood with respect to θout and θin , λin r u gives u dout u θout u = λout gu and din u θin u = λin gu . (16) To satisfy the constraints (15), we take λout r = λin r = 1 for all r, so that θout u = dout u and θin u = din u . Setting the partial derivative of the log-likelihood function with respect to ωrs to zero then gives ωrs = mrs r κin κout s . Degree Correction in Stochastic Block Models 19 B Another view of the ODC model Here we show that the oriented degree-corrected (ODC) model is a special case of the directed degree-corrected (DDC) model. Recall that the ODC model first generates an undirected graph according to the DC model with parameters θu and ωrs, and then orients each edge (u, v) from u to v with probability ρgu,gv . The number of directed edges from u to v is then Poisson-distributed as Auv ∼ Poi(θuθvωgu,gv ρgu,gv ) . But if we write then ω(cid:48) rs = ωrsρrs , Auv ∼ Poi(θuθvω(cid:48) ) . gu,gv u = θout Thus ODC is the special case of DDC where θin these parameters equal to their MLEs. We impose the constraint(cid:80) u = θu for all vertices u. For completeness, we check that the two models correspond when we set u:gu=r θu = r for all blocks r. Ignoring constants, the log-likelihood is then κr = κout r + κin log P (G θ, ω(cid:48), g) = (cid:88) du log θu + u rs (mrs log ω(cid:48) rs − κrκsω(cid:48) rs) , (cid:88) (17) (18) where du = dout u + din u . The MLEs for θu and ω(cid:48) θu = du , ω(cid:48) rs = Thus ω(cid:48) rs = ωrs ρrs where rs are then mrs κrκs . ωrs = ¯mrs κrκs and ρrs = mrs ¯mrs , recovering (11). C Bayesian estimation for DG models Bayesian inference focuses on posterior distributions of parameters rather than on point estimates. In hierarchical models like DG-DDC, the full Bayesian pos- terior of the θ parameters (omitting the other parameters g and ω) is P (θ G) = P (θ G, ψ) P (ψ G) dψ . Here we employ the Empirical Bayesian method, and use point estimates for the hyperparameters ψ, namely their MLEs ψ, P (G ψ) ψ = argmax ψ = argmax ψ P (G θ, ψ) P (θ ψ) dθ . (19) (cid:90) (cid:90) 20 Yaojia Zhu, Xiaoran Yan, Cristopher Moore With this approximation we have P (θ G) ≈ P (θ G, ψ) P (G θ) P (θ ψ) P (G ψ) P (G θ) P (θ ψ) (cid:82) P (G θ, ψ) P (θ ψ) dθ = = , (20) where we used Bayes’ rule in the second line. Computing the posterior P (θ G) is usually difficult, as the integral in the denominator of (20) is often intractable. However, with a clever choice of the prior distribution P (θ ψ), we can work out an analytic solution. It is called the conjugate prior of the likelihood term. We focus here on DG-DDC; the calculations for other degree-generated models are similar. and write X ∼ Γ (α, β), if its probability distribution is Say that a random variable X is Gamma-distributed with parameters α, β, f (x; α, β) = βα Γ (α) xα−1 e−βx . (cid:81) r =(cid:80) u (cid:81) (cid:81) uv Auv! P (G θout) = u(θin)din rs ωmrs rs In DG-DDC, the likelihood (3) can be written (where we have plugged in the MLEs for ω, and substituted κout u:gu=r dout u ) u exp(cid:0)−θout (cid:1) . u (θout u )dout (cid:89) u (21) If we assume that the θin and θout for each u are independent, this is propor- u + 1 and tional to a product of Gamma distributions with parameters α = dout β = 1 for each θout u . A natural conjugate prior for Gamma distributions is the Gamma distri- for each block r consist of a pair bution itself. Let the hyperparameters ψout (αout ), and consider the prior , βout r r r u ∼ Γ (αout θout gu , βout gu ) . That is, P (θout u ψout gu ) = gu (βout gu Γ (αout gu )αout ) (θout u )αout gu −1 exp(−βout gu θout u ) , Multiplying this prior by the likelihood (21) stays within the family of Gamma distributions, and simply updates the parameters: ) P (G θout) P (θout u G) ∝ P (θout ∝ (θout u ψout u )αout gu +dout gu u −1 exp(cid:0)−θout u (cid:0)βout gu + 1(cid:1)(cid:1) . Degree Correction in Stochastic Block Models 21 Thus the posterior distribution is u ∼ Γ(cid:0)αout gu θout + 1(cid:1) . + dout u , βout gu Note that if we use a uninformative prior, i.e., in the limit αout = 0, gu the Gamma prior reduces to a uniform prior. The maximum a posteriori (MAP) estimate of θout = 1 and βout gu is u and similarly for θin u , just as we obtained for the MLEs in (5). θout u = dout u , (22) However, our goal is to integrate over θ, not focus on its MAP estimate. So let us continue the Bayesian analysis. Assuming the θ parameters are independent, then their joint posterior is simply a product of their individual posteriors P (θout u G)P (θin u G) + 1(cid:1) f(cid:0)θin u ; αout gu + dout u , βout gu u ; αin gu + 1(cid:1) . + din u , βin gu (23) (cid:89) u (cid:89) P (θG) = f(cid:0)θout = u (cid:90) (cid:81) (cid:81) = = Then we can calculate the integral in (19) and (20) by the simple algebra: P (G θ, ψ)P (θ ψ) dθ = u + 1, 1) f (θin u f (θout u ; dout u f(cid:0)θout (cid:81) (cid:81) + 1(cid:1)αout (cid:0)βout gu gu u u ; αout gu + dout αout gu βin gu u (cid:0)βin u βout gu +dout gu P (G θ)P (θ ψ) u ; din u , βout gu αin P (θ G) u + 1, 1) f (θout + 1(cid:1) f(cid:0)θin gu Γ(cid:0)αout + 1(cid:1)αin gu u Γ (dout +din gu u + dout u ; αout gu u ; αin gu , βout gu + din u , βin gu (cid:1) Γ(cid:0)αin gu , βin gu ) f (θin u ; αin gu + 1(cid:1) (cid:1) u + 1) Γ(cid:0)αout + din u gu u + 1) Γ (din (24) ) (cid:1) Γ(cid:0)αin gu (cid:1) . Now that the dependence of the numerator and denominator on θ has cancelled out, the integral is a function only of the hyperparameters ψ, making it possible to do the point estimate of ψ in (19). In our case, optimizing for ψ requires some numeric techniques, but it is nonetheless doable. Empirical Bayesian solution not only gives better approximation to the orig- inal problem, it also make it possible to integrate prior knowledge if available. On top of that, because the posterior is now a direct function of the hyperpa- rameters ψ, we no longer have to worry about the Poisson noise when estimating ψ indirectly from degrees. On a final note, the above result only holds for Gamma priors. With any other prior, the integral may not be this simple. D Power-law distribution with upper bound In this section, we show that imposing an upper bound on our power-law distri- butions in order to ensure a certain average degree does not appreciably change 22 Yaojia Zhu, Xiaoran Yan, Cristopher Moore the procedure of [6] for estimating the exponent. Suppose x is distributed as a power law lower bound xmin, upper bound xmax, and exponent α > 0. Then p(x) = α − 1 min − x1−α x1−α max x−α, xmin ≤ x ≤ xmax . Given a random sample x = {x1, . . . , xn} drawn from this distribution indepen- dently, the likelihood function is α − 1 (cid:19)n n(cid:89) n(cid:89) α − 1 (cid:18) p(x) = min − x1−α x1−α max i=1 x−α i = min − x1−α x1−α max Thus, the log-likelihood is log p(x) = n(cid:0)log(α − 1) − log(cid:0)x1−α min − x1−α max x−α i . (cid:1)(cid:1) − α (cid:19) i=1 n(cid:88) i=1 log xi . − n(cid:88) i=1 log xi . (25) Taking the derivative with respect to α gives ∂ log p(x) ∂α = n α − 1 + min log xmin − x1−α x1−α min − x1−α x1−α max max log xmax (cid:18) 1 Setting (25) to zero, we get min log xmin − x1−α x1−α min − x1−α x1−α α − 1 max + 1 max log xmax (cid:80)n i=1 log xi n = . (26) If xmin = 1 and xmax → ∞, then solving (26) gives the MLE for α just as in (12).
1601.05834	2	1601	2016-10-05T06:14:18	Active Sensing of Social Networks	[ "cs.SI", "stat.ML" ]	This paper develops an active sensing method to estimate the relative weight (or trust) agents place on their neighbors' information in a social network. The model used for the regression is based on the steady state equation in the linear DeGroot model under the influence of stubborn agents, i.e., agents whose opinions are not influenced by their neighbors. This method can be viewed as a \emph{social RADAR}, where the stubborn agents excite the system and the latter can be estimated through the reverberation observed from the analysis of the agents' opinions. The social network sensing problem can be interpreted as a blind compressed sensing problem with a sparse measurement matrix. We prove that the network structure will be revealed when a sufficient number of stubborn agents independently influence a number of ordinary (non-stubborn) agents. We investigate the scenario with a deterministic or randomized DeGroot model and propose a consistent estimator of the steady states for the latter scenario. Simulation results on synthetic and real world networks support our findings.	cs.SI	cs	Active Sensing of Social Networks Hoi-To Wai, Student Member, IEEE, Anna Scaglione, Fellow, IEEE, and Amir Leshem, Senior Member, IEEE 1 6 1 0 2 t c O 5 ] I S . s c [ 2 v 4 3 8 5 0 . 1 0 6 1 : v i X r a Abstract-This paper develops an active sensing method to estimate the relative weight (or trust) agents place on their neighbors' information in a social network. The model used for the regression is based on the steady state equation in the linear DeGroot model under the influence of stubborn agents; i.e., agents whose opinions are not influenced by their neighbors. This method can be viewed as a social RADAR, where the stubborn agents excite the system and the latter can be estimated through the reverberation observed from the analysis of the agents' opinions. The social network sensing problem can be interpreted as a blind compressed sensing problem with a sparse measurement matrix. We prove that the network structure will be revealed when a sufficient number of stubborn agents independently influence a number of ordinary (non-stubborn) agents. We investigate the scenario with a deterministic or randomized DeGroot model and propose a consistent estimator of the steady states for the latter scenario. Simulation results on synthetic and real world networks support our findings. Index Terms- DeGroot model, opinion dynamics, social net- works, sparse recovery, system identification I. INTRODUCTION R Ecently, the study of networks has become a major research focus in many disciplines, where networks have been used to model systems from biology, physics to the social sciences [2]. From a signal processing perspective, the related research problems can be roughly categorized into analysis, control and sensing. While these are natural extensions of classical signal processing problems, the spatial properties due to the underlying network structure have yielded new insights and fostered the development of novel signal processing tools [3]–[5]. The network analysis problem has received much attention due to the emergence of 'big-data' from social networks, biological networks, etc. Since the networks to be analyzed consist of millions of vertices and edges, computationally efficient tools have been developed to extract low-dimensional structures, e.g., by detecting community structures [6], [7] and finding the centrality measures of the vertices [8]. Techniques in the related studies involve developing tools that run in (sub)- linear time with the size of the network; e.g., see [9]. Another problem of interest is known as network control, where the aim is to choose a subset of vertices that can provide full/partial control over the network. It was shown recently in [10] that the Kalman's classical control criterion is equivalent to finding a maximal matching on the network; see also [11], [12]. This material is based upon work supported by NSF CCF-1011811 and partially supported by ISF 903/13. A preliminary version of this work has been presented at IEEE CDC 2015, Osaka, Japan [1]. H.-T. Wai and A. Scaglione are with the School of Electrical, Com- puter and Energy Engineering, Arizona State University, Tempe, AZ 85281, USA. E-mails: {htwai,Anna.Scaglione}@asu.edu. A. Leshem is with Faculty of Engineering, Bar-Ilan University, Ramat-Gan, Israel. Email: [email protected] This work considers the social network sensing problem that has received relatively less attention in the community than the two aforementioned problems. We focus on social networks1 and our aim is to infer simultaneously the trust matrix and the network structure by observing the opinion dynamics. We model the opinion dynamics in the social network according to the DeGroot model [13]. In particular, at each time step, the opinions are updated by taking a convex combination of the neighboring opinions, weighted with respect to a trust matrix. Despite its simplicity, the DeGroot model has been widely popular in the social sciences; some experimental papers indicated that the model is able to capture the actual opinion dynamics, e.g., [14]–[18]. Notice that the DeGroot model is analogous to the Average Consensus Gossiping (ACG) model [19], [20] and drives the opinions of every agent in the network to consensus as time goes to infinity. Hence, in such situations the social network sensing method is required to track the individual opinion updates. This is the passive approach taken in previous works [14], [21]–[23]. As agents' interactions may occur asynchronously and at an unknown timescale, these approaches may be impractical due to the difficulty in precisely tracking the opinion evolution. An interesting extension in the study of opinion dynamics is to consider the effects of stubborn agents (or zealots), whose opinions remain unchanged throughout the network dynamics. Theoretical studies have focused on characterizing the steady- state of the opinion dynamics when stubborn agents are present [16], [24]–[29], developing techniques to effectively control the network and attain fastest convergence [29]–[32]. The model of stubborn agents has also been studied in the context of sensor network localization [33]. More recently, experimen- tal studies have provided evidence supporting the existence of stubborn agents in social networks, for instance, [15], [34] suggested that stubborn agents can be used to justify several opinion dynamic models for data collected from controlled experiments; [35] illustrated that the existence of stubborn agents is a plausible cause for the emergence of extreme opinions in social networks; [36] studied the controllability of opinions in real social networks using stubborn agents. The aim of this paper is to demonstrate an important consequence of the existence of stubborn agents. Specifically, we propose a social RADAR formulation through exploiting the special role of stubborn agents. As will be shown below, the stubborn agents will help expose the network structure through a set of steady state equations. The proposed model shows that the stubborn agents can play a role that is analogous to a traditional radar, which actively scans and probes the social network. 1 That said, the developed methodology is general and may be applied to other types of networks with similar dynamics models. 2 As the stubborn agents may constitute a small portion of the network, the set of steady state equation may be an undetermined system. To handle this issue, we formulate the network sensing problem as a sparse recovery problem using the popular (cid:96)0/(cid:96)1 minimization technique. Finally, we derive a low-complexity optimization method for social net- work sensing. The proposed method is applicable to large networks. An additional contribution of our investigation is a new recoverability result for a special case of blind com- pressive sensing; e.g., [37], [38]. In particular, we develop an identifiability condition for active sensing of social networks based on expander graph theories [39]. Compared to [39]– [42], our result is applicable when the sensing matrix is non- negative (instead of binary) and subject to small multiplicative perturbations. shows that The remainder of this paper is organized as follows. Sec- tion II introduces the DeGroot model for opinion dynamics and the effects of stubborn agents. Section III formulates the social network sensing problem as a sparse recovery problem and provides a sufficient condition for perfect recovery. Section IV further develops a fast optimization method for approximately solving the sparse recovery problem and the developed methods can be applied when the opinion dynamics is a time-varying system. Simulation results on both synthetic and a real network example conclude the paper, in Section V. Notation: We use boldfaced small/capital letters to denote vectors/matrices and [n] = {1, ..., n} to represent the index set from 1 to n and n ∈ N. Diag : Rn → Rn×n/diag : Rn×n → Rn to denote the diagonal operator that maps from a vector/square matrix to a square matrix/vector, ≥ to represent the element-wise inequality. For an index set Ω ⊆ [n], we use Ωc to denote its natural complement, e.g., Ωc = [n] \ Ω. A vector x is called k-sparse if (cid:107)x(cid:107)0 ≤ k. We denote off(W ) as the square matrix with only off-diagonal elements in W . II. OPINION DYNAMICS MODEL Consider a social network represented by a simple, con- nected weighted directed graph G = (V, E, W ), where V = [n] is the set of agents, E ⊆ V ×V is the network topology and W ∈ Rn×n denotes the trust matrix between agents. Notice that W can be asymmetric. The trust matrix satisfies W ≥ 0 and W 1 = 1, i.e., W is stochastic. Furthermore, W ii < 1 for all i and W ij > 0 if and only if ij ∈ E. We focus on the linear DeGroot model [13] for opinion dynamics. There are K different discussions in the social net- work and each discussion is indexed by k ∈ [K]. Let xi(t; k) be the opinion of the ith agent2 at time t ∈ N during the kth discussion. Using the intuition that individuals' opinions are influenced by opinions of the others, the DeGroot model postulates that the agents' opinions are updated according to the following random process: (cid:88) j∈Ni xi(t; k) = Wij(t; k)xj(t − 1; k), (1) 2 For example, the ith agent's opinion xi(t; k) may represent a probability distribution function of his/her stance on the discussion. While our discussion is focused on the case when xi(t; k) is a scalar, it should be noted that the techniques developed can be easily extended to the vector case; see [1]. which can also be written as x(t; k) = W (t; k)x(t − 1; k), where x(t; k) = (x1(t; k), . . . , xn(t; k))T , [W (t; k)]ij = Wij(t; k) and W (t; k) is non-negative and stochastic, i.e., W (t; k)1 = 1 for all t, s. See [13] and [43, Chapter 1] for a detailed description of the aforementioned model. We assume the following regarding the random matrix W (t; k). (2) Assumption 1 W (t; k) is an independently and identically distributed (i.i.d.) random matrix drawn from a distribution that satisfies E{W (t; k)} = W for all t ∈ N, k ∈ [K]. The agents' opinions can be observed as: x(t; k) = x(t; k) + n(t; k), (3) x(t; k) =a.s. 1cT (s)x(0; k), where n(t; k) is i.i.d. additive noise with zero-mean and bounded variance. Eq. (1) & (3) constitute a time-varying linear dynamical system. Since W encodes both the network topology and the trust strengths, the social network sens- ing problem is to infer W through the set of observations { x(t; k)}t∈T ,k∈[K], where T is the set of sampling instances. A well known fact in the distributed control literature [19] is that the agents' opinions in Eq. (1) attain consensus almost surely as t → ∞: lim t→∞ (4) for some c(s) ∈ Rn, i.e., xi(t; k) = xj(t; k) for all i, j as t → ∞. In other words, the information about the network structure vanishes in the steady state equation. The aforementioned property of the DeGroot model has prompted most works on social network sensing (or network reconstruction in general) to assume a complete/partial knowl- edge of T such that the opinion dynamics can be tracked. In particular, [14] assumes a static model such that W (t; k) = W and infers W by solving a least square problem; [21], [22] deals with a time-varying, non-linear dynamical system model and applies sparse recovery to infer W . The drawback of these methods is that they require knowing the discrete time indices for the observations made. This knowledge may be difficult to obtain in general. In contrast, the actual system states are updated with an unknown interaction rate and the interaction timing between agents is not observable in most practical scenarios. Our active sensing method relies on observing the steady state opinions; i.e., the opinions when t → ∞. The observa- tions are based on T such that min(t ∈ T ) (cid:29) 0 and are thus robust to errors in tracking of the discrete time dynamics. Our approach is to reconstruct the network via a set of stubborn agents, as described next. A. Stubborn Agents (a.k.a. zealots) Stubborn agents (a.k.a. zealots) are members whose opin- ions can not be swayed by others. If agent i is stubborn, then xi(t; k) = xi(0; k) for all t. The opinion dynamics of these agents can be characterized by the structure of their respective rows in the trust matrix: Definition 1 An agent i is stubborn if and only if its cor- responding row in the trust matrix W (t; k) is the canonical basis vector; i.e., for all j, (cid:40) Wij(t; k) = 1, if j = i, 0, otherwise, ∀ t, k (5) Note that stubborn agents are known to exist in social net- works, as discussed in the Introduction. (cid:18) Ins Suppose that there exists ns stubborn agents in the social network and the social RADAR is aware of their existence. Without loss of generality, we let Vs (cid:44) [ns] be the set of stubborn agents and Vr (cid:44) V \Vs be the set of ordinary agents. The trust matrix W and its realization can be written as the (cid:19) following block matrices: (cid:19) (cid:18) Ins W = , W (t; k) = , (6) where B(t; k) and D(t; k) are the partial matrices of W (t; k), note that E{B(t; k)} = B and E{D(t; k)} = D. Notice that B is the network between stubborn and ordinary agents, and D is the network among the ordinary agents themselves. B(t; k) D(t; k) 0 B D 0 We further impose the following assumptions: Assumption 2 The support of B, ΩB = {ij : Bij > 0} = E(Vr, Vs), is known. Moreover, each agent in Vr has non-zero trust in at least one agent in Vs. Assumption 3 The induced subgraph G[Vr] = (Vr, E(Vr)) is connected. the principal submatrix D of W satisfies We are interested in the steady state opinion resulting from Consequently, (cid:107)D(cid:107)2 < 1. (1) at t → ∞. (cid:44) Observation 1 (z(t; k), y(t; k))T ∈ Rn where z(t; k) ∈ Rns and y(t; k) ∈ Rn−ns are the opinions of stubborn agents and ordinary agents, respectively. Consider (1) by setting t → ∞, we have: x(t; k) [29], [33] Let E{x(t; k)x(0; k)} = W ∞x(0; k), lim t→∞ where (cid:18) (cid:19) . (7) (8) W ∞ = Ins 0 (I − D)−1B 0 lim t→∞ The expected opinions of ordinary agents at t → ∞ are: E{y(t; k)z(0; k)} = (I − D)−1Bz(0; k). As seen, the steady state opinions are dependent on the stubborn agents and the structure of the network. Furthermore, unlike the case without stubborn agents (cf. (4)), information about the network structure D, B will be retained in the steady state equation (9). Leveraging on Observation 1, the next section formulates a regression problem that estimates D, B from observing opinions that fit these steady state equations. (9) 3 Fig. 1. Data matrices used in the social network sensing problem (cf. (10)). III. SOCIAL NETWORK SENSING VIA STUBBORN AGENTS We now study the problem of social network sensing using stubborn agents. Instead of tracking the opinion evolution in the social networks similar to the passive methods [14], [21], [22], our method is based on collecting the steady-state opinions from K ≥ ns discussions. Define the data matrices: ∈ R(n−ns)×K, (10a) (10b) (cid:17) E{y(∞; 1)} ··· E{y(∞; K)} ∈ Rns×K. z(0; 1) ··· z(0; K) Y (cid:44)(cid:16) Z (cid:44)(cid:16) (cid:17) Notice that (9) implies that Y = (I − D)−1BZ. Fig. 1 illustrates the relationship between the data matrices/vectors used in the social network sensing problem. One often obtains a noisy estimate of Y ; i.e., Y = (I − D)−1BZ + N . (11) We relegate the discussion on obtaining Y when the agents' interactions are asynchronous to Section IV-B. From Eq. (11), a natural solution to estimate (B, D) is to obtain a tuple (B, D) that minimizes the square loss (cid:107) Y − F . However, the system equation (11) admits (I − D)−1BZ(cid:107)2 an implicit ambiguity: Lemma 1 Consider the pair of tuples, (B, D) and ( B, D), where B1 + D1 = 1 and the latter is defined as B = ΛB, off( D) = Λoff(D), diag( D) = 1 − Λ(B1 + off(D)1), (12a) (12b) for some diagonal matrix Λ > 0 such that diag( D) ≥ 0. Under (12), the equalities (I − D)−1B = (I − D)−1 B and B1 + D1 = 1 hold. The proof is relegated to Appendix A. Lemma 1 indicates that there are infinitely many tuples ( B, D) with the same product (I − D)−1 B. The diagonal entries of D; i.e., the magnitude of self trust, are ambiguous. In fact, the ambiguity described in Lemma 1 is due to the fact that the collected data Y , Z lacks information about the rate of social interaction. Social Network5101520253035404550−50510152025Stubborn AgentsZ2Rns⇥K5101520253035404550051015202530ProcessY=(ID)1BZS=(V,E,W)Y2Rn⇥KProcessBy Observation 1:z(0;k)4 In light of Lemma 1, we determine an invariant quantity among the ambiguous solutions. Define the equivalence class: (B, D) ∼ ( B, D) : ∃ Λ > 0 s.t. (12) holds, diag( D) ≥ 0. (13) A quick observation on (12) is that Λ modifies the magnitude of diag(D). This inspires us to resolve the ambiguity issue by imposing constraints on diag(D): Observation 2 The pair of relative trust matrices resulting from (B, D), denoted by the superscript (·)(cid:48): B(cid:48) = (I − Diag(c))Λ−1 off(D(cid:48)) = (I − Diag(c))Λ−1 s off(D), s B, diag(D(cid:48)) = c, (14) where Λs = I − Diag(diag(D)) and 0 ≤ c < 1, is unique for each equivalence class when c is fixed. In other words, if (B, D) ∼ ( B, D), their resultant pairs of relative trust matrices are equal, (B(cid:48), D(cid:48)) = ( B(cid:48), D(cid:48)). Notice that the pair of relative trust matrices is an important quantity for a social network since it contains the relative trust strengths and connectivity of the network. We are now ready to present the regression problems pertaining to recovering B(cid:48), D(cid:48). As we often have access to a superset S of the support of D; i.e., ΩD ⊆ S and ΩD denotes the support of D, we propose two different formulations when different knowledge on ΩD is accessible. Case 1: nearly full support information - When ΩD ⊆ S and S ≈ ΩD, the support of D is mostly known. We formulate the following least square problem: D,B (cid:107)(I − D) Y − BZ(cid:107)2 min s.t. B1 + D1 = 1, D ≥ 0, B ≥ 0, F PΩc B (B) = 0,PS c (D) = 0, diag(D) = c. The projection operator PΩ(·) is defined as: Aij, ij ∈ Ω, 0, otherwise, [PΩ(A)]ij = (cid:40) (15a) (15b) (15c) (16) where Ω ⊆ [n]×[m] is an index set for the matrix A ∈ Rn×m. Problem (15) is a constrained least-square problem that can be solved efficiently, e.g., using cvx [44]. Case 2: partial support information - When ΩD ⊆ S and S (cid:29) ΩD, the system equation (11) constitutes an undetermined system. This motivates us to exploit that D is sparse and consider the sparse recovery problem: (17a) D,B (cid:107)vec(D)(cid:107)0 min s.t. (cid:107)(I − D) Y − BZ(cid:107)2 B F ≤ , B1 + D1 = 1, (B) = 0, PS c (D) = 0, D ≥ 0, B ≥ 0, diag(D) = c, PΩc (17b) (17c) (17d) for some ≥ 0 and vec(D) denotes the vectorization of the matrix D. Note that Problem (17) is an (cid:96)0 minimization problem that is NP-hard to solve in general. In practice, the following convex problem is solved in lieu of (17): B,D (cid:107)vec(D)(cid:107)1 s.t. Eq. (17b) − (17d). min (18) Remark 1 A social network sensing problem similar to Case 1 has been studied in [31] and its identifiability conditions have been discussed therein. In fact, we emphasize that our focus in this paper is on the study of Case 2, which presents a more challenging problem due to the lack of support information. A. Identifiability Conditions for Social Network Sensing We derive the conditions for recovering the relative trust matrix (B(cid:48), D(cid:48)) resulting from (B, D) by solving (15) or (17). As a common practice in signal processing, the following analysis assumes noiseless measurements such that σ2 = 0. We set = 0 in (17b) and assume K ≥ ns. objective value of zero. Now, let us denote As σ2 = 0, the optimization problem (15) admits an optimal (19) A (cid:44)(cid:0) Y T ZT (cid:1) . (cid:18) di (cid:19) The identifiability condition for (15) and (17) can be analyzed by studying the linear system: yi = A , ∀ i, Eq. (15b) − (15c), bi (20) where yi is the ith column of Y , di, bi are the ith row of D, B, respectively. Note that the above condition is equivalent to (17b)–(17d) with = 0. In the following, we denote Si as the index sets restricted to the ith row of D and and Ωi B B. Case 1: nearly full support information - Due to the constraints in (15c), the number of unknowns in the right hand side of the first equality in (20) is Ωi B +Si− 1. From linear algebra, the following is straightforward to show: Proposition 1 The system (20) admits a unique solution if rank( A:,Si∪Ωi B ) ≥ Ωi B + Si − 1, ∀ i, (21) B . is a submatrix of A with the columns taken where A:,Si∪Ωi from the respective entries of Si and Ωi Similar result has also been reported in [31, Lemma 10]. Case 2: partial support information - As S (cid:29) ΩD, this case presents a more challenging scenario to analyze since (20) is an underdetermined system. In particular, (21) is often not satisfied. However, a sufficient condition can be given by exploiting the fact that (17) finds the sparsest solution: B Proposition 2 There is a unique optimal solution to (17) if for all Si ⊆ Si and Si ≤ 2Ωi ) ≥ Ωi D, we have B + 2Ωi rank( A:, Si∪Ωi D − 1, ∀ i, (22) B B . is a submatrix of A with the columns taken where A:, S∪Ωi from the respective entries of Si and Ωi Proposition 2 is equivalent to characterizing the spark of the matrix A; see [45]. Checking (22) is non-trivial and it is not clear how many stubborn agents are needed. We next show that using an optimized placement of stubborn agents, one can derive a sufficient condition for unique recovery using (17). B B. Optimized placement of stubborn agents We consider an optimized placement of stubborn agents when only partial support information is given. In other words, we design ΩB for achieving better sensing performance. This is possible in a controlled experiment where the stubborn agents are directly controlled to interact with a target group of ordinary agents. Note also that [25] has studied an optimal stubborn agent placement formulation for the voter's model opinion dynamics with a different objective. To fix ideas, the following discussion draws a connection between blind compressed sensing and our social network sensing problem. Consider the following equivalent form of (17b): (Y Z†)T = BT (I − D)−T , (23) (cid:48) where Z† denotes the psuedo-inverse of Z. By noting that Y Z† = (I − D(cid:48))−1B(cid:48), we have: T T B (cid:48) T ⇐⇒ B (cid:48) ⇐⇒ B (I − D(cid:48))−T = BT (I − D)−T (I − D(cid:48))−T (D(cid:48) − D)T = (B − B(cid:48))T (I − D(cid:48))−T (d(cid:48) i − di) = bi − b(cid:48) i, ∀ i, where d(cid:48) i and bi are the ith row of D(cid:48), D, B(cid:48) and B, respectively. Due to the constraint PS c (D) = 0 and is ni (cid:44) diag(D) = c, is the complement n − ns − si − 1, where si = S c of support set restricted to the ith row of D. the number of unknowns in di i and S c i, di, b(cid:48) (24) i T The matrix B (I − D(cid:48))−T can be regarded as a sensing matrix in (24). Note that because B(cid:48) is unknown, we have a blind compressed sensing problem, which was studied in [37], [38]. However, the scenarios considered there were different from ours since the zero values of the sensing matrix are partially known in that case. (cid:48) (cid:48) T (I + D(cid:48) + (D(cid:48))2 + . . .)T ; To study an identifiability condition for (24), we note that (cid:48) (I − D(cid:48))−T = B T i.e., B T . When the the sensing matrix is equal to a perturbed B perturbation is small, we could study B alone. Moreover, as the entries in B are unknown, it is desirable to consider a sparse construction to reduce the number of unknowns. (cid:48) As suggested in [41], [42], a good choice is to construct ΩB randomly such that each row in B has a constant number d of non-zero elements (independent of n(cid:48)i). We have the following sufficient condition: Theorem 1 Define supp(B) = {ij : Bij > 0}, bmin = minij∈supp(B) B(cid:48) ij, βi = ns/ni and β(cid:48)i = βi − d/ni. Let the support of B(cid:48) ∈ Rn×ns be constructed such that each row has d non-zero elements, selected randomly and independently. If ij, bmax = maxij∈supp(B) B(cid:48) (cid:110) H(αi) + β(cid:48)H(αi/β(cid:48)i) (cid:111) d > max 4, 1 + , αi log(β(cid:48)i/αi) and bmin(2d − 3) − 1 − 2bmax > 0, where H(x) is the binary entropy function, and (cid:107)d(cid:48) i(cid:107)0 ≤ αini/2 for all i, then as n−ns → ∞, there is a unique optimal (25) (26) 5 solution to (17) that (B(cid:63), D(cid:63)) = (B(cid:48), D(cid:48)) with probability one. Moreover, the failure probability is bounded as: Pr((B(cid:63), D(cid:63)) (cid:54)= (B(cid:48), D(cid:48))) (cid:19)4 d − 1 (cid:18) d βi ≤ max i 2 + O(ni ni 2−(d−1)(d−3)). (27) (cid:48) The proof of Theorem 1 is in Appendix B where the claim is T as erasure bits, proven by treating the unknown entries of B and showing that the sensing matrix with erasure corresponds to a high quality expander graph with high probability. To the best of our knowledge, Theorem 1 is a new recoverability result proven for blind compressed sensing problems. Condition (25) gives a guideline for choosing the number of stubborn agents ns employed. In fact, if we set α → 0 while keeping the ratio β/α constant, condition (25) can be approximated by β > α = 2p, where (cid:107)d(cid:48) i(cid:107)0 ≤ p(n − ns). Notice that in the limit as n − ns → ∞, if every ordinary agent in the sub-network that corresponds to D has an in- degree bounded by p(n − ns), then we only need: ns ≥ 2p(n − ns) (28) stubborn agents to perfectly reconstruct D. On the other hand, condition (26) indicates that the amount of relative trust in stubborn agents cannot be too small. This is reasonable in that the network sensing performance should depend on the influencing power of the stubborn agents. The effect of the known support is also reflected. In particular, an increase in si leads to a decrease in ni and an increase in βi. The maximum allowed sparsity αi is increased as a result. We conclude this subsection with the following remarks. Remark 2 When n is finite, the failure probability may grow with the size of ΩB, i.e., it is O(d5), coinciding with the intuition concerning the tradeoff between the size of ΩB and the sparse recovery performance. Although the failure probability vanishes as n − ns → ∞, the parameter d should be chosen judiciously in the design. Remark 3 Another important issue that affects the recover- ability of (17) is the degree distribution in G as the conditions in Theorem 1 requires the sparsity level of D to be small for every row. Given a fixed total number of edges, it is easier to recover a network with a concentrated degree distribution (e.g., the Watts-Strogatz network [46]) while a network with power law degree distribution (e.g., the Barabasi-Albert network [47]) is more difficult to recover. Remark 4 The requirement on the number of stubborn agents in Theorem 1 may appear to be restrictive. However, note that only the expected value B is considered in the model. Theoretically, one only need to employ a small number of stubborn agents that randomly wander in different positions of the social network and 'act' as agents with different opinion, thus achieving an effect similar to that of a synthetic aperture RADAR. This effectively creates a vast number of virtual stubborn agents from a small number of physically present stubborn agents. 6 IV. IMPLEMENTATIONS OF SOCIAL NETWORK SENSING In this section we discuss practical issues involved in imple- menting our network sensing method. First, we develop a fast algorithm for solving large-scale network sensing problems. Second, we consider a random opinion dynamics model and propose a consistent estimator for the steady state. A. Proximal Gradient Method for Network Sensing This subsection presents a practical implementation method for large-scale network sensing when n (cid:29) 0. We resort to a first order optimization method. The main idea is to employ a proximal gradient method [48] to the following problem: 4: 5: λ(cid:107)vec(D)(cid:107)1 + f (B, D) min D,B s.t. B ≥ 0, D ≥ 0, PS c(D) = 0, (B) = 0, diag(D) = c, PΩc B (29) (30) (31) where f (B, D) = (cid:107)(I − D) Y Z† − B(cid:107)2 F + γ(cid:107)B1 + D1 − 1(cid:107)2 and γ > 0. Note that Z† is the right pseudo-inverse of Z. Problem (29) can be seen as the Lagrangian relaxation of (18). Meanwhile, similar to the first case considered in Section III, we take λ = 0 when S ≈ ΩD. denote the feasible set of (29) as L: The last term in (29) is continuously differentiable. Let us 2 L ={(B, D) : B ≥ 0, D ≥ 0, PS c (D) = 0 PΩc B (B) = 0, diag(D) = c}. Let (cid:96) ∈ N be the iteration index of the proximal gradient method. At the (cid:96)th iteration, we perform the following update: (B(cid:96), D(cid:96)) = arg min (B,D)∈L αλ(cid:107)vec(D)(cid:107)1 where +(cid:107)B − B(cid:96)−1(cid:107)2 F + (cid:107)D − D(cid:96)−1(cid:107)2 D(cid:96) = D(cid:96) − α∇Df (B(cid:96), D)(cid:12)(cid:12)D=D(cid:96) B(cid:96) = B(cid:96) − α∇Bf (B, D(cid:96))(cid:12)(cid:12)B=B(cid:96), (32a) (32b) and α > 0 is a fixed step size such that α < 1/L, where L is the Lipschitz constant for the gradient of f. F Importantly, the proximal update (31) admits a closed form solution using the soft-thresholding operator: B ( B(cid:96)−1)} B(cid:96) = max{0,PΩc off(D(cid:96)) = PS c (soft_thαλ(off( D(cid:96)−1))), (33a) (33b) where soft_thλ(·) is a one-sided soft thresholding oper- ator [49] that applies element-wisely and soft_thλ(x) = u(x) max{0, x − λ}. To further accelerate the algorithm, we adopt an update rule similar to the fast iterative shrinkage (FISTA) algorithm in [49], as summarized in Algorithm 1. As seen, the per-iteration complexity is O(n2 + n · ns); i.e., it is linear in the number of variables. We conclude this subsection with a discussion on the convergence of Algorithm 1. As (29) is convex and f is con- tinuously differentiable, the proximal gradient method using (31) is guaranteed [49] to converge to an optimal solution of (29). Moreover, the convergence speed is O(1/(cid:96)2). Algorithm 1 FISTA algorithm for (29). 1: Initialize: (B0, D0) ∈ L, t0 = 0, (cid:96) = 1; 2: while convergence is not reached do 3: Compute the proximal gradient update direction: dB(cid:96) ← max{0,PΩc dD(cid:96) ← PS c(soft_thαλ(off( D(cid:96)−1))), ( B(cid:96)−1)} B (cid:113) where B(cid:96)−1, D(cid:96)−1 are given by (32). Update t(cid:96) ← (1 + Update the variables as: (cid:96)−1)/2. 1 + 4t2 Bk ← dB(cid:96) + off(D(cid:96)) ← dD(cid:96) + t(cid:96)−1 − 1 t(cid:96)−1 − 1 t(cid:96) t(cid:96) (dB(cid:96) − dB(cid:96)−1) (dD(cid:96) − dD(cid:96)−1), (cid:96) ← (cid:96) + 1. 6: 7: end while 8: Set diag(D(cid:96)) ← c. 9: Return: (B(cid:96), D(cid:96)). B. Random Opinion Dynamics So far, our method for network sensing only requires collect- ing the asymptotic states E{y(∞; k)x(0; k)}. Importantly, the data matrices Y and Z (cf. (10)) can be collected easily when the trust matrix W (t; k) is deterministic; i.e., W (t; k) = W for all t, k, and the observations are noiseless. For the case where D(t; k) is random, computing the expectation may be difficult since the latter is an average taken over an ensemble of the sample paths of {W (t; k)}∀t,∀k. This section considers adopting the proposed active sens- ing method to the case with randomized opinion dynamics, which captures the possible randomness in social interac- tions. Our idea is to propose a consistent estimator for E{y(∞; k)z(0; k)} using the opinions gathered from the same time series { x(t; k)}t∈Tk, where Tk ⊆ Z+ is now an arbitrary sampling set. Specifically, we show that the random process x(t; k) is ergodic. We first make an interesting observation pertaining to random opinion dynamics: Observation 3 [26, Theorem 2] [30] When ns ≥ 2 and under a random opinion dynamics model, the opinions will not converge; i.e., x(t; k) (cid:54)= x(t − 1; k) almost surely. Observation 3 suggests that a natural approach to computing the expectation E{y(∞; k)x(0; k)} is by taking averages over the temporal samples. We propose the following estimator: E{x(∞; k)x(0; k)} ≈ x(Tk; k) (cid:44) 1 Tk (cid:88) x(t; k), (36) t∈Tk the definition for x(t; k) where we recall from (3). Notice that E{y(∞; k)x(0; k)} can be retrieved from E{x(∞; k)x(0; k)} by taking the last n − ns elements of the latter. In order to compute the right hand side in (36), we only need to know the cardinality of Tk; i.e., the number of samples collected. Knowledge on the members in Tk is not required. Specifically, the temporal samples can be col- lected through random (and possibly noisy) sampling at time instances on the opinions. The following theorem characterizes the performance of (36): Theorem 2 Consider the estimator in (36) with a sampling set Tk. Denote x(∞; k) (cid:44) limt→∞ E{x(t; k)x(0; k)} = W ∞x(0; k) and assume that E{(cid:107)D(t; k)(cid:107)2} < 1. If To → ∞, then 1) the estimator (36) is unbiased: E{ x(Tk; k)z(0; k)} = x(∞; k). 2) the estimator (36) is asymptotically consistent: (37) lim Tk→∞ E{(cid:107) x(Tk; k) − x(∞; k)(cid:107)2 For the latter case, we have 2x(0; k)} = 0. (38) (cid:17) (39) 2x(0; k)} , E{(cid:107) x(Tk; k) − x(∞; k)(cid:107)2 λmin(cid:96) t(cid:96)+i−t(cid:96) (cid:16) Tk−1(cid:88) i=0 ≤ C(cid:48) Tk where C(cid:48) < ∞ is a constant and λ = λmax(D) < 1, i.e., the latter term is a geometric series with bounded sum. Note that a similar result to Theorem 2 was reported in [50]. Our result is specific to the case with stubborn agents, which allows us to find a precise characterization of the mean square error. The proof of Theorem 2 can be found in Appendix C. Remark 5 From (39), we observe that the upper bound on the mean square error can be optimized by maximizing mini,j,i(cid:54)=j ti − tj. Suppose that the samples x(Tk; k) have to be taken from a finite interval [Tmax] \ [To], Tmax < ∞ the best estimate can be obtained and Tk < ∞; here, by using sampling instances that are drawn uniformly from [Tmax] \ [To]. V. NUMERICAL RESULTS To validate our methods, we conducted several numerical simulations, reconstructing both synthetic networks and real networks. In this section, we focus on cases where the network dynamics model (2) is exact (while the measurement may be noisy),but emphasize the crucial importance of considering data collected from real networks, e.g., the online social net- works (e.g., Facebook, Twitter). The Monte-Carlo simulations were obtained by averaging over at least 100 simulation trials. We also set K = 2ns and c = 0 to respect the requirement K ≥ ns and for ease of comparison. A. Synthetic networks with noiseless measurement We evaluate the sensing performance on a synthetic network with noiseless measurement on the steady system state. In light of Theorem 1, for the placement of the stubborn agents, we randomly connect d stubborn agents to each ordinary agent. The first numerical example compares performance in re- covering D(cid:48) and B(cid:48) against the number of stubborn agents 7 Fig. 2. Comparing performance against the number of stubborn agents ns. (Left) The NMSE. (Right) The average support recovery error. In the legend, 'full' denotes the case with full support information; '(ps=0.1)' and '(d=5)' denotes the case where B is constructed as a random bipartite graph and a random d-regular bipartite graph, respectively. F /(cid:107)D(cid:48)(cid:107)2 ns. We fix the number of ordinary agents at n − ns = 60. The network G is constructed as an Erdos-Renyi (ER) graph with connectivity p = 0.1. Furthermore, the weights in W are generated uniformly first, and normalized to satisfy W 1 = 1 afterwards. As the problem size considered is moderate (n ≤ 100), the network reconstruction problem (18) is solved using cvx. We present the normalized mean square error (NMSE) under the above scenario in Fig. 2. The NMSE of D(cid:48) is defined as (cid:107) D−D(cid:48)(cid:107)2 F (and similarly for B(cid:48)). The NMSE against ns is shown for two cases: (i) solving (15) when S = ΩD; (ii) solving (18) when S = [n−ns]×[n−ns]. We also include the NMSE curve when B corresponds to a random ER bipartite graph with edge connectivity p = 0.1. The figure shows, first that only ns ≈ 20 stubborn agents are needed when the full support information is given. By contrast, we need ns ≈ 40 to attain a similar NMSE when there is no support information. Comparing the NMSE to d-regular/random graph construction for B shows that the recovery performance is significantly better when using the d-regular graph construction; e.g., if d = 5, the NMSE of D(cid:48) is less than 10−3 with ns ≥ 33. This implies that by inserting almost half the number of ordinary agents into the network, the social network structure can be revealed with high accuracy. This result is consistent with the Theorem 1, which predicts that when β ≥ 0.604, i.e., ns ≈ 36, perfect recovery can be achieved. The discrepancy between the simulation results and Theorem 1 is possibly due to the fact that we are solving (18) instead of (17). Moreover, in an ER graph, d(cid:48) i is only p(n− ns)-sparse on average, but Theorem 1 requires that every row of D(cid:48) to be p(n − ns)-sparse. In the second example, we examine the scenarios when G is constructed as the Barabasi-Albert (BA) graph with minimum degree m = 2 for each incoming node [47] or the Strogatz- Watts (SW) graph where each node is initially connected to b = 2 left and right nodes and the rewiring probability is p = 0.08 [46]. The results are shown in Fig. 3. It shows that the SW network can be recovered with high accuracy by using No. of stubborn agents (ns)1020304050MSE10-710-610-510-410-310-210-1100MSE of D (Full)MSE of B (Full)MSE of D (ps=0.1)MSE of B (ps=0.1)MSE of D (d=5)MSE of B (d=5)No. of stubborn agents (ns)1020304050Support Error0200400600800100012001400False Alarm (ps=0.1)Miss (ps=0.1)False Alarm (d=5)Miss (d=5)8 Fig. 3. Comparing performance against the number of stubborn agents ns with different network models. (Left) The NMSE. (Right) The average support recovery error. Fig. 4. Comparing the NMSE against the number of ordinary agents n− ns. Fix p = 0.08 and β is given by Theorem 1 as 0.528 (d = 5), 0.385 (d = 6) and 0.319 (d = 7). a much smaller number of stubborn agents than either the ER or BA networks. One possible explanation is that most of the vertices in SW have the same degree. The third numerical example examines the claim in The- orem 1. Recall that the latter requires n − ns → ∞ for its validity. We consider the ER graph as in the first example and fix the connectivity of G at p = 0.08 where the smallest β required by Theorem 1 are respectively 0.528 (d = 5), 0.385 (d = 6) and 0.319 (d = 7). We set ns = (cid:100)β(n − ns)(cid:101) and vary the number of ordinary agents n− ns to compare the NMSE. The NMSE comparison against n − ns can be found in Fig. 4. In all three cases tested (d = 5, 6, 7), there is a decreasing trend of the NMSE with n − ns, suggesting that the failure probability decreases with n− ns → ∞. Moreover, although d = 7 places the least requirement on β, it also has the highest probability of failure when n is finite, since the upper bound to failure probability grows with O(d5). The simulation results suggest that d needs to be chosen judiciously when deciding on the required number/placement of stubborn agents. Fig. 5. indices in Ωc W Comparing the NMSE against the percentage of known sparsity , i.e., S decreases when pknown increases . The next simulation example in Fig. 5 examines the case where a superset S of ΩD is known. In particular, we consider the ER network model with n− ns = 60 and connectivity p = 0.1 and compare the NMSE against the percentage of exposed . The figure shows that the network sensing performance Ωc D improves as pknown increases. When 40% of the support of D is exposed, employing ns = 28 stubborn agents yields a satisfactory NMSE of 10−3. B. Real world networks This subsection examines the performance of the proposed method applied to real network data. Specifically, we consider the facebook100 dataset [51] and focus on the medium- sized network example ReedCollege. The randomized opinion exchange model is based on the randomized broadcast gossip protocol in [52] with uniformly assigned trust weights. Out of the available agents, we picked ns = 180 agents with degrees closest to the median degree as the stubborn agents and removed the agents that are not adjacent to any of the stub- born agents. The selection of the stubborn agents is motivated by Theorem 1 as we require a moderate average degree for the resultant stubborn-to-nonstubborn agent network with better recovery guarantees. Our aim is to estimate the trust matrix D, which corresponds to the subgraph with n − ns = 666 ordinary agents, E = 13, 269 edges and mean degree 19.92. Note that the bipartite graph from stubborn agents to ordinary agents has a mean degree of 25.07. The opinion dynamics data Y is collected using the es- timator in Section IV-B, where we set Tk = 5 × 105 and the sampling instances are uniformly taken from the interval [105, 5 × 107]. We apply the FISTA algorithm developed in Section IV to approximately solve the network reconstruction problem (18), with λ = n × 10−12 and γ = 10−3/γ. The NMSE of the reconstructed D(cid:48) is 0.1035 after 4 × 104 iterations. The program has terminated in about 30 minutes on an IntelTM XeonTM server running MATLABTM 2014b. It is computationally infeasible to deploy generic solvers such as cvx as the number of variables involved is 563, 436. We compared the estimated social network in both macro- scopic and microscopic levels. Fig. 6 shows the true/estimated No. of stubborn agents (ns)253035404550MSE10-1010-810-610-410-2100MSE of D (BA)MSE of B (BA)MSE of D (SW)MSE of B (SW)No. of stubborn agents (ns)20253035404550Support Error050100150200250300350400False Alarm (BA)Miss (BA)False Alarm (SW)Miss (SW)2025300510152025303540No. of ordinary agents (n-ns)2030405060708090100MSE10-710-610-510-410-310-210-1100MSE of D (d=5)MSE of B (d=5)MSE of D (d=6)MSE of B (d=6)MSE of D (d=7)MSE of B (d=7)pknown00.10.20.30.40.50.60.70.8MSE10-1010-810-610-410-2MSE of D (ns = 24)MSE of B (ns = 24)MSE of D (ns = 28)MSE of B (ns = 28)MSE of D (ns = 32)MSE of B (ns = 32)9 Fig. 6. Comparing the social network of ReedCollege from the facebook100 dataset: (Left) the original network; (Right) the estimated network. a sparse recovery problem and a sufficient condition for perfect recovery was proven. In addition, a consistent estimator was also derived to handle the case where the network dynamics is random and a low complexity algorithm is proposed. Our simulation results on synthetic and real networks indicate that the network structure can be reconstructed with high accuracy when a sufficient number of stubborn agents is present. Ongoing research is focused on extending the active sens- ing method to nonlinear opinion dynamics models such as the Hegselmann and Krause model, working on real social network data collected from social media and combining this approach with the detection of stubborn agents. ACKNOWLEDGEMENT The authors are indebted to the anonymous reviewers for their invaluable comments to improve the current paper. APPENDIX A PROOF OF LEMMA 1 It is easy to check that: B1 + D1 = Λ(B1 + off(D)1) + diag( D) = 1, (40) where the last equality is due to B1 + D1 = 1. Furthermore, (I − D)−1 B = (I − Λoff(D) − Diag(diag( D)))−1ΛB. (41) As Diag(diag( D)) = I − Λ + ΛDiag(diag(D)), we have: (42) (I − D)−1 B = (I − D)−1B. APPENDIX B PROOF OF THEOREM 1 With a slight abuse of notations, in this appendix we assume that there are n ordinary agents and ns stubborn agents. In particular, the dimensions of the variables are D(cid:48) ∈ Rn×n and B(cid:48) ∈ Rn×ns. For simplicity, we set ni = n, αi = α and βi = β for all i. The proof of Theorem 1 is divided into two parts. The first part shows a sufficient condition for recovering (B(cid:48), D(cid:48)) using (17); and the second part shows that the sufficient condition holds with high probability as n → ∞. Fig. 7. Comparing the reconstructed network for the ReedCollege social network in facebook100 dataset. (Left) Original network. (Right) Reconstructed network. network plotted in gephi [53] using the 'Force Atlas 2' layout (with random initialization), where the edge weights are taken into account3. While it is impossible to compare every edges in the network, the figure gives a macroscopic view of the efficacy of the network reconstruction method. In partic- ular, using ns = 180 stubborn agents, the estimated network follows a similar topology as the original one. For instance, there are clearly two clusters in both the estimated and original network. Moreover, the relative roles for individual agents are matched in both networks. For example, agents {39, ..., 608} are found in the larger cluster, agents {378, ..., 663} are found at the boundary between the clusters and agents {43, ..., 404} are found in the smaller cluster, in both networks. Finally, in Fig. 7 we compare the estimated principal sub- matrix of D(cid:48) taken from the first 60 rows/columns, i.e., this corresponds to the social network between 60 agents. As seen, the original and estimated matrices appears to be similar to each other, both in terms of the support set and the weights on individual edges between the agents. VI. CONCLUSIONS In this paper, we considered the social network sensing problem using data collected from steady states of opinion dy- namics. The opinion dynamics is based on an extension of the linear DeGroot model with stubborn agents, where the latter plays a key role in exposing the network structure. Our main result is that the social network sensing problem can be cast as 3Readers are advised to read the figures on a color version of this paper. 5101520253035404550556051015202530354045505560510152025303540455055605101520253035404550556010 Fig. 8. Illustrating the properties of the expander graph. In the above example bipartite graph, if α = 1/3, δ is at most 3/4 since E(S, B) = 4 and N (S) = 3 when S is the first two vertices in the set of ordinary agents. Let d(v) denote the degree of a vertex v. Our proof relies on the following definition of an unbalanced expander graph: Definition 2 An (α, δ)-unbalanced expander graph is an A, B-bipartite graph (bigraph) with A = n,B = m with left degree bounded in [dl, du], i.e., d(vi) ∈ [dl, du] for all vi ∈ A, such that for any S ⊆ A with S ≤ αn, we have δE(S, B) ≤ N (S), where E(S, B) is the set of edges connected from S to B and N (S) = {vj ∈ B : ∃ vi ∈ S s.t. vjvi ∈ E} is the neighbor set of S in B. We imagine that A (B) is the set of ordinary (stubborn) agents and E(A, B) represents the connection between stubborn and ordinary agents; see the illustration in Fig. 8. We denote the collection of (α, δ)-unbalanced expander graphs by G(α, δ). Previous works [39]–[42] have shown that the expander graph structure allows for the construction of measurement matrices with good sparse recovery performance. We now proceed by showing the sufficient condition. De- note the support of bi − b(cid:48) B = d. Since is known a-priori, bi − b(cid:48) i is a sparse vector supported Ωi B on Ωi . We can thus treat the rows where bi is supported B on as 'erasure bits', which can be ignored. In particular, the following rows-deleted linear system can be deduced from the last line in (24): (cid:48) , where Ωi i as Ωi B B T (Ωi B )c(I − D(cid:48))−T (d(cid:48) i − di) = 0, (43) (cid:48) (cid:48) where B T . We prove the sufficient condition by deriving a Restricted )c is a d-rows-deleted matrix obtained from B T (Ωi B Isometry Property-1 (RIP-1) condition for A = BT and its perturbation A(I − D(cid:48))−T . We define amin = minij∈supp(A) Aij and amax = maxij∈supp(A) Aij and prove the following proposition: (ΩBi )c Proposition 3 Let n > m and A ∈ Rm×n be a non-negative matrix that has the same support as the adjacency matrix of an (α, δ)-unbalanced bipartite expander graph with bounded left degrees [dl, du]. Then A satisfies the RIP-1 property: (cid:0)aminδdl − amax(du − δdl)(cid:1) (cid:107)x(cid:107)1 ≤ (cid:107)Ax(cid:107)1 ≤ duamax(cid:107)x(cid:107)1, (44) for all k-sparse x such that k ≤ αn. Furthermore, we have (45) υ(cid:63) · (cid:107)x(cid:107)1 ≤ (cid:107)A(I − D(cid:48))−T x(cid:107)1, where υ(cid:63) = aminδdl − amax(du − δdl) − (1 − dlamin). Proof. The following proof is a generalization of [42, Ap- pendix D]. First of all, the upper bound in (44) follows from (cid:107)Ax(cid:107)1 ≤ (cid:107)A(cid:107)1,1(cid:107)x(cid:107)1, where (cid:107)A(cid:107)1,1 is the matrix norm induced by (cid:107) · (cid:107)1 on A [54], i.e., (cid:107)A(cid:107)1,1 = max 1≤j≤n m(cid:88) Aij. (46) i=1 Obviously we have (cid:107)A(cid:107)1,1 ≤ duamax. property, we observe that To prove the lower bound in (44), using the expander δdlS ≤ δE(S, B) ≤ N (S), (47) for all S ⊆ supp(x) = {i : xi (cid:54)= 0} and S ≤ αn. As a consequence of Hall's theorem [55], the bigraph induced by A contains δdl disjoint matchings for supp(x). We can thus decompose A as: A = AM + AC, (48) where the decomposition is based on dividing the support such that supp(AM ) ∩ supp(AC) = ∅. In particular, AM is supported on the δdl matchings for supp(x); i.e., by the matching property, each row of AM has at most one non-zero, and each column of AM has δdl non-zeros, and the remainder AC has at most du − δdl non-zeros per column. Applying the triangular inequality gives: (50) (cid:107)Ax(cid:107)1 ≥ (cid:107)AM x(cid:107)1 − (cid:107)ACx(cid:107)1, (49) since (cid:107)AM x(cid:107)1 ≥ aminδdl(cid:107)x(cid:107)1 and (cid:107)ACx(cid:107)1 ≤ amax(du − δdl)(cid:107)x(cid:107)1, this implies: (cid:0)aminδdl − amax(du − δdl)(cid:1) (cid:107)Ax(cid:107)1 ≥ (cid:107)x(cid:107)1. For the second part in the lemma, i.e., (45), note that: (cid:107)A(I−D(cid:48))−T x(cid:107)1 ≥ (cid:107)Ax(cid:107)1−(cid:107)AD(cid:48)T since A(I − D(cid:48))−T x = Ax + AD(cid:48)T quantity can be upper bounded by (cid:107)AD(cid:48)T ≤ (cid:107)A(cid:107)1,1(cid:107)D(cid:48)T ≤ duamax (cid:107)1,1(cid:107)(I − D(cid:48))−T(cid:107)1,1(cid:107)x(cid:107)1 (I − D(cid:48))−T x(cid:107)1 (cid:107)x(cid:107)1 ≤ (1 − dlamin)(cid:107)x(cid:107)1, (I−D(cid:48))−T x(cid:107)1, (51) (I − D(cid:48))−T x. The latter (52) (cid:107)D(cid:48)T (cid:107)1,1 1 − (cid:107)D(cid:48)T (cid:107)1,1 where in the second to last inequality, we used the property (cid:107)(I − C)−1(cid:107) ≤ 1/(1 − (cid:107)C(cid:107)) for any (cid:107)C(cid:107) < 1 [54]; and in the last inequality, we used the fact that 1 − duamax ≤ (cid:107)D(cid:48)T (cid:107)1,1 ≤ 1− dlamin (note that each row in D(cid:48) sums to at most 1 − dlamin and at least 1 − duamax). Combining (50), Q.E.D. (51) and (52) yields the desired inequality. A sufficient condition for (cid:96)0 recovery can be obtained by proving the following simple corollary: Corollary 1 Let the conditions from Proposition 3 on A holds. Suppose that both x1, x2 are (k/2)-sparse such that k ≤ αn and: A(I − D(cid:48))−T x1 = A(I − D(cid:48))−T x2, (53) Non-stubborn agentsStubborn agentsE(S,B)N(S)11 then x1 = x2 if υ(cid:63) = aminδdl − amax(du − δdl) − (1 − dlamin) > 0. (54) Proof. Observe that x1 − x2 is at most k-sparse, using Proposition 3, we have This implies that x1 = x2. υ(cid:63)(cid:107)x1 − x2(cid:107)1 ≤ (cid:107)A(I − D(cid:48))−T (x1 − x2)(cid:107)1 = 0. As d(cid:48) (55) Q.E.D. i is k/2-sparse, bmin ≤ amin and bmax ≥ amax, Eq. (26) and Corollary 1 guarantee that d(cid:48) is the unique solution out of all k/2-sparse vectors that di satisfies (43). This means that any di that satisfies (43) must be either d(cid:48) i or have (cid:107)di(cid:107)0 > (k/2). Since the optimization problem (17) finds the sparsest solution satisfying (43), we must have d(cid:63) i for i in (24) and we have all i. Furthermore, this implies b(cid:63) (B(cid:63), D(cid:63)) = (B, D(cid:48)). i = d(cid:48) i = b(cid:48) i (cid:48) T (Ωi B The second part of our proof shows that for all i, the support )c corresponds to an set of the d-rows-deleted matrix B (α, δ)-expander graph with high probability. Our plan is to first prove that the corresponding bipartite graph has a bounded degree r ∈ [d − 1, d] with high probability (w.h.p.), and then show that a randomly constructed bipartite with bounded degree r ∈ [d − 1, d] is also an expander graph w.h.p.. Now, let us observe the following proposition: Proposition 4 Let G be a random A, B-bigraph with A = n, B = ns = βn, constructed by randomly connecting d vertices from A to each vertex of B. All of the subgraphs G1, ..., Gn have left degree r ∈ [d−1, d] with high probability (as n → ∞) if each of these subgraphs is formed by randomly deleting d vertices from B in G. (cid:16) (cid:17) Proof. We lower bound the desired probability as follows: (cid:17) G1, ..., Gn = bipartite with (left) deg. r ∈ [d − 1, d] Pr (cid:16) i=1 (Gi = bipartite with min. deg. r < d − 1) ∪n = 1 − Pr (cid:16) (cid:17) ∪n ≥ 1 − n · Pr k=1 (d(vk) < d − 1, vk ∈ Ai, Ai ⊆ V (Gi)) ≥ 1 − n2 · Pr d(vk) < d − 1, vk ∈ Ai, Ai ⊆ V (Gi) (56) Note that the event described in the last term is equivalent to deleting at least 2 neighbors of vk ∈ Ai from B. As the neighbors of A are also randomly selected, the latter probability can be upper bounded by: (cid:16) = Pr(cid:0)d(vk) = 0 ∪ ··· ∪ d(vk) = d − 2(cid:1) (cid:18) d d(vk) < d − 1, vk ∈ Ai, Ai ⊆ V (Gi) (cid:18) d2 (cid:17) (cid:19)4 (cid:19)2 (cid:16) (57) Pr ≤ (d − 1) · (βn)2 = (d − 1) · βn , Plugging this back into (56) yields the desired result. Q.E.D. The proof of Theorem 1 is completed by the proposition: Proposition 5 Let G be a random A, B-bigraph with A = n, B = β(cid:48)n = ns − d, constructed by randomly connecting r ∈ [d − 1, d] vertices from A to each vertex of B. Then G is an (α, 1− 1/(d− 1))-expander graph with high probability if d ≥ 4, α < β(cid:48) and d−1 > (H(α)+β(cid:48)H(α/β(cid:48)))/α log(β(cid:48)/α). (cid:17) 1 (cid:16) (cid:16) r r ≤ d (58) (cid:17) (cid:17) (cid:19) r=d−1 Ei1,...,ir (cid:91) Pr(E1,...,r). (cid:18)n Pr ≤ Pr d−1≤r≤αn,1≤i1<i2<···<ir G /∈ G(α, 1 − 1/(d − 1)) Proof. The following proof is similar in flavor to the proof of [42, Proposition 1], with the additional complexity that the left degree is variable. For simplicity, we denote A as the adjacency matrix of G and let Ei1,...,ir be the event such that A:,i1,...,ir contains at least m − r + 1 zero rows, where A:,i1,...,ir is the submatrix formed by choosing the {i1, ..., ir} columns. Note that if r ≤ αn and Ei1,...,ir occurs, G /∈ G(α, 1−1/(d−1)) since (1−1/(d−1))E({i1, ..., ir}) ≥ r > r−1 = N ({{i1, ..., ir}). The failure probability can thus be upper bounded as: (cid:16) (cid:16) αn(cid:88) (cid:0) βn (cid:1)s possible sub-matrices to choose from. Now, (cid:1)r−s(cid:0)βn Suppose that there are r − s columns with d − 1 non-zero entries and s columns with d non-zero entries; hence we have There are at most (cid:0) r (cid:1)s possible configurations and (cid:1)r−s(cid:0)r d−1 a necessary condition for E1,...,r is such that all the non- (cid:1) such sub-sub-matrices. For this case, we obtain the upper (cid:0)βn zero entries are contained in a sub-sub-matrix of size r × r. (cid:1)s (cid:1)r−s(cid:0)r (cid:1)(cid:0) r (cid:0)β (cid:1)s (cid:1)r−s(cid:0)β(cid:48)n (cid:0) β(cid:48)n (cid:19)(r−s)(d−1) (cid:18)β(cid:48)n (cid:18) r (cid:18) r (cid:19) (cid:19)(r−s)(d−1)+sd (cid:18)β(cid:48)n (cid:19) (cid:18) r where we used the fact that (cid:0)r (cid:1) (cid:1)/(cid:0)m (cid:80)r (cid:32)(cid:18) r (cid:18)β(cid:48)n (cid:19)r(d−1) (cid:19)r(d−1)+1 (cid:18) r (cid:32)(cid:18) r (cid:18) r (cid:19)r(d−1) (cid:18)β(cid:48)n (cid:19)r(d−1) (cid:19) (cid:18)β(cid:48)n (cid:18) r (59) ≤ (r/m)d if r < m. Taking the union bound for all configurations s ∈ [0, r] gives: Pr(E1,...,r) ≤ ≤ Pr(E1,...,r, fix s) ≤ (cid:19)rd(cid:33) (cid:18) r (cid:19)rd+1(cid:33) s=0 Pr(E1,...,r, fix s) 1 − r/(β(cid:48)n) (cid:19)sd + ··· + (cid:19) (cid:19) β(cid:48)n 1 bound: ≤ = β(cid:48)n β(cid:48)n β(cid:48)n β(cid:48)n β(cid:48)n β(cid:48)n β(cid:48)n r r r r d−1 d−1 d−1 d d d (cid:48) n r d · · − · · + · 1 − α/β(cid:48) r · β(cid:48)n (60) The second equality is due to the geometric series and the last inequality is due to r ≤ αn. We thus have: = < d (cid:17) (cid:19)(cid:18)β(cid:48)n (cid:18)n Pr G /∈ G(α, 1 − 1/(d − 1)) αn(cid:88) r=d−1 ≤ 1 1 − α β(cid:48) r r β(cid:48)n (cid:19)(cid:18) r (cid:19)r(d−1) (61) The remainder of the proof follows from that of [42]; i.e., Lemma A.1 and A.2, through replacing d by d − 1. In particular, we can show that Pr G /∈ G(α, 1 − 1/(d − 1)) ≤ O(n1−(d−1)(d−3)) (62) (cid:17) The desired result is achieved by observing D(t) . . . D(s) → 0 as t − s → ∞. Lastly, the almost sure boundedness of B(s, t) can be obtained from the fact that Φ(s, t) is stochastic. Q.E.D. We consider the following chain: (70) (71) (72) 12 if d− 1 > (H(α) + β(cid:48)H(α/β(cid:48)))/α log(β(cid:48)/α). This completes Q.E.D. the proof. Combining Proposition 4 & 5 indicates that the d-rows- deleted sensing matrix B )c corresponds to an (α, 1 − 1/(d−1))-expander graph with high probability. Therefore, the conclusion in Corollary 1 follows by setting δ = 1−1/(d−1). Moreover, by applying the union bound, the probability of failure is upper bounded as: T (Ωi B (cid:48) Pr(F ail) ≤ n2 + O(n2−(d−1)(d−3)), (63) (cid:19)4 d − 1 (cid:18) d β which vanishes as n → ∞. APPENDIX C PROOF OF THEOREM 2 To simplify the notations, in this section, we drop the dependence on the discussion index k for the opinion vectors x(t; k) and the trust matrices W (t; k). We first prove that the estimator is unbiased. Consider the following chain: (cid:88) (cid:88) E{ x(Tk)x(0)} = W E{ x(ti)x(0)} 1 Tk tix(0) = W ∞x(0), ti∈Tk = 1 Tk ti∈Tk where we used the fact that To → ∞ and ti ≥ To for all ti in the last equality. Next, we prove that the estimator is asymptotically consis- tent, i.e., (38). Without loss of generality, we let t1 < t2 < as the sampling instances. The following shorthand . . . < tTk notation will be useful: Φ(s, t) (cid:44) W (t)W (t − 1) . . . W (s + 1)W (s), (65) where t ≥ s and Φ(s, t) is a random matrix. Our proof involves the following lemma: Lemma 2 When t − s → ∞, the random matrix Φ(s, t) converges almost surely to the following: (cid:18) (cid:19) where B(s, t) = (cid:80)t t−s→∞ lim Φ(s, t) = B(s, t) is bounded almost surely. I 0 B(s, t) 0 , (66) q=s(D(t) . . . D(q))B(q). Moreover, Proof: We first establish the almost sure convergence of D(t)D(t − 1) . . . D(s) to 0. Define β(s, t) (cid:44) (cid:107)D(t)D(t − 1) . . . D(s)(cid:107)2, and observe the following chain E{β(s, t)β(s, t − 1), ..., β(s, s)} = E{(cid:107)D(t)D(t − 1) . . . D(s)(cid:107)2β(s, t − 1)} ≤ E{(cid:107)D(t)(cid:107)2(cid:107)D(t − 1) . . . D(s)(cid:107)2β(s, t − 1)} = E{(cid:107)D(t)(cid:107)2}β(s, t − 1) ≤ cβ(s, t − 1), where c = (cid:107)D(cid:107)2 < 1. The almost sure convergence of β(s, t) follows from [56, Lemma 7]. Now, expanding the multiplication (65) yields: Φ(s, t) = I 0 B(s, t) D(t) . . . D(s) . (69) (cid:18) (cid:19) (67) (68) Recall that x(ti) = x(ti) + n(ti) and the noise term n(ti) is independent of W (t) for all t. The above expression reduces to: (cid:111) . (cid:111) (cid:111) E Tk = E ti∈Tk Tk (cid:80) 2x(0) ti∈Tk +E 2x(0)} = 2x(0) n(ti) (cid:110)(cid:13)(cid:13)(cid:13) 1 (cid:110)(cid:13)(cid:13)(cid:13) 1 (cid:88) E{(cid:107) x(Tk) − x(∞)(cid:107)2 (cid:0) x(ti) − x(∞)(cid:1)(cid:13)(cid:13)(cid:13)2 (cid:0)x(ti) − x(∞)(cid:1)(cid:13)(cid:13)(cid:13)2 (cid:13)(cid:13)(cid:13)2 (cid:110)(cid:13)(cid:13)(cid:13) 1 (cid:80) (cid:110)(cid:13)(cid:13)(cid:13) 1 (cid:0)x(ti) − x(∞)(cid:1)(cid:13)(cid:13)(cid:13)2 (cid:111) (cid:88) (cid:110)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)2 (cid:0)Φ(0, ti) − W ∞(cid:1)x(0) Tk 1 Tk2 E(cid:8)Tr(cid:0)Ξx(0)x(0)T(cid:1)(cid:9), Tk2 E (cid:88) (cid:0)Φ(0, tj) − W ∞(cid:1)T (cid:88) ti∈Tk ti∈Tk Tk 1 = = 2 2 . tj∈Tk ti∈Tk where Ξ = It is easy to check that the latter term vanishes when Tk → ∞. We thus focus on the former term, which gives (cid:111) 2x(0) ti∈Tk E (64) (cid:0)Φ(0, ti) − W ∞(cid:1). (73) Expanding the above product yields two groups of terms - when ti = tj and when ti (cid:54)= tj. When ti = tj, using To → ∞ and Lemma 2, it is straightforward to show that: (cid:107)E(cid:8)(cid:0)Φ(0, ti) − W ∞(cid:1)T(cid:0)Φ(0, ti) − W ∞(cid:1)(cid:9) (74) for some constant C < ∞. As a matter of fact, we observe that the above term will not vanish at all. This is due to Observation 3, the random matrix Φ(0, ti) does not converge in mean square sense. (cid:107) ≤ C, For the latter case, we assume tj > ti. We have (cid:0)Φ(0, tj) − W ∞(cid:1)T(cid:0)Φ(0, ti) − W ∞(cid:1) =(cid:0)Φ(ti + 1, tj)Φ(0, ti) − W ∞(cid:1)T(cid:0)Φ(0, ti) − W ∞(cid:1). tj−ti(cid:0)Φ(0, ti) − W ∞(cid:1)(cid:9), E(cid:8)(cid:0)Φ(0, ti) − W ∞(cid:1)T Taking expectation of the above term gives: (75) (76) where we used the fact that Φ(ti + 1, tj) is independent of the other random variables in the expression and W ∞W = W ∞ for any (cid:96) ≥ 0. Now, note that tj−ti = W ∞ + O(λtj−ti ), (77) for some 0 < λ (cid:44) λmax(D) < 1. This is due to the fact that D is sub-stochastic. As To → ∞ and by invoking Lemma 2, the matrix (Φ(0, ti) − W ∞) has almost surely only non-empty entries W W (cid:96) Combining these results, we can show in the lower left block. Carrying out the block matrix multi- plications and using the boundedless of Φ(0, ti) gives (cid:13)(cid:13)E(cid:8)(cid:0)Φ(0, tj) − W ∞(cid:1)T(cid:0)Φ(0, ti) − W ∞(cid:1)(cid:9)(cid:13)(cid:13) ≤ O(λtj−ti). E(cid:8)Tr(cid:0)Ξx(0)x(0)T(cid:1)(cid:9) (cid:17) summable geometric series (cid:80)Tk−1 (79) for some C(cid:48) < ∞. Notice that min(cid:96) t(cid:96)+i − t(cid:96) ≥ i and the terms inside the bracket can be upper bounded by the λi, since λ < 1. Conse- quently, the mean square error goes to zero as Tk → ∞. The estimator (36) is consistent. (cid:16) Tk−1(cid:88) λmink tk+i−tk C(cid:48) Tk Tk2 (78) ≤ i=0 i=0 , REFERENCES [1] H.-T. Wai, A. Scaglione, and A. Leshem, "The social system identification problem," in Proc CDC, Dec. 2015, pp. 406–411. [Online]. Available: http://arxiv.org/abs/1503.07288 [2] M. O. Jackson, Social and Economic Networks. Princeton, NJ, USA: Princeton University Press, 2008. [3] A. H. Sayed, S.-y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, "Diffusion strategies for adaptation and learning over networks: an examination of distributed strategies and network behavior," IEEE Signal Process. Mag., vol. 30, no. 3, pp. 155–171, May 2013. [4] A. Anis, A. Gadde, and A. Ortega, "Towards a sampling theorem for signals on arbitrary graphs," in Proc ICASSP, 2014, pp. 3864–3868. [5] A. Sandryhaila and J. M. Moura, "Big Data Analysis with Signal Processing on Graphs: Representation and processing of massive data sets with irregular structure," IEEE Signal Process. Mag., vol. 31, no. 5, pp. 80–90, 2014. [6] A. Clauset, M. Newman, and C. Moore, "Finding community structure in very large networks," Physical Review E, vol. 70, pp. 1–6, 2004. [7] S. Fortunato, "Community detection in graphs," Physics Reports, vol. 486, pp. 75–174, 2010. [8] L. Page, S. Brin, R. Motwani, and T. Winograd, "The pagerank citation ranking: Bringing order to the web." technical report, 1999. [9] H. Ishii and R. Tempo, "The pagerank problem, multiagent consensus, and web aggregation: A systems and control viewpoint," IEEE Control Systems Magazine, vol. 34, pp. 34–53, 2014. [10] Y.-Y. Liu, J.-J. Slotine, and A.-L. Barab´asi, "Controllability of complex networks." Nature, vol. 473, pp. 167–173, 2011. [11] --, "Observability of complex systems." PNAS, vol. 110, no. 7, pp. 2460–5, 2013. [12] M. Doostmohammadian and U. A. Khan, "Graph-theoretic distributed inference in social networks," IEEE J. Sel. Topics Signal Process., vol. 8, pp. 613–623, 2014. [13] M. DeGroot, "Reaching a consensus," in Journal of American Statistcal Association, vol. 69, 1974, pp. 118–121. [14] A. De, S. Bhattacharya, P. Bhattacharya, N. Ganguly, and S. Chakrabarti, "Learning a linear influence model from transient opinion dynamics," Proc CIKM, pp. 401–410, 2014. [15] A. Das, S. Gollapudi, and K. Munagala, "Modeling opinion dynamics in social networks," in Proc WSDM, 2014, pp. 403–412. [16] P. Jia, A. Mirtabatabaei, N. E. Friedkin, and F. Bullo, "Opinion Dynam- ics and the Evolution of Social Power in Influence Networks," SIAM Review, pp. 1–27, 2013. [17] A. G. Chandrasekhar, H. Larreguy, and J. P. Xandri, "Testing models of social learning on networks: evidence from a framed field experiment," Working Paper, 2012. [18] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in social networks," Dynamic Games and Applications, vol. 1, no. 1, pp. 3–49, 2011. [19] V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis, "Convergence in multiagent coordination, consensus, and flocking," in Proc CDC-ECC '05, vol. 2005, 2005, pp. 2996–3000. [20] L. Xiao, S. Boyd, and S. Kim, "Distributed average consensus with least-mean-square deviation," in Journal of Parallel and Distributed Computing, vol. 67, 2007, pp. 33–46. 13 [21] M. Timme, "Revealing network connectivity from response dynamics," Physical Review Letters, vol. 98, no. 22, pp. 1–4, 2007. [22] W.-X. Wang, Y.-C. Lai, C. Grebogi, and J. Ye, "Network Reconstruction Based on Evolutionary-Game Data via Compressive Sensing," Physical Review X, vol. 1, no. 2, pp. 1–7, 2011. [23] J. Mei and J. M. F. Moura, "Signal Processing on Graphs: Modeling Causal Relations in Big Data," no. 412, pp. 1–22, 2015. [Online]. Available: http://arxiv.org/abs/1503.0017 [24] M. E. Yildiz, D. Acemoglu, A. Ozdaglar, A. Saberi, and A. Scaglione, "Discrete opinion dynamics with stubborn agents," SSRN eLibrary, 2011. [25] M. E. Yildiz, A. Ozdaglar, D. Acemoglu, A. Saberi, and A. Scaglione, "Binary opinion dynamics with stubborn agents," ACM Trans. Econ. Comput., vol. 1, no. 4, p. 19, 2013. [26] D. Acemoglu, G. Como, F. Fagnani, and A. Ozdaglar, "Opinion Fluctua- tions and Disagreement in Social Networks," Mathematics of Operations Research, vol. 38, no. 1, pp. 1–27, Feb. 2013. [27] M. Mobilia, "Does a single zealot affect an infinite group of voters ?" Physical Review Letters, July 2003. [28] A. Waagen, G. Verma, K. Chan, A. Swami, and R. D'Souza, "Effect of zealotry in high-dimensional opinion dynamics models," Physical Review E, Feb. 2015. [29] M. E. Yildiz and A. Scaglione, "Computing along routes via gossiping," IEEE Trans. on Signal Process., vol. 58, no. 6, pp. 3313–3327, 2010. [30] W. Ben-Ameur, P. Bianchi, and J. Jakubowicz, "Robust Average Consen- sus using Total Variation Gossip Algorithm," in VALUETOOLS, 2012, pp. 99–106. [31] U. A. Khan, S. Kar, and J. M. F. Moura, "Higher dimensional consensus: Learning in large-scale networks," IEEE Trans. Signal Process., vol. 58, no. 5, pp. 2836–2849, 2010. [32] T. Wang, H. Krim, and Y. Viniotis, "Analysis and Control of Beliefs in Social Networks," IEEE Trans. on Signal Process., vol. 62, no. 21, pp. 5552–5564, 2014. [33] U. A. Khan, S. Kar, and J. M. F. Moura, "Distributed sensor localization in random environments using minimal number of anchor nodes," IEEE Trans. Signal Process., vol. 57, no. 5, pp. 2000–2016, May 2009. [34] M. Moussad, J. E. Kammer, P. P. Analytis, and H. Neth, "Social influence and the collective dynamics of opinion formation," PLoS ONE, vol. 8, no. 11, November 2013. [35] M. Ramos, J. Shao, S. D. S. Reis, C. Anteneodo, J. S. A. Jr, S. Havlin, and H. A. Makse, "How does public opinion become extreme?" Sci. Rep., no. 10032, May 2015. [36] C. J. Kuhlman, V. S. A. Kumar, and S. S. Ravi, "Controlling opinion bias in online social networks," in Proc. WebSci, 2012, pp. 165–174. [37] S. Gleichman and Y. C. Eldar, "Blind compressed sensing," IEEE Trans. Inf. Theory, vol. 57, no. 10, pp. 6958–6975, 2011. [38] C. Studer and R. G. Baraniuk, "Dictionary learning from sparsely corrupted or compressed signals," in Proc ICASSP, no. Dl, 2012, pp. 3341–3344. [39] A. Gilbert and P. Indyk, "Sparse recovery using sparse matrices," Proceedings of the IEEE, vol. 98, no. 6, pp. 937–947, 2010. [40] R. Berinde, A. C. Gilbert, P. Indyk, H. Karloff, and M. J. Strauss, "Combining geometry and combinatorics: A unified approach to sparse signal recovery," in 46th Annual Allerton Conference on Communication, Control, and Computing, 2008, pp. 798–805. [41] M. Wang, W. Xu, and A. Tang, "A unique "nonnegative" solution to an underdetermined system: From vectors to matrices," IEEE Trans. Signal Process., vol. 59, no. 3, pp. 1007–1016, 2011. [42] M. A. Khajehnejad, A. G. Dimakis, W. Xu, and B. Hassibi, "Sparse recovery of nonnegative signals with minimal expansion," IEEE Trans. Signal Process., vol. 59, no. 1, pp. 196–208, 2011. [43] N. E. Friedkin and E. C. Johnsen, Social Influence Network Theory: Cambridge A Sociological Examination of Small Group Dynamics. University Press, 2011. [44] M. Grant and S. Boyd, "CVX: Matlab software for disciplined convex programming, version 2.1," http://cvxr.com/cvx, Mar. 2014. [45] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. New York, NY, USA: Cambridge University Press, 2014. [46] D. J. Watts and S. H. Strogatz, "Collective dynamics of 'small-world' networks." Nature, vol. 393, no. 6684, pp. 440–442, 1998. [47] R. Albert and A. L. Barab´asi, "Statistical mechanics of complex net- works," Reviews of Modern Physics, vol. 74, no. 1, pp. 47–97, 2002. [48] N. Parikh and S. Boyd, "Proximal Algorithms," Foundations and Trends in Optimization, vol. 1, no. 3, pp. 123–231, 2014. [49] A. Beck and M. Teboulle, "A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems," SIAM Journal on Imaging Sciences, vol. 2, no. 1, pp. 183–202, 2009. 14 [50] C. Ravazzi, P. Frasca, R. Tempo, and H. Ishii, "Ergodic Randomized Algorithms and Dynamics over Networks," IEEE Trans. Control of Network Sys., pp. 1–11, to appear. [51] A. L. Traud, P. J. Mucha, and M. A. Porter, "Social structure of Facebook networks," Physica A: Statistical Mechanics and its Applications, vol. 391, no. 16, pp. 4165–4180, 2012. [52] T. C. Aysal, M. E. Yildiz, A. D. Sarwate, and A. Scaglione, "Broadcast gossip algorithms for consensus," IEEE Trans. Signal Process., vol. 57, no. 7, pp. 2748–2761, 2009. [53] M. Bastian, S. Heymann, and M. Jacomy, "Gephi: An open source networks," 2009. [Online]. Available: http://www.aaai.org/ocs/index.php/ICWSM/ 09/paper/view/154 and manipulating exploring software for [54] R. A. Horn and C. R. Johnson, Eds., Matrix Analysis. Cambridge University Press, 1986. [55] D. B. West, Introduction to Graph Theory, 2nd ed. Prentice Hall, [56] B. Polyak, Introduction to Optimization. New York: Optimization September 2000. Software, Inc., 1987.
1712.10195	4	1712	2019-04-16T05:21:03	Growing Attributed Networks through Local Processes	[ "cs.SI" ]	This paper proposes an attributed network growth model. Despite the knowledge that individuals use limited resources to form connections to similar others, we lack an understanding of how local and resource-constrained mechanisms explain the emergence of rich structural properties found in real-world networks. We make three contributions. First, we propose a parsimonious and accurate model of attributed network growth that jointly explains the emergence of in-degree distributions, local clustering, clustering-degree relationship and attribute mixing patterns. Second, our model is based on biased random walks and uses local processes to form edges without recourse to global network information. Third, we account for multiple sociological phenomena: bounded rationality, structural constraints, triadic closure, attribute homophily, and preferential attachment. Our experiments indicate that the proposed Attributed Random Walk (ARW) model accurately preserves network structure and attribute mixing patterns of six real-world networks; it improves upon the performance of eight state-of-the-art models by a statistically significant margin of 2.5-10x.	cs.SI	cs	Growing Attributed Networks through Local Processes Harshay Shah, Suhansanu Kumar, Hari Sundaram Department of Computer Science University of Illinois at Urbana-Champaign {hrshah4,skumar56,hs1}@illinois.edu 9 1 0 2 r p A 6 1 ] I S . s c [ 4 v 5 9 1 0 1 . 2 1 7 1 : v i X r a ABSTRACT This paper proposes an attributed network growth model. Despite the knowledge that individuals use limited resources to form con- nections to similar others, we lack an understanding of how local and resource-constrained mechanisms explain the emergence of rich structural properties found in real-world networks. We make three contributions. First, we propose an interpretable and accurate model of attributed network growth that jointly explains the emer- gence of in-degree distribution, local clustering, clustering-degree relationship and attribute mixing patterns. Second, we make use of biased random walks to develop a model that forms edges lo- cally, without recourse to global information. Third, we account for multiple sociological phenomena -- bounded rationality; structural constraints; triadic closure; attribute homophily; preferential attach- ment. We explore the parameter space of the proposed Attributed Network Growth (ARW) to show each model parameter intuitively modulates network structure. Our experiments show that ARW ac- curately preserves network structure and attribute mixing patterns of six real-world networks; it improves upon the performance of eight well-known models by a significant margin of 2.5 -- 10×. 1 INTRODUCTION We present a network growth model that explains how distinct structural properties of attributed networks can emerge from local edge formation processes. In real-world networks, individuals tend to form edges despite limited information and partial network ac- cess. Moreover, phenomena such as triadic closure and homophily simultaneously influence individuals' decisions to form connections. Over time, these decisions cumulatively shape real-world networks to exhibit rich structural properties: heavy-tailed in-degree distri- bution, skewed local clustering and homophilic mixing patterns. However, we lack an understanding of local, resource-constrained mechanisms that incorporate sociological factors to explain the emergence of rich structural properties. Classic models of network growth tend to make unrealistic as- sumptions about how individuals form edges. Consider a simple stylized example: the process of finding a set of papers to cite when writing an article. In preferential attachment [3] or fitness [5, 10, 49] based models, a node making m citations would pick papers from the entire network in proportion to their in-degree or fitness re- spectively. This process assumes that individuals possess complete knowledge of in-degree or fitness of every node in the network. An equivalent formulation -- vertex copying [26] -- induces prefer- ential attachment: for every citation, a node would pick a paper uniformly at random from all papers, and either cite it or copy its citations. Notice that the copying mechanism assumes individuals have complete access to the network and forms each edge inde- pendently. Although these models explain the emergence of power Figure 1: The figure shows how our proposed model of an Attributed Random Walk (ARW) accurately preserves local clustering and assor- tativity; we contrast with a non-attributed growth model [20] to un- derscore the importance of using attributes for network growth. law degree distributions, they are unrealistic: they require global knowledge (e.g., preferential attachment requires knowledge of the global in-degree distribution) or global access (e.g., vertex copying requires random access to all nodes). Additionally, these models do not account for the fact that many networks are attributed (e.g., a paper is published at a venue; a Facebook user may use gender, political interests to describe them) and that assortative mixing is an important network characteristic [38]. Recent papers tackle resource constraints [35, 50, 52] as well as nodal attributes [12, 17]. However, the former disregard attributes and the latter do not provide a realistic representation of edge formation under resource constraints. Furthermore, both sets of models do not jointly preserve multiple structural properties. De- veloping an interpretable and accurate model of attributed network growth that accounts for observed sociological phenomena is non- trivial. Accurate network growth models are useful for synthesizing networks as well as to extrapolate existing real-world networks. We aim to develop a growth model that accounts for resource constraints and sociological phenomena influencing edge forma- tion in addition to preserving global network structure. We make three key contributions. First, we propose a simple and accurate model of attributed network growth. Second, our model is based on local processes to form edges, without recourse to global net- work information. Third, our model unifies multiple sociological phenomena -- bounded rationality; structural constraints; triadic closure; attribute homophily; preferential attachment -- to jointly model global network structure and attribute mixing patterns. The proposed model -- Attributed Random Walk (ARW) -- jointly explains the emergence of in-degree distribution, local cluster- ing, clustering-degree relationship and attribute mixing patterns through a resource constrained mechanism based on random walks (see Figure 1). In particular, the model relies entirely on local infor- mation to grow the network, without access to information of all Assortativity 0.657Average Local Clustering 0.252Attributed Random Walk (ARW) ModelAssortativity 0.002Average Local Clustering 0.073Holme-Kim (HK) ModelAmerican Physical Society (APS) DatasetAssortativity 0.692Average Local Clustering 0.237nodes. In ARW, incoming nodes select a seed node based on attribute similarity and initiate a biased random walk: at each step of the walk, the incoming node either jumps back to its seed or chooses an outgoing link or incoming link to visit another node; it links to each visited node with some probability and halts after it has exhausted its budget to form connections. Our experiments on six large-scale network datasets indicate that the proposed growth model outperforms eight state-of-the-art network growth models by a statistically significant margin of 2.5 -- 10×. Furthermore, we an- alyze the parameter space of the model to show how each parameter intuitively controls one or more key structural properties. The rest of the paper is organized as follows. We begin by defin- ing the problem statement in Section 2. In Section 3, we outline six network datasets, describe key structural properties of real- world networks and discuss insights from sociological studies. Then, in Section 4, we describe the network growth model. We follow by presenting experiments in Section 5, analysis of assortative mixing in Section 6 and discussion in Section 7. 2 PROBLEM STATEMENT Consider an attributed directed network G = (V , E, B), where V & E are sets of nodes & edges and each node has an attribute value b ∈ B. The goal is to develop a directed network growth model that preserves structural and attribute based properties observed in G. The growth model should be normative, accurate and parsimonious: (1) Normative: The model should account for multiple soci- ological phenomena that influence how individuals form edges under constraints of limited global information and partial network access. (2) Accurate: The model should preserve key structural and at- tribute based properties: degree distribution, local clustering, degree-clustering relationship and attribute mixing patterns. (3) Parsimonious: The model should be able to generate net- works with tunable structural properties, while having few parameters. common assumptions in network modeling. Finally, we discuss the role of structural proximity in edge formation. 3.1 Datasets We consider six citation networks of different scales (size, time) from diverse sources: research articles, utility patents and judicial cases. Table 1 lists their summary statistics and global network properties. Three of the six datasets are attributed networks; that is, each node has a categorical attribute value. We focus on citation networks for two reasons. First, since nodes in citation networks form all outgoing edges to existing nodes at the time of joining the network, these datasets provide a clean basis to study edge formation in attributed networks. Second, the node- level, temporal information in datasets that span long time periods (e.g. USSC) enables us to study structural properties at different time stages via network snapshots. Next, we study the structural and attribute properties of these networks. 3.2 Global Network Properties Compact statistical descriptors of global network properties [36] such as degree distribution, local clustering, and attribute assorta- tivity quantify the extent to which local edge formation phenomena shape global network structure. Degree distribution: Real-world networks tend to exhibit heavy tailed degree distributions in which a small but significant fraction of the nodes turn into high-degree hubs. We observe that Log- normal fits, with parameters listed in Table 1, well describe the in-degree distributions, consistent with Broido and Clauset's [9] observation that scale-free, real-world networks are rare. Local Clustering: Real-world networks exhibit high local clus- tering (LCC), as shown in Table 1. Local clustering can arise from Next, we present empirical analysis on real-world datasets to motivate our attributed random walk model. 3 EMPIRICAL ANALYSIS We begin by describing six large-scale network datasets that we use in our analysis and experiments. Then, we describe global network properties, insights from empirical studies in the Social Science and Description U.S. Supreme Court cases ArXiv Physics manuscripts Figure 2: Local clustering in real-world networks have common characteristics: skewed local clustering distribution (left subplot) and a negatively correlated relationship between in-degree and av- erage local clustering (right subplot). LN (µ, σ) (1.19, 1.18) (1.32, 1.41) (1.78, 0.96) (1.93, 1.38) (1.62, 1.20) (1.10, 1.01) Network AA r - USSC [14] - HEP-PH [15] - Semantic [2] Academic Search Engine 0.07 ACL [42] 0.44 APS [1] Patents [29] U.S. NBER patents 0.72 Table 1: Summary statistics & global properties of six network datasets: V nodes join the networks and form edges E over time period T . In attributed networks, each node has a categorical attribute value that belongs to set A of size A. The networks exhibit lognormal (LN) in-degree distribution with mean µ and standard deviation σ , high average local clustering (LCC) & attribute assortativity (AA) coefficient and densify over time with power law (DPL) exponent α. DPL α Avg. LCC 0.12 0.12 0.06 0.07 0.11 0.04 2.32 1.67 1.58 1.43 1.26 1.94 E 216,738 421,533 59,079,055 115,311 6,967,873 16,522,438 T 1754-2002 1992-2002 1991-2016 1965-2016 1893-2015 1975-1999 V 30,288 34,546 7,706,506 18,665 577,046 3,923,922 NLP papers Physics journals A, A - - - venue, 50 journal, 13 category, 6 2 0.00.10.20.30.40.5Localclusteringcoecientc105104103102101100Probabilityp(c)LocalClusteringDistributionDatasetsAPSPatentsSemantic101102103Indegreek0.000.040.080.120.16Averagelocalclusteringc(k)Indegree-ClusteringRelationshipDatasetsHepPHACLUSSCTightly knit neighborhoodStar like neighborhoodFigure 3: Attributed networks ACL, APS and Patents exhibit ho- mophily w.r.t attributes Venue (r = 0.07), Journal (r = 0.44) and Category (r = 0.72) respectively. triadic closure [37, 44], where nodes with common neighbor(s) have an increased likelihood of forming a connection. The coefficient of node i equals the probability with which two randomly chosen neighbors of the node i are connected. In directed networks, the neighborhood of a node i can refer to the nodes that link to i, nodes that i links to or both. We define the neighborhood to be the set of all nodes that link to node i. In Figure 2, we show that (a) average local clustering is not a representative statistic of the skewed lo- cal clustering distributions and (b) real-world networks exhibit a negative correlation between in-degree and clustering. That is, low in-degree nodes have small, tightly knit neighborhoods and high in-degree nodes tend have large, star-shaped neighborhoods. Homophily: Attributed networks tend to exhibit homophily [32], the phenomenon where similar nodes are more likely to be con- nected than dissimilar nodes. The assortativity coefficient [38] r ∈ [−1, 1], quantifies the level of homophily in an attributed net- work. Intuitively, assortativity compares the observed fraction of edges between nodes with the same attribute value to the expected fraction of edges between nodes with same attribute value if the edges were rewired randomly. In Figure 3, we show that attributed networks ACL, APS and Patents exhibit varying level of homophily with assortativity coefficient ranging from 0.07 to 0.72. Increasing Out-degree over Time: The out-degree of nodes that join real-world networks tends to increase as functions of net- work size and time. This phenomenon densifies networks and can shrink effective diameter over time. Densification tends to exhibit a power law relationship [29] between the number of edges e(t) and nodes n(t) at time t: e(t) ∝ n(t)α . Table 1 lists the densification power law (DPL) exponent α of the network datasets. To summarize, citation networks tend to be homophilic networks that undergo accelerated network growth and exhibit regularities in structural properties: heavy tailed in-degree distribution, skewed local clustering distribution, negatively correlated degree-clustering relationship, and varying attribute mixing patterns. 3.3 Insights from Sociological Studies Sociological studies on network formation seek to explain how individuals form edges in real-world networks. Interplay of Triadic Closure and Homophily: Empirical stud- ies [6, 25] that analyze the interplay between triadic closure and homophily indicate that both structural proximity and homophily are statistically significant factors that simultaneously influence edge formation. Homophilic preferences [32] induce edges between similar nodes, whereas structural factors such as network distance limit edge formation to proximate nodes (e.g. friend of a friend). Bounded Rationality: Extensive work [16, 30, 46] on decision making shows that individuals are boundedly rational actors; con- straints such as limited information, cognitive capacity and time impact decision making. This suggests that resource-constrained individuals that join networks are likely to employ simple rules to form edges using limited information and partial network access. Current preferential attachment and fitness-based models [3, 13, 47] make two assumptions that are at variance with these findings. First, by assuming that successive edge formations are independent, these models disregard the effect of triadic closure and structural proximity. Second, as discussed in section 1, these models require complete network access or knowledge of node-level properties. Insights from sociological studies indicate that edge formation in real-world networks comprises biases towards nodes that are similar, well-connected or structurally proximate. Coupled with em- pirical analyses, it also motivates the need to model how resource- constrained edge formation processes collectively shape global network properties of large-scale networks over time. 3.4 Proximity-biased Edge Formation We investigate the effect of structural proximity on edge formation in real-world networks. Prior work [25] shows that the probability of edge formation in social networks decreases as a function of net- work distance. Indeed, triadic closure explains how individuals form additional edges to proximate nodes (e.g. friend of friend) over time. However, we lack a concrete understanding of the extent to which structural proximity influences edge formation in bibliographic networks, wherein incoming nodes form all edges at the time of joining the network. In Figure 4, we show how high structural proximity among incoming (shown in blue) node's (shown in red) connections in the Hep-PH dataset hints at edge formation processes biased towards proximate nodes in the same local neighborhood. Figure 4: Proximity-biased edge formation. The diagram and prox- imity distributions collectively indicate how edge formation in real- world networks are biased towards structurally proximate nodes in the same locality. We rely on network snapshots and node arrival sequence to estimate a statistic based on path length that measures structural proximity between nodes' connections. Consider an incoming node u that forms edges to nodes in N(u). To measure the proximity between node u's connections, we compute the average pairwise shortest path distance between the connections in the network snapshot immediately preceding node u's arrival. The right subplot in Figure 4 compares the proximity statistic distribution of the Hep-PH dataset to two null models: uniform and configuration. In the uniform model, incoming nodes form connections to existing nodes uniformly at random, whereas the 3 ACLVenuesACLVenuesACLVenue,r=0.07PREPRDPRAPRABPRCPRBPRPRFPRAPRLPRSTABPRFPRSTPERPREPRDPRAPRABPRCPRBPRPRFPRAPRLPRSTABPRFPRSTPERAPSJournal,r=0.44ChemCompMedElecMechMiscChemCompMedElecMechMiscPatentsCategory,r=0.720.00.20.40.60.81.01234567Averageshortestpathlengthbetweencitednodes0.00.30.60.91.21.51.82.1DensityMeasuringlocalityinHep-PHdatasetHep-PHdatasetConfigurationnullmodelUniformnullmodelNew nodeEdge formation in Hep-PH network Measuring Structural Proximity in Hep-PH datasetAverage Pairwise Path Length between ConnectionsDensityconfiguration model randomly rewire all edges in Hep-PH while pre- serving the out-degree and in-degree distributions. We first observe that the connections of incoming nodes in the uniform null model are structurally distant from each other on average. Although the presence of hubs in the configuration model considerably decreases the distance between nodes' connections, it does not explain why the majority of connections in Hep-PH are either connected directly or via an intermediate node. The disparity between the observed and null distributions suggests that structural proximity between connections is intrinsic to edge formation in real-world networks. To summarize, empirical analyses and insights from the Social Sciences motivate the need to model how resource-constrained edge formation processes collectively shape well-defined global network properties of large-scale networks over time. 4 ATTRIBUTED RANDOM WALK MODEL We propose an Attributed Random Walk (ARW) model to explain the emergence of key structural properties of real-world networks through entirely local edge formation mechanisms. Consider a stylized example of how a researcher might go about finding relevant papers to cite. First, the researcher broadly identi- fies one or more relevant papers, possibly with the help of external information (e.g. Google Scholar). These initial set of papers act as seed nodes. Then, acting under time and information constraints, she will examine papers cited by the seed and papers that cite the seed. Thus, she navigates a chain of backward and forward refer- ences to identify similar, relevant papers. Next, through careful analysis, she will cite a subset of these papers. Similarly, users in online social networks might form new friendships by navigating their social circle (e.g., friends of friends) to find similar others. ARW grows a directed network as new nodes join the network. The mechanism is motivated by the stylized example: an incoming node selects a seed node and initiates a random walk to explore the network by navigating through neighborhoods of existing nodes. It halts the random walk after connecting to a few visited nodes. In this section, we describe the edge formation mechanisms underlying ARW, explain how ARW unifies multiple sociological phe- nomena, discuss model interpretability and summarize methods required to fit ARW to network data. 4.1 Model Description The Attributed Random Walk (ARW) model grows a directed network { Gt }T t =1 in T time steps. More formally, at every discrete time step t, a new node u, with attribute value B(u), joins the network Gt . After joining the network, node u forms m(t) edges to existing nodes. The edge formation mechanism consists of two components: Select-Seed and Random-Walk. As shown in Figure 5, an incom- ing node u with attribute value B(u) that joins the network at time t first selects a seed node using Select-Seed. Select-Seed accounts for homophilic preferences of incoming nodes using parameters psame and pdiff to tune attribute preferences. In Figure 5, node u selects a seed node and initiates a random walk using Random-Walk to form m(t) links: Select-Seed (1) With probability psame/psame+pdiff, randomly select a seed node from existing nodes that have the same attribute value, B(u). (2) Otherwise, with probability pdiff/psame+pdiff, randomly select a seed node from existing nodes that do not have the same attribute value, B(u). The Random-Walk mechanism consists of four parameters: attribute-based parameters psame & pdiff model edge formation de- cisions and the jump parameter pjump & out-link parameter pout characterize random walk traversals: Random-Walk (1) At each step of the walk, new node u visits node vi. • If B(u) = B(vi), u links to vi with probability psame • Otherwise, u links to vi with probability pdiff (2) Then, with probability pjump, u jumps back to seed su. (3) Otherwise, with probability 1 − pjump, u continues to walk. It picks an outgoing edge with prob. pout or an in- coming edge with prob. 1 − pout to visit a neighbor of vi. (4) Steps 1-3 are repeated until u links to m(t) nodes. When attribute data is absent, ARW simplifies further. A single link parameter plink replaces both attribute parameters psame & pdiff. Select-Seed reduces to uniform seed selection and in Random- Walk, the probability of linking to visited nodes equals plink. Note that ARW has two exogenous parameters: the out-degree m(t) and attribute B(u) of incoming nodes. The attribute distribu- tion varies with time as new attribute values (e.g., journals) crop up, necessitating an exogenous parameter. The parameter m(t) is the mean-field value of out-degree m at time t in the observed net- work. While it is straightforward to model m(t) endogenously by incorporating a densification power-law DPL exponent, exogenous factors (e.g., venue, topic) may influence node out-degree. Next, we explain how each parameter is necessary to conform to normative behavior of individuals in evolving networks. 4.2 ARW and Normative Behavior The Attributed Random Walk model unifies multiple sociological phenomena into its edge formation mechanisms. Phenomenon 1. (Limited Resources) Individuals are boundedly ra- tional [16, 30, 46] actors that form edges under constraints of limited information and partial network access. As shown in Figure 5, Random-Walk only requires information only about the 1-hop neighborhood of a few visited nodes, thereby accounting for the constraints of limited information and partial network access. Phenomenon 2. (Structural Constraints) Network distance act as a constraint that limits long-range connections. [25] We incorporate structural constraints using pjump, the probability with which a new node jumps back to its seed node after every step of the random walk. This implies that the probability with which the new node is at most k steps from its seed node is (1 − pjump)k; 4 Figure 5: Edge formation in ARW: consider an incoming node u with outdegree m = 3 and attribute value B(u) = red ∈ {red, green}. In fig. 5a, u joins the network and selects seed va via Select-Seed. Then, in fig. 5b, u initiates a Random-Walk and traverses from va to vb to vc . Finally, u jumps back to its seed va and restarts the walk, as shown in fig. 5c. Node u halts the random walk after linking to va, vc & vd . as a result, pjump controls the extent to which nodes' random walks explore the network to form edges. Phenomenon 3. (Triadic Closure) Nodes with common neighbors have an increased likelihood of forming a connection. [44] When attribute data is absent, ARW controls the effect of triadic closure on link formation using plink because with probability pro- portional to p2 link, a new node u closes a triad through its random traversal by linking to both, a visited node and its neighbor, Phenomenon 4. (Attribute Homophily) Nodes that have similar attributes are more likely to form a connection. [32] The attribute parameters psame and pdiff modulate attribute assor- tativity. When psame > pdiff, nodes are more likely to connect if they share the same attribute value, thereby resulting in a homophilic network over time. Similarly, psame < pdiff and psame = pdiff make edge formation heterophilic and attribute agnostic respectively. Phenomenon 5. (Preferential Attachment) Nodes tend to link to high degree nodes that have more visibility. [3] heavy-tailed distributions. In Figure 6, we observe that increasing pout from 0.2 to 0.8 shifts probability mass from average degree nodes (B) to hubs (C) and low degree nodes (A). As a result, pout controls the extent to which hubs skew the in-degree distribution. Similarly, local clustering increases as a function of plink because plink implicitly controls the rate at which new nodes close triads by linking to adjacent nodes in their random walks. We use contour plots to visualize how(pdiff, psame) and(plink, pjump) alter attribute assortativity and average path length. As shown in the left subplot of Figure 7, psame − pdiff tunes the extent to which attributes influence edge formation. Increasing psame−pdiff increases attribute assortativity by amplifying nodes' propensity to link to similar nodes, which subsequently increases the fraction of edges between similar nodes. More importantly, when psame−pdiff remains ARW controls preferential attachment by adding structural bias to the random walk traversal using outlink parameter pout, instead of relying on the global degree distribution. Random walks that traverse outgoing edges only (i.e., pout = 1) eventually visit old nodes that tend to have high in-degree. Similarly, random walks that traverse incoming edges only (i.e., pout = 0) visit recently joined nodes that tend to have low indegree. As a result, we use pout to adjust bias towards node degree. To summarize: ARW incorporates five well-known sociological phenomena -- limited resources; structural constraints; triadic clo- sure; attribute homophily; preferential attachment -- into a single edge formation mechanism based on random walks. 4.3 Model Interpretability ARW parameters intuitively shape key structural properties: in-degree distribution, local clustering, path length and attribute assortativity. In order to understand how global network properties vary as functions of ARW parameters, we explore the parameter space of the model. As described in Subsection 4.1, ARW uses two parameteriza- tions to model networks with or without attribute data. We analyze network structure and attribute assortativity using (plink, pjump, pout) and (psame, pdiff, pjump, pout) respectively. Figure 6 illustrates how in-degree and local clustering depend on pout and plink. Increasing pout steers random walks towards older nodes that tend to have higher in-degree. Over time, as more nodes join the network, initial differences in degree amplify, resulting in Figure 6: Effect of pout and plink on in-degree and local clustering. The left and right subplots show how increasing pout and plink yield in-degree distributions with heavier tails and neighbors with higher clustering by adding structural bias towards well-connected and proximate nodes respectively. Figure 7: Effect of (psame, pdiff) and (plink, pjump) on attribute assorta- tivity and average path length. In the left, we observe that increas- ing psame − pdiff leads to higher assortativity by making edge forma- tion more homophilic. Similarly, decreasing pjump while increasing pout shortens average path length by enabling incoming nodes to form connections in disparate regions of the network. 5 321SEEDvcvbvauNEW NODE53421SEEDvcvbvavduNEW NODE1SEEDvauNEW NODEu3a Select seed3b Initiate random walk3c Jump back to seed100101102103In-degreek104103102101Probabilityp(k)E↵ectofpoutonIn-degreepout=0.8pout=0.5pout=0.2101102103In-degreek0.0000.0250.0500.0750.1000.1250.150Averagelocalclustering¯c(k)E↵ectofplinkonLocalClusteringplink=0.8plink=0.5plink=0.20.10.20.30.40.50.60.70.80.9pdiff0.10.20.30.40.50.60.70.80.9psameE↵ectof(psame,pdiff)onAssortativityr1.000.750.500.250.000.250.500.751.00r0.10.20.30.40.50.60.70.80.9pout0.10.20.30.40.50.60.70.80.9pjumpE↵ectof(pjump,pout)onAvg.PathLength¯p2.562.883.203.523.844.164.484.805.12¯pconstant, increasing (psame, pdiff) raises local clustering without al- tering attribute assortativity. In the right subplot, we observe that increasing pout while decreasing pjump shortens the average path length. This is because low values of pjump do not restrict incom- ing nodes to the local neighborhood of their seed nodes, thereby allowing incoming nodes to visit and form edges to nodes that are structurally distant. Additionally, increasing pout results in greater number of hubs, which in turn act as intermediate nodes to connect nodes via short path lengths. Thus, ARW unifies multiple sociological phenomena at the local level as well as intuitively controls key global network properties. 4.4 Model Fitting We now briefly describe methods to estimate model parameters, initialize G, densify G over time and sample nodes' attribute values. Parameter Estimation. We use grid search to estimate the four pa- rameters using evaluation metrics and selection criterion described in Subsection 5.1. Initialization. ARW is sensitive to a large number of weakly con- nected components (WCCs) in initial network G0 because incoming nodes only form edges to nodes in the same WCC. To ensure that G0 is weakly connected, we perform an undirected breadth-first search on the observed, to-be-fitted network G that starts from the oldest node and halts after visiting 0.1% of the nodes. The initial network G0 is the small WCC induced from the set of visited nodes. Node Out-degree. Node out-degree increases non-linearly over time in real-world networks. We coarsely mirror the growth rate of observed network G as follows. Each incoming node u that joins G at time t corresponds to some node that joins the observed network G in year y(t); the number of edges m(t) that u forms is equal to the average out-degree of nodes that join G in year y(t). Sampling Attribute Values. The distribution over nodal attribute values Pg(B) tends to change over time. The change in the attribute distribution over time is an exogenous factor and varies for every network. Therefore, we sample the attribute value B(u) of node u, that joins G at time t, from Pg(B year = y(t)), the observed attribute distribution conditioned on the year of arrival of node u. To summarize, ARW intuitively describes how individuals form edges under resource constraints. ARW uses four parameters -- psame, pdiff, pjump, pout -- to incorporate individuals' biases towards similar, proximate and high degree nodes. Next, we discuss our experi- ments on the performance of ARW in accurately preserving multiple structural and attribute properties of real networks. 5 MODELING NETWORK STRUCTURE In this section, we evaluate ARW's performance in preserving real- world network structure relative to well-known growth models. 5.1 Setup In this subsection, we introduce eight representative growth models and describe evaluation metrics used to fit models to the datasets. State-of-the-art Growth Models. We compare ARW to eight state- of-the-art growth models representative of the key edge formation mechanisms: preferential attachment, fitness, triangle closing and random walks. Two of the eight models account for attribute ho- mophily and preserve attribute mixing patterns, as listed below: 6 (1) Dorogovtsev-Mendes-Samukhin model [13] (DMS) is a pref- erential attachment model that generates directed scale-free graphs. In this model, the probability of linking to a node is proportional to the sum of its in-degree and "initial attractiveness." (2) Relay Linking model [47] (RL) comprises preferential attach- ment models for directed networks that use relay linking to model node popularity over time. We use the iterated preferential relay- cite (IPRC) variant, which best fits real-world network properties. (3) Kim-Altmann model [23] (KA) is a fitness-based model that defines fitness as the product of degree and attribute similarity. It generates attributed networks with assortative mixing and power law degree distribution. To generate directed networks, we modify KA to form directed edges to nodes in proportion to their in-degree. (4) Holme-Kim model [20] (HK) is a preferential attachment model that generates scale-free, clustered, undirected networks using a triangle-closing mechanism. To generate directed networks, we modify HK to form directed edges to nodes in proportion to their in- degree and close triangles in their undirected 1-hop neighborhood. (5) Social Attribute Network model [17] (SAN) generates scale- free, clustered, attributed networks via attribute-augmented prefer- ential attachment and triangle closing processes. We modify SAN to create directed edges and thereby produce directed networks. (6) Herera-Zufiria model [43] (SK) is a random walk model that generates scale-free, undirected networks with tunable average clustering. In order to generate directed networks, we allow the random walk mechanism in SK to traverse edges in any direction. (7) Saramaki-Kaski [19] (HZ) is a random walk model that gen- erates scale-free networks with tunable average local clustering. To generate directed networks, we modify HZ to allow its random walk mechanism to traverse edges in any direction. (8) Forest Fire model [29] (FF) is a recursive random walk model that can generate directed networks with shrinking diameter over time, heavy-tailed degree distributions and high clustering. Ensuring Fair Comparison. To ensure fair comparison, we modify existing models in three ways. First, for DMS, SAN, KA do not have an explicitly defined initial graph, so we use initialization method used for ARW, described in subsection 4.4. Second, we extend models that use constant node outdegree m by increasing outdegree over time m(t) using the method described in subsection 4.4. In the absence of model-specific parameter estimation methods, we use grid search to estimate the parameters of every network model, including ARW, using evaluation metrics and selection criterion described below. Evaluation Metrics. We evaluate the network model fit by com- paring four structural properties of G & G: degree distribution, local clustering distribution, degree-clustering relationship and attribute assortativity. We use Kolmogorov-Smirnov (KS) statistic to compare in-degree & local clustering distributions. We compare the degree- clustering relationship in G and G using Weighted Relative Error (WRE), which aggregates the relative error between the average lo- cal clustering c(k) and c(k) of nodes with in-degree k in G and G respectively; The relative error between c(k) and c(k) is weighted in proportion to the number of nodes with in-degree k in G. Jointly preserving multiple structural properties is a multi-objective optimization problem; model parameters that accurately preserve the degree distribution (i.e. low KS statistic) may not preserve the clustering distribution. Therefore, for each model, the selection ) or α = 0.001 ( ) respectively. models in accurately modeling degree distribution (Figure 8A) be- cause its "initial attractiveness" parameter can be tuned to adjust preference towards low degree nodes. Unlike KA, however, DMS cannot preserve global assortativity. However, by assuming that successive edge formations are independent, both models disregard triadic closure and local clustering. (Figure 8B & Figure 8C). Figure 8: Modeling network structure. We assess the extent to which network models fit key structural properties of six real-world networks. Tables 5A, 5B and 5C measure the accuracy of eight models in fitting the in-degree distribution, local clustering distribution, in-degree & clustering relationship respectively and global attribute assortativity. Existing models tend to underperform because they either disregard the effect of factors such as triadic closure and/or homophily or are unable to generate networks with varying structural properties. Our model, ARW, jointly preserves all three properties accurately and often performs considerably better than existing models: the cells are shaded gray or dark gray if the proposed model ARW performs better at significance level α = 0.01 ( criterion for the grid search parameter estimation method chooses the model parameters that minimizes the ℓ2-norm of the aforemen- tioned evaluation metrics. Since the metrics have different scales, we normalize the metrics before computing the ℓ2-norm to prevent unwanted bias towards any particular metric. We note that the parameter sensitivity of the Forest Fire (FF) model necessitates a manually guided grid search method. 5.2 Results Now, we evaluate the performance of ARW relative to eight well- known existing models on the datasets introduced in Subsection 3.1. Figure 8 tabulates the evaluation metrics for every pair of model and dataset. These metrics measure the accuracy with which the fitted models preserve key global network properties: degree distribution, local clustering distribution, and indegree-clustering relationship. To evaluate the performance of these models, we first fit each model to all network datasets G in Subsection 3.1. Thereafter, we compare the structural properties of network dataset G and net- work G generated by the fitted model using evaluation metrics in Subsection 5.1. We average out fluctuations in G over 100 runs. We use one-sided permutation tests [18] to evaluate the rela- tive performance of ARW. If ARW performs better than a model on a dataset with significance level α = 0.01 or α = 0.001, the cor- responding cells in Figure 8 are shaded gray ( ) or dark gray ( ) respectively. We also group models that have similar edge formation mechanisms by color-coding the corresponding rows in Figure 8. We use green ticks in Figure 8 to annotate models that preserve assortativity up to two decimal places. Triangle Closing Models: HK and SAN are preferential attach- ment models that use triangle closing mechanisms to generate scale-free networks with high average local clustering. While tri- angle closing leads to considerable improvement over DMS and KA in modeling local clustering, HK and SAN are not flexible enough to preserve local clustering in all datasets (see Figure 8B & Figure 8C). Existing random walk models: FF, SK, and HZ cannot accu- rately preserve structural properties of real-world network datasets. The recursive approach in FF considerably overestimates local clus- tering. because nodes perform a probabilistic breadth-first search and link to all visited/burned nodes. SK and HZ can control local clustering to some extent, as nodes perform a single random walk and link to each visited node with tunable probability µ. However, both models lack control over the in-degree distribution. Further- more, existing random walk models disregard attribute homophily and do not account for attribute mixing patterns. Attributed Random Walk model: Figure 8 clearly indicates the effectiveness of ARW in jointly preserving multiple global net- work properties. ARW can generate networks with tunable in-degree distribution by adjusting nodes' bias towards high degree nodes using pout. As a result, ARW accurately preserves in-degree distribu- tions (Figure 8A), often significantly better than all models except DMS. Similarly, ARW matches the local clustering distribution (Fig- ure 8B) and in-degree & clustering relationship (Figure 8C) with high accuracy using pjump and plink. Similarly, ARW preserves attribute assortativity using the attribute parameters psame and pdiff. Barring one to two datasets, ARW preserves all three properties significantly better (α < 0.001) than existing random walk models. Figure 8 shows that existing models fail to jointly preserve multiple structural properties in an accurate manner. This is because existing models either disregard important mechanisms such as triadic closure and homophily or are not flexible enough to generate networks with varying structural properties. Preferential attachment models: DMS, RL and KA preserve in- degree distributions but disregard clustering. DMS outperforms other 7 USSCHepPHSemanticACLAPSPatents1.001.001.001.001.001.001.001.001.001.001.001.000.991.001.000.991.001.000.590.740.080.250.730.170.130.140.340.310.151.281.641.740.544.110.150.730.140.460.740.410.510.380.180.450.210.220.510.040.140.100.050.130.080.08C:I\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}C\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}R\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}(WRE)USSCHepPHSemanticACLAPSPatents0.800.820.560.630.830.500.800.820.560.630.820.500.790.820.560.620.830.500.390.550.150.080.520.050.120.050.120.160.050.190.530.540.330.690.190.400.150.290.260.340.340.110.080.290.100.070.340.030.080.040.050.050.050.09B:L\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}C\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}D\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}(KSS\x{FFFF}\x{FFFF}\x{FFFF})ARWDMSKAHKSANFFSKHZRLAssortativityUSSCHepPHSemanticACLAPSPatents0.030.030.050.090.040.020.110.190.220.260.130.060.120.120.170.150.070.150.110.190.220.260.130.050.120.180.190.240.110.050.160.170.140.120.460.320.190.220.250.270.130.130.180.220.230.260.130.130.070.060.070.090.070.08A:I\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}D\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}(KSS\x{FFFF}\x{FFFF}\x{FFFF})p < 0.001p < 0.01Significance level PREFERENTIAL ATTACHMENTRANDOM WALKTRIANGLE CLOSINGFigure 9: Performance of ARW in accurately preserving key global structural properties of the APS network dataset relative to state-of-the- art, representative network models. Existing models such as DMS and HK cannot preserve high local clustering. Moreover, the triangle closing mechanism in SAN incurs high Weighted Relative Error (WRE) because it cannot explain why low in-degree nodes have high local clustering. ARW outperforms existing network models in jointly preserving all three structural properties, in addition to attribute mixing patterns. To summarize, ARW unifies five sociological phenomena into a single mechanism to jointly preserve real-world network structure. 6 MODELING LOCAL MIXING PATTERNS The global assortativity coefficient quantifies the average propen- sity of links between similar nodes. However, global assortativity is not a representative summary statistic of heterogeneous mixing patterns observed in large-scale networks [40]. Furthermore, it does not quantify anomalous mixing patterns and fails to measure how mixing varies across a network. We use local assortativity [40] to measure varying mixing pat- terns in an attributed network G = (V , E, B) with attribute values B = {b1...bk}. Unlike global assortativity that counts all edges between similar nodes, local assortativity of node i, rlocal(i), cap- tures attribute mixing patterns in the neighborhood of node i us- ing a proximity-biased weight distribution wi. The distribution wi reweighs edges between similar nodes based on proximity to node i. As Peel et al. [40] indicate, there are multiple ways to define node i's weight distribution wi other than the prescribed personalized pagerank weight distribution, which is prohibitively expensive to compute for all nodes in large graphs. We define wi as a uniform distribution over N2(i), the two-hop local neighborhood of node i, to allow for a highly efficient local assortativity calculation. In- tuitively, rlocal(i) compares the observed fraction of edges between similar nodes in the local neighborhood of node i to the expected fraction if the edges are randomly rewired. Figure 10: ARW outperforms existing network models in jointly pre- serving key structural properties -- in-degree distribution, local clus- tering distribution and degree-clustering relationship -- by a signif- icant margin of 2.5x-10x. 8 Figure 11: Local assortativity distributions of attributed networks ACL, APS and Patents reveal anomalous, skewed and heterophilic lo- cal mixing patterns. ARW accurately preserves local assortativity, but does not account for anomalous mixing patterns. As shown in Figure 11, local assortativity distributions of ACL, APS and Patents reveal anomalous, skewed and heterophilic local mixing patterns that are not inferred via global assortativity. Our model ARW can preserve diverse local assortativity distributions with high accuracy even though nodes share the same attribute parameters psame and pdiff. This is because, in addition to sampling attributes conditioned on time, ARW incorporates multiple sources of stochasticity through its edge formation mechanism. As a result, incoming nodes with fixed homophilic preferences can position themselves in neighborhoods with variable local assortativity by (a) selecting a seed node in a region with too few (or too many) similar nodes or (b) exhausting all its links before visiting similar (or dissimilar) nodes. We note that ARW is not expressive enough to model anomalous mixing patterns; richer mechanisms such as sampling psame or pdiff from a mixture of Bernoullis are necessary to account for anomalous mixing patterns. 7 DISCUSSION In this section, we discuss weaknesses of triangle closing mecha- nisms, the effect of out-degree on network diameter and limitations & potential modifications of our model ARW. 7.1 Dissecting the Triangle Closing Mechanism A set of network models (e.g., SAN [17] & HK [20]) use triangle closing mechanisms to generate networks with varying average local clustering. However, our experimental results in Subsection 5.2 show that models that rely on triangle closing cannot explain local clustering distribution or bivariate degree-clustering relationship 100101102103104In-degreek106105104103102101Probabilityp(k)In-degreedistributionDatasetARW,KS:0.05SAN,KS:0.11HZ,KS:0.13DMS,KS:0.0495thpercentile0.00.10.20.30.4Localclusteringcoefficientc105104103102101100Probabilityp(c)LocalClusteringdistributionDatasetARW,KS:0.02SAN,KS:0.04HZ,KS:0.31DMS,KS:0.81101102103104In-degreek0.0000.0250.0500.0750.1000.1250.150Averagelocalclustering¯c(k)Indegree-ClusteringrelationshipDatasetARW,WRE:0.08SAN,WRE:0.15HZ,WRE:0.51DMS,WRE:1.0095thpercentile0.00.10.20.30.40.50.60.70.80.91.01.11.2`2-normofnormalizedevaluationmetrics(lowerisbe\x{FFFF}er)JointlyModelingMultipleStructuralPropertiesPreferentialA\x{FFFF}achmentPref.A\x{FFFF}achment&TriangleClosingRandomWalkA\x{FFFF}ributedNetworkModelsDMSKAHKSANFFSKHZARWRLvariance−0.20.00.20.40.60.81.0LocalAssortativityri0.000.020.040.060.080.100.120.14Probabilityp(ri)ACL(Venue),r=0.07DataARW,KS0.08−0.20.00.20.40.60.81.0LocalAssortativityri0.000.020.040.060.080.100.12APS(Journal),r=0.44DataARW(KS0.05)−0.20.00.20.40.60.81.0LocalAssortativityri0.000.050.100.150.20Patents(Category),r=0.72DataARW,KS0.04accurately. To understand why, we examine the degree-clustering relationship in the APS network in Figure 12. Figure 12 reveals that models based on triangle closing mecha- nisms, SAN and HK, considerably underestimate the local clustering of nodes that have low in-degree. This is because incoming nodes in SAN and HK tend to close triangles in the neighborhood of high in-degree nodes to which they connect via preferential attachment. Local clustering plateaus as in-degree decreases because triangle closing along with preferential attachment fail to form connections in neighborhoods of low in-degree nodes. In contrast, ARW accu- rately models the degree-clustering relationship because incoming nodes initiate random walks and close triangles in neighborhoods of low in-degree seed nodes chosen via Select-Seed. 7.2 Effect of Out-degree on Network Diameter Extensive analyses [21, 29, 31] on evolving real-world networks reveal two key temporal properties: network densification and diameter shrinkage over time. Growth models can be adjusted to densify networks over time by allowing node out-degree to increase super-linearly as a function of network size. However, we lack a concrete understanding of existing edge formation mechanisms' inability to preserve diameter shrinkage. Through our analysis, we observe that the out-degree sequence of incoming nodes in network models has a significant impact on effective diameter over time. Figure 13 illustrates the effective diameter of network models fitted to Hep-PH as a function of node out-degree sequence and Figure 12: Triangle closing mechanisms used in SAN HK fail to model average local clustering of low in-degree nodes. In contrast, to accu- rately preserve local clustering, ARW uses random walks to visit low in-degree nodes and close triangles in their neighborhoods . Figure 13: Effect of out-degree on effective diameter. The left sub- plot shows that increasing out-degree over time leads to diameter shrinkage for all models. The right subplot shows that constant out- degree sequence does not account for diameter shrinkage and con- sistently underestimates effective diameter of the Hep-PH network. 9 over time. However, when the out-degree ¯m = n−1 time. By increasing the out-degree m(t) over time using the method described in subsection 4.4, network models representative of key edge formation mechanisms -- ARW, FF, DMS & SAN -- generate net- works that exhibit diameter shrinkage. In particular, FF 1 and ARW mirror the observed rate at which the effective diameter shrinks i m(i) of in- coming nodes is constant, fitted networks, including Forest Fire (FF), cannot preserve shrinking diameter; the effective diameter of the fitted models remain consistently lower than that of Hep-PH. Increasing out-degree over time can effectively incorporate diam- eter shrinkage in all representative network models. This phenom- enon is best understood through the simple Uniform null model, in which incoming nodes form edges to existing nodes chosen uniformly at random. In the Uniform model, nodes with higher out-degree have a greater probability of linking to existing nodes that are structurally distant from each other. Consequently, over time, incoming nodes with higher out-degree are more likely to bridge distant regions of the network, reduce path length between existing nodes and subsequently shrink effective diameter. To summarize, our analysis indicates how increasing out-degree over time enables existing models that rely on different edge for- mation processes to account for diameter shrinkage. 7.3 ARW Limitations We discuss three limitations of ARW. First, we consider only bibli- ographic network datasets in which nodes form all edges at the time of joining. This allows us to analyze edge formation in the absence of confounding edge processes such as edge deletion and edge creation between existing nodes. We plan to extend ARW to handle social networks, where individuals can form edges at any time. One potential way is to incorporate random walks that pause and resume intermittently, thus allowing for older nodes to connect with more recent arrivals. Second, the out-degree m(t) of incom- ing nodes in ARW rely on the observed out-degree sequence, which might be unavailable in datasets without fine-grained temporal data. In this case, ARW can be adapted to rely on the prescribed range of densification exponent αDPL [29] in real-world networks. Since e(t) = m(t)n(t), the power law relationship e(t) ∝ n(t)αDPL between number of edges e(t) and nodes n(t) at time t implies that out-degree m(t) must be proportional to n(t)αDPL−1. Third, ARW fo- cuses on modeling networks in which nodes have a single attribute. The difficulty in incorporating multiple attributes into the edge formation mechanism rests on how we measure attribute similarity. If two nodes are similar only when all their attribute values are identical, we can simply create a new categorical attribute that encodes all multiple attribute combinations and then directly ap- ply ARW. Additional analysis is necessary to identify definitions of attribute similarity that best describe how multiple attributes influence individuals' edge formation processes. 8 RELATED WORK Network growth models seek to explain a subset of structural prop- erties observed in real networks. Note that, unlike growth models, statistical models of network replication [28, 41] do not model how 1FF inherently increases out-degree over time because incoming nodes "burn" through the network for duration in proportion to the network size. 101102103104indegreek0.000.050.100.15averagelocalclustering¯c(k)Indegreevs.LocalClusteringinAPSdatasetDatasetARW,WRE:0.07SAN,WRE:0.14HK,WRE:0.73Triangle closing underestimates local clustering of low degree nodes19941996199820002002Year23456789101112EffectiveDiameterIncreasingOut-degreeoverTimem(t)19941996199820002002Year23456789101112ConstantOut-degree¯mHep-PHARWSANDMSFFUniformOut-degree&EffectiveDiameter(95thpercentile)networks grow over time and are not relevant to our work. Below, we discuss relevant and recent work on modeling network growth. Preferential Attachment & Fitness: In preferential attach- ment and fitness-based models [4, 5, 10, 33], a new node u links to an existing node v with probability proportional to the attachment function f (kv), a function of either degree kv or fitness ϕv of node v. For instance, linear preferential attachment functions [3, 13, 26] lead to power law degree distributions and small diameter [8] and attachment functions of degree & node age [49] can preserve re- alistic temporal dynamics. Extensions of preferential attachment [35, 50, 52] that incorporate resource constraints disregard network properties other than power law degree distribution and small di- ameter. Additional mechanisms are necessary to explain network properties such as clustering and attribute mixing patterns. Triangle Closing: A set of models [20, 24, 27] incorporate tri- adic closure using triangle closing mechanisms, which increase average local clustering by forming edges between nodes with one or more common neighbors. However, as explained in Sub- section 7.1, models based on preferential attachment and triangle closing do not preserve the local clustering of low degree nodes. Attributed network models: These models [12, 17, 22, 53] ac- count for the effect of attribute homophily on edge formation and preserve mixing patterns. Existing models can be broadly catego- rized as (a) fitness-based model that define fitness as a function of attribute similarity and (b) microscopic models of network evolu- tion that require complete temporal information about edge arrivals & deletion. Our experiment results in Subsection 5.2 show that well- known attributed network models SAN and KA preserve assortative mixing patterns, degree distribution to some extent, but not local clustering and degree-clustering correlation. Random walk models: First introduced by Vazquez [48], ran- dom walk models are inherently local. Models [7] in which new nodes only link to terminal nodes of short random walks generate networks with power law degree distributions [11] and small diam- eter [34] but do not preserve clustering. Models such as SK [43] and HZ [19], in which new nodes probabilistically link to each visited nodes incorporate triadic closure but are not flexible enough to preserve skewed local clustering of real-world networks, as shown in Subsection 5.2. We also observe that recursive random walk models such as FF [29] preserve temporal properties such as shrink- ing diameter but considerably overestimate local clustering and degree-clustering relationship of real-world networks. Furthermore, existing random walk models disregard the effect of homophily and do not model attribute mixing patterns. To summarize, existing models do not explain how resource constrained and local processes jointly preserve multiple global network properties of attributed networks. 9 CONCLUSION In this paper, we proposed a simple, interpretable model of attrib- uted network growth. ARW grows a directed network in the follow- ing manner: an incoming node selects a seed node based on attribute similarity, initiates a biased random walk to explore the network by navigating through neighborhoods of existing nodes, and halts the random walk after connecting to a few visited nodes. To the best of our knowledge, ARW is the first model that unifies multiple soci- ological phenomena -- bounded rationality; structural constraints; 10 triadic closure; attribute homophily; preferential attachment -- into a single local process to model global network structure and attribute mixing patterns. We explored the parameter space of the model to show how each parameter intuitively controls one or more key structural properties. Our experiments on six large-scale citation networks showed that ARW outperforms relevant and recent existing models by a statistically significant factor of 2.5 -- 10×. We plan to extend the ARW model in two directions: modeling undirected, social networks, and analyzing the effect of attribute homophily on the formation of temporal motifs [39] REFERENCES [1] [n. d.]. APS Datasets for Research. ([n. d.]). https://journals.aps.org/datasets [2] Waleed Ammar, Dirk Groeneveld, Chandra Bhagavatula, Iz Beltagy, Miles Craw- ford, Doug Downey, Jason Dunkelberger, Ahmed Elgohary, Sergey Feldman, Vu Ha, Rodney Kinney, Sebastian Kohlmeier, Kyle Lo, Tyler Murray, Hsu-Han Ooi, Matthew Peters, Joanna Power, Sam Skjonsberg, Lucy Lu Wang, Chris Wilhelm, Zheng Yuan, Madeleine van Zuylen, and Oren Etzioni. 2018. Construction of the Literature Graph in Semantic Scholar. In NAACL. [3] Albert-László Barabási and Réka Albert. 1999. Emergence of scaling in random networks. science 286, 5439 (1999), 509 -- 512. [4] Michael Bell, Supun Perera, Mahendrarajah Piraveenan, Michiel Bliemer, Tanya Latty, and Chris Reid. 2017. Network growth models: A behavioural basis for attachment proportional to fitness. Scientific Reports 7 (2017), 42431. [5] Ginestra Bianconi and Albert-László Barabási. 2001. Bose-Einstein condensation in complex networks. Physical review letters 86, 24 (2001), 5632. [6] Per Block and Thomas Grund. 2014. Multidimensional homophily in friendship networks. Network Science 2, 2 (2014), 189 -- 212. [7] Avrim Blum, TH Hubert Chan, and Mugizi Robert Rwebangira. 2006. A random- surfer web-graph model. In 2006 Proceedings of the Third Workshop on Analytic Algorithmics and Combinatorics (ANALCO). SIAM, 238 -- 246. [8] Béla Bollobás and Oliver Riordan. 2004. The diameter of a scale-free random graph. Combinatorica 24, 1 (2004), 5 -- 34. [9] Anna D Broido and Aaron Clauset. 2018. Scale-free networks are rare. arXiv preprint arXiv:1801.03400 (2018). [10] Guido Caldarelli, Andrea Capocci, Paolo De Los Rios, and Miguel A Munoz. 2002. Scale-free networks from varying vertex intrinsic fitness. Physical review letters 89, 25 (2002), 258702. [11] Prasad Chebolu and Páll Melsted. 2008. PageRank and the random surfer model. In Proceedings of the nineteenth annual ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied Mathematics, 1010 -- 1018. [12] Maurício L de Almeida, Gabriel A Mendes, G Madras Viswanathan, and Luciano R da Silva. 2013. Scale-free homophilic network. The European Physical Journal B 86, 2 (2013), 38. [13] Sergey N Dorogovtsev, Jose Ferreira F Mendes, and Alexander N Samukhin. 2000. Structure of Growing Networks: Exact Solution of the Barabási -- Albert's Model. arXiv preprint cond-mat/0004434 (2000). [14] James H Fowler and Sangick Jeon. 2008. The authority of Supreme Court prece- dent. Social networks 30, 1 (2008), 16 -- 30. [15] Johannes Gehrke, Paul Ginsparg, and Jon Kleinberg. 2003. Overview of the 2003 KDD Cup. ACM SIGKDD Explorations Newsletter 5, 2 (2003), 149 -- 151. [16] Gerd Gigerenzer and Daniel G Goldstein. 1996. Reasoning the fast and frugal way: models of bounded rationality. Psychological review 103, 4 (1996), 650. [17] Neil Zhenqiang Gong, Wenchang Xu, Ling Huang, Prateek Mittal, Emil Stefanov, Vyas Sekar, and Dawn Song. 2012. Evolution of social-attribute networks: mea- surements, modeling, and implications using google+. In Proceedings of the 2012 Internet Measurement Conference. ACM, 131 -- 144. [18] Phillip Good. 2013. Permutation tests: a practical guide to resampling methods for testing hypotheses. Springer Science & Business Media. [19] Carlos Herrera and Pedro J Zufiria. 2011. Generating scale-free networks with adjustable clustering coefficient via random walks. In Network Science Workshop (NSW), 2011 IEEE. IEEE, 167 -- 172. [20] Petter Holme and Beom Jun Kim. 2002. Growing scale-free networks with tunable clustering. Physical review E 65, 2 (2002), 026107. [21] Haibo Hu and Xiaofan Wang. 2009. Evolution of a large online social network. Physics Letters A 373, 12-13 (2009), 1105 -- 1110. [22] Fariba Karimi, Mathieu Génois, Claudia Wagner, Philipp Singer, and Markus Strohmaier. 2017. Visibility of minorities in social networks. arXiv preprint arXiv:1702.00150 (2017). [23] Kibae Kim and Jörn Altmann. 2017. Effect of homophily on network formation. Communications in Nonlinear Science and Numerical Simulation 44 (2017), 482 -- 494. [24] Konstantin Klemm and Victor M Eguiluz. 2002. Highly clustered scale-free networks. Physical Review E 65, 3 (2002), 036123. [25] Gueorgi Kossinets and Duncan J. Watts. 2009. Origins of Homophily in an Evolving Social Network. Amer. J. Sociology 115 (2009), 405 -- 450. http://www. journals.uchicago.edu/doi/abs/10.1086/599247 [26] Ravi Kumar, Prabhakar Raghavan, Sridhar Rajagopalan, D Sivakumar, Andrew Tomkins, and Eli Upfal. 2000. Stochastic models for the web graph. In Foundations of Computer Science, 2000. Proceedings. 41st Annual Symposium on. IEEE, 57 -- 65. [27] Jure Leskovec, Lars Backstrom, Ravi Kumar, and Andrew Tomkins. 2008. Mi- croscopic evolution of social networks. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 462 -- 470. [28] Jure Leskovec, Deepayan Chakrabarti, Jon Kleinberg, Christos Faloutsos, and Zoubin Ghahramani. 2010. Kronecker graphs: An approach to modeling networks. Journal of Machine Learning Research 11, Feb (2010), 985 -- 1042. [29] Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. 2005. Graphs over time: densification laws, shrinking diameters and possible explanations. In Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. ACM, 177 -- 187. [30] Barton L Lipman. 1995. Information processing and bounded rationality: a survey. Canadian Journal of Economics (1995), 42 -- 67. [31] Mary McGlohon, Leman Akoglu, and Christos Faloutsos. 2011. Statistical prop- erties of social networks. In Social network data analytics. Springer, 17 -- 42. [32] Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather: Homophily in social networks. Annual review of sociology 27, 1 (2001), 415 -- 444. [33] Matúš Medo, Giulio Cimini, and Stanislao Gualdi. 2011. Temporal effects in the growth of networks. Physical review letters 107, 23 (2011), 238701. [34] Abbas Mehrabian and Nick Wormald. 2016. Itś a small world for random surfers. Algorithmica 76, 2 (2016), 344 -- 380. [35] Stefano Mossa, Marc Barthelemy, H Eugene Stanley, and Luis A Nunes Amaral. 2002. Truncation of power law behavior in scale-free network models due to information filtering. Physical Review Letters 88, 13 (2002), 138701. [36] Mark Newman. 2010. Networks: an introduction. Oxford university press. [37] Mark EJ Newman. 2001. Clustering and preferential attachment in growing networks. Physical review E 64, 2 (2001), 025102. [38] Mark EJ Newman. 2002. Assortative mixing in networks. Physical review letters 89, 20 (2002), 208701. [39] Ashwin Paranjape, Austin R Benson, and Jure Leskovec. 2017. Motifs in temporal networks. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining. ACM, 601 -- 610. [40] Leto Peel, Jean-Charles Delvenne, and Renaud Lambiotte. 2018. Multiscale mixing patterns in networks. Proceedings of the National Academy of Sciences 115, 16 (2018), 4057 -- 4062. [41] Joseph J Pfeiffer III, Sebastian Moreno, Timothy La Fond, Jennifer Neville, and Brian Gallagher. 2014. Attributed graph models: Modeling network structure with correlated attributes. In Proceedings of the 23rd international conference on World wide web. ACM, 831 -- 842. [42] Dragomir R. Radev, Pradeep Muthukrishnan, Vahed Qazvinian, and Amjad Abu- Jbara. 2013. The ACL anthology network corpus. Language Resources and Evaluation (2013), 1 -- 26. https://doi.org/10.1007/s10579-012-9211-2 [43] Jari Saramäki and Kimmo Kaski. 2004. Scale-free networks generated by random walkers. Physica A: Statistical Mechanics and its Applications 341 (2004), 80 -- 86. [44] Georg Simmel. 1950. The sociology of georg simmel. Vol. 92892. Simon and [45] Herbert A Simon. 1955. On a class of skew distribution functions. Biometrika 42, Schuster. 3/4 (1955), 425 -- 440. 1, 1 (1972), 161 -- 176. [46] Herbert A Simon. 1972. Theories of bounded rationality. Decision and organization [47] Mayank Singh, Rajdeep Sarkar, Pawan Goyal, Animesh Mukherjee, and Soumen Chakrabarti. 2017. Relay-linking models for prominence and obsolescence in evolving networks. In Proceedings of the 23rd ACM SIGKDD International Confer- ence on Knowledge Discovery and Data Mining. ACM, 1077 -- 1086. [48] Alexei Vazquez. 2000. Knowing a network by walking on it: emergence of scaling. arXiv preprint cond-mat/0006132 (2000). [49] Dashun Wang, Chaoming Song, and Albert-László Barabási. 2013. Quantifying long-term scientific impact. Science 342, 6154 (2013), 127 -- 132. [50] Li-Na Wang, Jin-Li Guo, Han-Xin Yang, and Tao Zhou. 2009. Local preferential attachment model for hierarchical networks. Physica A: Statistical Mechanics and its Applications 388, 8 (2009), 1713 -- 1720. [51] Hao Yin, Austin R Benson, and Jure Leskovec. 2018. Higher-order clustering in networks. Physical Review E 97, 5 (2018), 052306. [52] Jianyang Zeng, Wen-Jing Hsu, and Suiping Zhou. 2005. Construction of scale-free networks with partial information. Lecture notes in computer science 3595 (2005), 146. [53] Elena Zheleva, Hossam Sharara, and Lise Getoor. 2009. Co-evolution of social and affiliation networks. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 1007 -- 1016. 11
1605.04635	2	1605	2017-02-20T08:20:52	Cumulative Activation in Social Networks	[ "cs.SI" ]	Most studies on influence maximization focus on one-shot propagation, i.e. the influence is propagated from seed users only once following a probabilistic diffusion model and users' activation are determined via single cascade. In reality it is often the case that a user needs to be cumulatively impacted by receiving enough pieces of information propagated to her before she makes the final purchase decision. In this paper we model such cumulative activation as the following process: first multiple pieces of information are propagated independently in the social network following the classical independent cascade model, then the user will be activated (and adopt the product) if the cumulative pieces of information she received reaches her cumulative activation threshold. Two optimization problems are investigated under this framework: seed minimization with cumulative activation (SM-CA), which asks how to select a seed set with minimum size such that the number of cumulatively active nodes reaches a given requirement $\eta$; influence maximization with cumulative activation (IM-CA), which asks how to choose a seed set with fixed budget to maximize the number of cumulatively active nodes. For SM-CA problem, we design a greedy algorithm that yields a bicriteria $O(\ln n)$-approximation when $\eta=n$, where $n$ is the number of nodes in the network. For both SM-CA problem with $\eta<n$ and IM-CA problem, we prove strong inapproximability results. Despite the hardness results, we propose two efficient heuristic algorithms for SM-CA and IM-CA respectively based on the reverse reachable set approach. Experimental results on different real-world social networks show that our algorithms significantly outperform baseline algorithms.	cs.SI	cs	Cumulative Activation in Social Networks Xiaohan Shan Institute of Computing Technology, CAS [email protected] Wei Chen Microsoft [email protected] Qiang Li Institute of Computing Technology, CAS [email protected] Xiaoming Sun Institute of Computing Technology, CAS [email protected] Jialin Zhang Institute of Computing Technology, CAS [email protected] 7 1 0 2 b e F 0 2 ] I S . s c [ 2 v 5 3 6 4 0 . 5 0 6 1 : v i X r a ABSTRACT Most studies on influence maximization focus on one-shot propagation, i.e. the influence is propagated from seed users only once following a probabilistic diffusion model and users' activation are determined via single cascade. In reality it is often the case that a user needs to be cumulatively im- pacted by receiving enough pieces of information propagated to her before she makes the final purchase decision. In this paper we model such cumulative activation as the follow- ing process: first multiple pieces of information are prop- agated independently in the social network following the classical independent cascade model, then the user will be activated (and adopt the product) if the cumulative pieces of information she received reaches her cumulative activa- tion threshold. Two optimization problems are investigated under this framework: seed minimization with cumulative activation (SM-CA), which asks how to select a seed set with minimum size such that the number of cumulatively active nodes reaches a given requirement η; influence max- imization with cumulative activation (IM-CA), which asks how to choose a seed set with fixed budget to maximize the number of cumulatively active nodes. For SM-CA problem, we design a greedy algorithm that yields a bicriteria O(ln n)- approximation when η = n, where n is the number of nodes in the network. For both SM-CA problem with η < n and IM-CA problem, we prove strong inapproximability results. Despite the hardness results, we propose two efficient heuris- tic algorithms for SM-CA and IM-CA respectively based on the reverse reachable set approach. Experimental results on different real-world social networks show that our algorithms significantly outperform baseline algorithms. Keywords social networks; independent cascade model; cumulative ac- tivation; influence maximization; seed minimization 1. INTRODUCTION With the wide popularity of social media and social net- work sites such as Facebook, Twitter, WeChat, etc., social networks have become a powerful platform for spreading in- formation, ideas and products among individuals. In partic- ular, product marketing through social networks can attract large number of customers. Motivated by this background, influence diffusion in so- cial networks has been extensively studied (cf. [11, 20, 7]). However, most of previous works only consider the influence after one-shot propagation -- influence propagates from the seed users only once and user activation or adoption is fully determined after this single cascade. In contrast, in the real world, people often make decisions after they have cumu- lated many pieces of information about a new technology, new product, etc., and these different pieces of information are propagated in the network independently as different information cascades. Consider the following scenario: A company is going to launch a new version (named as V11 for convenience) of their product with many new features, but most people are not familiar with these new features. Thus, it is often ben- eficial that the company conduct a series of advertisement and marketing campaigns covering different features of the product. An effective way of marketing in a social network is to select influential users as seeds to initiate the informa- tion cascades of these campaigns. From potential customers' perspective, when they receive the first piece of information about V11 from their friends, they may find it interesting and forward it to their friends, but this may not necessarily lead to their purchase actions. Later they may receive and be impacted by further information about V11, and when they are impacted by enough pieces of information cascades, they may finally decide to buy the new product. We model the above behavior by an integrated process consisting of two phases: (a) repeated information cascades, and (b) threshold-based user adoptions. First, there are multiple information cascades about multiple pieces of pro- duction information in the network. We model information cascades by the classical independent cascade (IC) model [20]: A social network is modeled as a weighted directed graph, with an influence probability as the weight on every edge. Initially, some nodes are selected as seeds and become active, and all other nodes are inactive. At each step, newly activated nodes have one chance to influence each of their inactive out-neighbors with the success probability given on the edge. Independent cascade model is suitable to model simple contagions [5, 7] such as virus and information propa- gation, and thus we adopt it to model information cascades in the first phase. We consider multiple pieces of produc- tion information propagates independently following the IC model. For the second phase, we assume that there is a threshold for each user, who will adopt the product if the amount of information that she receives in the first phase exceeds her threshold. We measure the amount of informa- tion a user received as the fraction of information cascades that reaches the user, which is equivalent to the probability of the user being activated in an information cascade. A node is cumulatively activated if this probability exceeds the threshold. We refer to this model as the cumulative activa- tion (CA) model. Given the above cumulative activation model, the com- pany may face one of the following two objectives: either the company has a fixed budget to activate the seed nodes, and wants to maximize the number of cumulative active nodes, or the company needs to reach a predetermined number of cumulative active nodes, and wants to minimize the number of seeds. We formulate the above scenarios as the following two optimization problems: Seed minimization with cumulative activation (SM-CA) and influence maximization with cu- mulative activation (IM-CA). Given a directed graph with a probability on each edge and a threshold for each node, an activation requirement η and a budget k, the SM-CA prob- lem is to find a seed set with minimum size such that the number of cumulatively activated nodes is at least η. The IM-CA problem is to find a seed set with k nodes such that the number of cumulatively activated nodes is maximum. Let ρ(S) denote the number of cumulative activated nodes given seed set S. We first show that set function ρ(·) is not submodular, which means unlike most of the current studies, we cannot guarantee the approximation ratio by using the greedy algorithm directly. For SM-CA problem, we consider the case η = n and η < n separately, where n is the number of nodes in the net- work and η is the activation requirement. The complexity results of these two cases are quite different. When η = n, we show while it is NP-hard to approximate SM-CA problem within factor (1 − ε) ln n for any ε > 0, we can achieve a bi- criteria O(ln n)-approximation. Our technique is to replace the nonsubmodular ρ(S) with a submodular surrogate func- tion f (S), and show that the set of feasible solutions to the original SM-CA problem with constraint η = n is exactly the same as the set of feasible solutions for f (S) to assume its maximum value, and then we can apply the greedy algorithm to the surrogate f (S) instead of ρ(S) to provide the theo- retical guarantee. When η < n, we construct a reduction from the densest k-subgraph problem to SM-CA problem and show that SM-CA problem cannot be approximated within factor 1√ nδ/2 if the densest k-subgraph problem cannot be approximated within nδ, for any δ > 0, which is commonly acknowledged as a hard problem for some small δ. 6 For IM-CA problem, we construct a reduction from the Set Cover problem and prove that it is NP-hard to approxi- mate IM-CA problem within a factor of n1−ε for any ε > 0. Despite the approximation hardness on the SM-CA prob- lem with η < n and the IM-CA problem, we may still need practical solutions for them. For this purpose, we propose some heuristic algorithms, which utilize the state-of-the-art approach in influence maximization, namely the Reverse Reachable Set (RR set) approach [3, 26, 30, 29], to improve the efficiency of the algorithms comparing to the old greedy algorithms based on naive Monte Carlo simulations. Finally, we conduct experiments on three real-world so- cial networks to test the performance of our algorithms. Our results demonstrate that one heuristic algorithm pro- posed consistently out-performs all other algorithms under comparison in all test cases and clearly stands out as the winning choice for both the SM-CA and IM-CA problems. To summarize, our contributions include: (a) we propose the seed minimization and influence maximization problem under cumulative activation (SM-CA problem and IM-CA problem respectively), which is a reasonable model for pur- chasing behavior of customers exposed to repeated informa- tion cascades; (b)we design an O(ln n) approximate algo- rithm for SM-CA problem when η = n; (c) we show strong hardness results for SM-CA problem with η < n and IM- CA problem; (d) we propose efficient heuristic algorithms and validate them through extensive experiments on real- world datasets and conclude that one heuristic is the best choice for both SM-CA and IM-CA problems. 1.1 Related Work The classical influence maximization problem is to find a seed set of at most k nodes to maximize the expected num- ber of active nodes. It is first studied as an algorithmic problem by Domingos and Richardson [11] and Richardson and Domingos [27]. Kempe et al. [20] first formulate the problem as a discrete optimization problem. They summa- rize the independent cascade model and the linear threshold model, and obtain approximation algorithms for influence maximization by applying submodular function maximiza- tion. Extensive studies follow their approach and provide more efficient algorithms [10, 9, 23]. Leskovec et al. [23] present a "lazy-forward" optimization method in selecting new seeds, which greatly reduce the number of influence spread evaluations. Chen et al. [10, 9] propose scalable algorithms which are faster than the greedy algorithms pro- posed in [21]. Recently, Borgs et al. [3] and Tang et al. [29, 30] and Nguyen et al. [26] propose a series of more ef- fective algorithms for influence maximization in large social networks that both has theoretical guarantee and practical efficiency. The approach is based on the "Reverse Reachable Set" idea first proposed in [3]. Another aspect of influence problem is seed set minimiza- tion, Chen [6] studies the seed minimization problem under the fix threshold model and shows some strong negative re- sults for this model. Long et al. [24] also study independent cascade model and linear threshold model from a minimiza- tion perspective. In [15], Goyal et al. study the problem of finding the minimum size of seed set such that the ex- pected number of active nodes reaches a given threshold, they provide a bicriteria approximation algorithm for this problem. Zhang et al. [32] study the seed set minimization problem with probabilistic coverage guarantee, and design an approximation algorithm for this problem. He et al. [19] study positive influence model under single-step activation and propose an approximation algorithm. Note that, the work in [19] is a special case of our work. Beyond influence maximization and seed minimization, another interesting direction is the learning of social influ- ence over real online social network data set, e.g. influence learning in blogspace [17] and academic collaboration net- work [28]. Most early studies on influence maximization and influ- ence learning are summarized in the monograph [7]. How- ever, almost all the existing studies consider only node acti- vation after a single information or influence cascade. Our work differentiate with all these studies on this important aspect, as discussed in the introduction. Paper organization. We formally define the diffusion model and the optimization problems SM-CA and IM-CA in Sec- tion 2. The approximation algorithms and hardness results of these two problems are proposed in Section 3, including a greedy algorithm for SM-CA problem with η = n in section 3.1.1, the hardness result of SM-CA problem with η < n in Section 3.1.2 and the inapproximate result of IM-CA prob- lem in Section 3.2. In Section 4, we present two heuristic algorithms for SM-CA problem and two heuristic algorithms for IM-CA problem. Section 5 shows our experimental re- sults on real-world datasets. We summarize the paper with some further directions in Section 6. 2. MODEL AND PROBLEM DEFINITIONS Our social network is defined on a directed graph G = (V, E), where V is the set of nodes representing individu- als and E is the set of directed edges representing social ties between pairs of individuals. Each edge e = (u, v) ∈ E is as- sociated with an influence probability puv, which represents the probability that u influences v. The entire activation process consists of information dif- fusion process and node activation. The information dif- fusion process follows the independent cascade (IC) model proposed by Kempe et al. [20]. In the IC model, discrete time steps t = 0, 1, 2,··· are used to model the diffusion process. Each node in G has two states: inactive or active, At step 0, a subset S ⊆ V is selected as seed set and nodes in S are active directly, while nodes not in S are inactive. For any step t ≥ 1, if a node u is newly active at step t − 1, then u has a single chance to influence each of its inactive out-neighbor v with independent probability puv to make v active. Once a node becomes active, it will never return to the inactive state. The diffusion process stops when there is no new active nodes at a time step. The above basic IC model describe the diffusion of one piece of information, but actually there could be many pieces of information about a product being propagated in the net- work, all following the same IC model. Users' final produc- tion adoption is based on cumulative information collected, which we refer to as cumulative activation (CA) and is de- scribed below, and it is different from the user becoming active for one piece of information specified above in the IC model. Let Pu(S) be the probability that u becomes active after an information cascade starting from the seed set S. Since Pu(S) also represents the fraction of information ac- cepted by u in multiple cascades, we use Pu(S) to define cumulative activation: Suppose that each node u ∈ V has an activation threshold τu ∈ (0, 1], then u becomes cumula- tively active if Pu(S) ≥ τu. Given a target set U ⊆ V and a seed set S, let ρU (S) be the number of cumulatively active nodes in U from seed set S. When U = V , we omit the subscript U and use ρ(S) directly. We consider two optimization problems under cumula- tive activation, seed minimization with cumulative activa- tion (SM-CA) and influence maximization with cumulative activation (IM-CA). SM-CA aims at finding a seed set S with minimum size such that there are at least η (η ≤ n) nodes in the target set become cumulatively active. IM-CA is the problem of finding a seed set of size k to maximize the number of cumulatively active nodes in the target set. The formal definitions are as follows. Definition 1 (Seed minimization with cumulative ac- tivation) In the seed minimization with cumulative activa- tion (SM-CA) problem, the input includes a directed graph G = (V, E) with V = n, E = m, an influence probability vector P = {puv : puv ∈ [0, 1], (u, v) ∈ E}, a target set U ⊆ V , an activation threshold τu ∈ (0, 1] for each node u ∈ U and a coverage requirement η ≤ U. Our goal is to find the minimum size seed set S∗ ⊆ V such that at least η nodes in U can be cumulatively activated, that is, S∗ = arg min S:ρU (S)≥η S. Definition 2 (Influence maximization with cumula- tive activation) In the influence maximization with cumu- lative activation (IM-CA) problem, the input includes a di- rected graph G = (V, E) with V = n, E = m, an influence probability vector P = {puv : puv ∈ [0, 1], (u, v) ∈ E}, a tar- get set U ⊆ V , an activation threshold τu ∈ (0, 1] for each node u and a size budget k ≤ n. Our goal is to find a seed set S∗ ⊆ V of size k such that the number of cumulatively active nodes in U is maximized, that is, S∗ = arg max S:S=k ρU (S). 2.1 Equivalence to Frequency-based Definition Suppose there are N diffusions, which lead to final cumu- lative activation. Intuitively, a node u ∈ V becomes cumu- lative activated when the number of times that u becomes influenced in these N diffusions is larger than a threshold. Formally, given a seed set S and a node u ∈ V , let X i u be a random variable defined as follows: X i u(S) = 1 if u is influ- enced in the i-th diffusion and X i u(S) = 0 otherwise. Thus, u(S) denotes the number of times that u becomes influenced after N diffusions, and Xu(S)/N is the influence frequency of u. By Hoeffding's inequality, we show the relationship between pu(S) > τu and Xu(S)/N ≥ τu in Lemma 1. Xu(S) = (cid:80)N i=1 X i Lemma 1. Given a seed set S, a node u ∈ V and a large enough N , (a) if pu(S) > τu, then Pr(Xu(S)/N ≥ τu) = 1 − o(1); and (b) if pu(S) < τu, then Pr(Xu(S)/N ≤ τu) = 1− o(1), where o(1) is asymptotic to the number of diffusions N . Proof. It is obvious that the expectation of Xu(S) is u(S)] = N pu(S). When Pu(S) < τu, i=1 E[X i E[Xu(S)] =(cid:80)N by Hoeffding's inequality, we have: Pr(Xu(S)/N ≥ τu) = Pr(Xu(S) − E[Xu(S)] ≥ N τu − E[Xu(S)]) ≤ exp(−2N (N τu − E[Xu(S)])2) = exp(−2N 3(τu − Pu(S))2) = o(1). Thus, when N is large enough, Xu(S)/N ≤ τu is a high probability event if pu(S) < τu. Similarly, Xu(S)/N ≥ τu is a high probability event if pu(S) > τu. Based on Lemma 1, the formal definition of cumulative activation is consistent with our motivation. 2.2 Comparison with IC and LT Models We first explain the differences between the CA model the IC model. CA model uses IC model as information cas- cades in its first stage, and thus the main difference is at the determination of which nodes are finally activated, or simply at the objective function. This is clearly illustrated by the simple example in Figure 1(a), which shows a five- node graph with edge probabilities shown next to edges. In the IC model, it is clear that the influence spread of v1 and v2 are the same: σ({v1}) = σ({v2}) = 2.2. For the CA model, if every node has the activation threshold as 0.3, then ρ({v1}) = 2 and ρ({v2}) = 4, because v1 can only acti- vate itself and u1, while v2 can activate itself and u1, u2, u3. If the activation threshold of every node is increased to 0.6, then ρ({v1}) = 2 and ρ({v2}) = 1. Therefore, if we want to select one seed in the influence maximization task, for IC model either v1 or v2 is fine, but for the CA model v1 or v2 is selected based on different threshold values. This means the influence maximization task under CA model is differ- ent from the task under the IC model. The above example also provides the intuition that the influence maximization task under the IC model focuses on the average effect of the influence, while the task under the CA model may need to select either nodes that has wide but average influence (e.g. v2) or nodes with concentrated influence (e.g. v1) based on the threshold setting. threshold (LT) model proposed in [20]. each edge (u, v) has a weight wuv ∈ [0, 1] with (cid:80) We next distinguish our CA model with the popular linear In the LT model, u wuv ≤ 1 (wuv = 0 if (u, v) is not an edge). Each node v has a threshold θv, which is drawn from [0, 1] uniformly at random before the propagation starts. Then, starting from the seed set S, an inactive node v becomes active at time t ≥ 1 if any only if the total weights from its active in-neighbors exceeds wuv ≥ θv, where St is the set of v's threshold: (cid:80) u∈St−1 active nodes at time t with S0 = S. Despite the superficial similarity on using thresholding to model user adoption behavior, the two models are quite dif- ferent. One key difference is that in the LT model what is being propagated are the user adoption behavior, while in the CA model, what is being propagated are multiple pieces of information about a product, and user's adoption in the end is based on the information received. This is actually the difference between CA and most other models on influ- ence diffusion, as discussed in the introduction. This further leads to a specific difference between LT and CA: the thresh- old in the LT model is on the number of friends who already adopt a product, while the threshold in the CA model is on the fraction of information cascades that reach a user. Fi- nally, in LT the threshold θv is a random number in [0, 1], making the influence spread objective function submodular, while in CA the threshold τv is fixed as an input, causing the objective function ρ(S) not submodular, as discussed in the next section. approximation ratio by using the greedy algorithm [25], and indeed most of the existing work on social influence takes advantage of this nature (e.g. [3, 8, 15, 20]). Unfortunately, our objective function ρU (S) is monotone but not submodular in general as shown in the example be- low, which makes our problems much harder. Example 1. (Figure 1(b)) Suppose G is a bipartite graph and the influence probability on each edge is 1/2, the acti- vation thresholds are: τu = 7/8, τa = τb = τc = 1. Let U = {u}. Let S = {a} and T = {a, b}, then, ρU (S) = 0, ρU (T ) = 0, ρU (S ∪ {c}) = 0, ρU (T ∪ {c}) = 1. Thus, ρU (S ∪{c})− ρU (S) = 0 < ρU (T ∪{c})− ρU (T ) = 1, imply- ing that ρU (S) is not submodular. We further remark that in this example, if we set U = V , the function ρ(S) = ρV (S) is still not submodular. In the rest part of this section, we consider how to design approximation algorithms for SM-CA problem and IM-CA problem as well as the computational complexity of them. 3.1 Seed minimization with cumulative acti- vation (SM-CA) problem In this section, we study SM-CA problem. We first show the hardness result of SM-CA problem in Theorem 1. Theorem 1. SM-CA problem is NP-hard. Moreover, SM- CA problem cannot be approximated within (1 − ε) ln η in polynomial time unless N P ⊆ DT IM E(nO(log log n)), ∀ ε > 0. Proof. We construct a reduction from the Partial Set Cover (PSC) problem. An instance of PSC problem IP SC = (U,S, η) consists of a ground set U and a family of subsets S ⊆ 2U and a coverage requirement η ≤ U. The objective is to find a subcollection C ⊆ S such that (cid:83)C ≥ η and C is minimized. Given any instance of PSC problem IP SC = (U,S, η), we construct an instance of SM-CA problem ISM−CA = (G, η) as follows. The PSC instance is reduced to a bipartite graph G = (V, W, E), in which each node v ∈ V corresponds to a subset Cv ∈ S one to one, each node w ∈ W corresponds to an element uw ∈ U one to one, there is an edge (v, w) ∈ E if and only if uw ∈ Cv. The target set is W , the influence probability on each edge is 1, the activation thresholdof each node in W is 1. The activation requirement is η, which is the same with the coverage requirement in the PSC instance. Based on the above construction, it is easy to check that the objective function values for any given S in the instances ISM−CA and IP SC are always the same. Which means the two problems should have the same approximation ratio. For the PSC problem, Feige showed that it cannot be ap- proximated within a factor of (1− ε) ln η (∀ε > 0) in polyno- mial time unless N P ⊆ DT IM E(nO(log log n)) [13]. There- fore, SM-CA problem has the same computational complex- ity. 3. ALGORITHMS AND HARDNESS RESULTS In this section, we provide algorithmic as well as hardness results for SM-CA problem and IM-CA problem. A set function f : 2V → R is monotone if f (S) ≤ f (T ) for all S ⊆ T ⊆ V , and submodular if f (S ∪ {w}) − f (S) ≥ f (T ∪ {w}) − f (T ) for all S ⊆ T ⊆ V and w /∈ T . It is well known that monotone submodular functions leads to a good Based on the hardness result of SM-CA problem, our next goal is to design an algorithm with approximation ratio close to ln η. Surprisely, it turns out that the results are quite different between "activating all nodes" (η = n) and "partial activation" (η < n), as we discuss separately below. 3.1.1 SM-CA problem with η = n (a) CA & IC (b) nonsubmoduarity (c) η < n Figure 1: Figures for understanding the model When η = n, we can design an algorithm with a bicriteria O(ln n)-approximation, even though the objective function is not submodular. The key idea to our solution is to find a submodular function f (S) as the surrogate for the original nonsubmodular ρ(S), as the following lemma specifies. to the SM-CA problem if and only if f (S) =(cid:80) Lemma 2. When η = n, a seed set S is a feasible solution u∈V τu, where f (S) is a surrogate function defined as: f (S) =(cid:80) u∈V min{Pu(S), τu}. forward. Proof. If S is a feasible solution to SM-CA, that is, ρ(S) = n = η, then every node u satisfies Pu(S) ≥ τu, u∈V τu. The only-if part is also straight- and thus f (S) =(cid:80) set size for the constraint of f (S) =(cid:80) The above lemma implies that minimizing seed set size for the constraint of ρ(S) = n is the same as minimizing the seed u∈V τu. The reason we want to switch the minimization problem on the surrogate function f (S) is because it is submodular, as pointed out by Lemma 3. We remark that in [12], Farajtabar et al. study an objective function in the similar form as f (S) in the continuous-time influence model, but the interpretation of τ is the cap on user activity intensity in [12] rather than the activation threshold. Lemma 3. The surrogate function f (S) is monotone and submodular. Proof sketch. It is obvious that f (S) is monotone. For submodularity, following [20] we know that Pu(S) is sub- modular. Then it is easy to check that the minium of a submodular function and a constant is still submodular, and the simple summation of submodular functions is also sub- modular. hard [9], where σ(S) =(cid:80) is also #P-hard since σ(S) =(cid:80) Having the submodularity, we can design a greedy algorithm guided by f (S). But like most work in the IC model, we cannot avoid the problem of computing f (S). It has been shown that exactly computing σ(S) in the IC model is #P- u∈V Pu(S) is the expected number of active nodes given the seed set S. Thus, computing f (S) u∈V Pu(S) = f (S) if we set τu = 1 for all u ∈ V . In this section, we use Monte Carlo simulation to estimate f (S). A more efficient method will be discussed in Section 4.2. Algorithm 1 shows the procedure of the Monte Carlo method. Given a seed set S and a node u, Algorithm 1 simulates the diffusion process from S for R runs, and uses the frequency that u has been influenced as the estimation of Pu(S). Then we can obtain the estimation of f (S) directly by a truncation tu = tu + 1 f (S): the estimation of f (S) simulate IC diffusion from seed set S if u is activated then Algorithm 1 Estimate f (S) by Monte Carlo Input: G = (V, E), {puv}(u, v)∈E,{τu}u∈V , U, S, R Output: 1: f (S) = 0; 2: Pu(S) = 0; tu = 0 for all u ∈ V 3: for i = 1 to R do 4: 5: 6: 7: end if 8: end for 9: for u ∈ U do 10: 11: 12: 13: 14: end if 15: 16: end for 17: return f (S) Pu(S) = tu/R if Pu(S) ≥ τu then f (S) = f (S) + τu f (S) = f (S) + Pu(S) else Algorithm 2 Greedy algorithm for SM-CA with η = n Input: G = (V, E), {puv}(u, v)∈E,{τu}u∈V , U , ε Output: 1: S = ∅, f (S) = 0 Seed set S 2: while f (S) <(cid:80) u∈V τu − ε do choose v = arg maxu∈V \S [ f (S ∪ {u}) − f (S)] S = S ∪ {v} 3: 4: 5: end while 6: return S operation. The estimations of Pu(S) and f (S) are denoted by Pu(S) and f (S) respectively. The accuracy of the estimate f (S) depends on the number of simulation runs R, as rigorously specified by the following lemma. Lemma 4. For any seed set S, suppose f (S) is the esti- mate of f (S) output by Algorithm 1, then ∀ γ > 0, δ > 0, Pr( f (S)−f (S) ≤ γ) ≥ 1−1/nδ if R ≥ (n2 ln(2nδ+1))/2γ2. Proof. For each node u, Let Xu = (cid:80)R tions. Thus, Xu = R · Pu(S) and E[Xu] =(cid:80)R u , where X (i) u = 1 if u is influ- enced in the i-th simulation and X (i) u = 0 otherwise. Then Xu is the number of times that u is active after R simula- u ] = R · Pu(S). By Hoeffding's inequality and the condition R ≥ u is a random variable defined as X (i) i=1 E[X (i) i=1 X (i) (n2 ln(2nδ+1))/2γ2, for any constant γ > 0 and δ > 0, Pr( Pu(S) − Pu(S) ≥ γ/n) = Pr(Xu − E[Xu] ≥ Rγ/n) ≤ 2 exp ≤ 1 (cid:33) (cid:32) − 2( Rγ n )2 R always holds. nδ+1 . We next show that f (S)−f (S) ≤(cid:80) min{ Pu(S), τu} −(cid:88) (cid:12)(cid:12)(cid:12)min{ Pu(S), τu} − min{Pu(S), τu}(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88) f (S) − f (S) = u∈V u∈V u∈V Pu(S)−Pu(S) min{Pu(S), τu} ≤ (cid:88) ≤ (cid:88) u∈V u∈V Pu(S) − Pu(S). Then, we have (cid:88) u∈V Pr( f (S) − f (S) ≤ γ) ≥ Pr( Pu(S) − Pu(S) ≤ γ) ≥ Pr(∀u ∈ V, Pu(S) − Pu(S) ≤ γ/n) = 1 − Pr(∃u ∈ V, Pu(S) − Pu(S) ≥ γ/n) Pr( Pu(S) − Pu(S) ≥ γ/n) ≥ 1 − 1 nδ . ≥ 1 −(cid:88) u∈V Having the estimation algorithm of f (S), we show our greedy algorithm for SM-CA problem with η = n in Algorithm 2. f (S)]. The algorithm ends when f (S) ≥(cid:80) Algorithm 2 starts from an empty seed set S. At each iteration, it adds one node v providing the largest marginal [ f (S ∪ {u}) − increment to f (S) into S, i.e., v = arg max u∈V τu − ε and outputs S as the selected seed set. Goyal et al. proved the performance guarantee for the greedy algorithm when f (S) is monotone and submodular [15]. u∈V Theorem 2 ([15]) Let G = (V, E) be a social graph and f (·) be a nonnegative, monotone and submodular function defined on 2V . Given a threshold 0 < η ≤ f (V ), let S∗ ⊆ V be a subset with minimum size such that f (S∗) ≥ η, and S be the greedy solution using a (1 − γ)-approximate function f (·) with the stopping criteria f (S) ≥ η − ε. Then, there exists a γ such that for any ϕ > 0 and ε > 0, S ≤ S∗(1 + ϕ)(1 + ln(η/ε)) with high probability. Now we can conclude the approximation ratio of Algorithm 2 based on Lemmas 2 -- 4 and Theorem 2. rithm 2 ends when f (S) ≥ (cid:80) Theorem 3. When η = n, for any φ > 0, ε > 0, Algo- u∈V τu − ε and approximates ) SM-CA problem within a factor of (1 + φ) · (1 + ln with high probability. u∈V τu (cid:80) ε SM-CA problem with η < n 3.1.2 When η < n, the surrogate function f (S) does not enjoy the property in Lemma 2 any more, and thus the problem becomes more difficult. We use the following example to explain this phenomenon. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Example 2. (Figure 1(c)). Suppose the influence proba- bility on each edge from S1 is 0.5 and each edge from S2 is 1. The activation threshold of each node is 1, and η = 3. Then f (S1) = 5, f (S2) = 4, but ρ(S1) = 2, ρ(S2) = 4. Thus, S1 is not a feasible solution even though f (S1) is large enough. This simple example shows that too many "small active prob- ability" nodes may mislead f (S) causing it to diverge signif- icantly from ρ(S). Now we show the hardness result of SM-CA problem with η < n. Our analysis is based on the hardness of the densest k-subgraph (DkS) problem [14]. An instance of DkS problem consists of an undirected graph G = (V, E) and a parameter k < n, where n = V . The objective is to find a subset V (cid:48) ⊆ V of cardinality k such that the number of edges with both endpoints in V (cid:48) is maximized. The first polynomial time approximation algorithm for DkS problem is given by Feige et al. in 2001 [14] with the approximation ratio O(n1/3). This result was improved to O(n1/4+ε) (for any ε > 0) by Bhaskara et al. [2] in 2010 and this is the currently the best known guarantee. For the hardness of DkS problem, Khot [22] proved that the DkS problem does not admit PTAS under the assumption that NP problems does not have sub-exponential time random- ized algorithms. The exact complexity of approximating DkS problem is still open, but it is widely believed that DkS problem can only be approximated within polynomial ratio. Partially borrowing the idea in [18], we can prove a hard- ness result for SM-CA problem with η < n based on the hardness of DkS problem. Theorem 4. When η < n, SM-CA problem cannot be nδ/2 if DkS problem cannot be ap- approximated within 1√ proximated within nδ, for any δ > 0. 6 Proof. Suppose there is a polynomial time approxima- tion algorithm A with performance ratio r for SM-CA with η < n, we design an algorithm for DkS problem based on A, which has approximation ratio 6r2, hence the theorem follows. Given any instance of DkS problem on graph G = (V, E), construct an instance (denoted by SM-CA-I) of SM-CA prob- lem as follows. It is defined on a one-way bipartite graph G(cid:48) = (V (cid:48) = V1 ∪ V2, E(cid:48)), where V1 = V, V2 = E, the di- rected edge set E(cid:48) = {(v, e) : ∀ v ∈ V1, e ∈ V2, and v is one of the endpoints of e in E}. The probability on each edge e(cid:48) = (v, e) is pve = 1/2. The target set U = V1 ∪ V2, for each node e ∈ V2, τe = 3/4 and for each node v ∈ V1, τv = 1. For any k, let η = η(k) be the maximum threshold requirement for which A outputs a solution for SM-CA with k nodes. That is to say, A outputs a seed set with k nodes if the threshold is η(k) and at least k +1 nodes if the threshold is larger than η(k). 1 It is clearly that, in SM-CA-I, nodes in V2 are no better than nodes in V1 as candidates of seed since the target set is the set of all nodes, select a node in V2 can only activate itself, but a node in V1 may help to activate nodes in V2. So here we assume that all seeds selected by algorithm A are from V1. Since for each edge (v, e) ∈ E(cid:48), pve = 1/2 and for each node e ∈ V2, τe = 3/4, an easy probability calculation 1For any k, η(k) can be computed efficiently by using algo- rithm A and linear search. implies that a node e ∈ V2 can be cumulatively activated if and only if both endpoints of e are selected as seeds. Suppose the seed set of SM-CA-I with parameter η = η(k) computed by algorithm A is S(cid:48), then we can use the corresponding node set S in graph G as an approximate solution of the DkS problem. Indeed, we have S = k. Since in SM-CA-I S(cid:48) cumulatively activates at least η nodes, only k of them are in V1, so at least η − k nodes are cumulatively activated in V2. Therefore, in graph G the number of edges induced by S is at least η − k. Without loss of generality, we can assume η ≥ k + (cid:98)k/2(cid:99), this is because we can easily choose k nodes from V1 to cumulatively active (cid:98)k/2(cid:99) nodes in V2. It is easy to check that η − k ≥ 1 Suppose the optimal solution of DkS problem contains opt edges, then it is sufficient to show opt ≤ 2(η−2)r2. Indeed, if we can prove opt ≤ 2(η−2)r2, then we have opt ≤ 6(η−k)r2, which means there is a 6r2-approximate algorithm for the DkS problem. In SM-CA-I, based on the choice of η and the fact that A is a r-approximate algorithm, any seed set with size (cid:98)k/r(cid:99) can cumulatively activate at most η nodes. Thus, at most η − (cid:98)k/r(cid:99) nodes in V2 can be cumulatively activated by any (cid:98)k/r(cid:99) seeds in V1. This is equivalent to the fact that there are at most η−(cid:98)k/r(cid:99) edges induced by any (cid:98)k/r(cid:99) vertexes in G. Thus, for any T ⊆ V with T = k, all possible subset of (cid:0) k(cid:98)k/r(cid:99) edges and each edge is counted exactly(cid:0) k−2 (cid:1) vertexes in T can induce at most (η − (cid:98)k/r(cid:99))(cid:0) k(cid:98)k/r(cid:99) (cid:1) (cid:1) times. So, 3 (η − 2). (cid:98)k/r(cid:99)−2 if k > 2r, the total number of edges induced by T is at most ≤ r2(η − (cid:98)k/r(cid:99)) k−1 k−r < 2(η − 2)r2. (η−(cid:98)k/r(cid:99))( k(cid:98)k/r(cid:99)) ( k−2 (cid:98)k/r(cid:99)−2) if k ≤ 2r, then opt ≤ (cid:0)k (cid:1) ≤ k2 2 ≤ 2r2. By the arbitrary chosen of T , we have opt ≤ 2(η − 2)r2 and this completes the proof. 2 We remark that when η = n, it corresponds to the case of k = n in the DkS problem, which has a trivial solution and makes the theorem statement vacuously true. Thus we add η < n just to emphasize that the theorem is only useful when η < n. 3.2 Influence maximization with cumulative ac- tivation (IM-CA) problem In IM-CA problem, we prove a strong inapproximability result even when the base graph is a bipartite graph. Theorem 5. For any ε > 0, it is NP-hard to approximate IM-CA problem within a factor of N 1−ε, where N is the input size. Proof. Similar to the proof of inapproximability result in [20], We construct a reduction from SET COVER prob- lem. The input of the SET COVER problem includes a ground set W = {w1, w2, . . . , wn}, a collection of sub- sets S1, S2, . . . , Sm ⊂ W , and a positive integer k < m. The question is whether there exists k subsets whose union is W . Given an instance of the set cover problem, we construct an instance of IM-CA problem as follows: There are three types of nodes, set nodes, element nodes, and dummy nodes. There is a set node u corresponding to each set, an ele- ment node v corresponding to each element, and a directed edge (u, v) with activation probability puv = 1 if the el- ement represented by v is belong to the set represented (cid:101) + (cid:100) log m by u and puv = 0 otherwise. There are nc dummy nodes log n (cid:101) + 1), and there x1, x2,··· , xnc (where c = 2(cid:100) 1 is a directed edge (v, x) for each v and x. The activation probability on (v, x) is pvx = 1/2. The activation thresh- olds of set nodes, element nodes and dummy nodes are τu = τv = 1, τx = 1 − 1 2n , respectively. The budget of the size of a seed set is k and the target set is all nodes. Notice that the input size of our IM-CA problem is N = nc +n+m, so N 1− < 2nc N ≤ nc n+k . Under our construction, if there exists a collection of k sets covering all elements in W for SET COVER problem, then in IM-CA problem, the node set corresponding to the col- lection denoted by C will cumulatively activate all element nodes and all dummy nodes. In total, there will be nc +n+k nodes become cumulatively active. On the other hand, let's consider the case if there is no set cover of size k. Again we can assume all the seeds are selected from set nodes, since as a candidate for seeds, set nodes are more efficient than element nodes and dummy nodes. Thus, if there is no set cover of size k, then we cannot find k seeds which activate all the element nodes, hence none of the dummy notes are ac- tivated. Therefore, the total number of nodes cumulatively activated are no more than n + k. It follows that if a poly- nomial algorithm can approximate IM-CA problem within N 1−, then we can answer the decision problem of the SET COVER problem in polynomial time, this is impossible un- der the assumption P(cid:54)= NP. 4. EFFICIENT HEURISTIC ALGORITHMS In Section 3, we prove that both SM-CA with η < n and IM-CA are hard to approximate. Despite this difficulty, in this section we present efficient heuristic algorithms based on the greedy strategy, in order to tackle the problem in practice. We first show the outline of our greedy strategies in Section 4.1. In Section 4.2, we adopt an efficient method to design scalability greedy algorithms. 4.1 Greedy Strategies In this section, we introduce two possible greedy strategies additional parameter c. A large c reduces the difficulty of lifting Pu(S) over the threshold τu when it is getting close to τu, but it continu- ously rewards Pu(S) above τu, while a c close to 1 has the reverse effect. Essentially c balances between the truncated surrogate f (S) (when c = 1) and the expected influence function σ(S) (when c is large). Thus, our first greedy strat- egy is to use F (S) with a proper tuned c as the greedy ob- jective, and we call it the balanced truncation greedy (BTG) strategy. function f (S) =(cid:80) for SM-CA problem and IM-CA problem. From Section 3.1.1, we know that greedy by the surrogate u∈V min{Pu(S), τu} can guarantee good approximation ratio for SM-CA problem with η = n. Thus, intuitively we could adopt f (S) as our surrogate objective even when η < n and apply the greedy strategy based on f (S). However, our initial experiments demonstrate that directly adopting f (S) is less effective, especially when seed set size is relative small. We believe that this is because greedy on f (S) would prefer larger increment of Pu(S) far below τu over smaller increment of Pu(S) close to τu, but the latter actually provides new cumulative activations. To guide seed selection towards the latter case, we generalize u∈V min{Pu(S), cτu} by introducing an f (S) to F (S) = (cid:80) The second strategy is to apply greedy on the objective function ρ(S) directly. That is, we select the node with the largest increment to ρ(S) in each step. However, since ρ(S) is a discrete rounding function, there could be many nodes having the same effect (or no effect at all) in any step. For tie-breaking, we select nodes according to f (S), which is equal to σ(S) under this situation. In summary, the second strategy preferentially selects nodes promoting ρ(S) most, then chooses the node contributing to f (S) most among nodes having the same promotion to ρ(S). In this strategy, the objective function (i.e. ρ(S)) plays a dominant role in selecting seeds. We call it the activation dominance greedy (ADG) strategy. During the process of greedy algorithms, we need to esti- mate Pu(S) for each node u ∈ V . It will be very expensive if we do this estimation by Monte Carlo simulations. Spe- cially, by Lemma 4, we need to simulate O(n2 ln n) times to guarantee the accuracy, each simulation takes O(m) time in the worst case. Thus, it takes O(n2m ln n) for each node u to estimate Pu(S). To improve the efficiency, we adopt a new approach named reverse reachable ret (RR set), as we describe in the next section. 4.2 Greedy Algorithms Based on RR Set In this section, we present our efficient algorithms based on RR set. We first introduce the background of RR set. RR set was first proposed by Borgs et al. in 2014 [3] to provide the first near-linear-time algorithm for the classical influence maximization problem in [20]. The approach is further optimized later in a series of follow-up work [29, 30, 26]. The definition of RR set is as follows: Definition 3 (Reverse Reachable Set) Let u be a node in G, and g be a random graph obtained by independently removing each edge e = (v, w) in G with probability 1 − pe. The reverse reachable set (RR set) for u is the set of nodes in g that can reach u. Borgs et al. established a crucial connection between RR set and the influence propagation process on G. We restate- ment it in Lemma 5. Lemma 5 ([3]) Let S be a seed set and u be a fixed node. Suppose Ru is an RR set for u generated from G, then Pu(S) equals the probability that S overlaps with Ru, that is, Pu(S) = Pr(S ∩ Ru (cid:54)= ∅). Now we introduce our new method to estimate Pu(S) for each node u ∈ V . We first generate θ RR sets for u independently. Let Ru be the collection of all generated RR sets for u. For any node set S, let FRu (S) be the fraction of RR sets in Ru overlapping with S. That is, FRu (S) (cid:44) {Ru ∈ Ru : Ru ∩ S (cid:54)= ∅}/θ. Then for any u ∈ V , we use FRu (S) as the estimation of Pu(S). We can prove that we can bound the estimation error if θ is large enough. Lemma 6. For any ε > 0, if θ satisfies θ ≥ ln(2n)/2ε2, then for each node u ∈ V , Pr[FRu (S) − Pu(S) ≥ ε] ≤ n−1. Proof. Let X (cid:44) θFRu (S), then X is the number of RR sets in Ru overlapping with S. Moreover, X can be regarded as the sum of θ i.i.d. Bernoulli variables. Specifically, let Seed set S Algorithm 3 Framework of greedy algorithm for IM-CA problem Input: G = (V, E), {puv}(u, v)∈E,{τu}u∈V , k, θ Output: 1: set S = ∅ 2: generate θ RR sets for each node u ∈ V : {Ru}u∈V 3: set req(u) = τuθ for each node u ∈ V 4: for j = 1 to k do 5: 6: 7: 8: 9: 10: 11: 12: 13: end for 14: return S x = SS(G,{puv}(u,v)∈E,{req(u)}u∈V , {Ru}u∈V ) /SS is a general term of SSBT and SSAD/ S = S ∪ {x} remove all RR Sets containing x for each u in V do rem(u): the number of RR Sets removed from Ru req(u) = req(u) − rem(u) end for X =(cid:80)θ have E[X] = (cid:80)θ i=1 Xi where Xi = 1 if S overlaps with the i-th RR set in Ru and Xi = 0 otherwise. Based on Lemma 5, we i=1 E[Xi] = θPu(S). By the Hoeffding's inequality and the condition θ ≥ ln(2n)/2ε2, we have Pr(FRu (S) − Pu(S) ≥ ε) = Pr(θFRu (S) − θPu(S) ≥ θε) = Pr(X − E[X] ≥ θε) ≤ 2 exp(−2(θε)2/θ) ≤ n −1. two greedy functions: F (S) = (cid:80) ρ(S) =(cid:80) We now present our greedy algorithms. Recall that we use u∈V min{Pu(S), cτu} and I{Pu(S) ≥ τu}, where I is the indicator func- u∈V tion. In order to make it easier to understand, we describe the processes of selecting seeds in subprograms. We first present the framework of the whole greedy algorithm for IM-CA problem in Algorithm 3. In Algorithm 3, we first initialize for the seed set S (line 1). Then we generate θ RR sets for each node u in V . Let Ru be the collection of RR sets for u. In line 3, req(u) is the requirement of node u ∈ V , which is the number of RR sets in Ru that needs to be hit by a seed set so that u can become cumulatively active. We say a set S hits an RR set R if S ∩ R (cid:54)= ∅. Based on Lemma 6, u is cumulatively active only if there are at least θτu RR sets in Ru hit by the seed set. Thus, we set req(u) = τuθ for each node in u ∈ V . At each step, we add a new node x into the current seed set (line 5). After x is selected, we need to remove all RR sets containing x and update the requirements for all nodes. The algorithm ends when S = k. Note that Algorithm 3 needs to call the seed selecting procedures (line 5). Here, SS(·) is a general term for our two subprograms SSBT (Selecting Seeds via Balanced Trun- cation strategy) and SSAD (Selecting Seeds via Activation Dominance strategy). Specifically, SSBT (Procedure 4) is the subprogram that selects one node with the largest marginal increment to F (S) into the current seed set S. SSAD (Proce- dure 5) is the subprogram selecting the node with the largest marginal increment to ρ(S), with tie-breaking on f (S). The algorithm calling SSBT is named as BTG-IM-CA (Balanced Truncation Greedy algorithm for IM-CA problem) and the algorithm calling SSAD is named as ADG-IM-CA (Activa- tion Dominance Greedy algorithm for IM-CA problem ). for each node v ∈(cid:83)Ru do Procedure 4 SSBT: Selecting Seeds via Balanced Trunca- tion strategy Input: G = (V, E), {puv}(u,v)∈E, {req(u)}u∈V , {Ru}u∈V Output: a new seed 1: set inc(v) = 0 for all v ∈ V 2: for each node u ∈ V and req(u) > 0 do 3: 4: 5: 6: 7: end for 8: select x = arg maxv inc(v) 9: return x /compute the marginal increment of v/ inc(v) = inc(v) + min{overlap(v, Ru), c · req(u)} end for for v ∈(cid:83)Ru do /compute the marginal increment of v/ inc(v) = inc(v) + I{overlap(v, Ru) ≥ req(u)} Procedure 5 SSAD: Selecting Seeds via Activation Domi- nance strategy Input: G = (V, E), {puv}(u,v)∈E, {req(u)}u∈V , {Ru}u∈V Output: a new seed 1: set inc(v) = 0 for all v ∈ V 2: for each node u ∈ V and req(u) > 0 do 3: 4: 5: 6: 7: end for 8: /select one better node from nodes with the largest 9: for each node u ∈ V and req(u) > 0 do 10: 11: 12: 13: end for 14: select x = arg maxv inc(v) 15: return x for v ∈(cid:83)Ru with the largest inc(v) values do inc(v) = inc(v) + min{overlap(v, Ru), req(u)} marginal increment/ end for end for Now we describe our two subprograms SSBT and SSAD. We first introduce SSBT in Procedure 4. Let inc(v) be the value of the marginal increment generated by any node v ∈ V , overlap(v,Ru) be the number of RR sets in Ru over- lapping with node v. In the main loop of SSBT, we select the node providing the largest marginal increment to F (S). To this end, for each node v ∈ V , we compute the marginal increment of v to all nodes which are not cumulatively active yet. Based on Lemma 5, the marginal increment of a node v to node u can be measured by min{overlap(v, Ru), c · req(u)}. Summing up the increments of v on all not-yet cu- mulatively active nodes, we can obtain inc(v) (see details in lines 2 - 7). Then, we choose the node with the maximum inc(v). Another greedy strategy is shown in SSAD (Procedure 5). In this procedure, we first find nodes with the largest marginal increment to ρ(S). For any node v, the marginal increasing inc(v) can be denoted by: u: req(u)>0 I{overlap(v, Ru) ≥ req(u)} (cid:80) There may be many nodes with the same value of inc(v) due to the truncation operation of ρ(S). To break the tie, we choose the node with the maximum marginal increase to f (S) among all nodes with the largest inc value (lines 9 - 13). The framework of the whole greedy algorithm for SM-CA problem follows the same structure with Algorithm 3. The only difference between these two algorithms is the stop- ping condition. For SM-CA, the algorithm stops when the number of cumulatively active nodes is no less than η. The corresponding algorithms of SM-CA problem are named as BTG-SM-CA (Balanced Truncation Greedy algorithm for SM-CA problem) and ADG-SM-CA (Activation Dominance Greedy algorithm for SM-CA problem). Other details of the algorithms are essentially the same as the algorithms for IM-CA and are thus omitted. Time complexity. Now we analyze the time complexity of BTG-IM-CA and ADG-IM-CA. Let EP T be the expected sum of in-degrees of all nodes in a random RR set, which is the same as the expected time of generating an RR set. Thus, the total expected time of the generation is O(nθ · EP T ). By Lemma 6, θ = ln(2n)/2ε2 is enough for accuracy. Thus, the expected generation time is O(n ln n · EP T /ε2). Besides the generation time, the main time cost depends on SS(·) since other operations only take time O(n). For each node u ∈ V , let EP T Vu be the expected number of u∈V EP T Vu. Then, both SSBT and SSAD takes time O(n · EP T V ) in expectation. Hence, the expected time complexity of both BTG-IM-CA and ADG-IM-CA is O(n(kEP T V + ln n · EP T /ε2)). nodes in(cid:83)Ru and EP T V = 1 (cid:80) n At each step of BTG-SM-CA and ADG-SM-CA , the num- ber of cumulatively active nodes increases at least 1 since the selected seed contributes 1 to ρ(·), which means the times of the outer loop is at most η. Thus, the expected time cost of BTG-SM-CA and ADG-SM-CA is O(n(ηEP T V + ln n · EP T /ε2)). 5. EXPERIMENTS In order to test the performance of our heuristic algo- rithms, we conduct experiments on real social networks. Our experiments are run on a machine with a 2.4GHz Intel(R) Xeon(R) E5-2670 CPU, 2 processors (16 cores), 64GB mem- ory and Red Hat Enterprise Linux Server release 6.3 (64bit). All algorithms tested in this paper are written in C++ and compiled with g++ 4.8.4. 5.1 Experiment setup Datasets. We use three real-world networks in our ex- periments: Flixster, NetPHY and DBLP. Table 1 shows the datasets used in our experiments, in which AOD denotes the average out degree of a dataset. The first network is Flixster, Name Flixster NetPHY DBLP Table 1: Datasets Type # Node #Edge 174K 348K undirected 2M undirected 29K 37K 655k directed AOD 6.0 18.8 6.1 which is an American movie rating social site for discovering new movies. In the Flixster graph, each node represents a user and a directed edge e = (u, v) represents that u and v rate the same movie and v rates the movie shortly after u. We simply use one specific topic in this network with 29357 nodes and 174939 directed edges. And we learn the active probabilities on edges by using the Topic-aware Inde- pendent Cascade Model presented in [1]. The mean of edge probabilities is 0.118 and the standard deviation is 0.025. The second one, called NetPHY, is the same as the one used in [9, 15, 16]. It is an academic collaboration network extracted from the "Physics" section from arXiv(http:// www. arXiv.org). The nodes in NetPHY are authors and undi- rected edges represent coauthorship relations. We use data from year 1991 to year 2003 which includes 37154 nodes and 348322 edges. The influence probabilities on edges are as- signed by weighted cascade model [20]. Specifically, for each edge (u, v) ∈ E, we set puv = c(u, v)/d(v), where d(v) is the number of published papers of author v and c(u, v) is the number of papers that u and v collaborated. In this network, the mean of edge probabilities is 0.107 and the standard deviation is 0.025. The last one is a larger collaboration network DBLP main- tained by Michael Ley (654628 nodes and 2056186 edges). The method of generating edge probability is the same as that in NetPHY. We follow the TRIVALENCY model [31] to assign edge probabilities: On every edge e = (u, v), we uniformly select a probability from the set {0.1, 0.01, 0.001} at random, which corresponds to high, medium and low in- fluences. Under the above method, the mean of edge prob- abilities is 0.069 and the standard deviation is 0.002. Algorithms. We test our algorithms using both BTG and ATG strategies. For comparison, we use the following baseline algorithms. • TIM+: TIM+ is a greedy algorithm presented in [30]. The basic greedy rule is to choose the node covers the maximum number of Random RR sets (see more de- tails in [30]). • High-degree: High-degree generates seed set sequence by the decreasing order of the out-degree of nodes. It is popular to consider high degree nodes as influential nodes in social and other networks. e = (v, u) is puv/(cid:80) • PageRank: It is the popular algorithm used for rank- ing web pages [4]. The transition probability on edge w:(w,v)∈E pwv. In the PageRank al- gorithm, higher puv indicates u is more influential to v and thus v should vote u higher. This is the reason we relate puv with the transition probability of the reverse link (v, u), which is the same as in earlier studies such as [31]. We use 0.15 as the restart probability and use the power method to compute the PageRank values. The stopping criteria of computing PageRank values is when two consecutive iterations are different for at most 10−4 in L1 norm. We select seeds by decreasing order of the PageRank values. • Random: As a trivial baseline, Random selects seeds sequence in random order. For all the above algorithms, we use the same Monte Carlo method to compute the number of cumulative active nodes. The stopping criteria of IM-CA problem and SM-CA prob- lem are the number of seeds is k and the number of cumu- lative active nodes exceeds η, respectively. Parameters. In our experiment, we take two simplified operations: First, we set parameters θ = 1000, ε = 0.1 in Flixster and NetPHY, this setting is enough for the require- ment in Lemma 6. In DBLP, we set θ = 500, in practice, this is good enough for illustrating our results. We also sup- pose that all nodes have the same activation threshold τ . In our experiments, τ ranges from 0.1 to 0.9. In order to determine a proper value of parameter c for the BTG strategy, we fix k = 500 on Flixster, then we imple- ment BTG-IM-CA on different τ and c. Figure 2 shows the result. In order to present more clearly, we set the ordinate as the number of cumulatively active nodes minus the cor- responding number under c = 1. In general, we can see that τ greater than 1 yields better result, and τ between 1.6 and 1.8 provides close-to-best results in all cases. In the rest, we choose c = 1.7 for all tests (except tests on SM-CA, which involves a large number of seeds). Tests on other datasets yield similar results. 5.2 Experiment results Experiment results on spreading performance of IM-CA problem. Figure 3, Figure 4 and Figure 5 show the comparison of different algorithms on Flixster, NetPHY and DBLP respectively. These figures reflect the spreading performances with different τ , varying k from 1 to 500. Figures on these three datasets reflect some common fea- tures. Firstly, the performances of BTG-IM-CA and TIM+ are similar when τ is large. This is because when c = 1.7, In this case, BTG-IM-CA is exactly cτ > 1 if τ > 0.59. the same as TIM+. While τ is smaller, BTG-IM-CA is bet- ter than TIM+ sometimes. For example, on NetPHY with τ = 0.1 and k = 500 (Figure 4(a)) BTG-IM-CA cumulatively activates nodes with size 7.6% more than those cumulatively activated by TIM+. Secondly, ADG-IM-CA performs similar to TIM+ and BTG- IM-CA when τ is small. However, the curves of ADG-IM-CA and TIM+ become separated with the increase of τ on all datasets. ADG-IM-CA outperforms all other algorithms sig- nificantly when τ is large. In particular, on NetPHY with τ = 0.7 and k = 500 (Figure 4(d)), the number of nodes cumulatively activated by ADG-IM-CA is 90.8% more than that by TIM+, 89.5% more than BTG-IM-CA , 88.8% more than PageRank, 190.2% more than High-degree, 227.0% more than Random. For the two million edges DBLP dataset, when τ = 0.7 and k = 500 (Figure 5(d)), ADG-IM-CA cumu- latively activates nodes with size 162.0% more than those cu- mulatively activated by TIM+, 162.0% more than BTG-IM- CA , 157.0% more than PageRank, 246.7% more than High- degree, 801.1% more than Random. We think this feature is mainly because the cumulative activation is easy when τ is small, thus, the seed set generated by TIM+ is likely to cumulatively activate enough nodes. However, when τ is large, most nodes are not easy to be cumulatively activated. In this case, selecting seeds directly contribute to ρ(S) may be more effective. Experiment results on the activa- tion threshold τ . To see the change of cumulative activa- tion influence size with the increase of parameter τ , we also conduct experiments on different τ on NetPHY (Figure 6). From these figures, we observe that ADG-IM-CA is the best algorithm for all settings of τ and seed set size k. With the increase of τ , in all algorithms, the size of the cumulatively active nodes decreases rapidly. This is consistent with the previous description. Experiment results on spreading performance of SM-CA problem. For seed minimization problem, al- gorithms Random, High-degree and PageRank will output a very large seed set to meet the target requirement η and thus are very ineffective. This is already demonstrated in the pre- vious test results, and thus we only focus on the performance comparison of TIM+, BTG-SM-CA and ADG-SM-CA for the Figure 2: Results of c on Flixster (a) Flixster τ = 0.1 (b) Flixster τ = 0.3 (c) Flixster τ = 0.5 (d) Flixster τ = 0.7 (e) Flixster τ = 0.9 Figure 3: Results of IM-CA on Flixster SM-CA problem. We also clarify that when η is small, the results of SM-CA problem can be reflected by the results of IM-CA problem since we adopt the same strategies for these two problems. Thus, we only present the result on NetPHY with large enough values of η (Figure 7). Furthermore, for large η, we notice that setting c > 1 for the BTG strategy is no longer beneficial, perhaps because with a large number of seeds, the needs to penalize over-the-top influence (setting c = 1) out-weighs the need of compensating near-the-top influence (setting c > 1). It also coincides with Theorem 3 for the case of η = n, so we set c = 1 for this test. From Figure 7, we can see that ADG-SM-CA outperforms TIM+ and BTG-SM-CA significantly. Running Time. We compare the running times of ADG- IM-CA , BTG-IM-CA and TIM+ in Table 2. We set τ = 0.3 and k = 500 on all datasets. The results indicate that TIM+ runs faster than our ADG and BTG strategies, although for the two relatively small datasets the gap is not that much. We believe that this is because TIM+ employs an estimation for the optimal influence spread with k seeds for the influence maximization task, but we are not able to do so for the IM- CA task because we need to estimate individual nodes active probability Pu(S). Nevertheless, our algorithm can still be scaled to the large graph of DBLP with millions of edges. Further improving the efficiency while preserving the same level of quality is a future work item. Table 2: Running Time (τ = 0.3, k = 500) Flixster NetPHY DBLP TIM+ ADG-IM-CA 39s 54s 509s 87s 112s 8865s BTG-IM-CA 138s 142s 8685s Conclusion. From these experiment results, we conclude that ADG consistently provides the best performance cross all test cases, and thus we propose ADG-IM-CA and ADG- SM-CA as our solution to the IM-CA and SM-CA problems, respectively. The BTG strategy performs well in some cases (such as small τ for IM-CA), but its performance is not sta- ble, and it also requires tuning c for different cases, and thus (a) NetPHY τ = 0.1 (b) NetPHY τ = 0.3 (c) NetPHY τ = 0.5 (d) NetPHY τ = 0.7 (e) NetPHY τ = 0.9 Figure 4: Results of IM-CA on NetPHY (a) DBLP τ = 0.1 (b) DBLP τ = 0.3 (c) DBLP τ = 0.5 (d) DBLP τ = 0.7 (e) DBLP τ = 0.9 Figure 5: Results of IM-CA on DBLP is less desirable than the ADG strategy. TIM+ has similar performance as the BTG strategy, but is even more unstable across the tests, and thus is not competitive comparing to ADG for both the SM-CA and IM-CA problems. 6. CONCLUSION AND FUTURE WORK In this paper, we propose the cumulative activation in- fluence model to reflect realistic scenarios where user adop- tion is based on repeated exposure to multiple information cascades in the network, which is different from the most existing study where user adoption is only based on a single cascade. We study both the seed minimization and influence maximization problems in the cumulative activation setting, providing both theoretical hardness results and approxima- tion algorithms, and further propose efficient heuristic so- lutions despite the theoretical hardness result. Our experi- mental results demonstrate the effectiveness of our proposed solutions. Our current study focuses on dealing with multiple in- formation cascades for product adoption, but there are cer- tainly many chances to further elaborate the model, such as mixing the information cascades and adoption cascades and study the impact to the optimization problems. Another (a) NetPHY k = 240 maximization in social networks. In KDD, pages 199 -- 208. ACM, 2009. [11] P. Domingos and M. Richardson. Mining the network value of customers. In KDD, pages 57 -- 66. ACM, 2001. [12] M. Farajtabar, N. Du, M. Gomez-Rodriguez, I. Valera, H. Zha, and L. Song. Shaping social activity by incentivizing users. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, NIPS, pages 2474 -- 2482. Curran Associates, Inc., 2014. [13] U. Feige. A threshold of ln n for approximating set cover. J. ACM, 45(4):634 -- 652, 1998. [14] U. Feige, G. Kortsarz, and D. Peleg. The dense k-subgraph problem. Algorithmica, 29(3):410 -- 421, 2001. [15] A. Goyal, F. Bonchi, L. V. Lakshmanan, and S. Venkatasubramanian. On minimizing budget and time in influence propagation over social networks. Social Network Analysis and Mining, pages 1 -- 14, 2012. (b) NetPHY k = 500 [16] A. Goyal, W. Lu, and L. V. S. Lakshmanan. Figure 6: Results of activation threshold τ on NetPHY direction is to circumvent the theoretical hardness for gen- eral graphs due to nonsubmodularity and develop efficient algorithms with guarantees on the class of graphs that are closer to real-world graphs. 7. REFERENCES [1] N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influence propagation models. In Data Mining (ICDM), 2012 IEEE 12th International Conference on, pages 81 -- 90. IEEE, 2012. [2] A. Bhaskara, C. Moses, E. Chlamt´ac, and U. Feige. Detecting high log-densities: an o(n1/4) approximation for densest k-subgraph. In STOC, pages 201 -- 210. ACM, 2010. [3] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier. Maximizing social influence in nearly optimal time. In SODA, pages 946 -- 957. ACM-SIAM, 2014. [4] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. Computer networks and ISDN systems, 30(1):107 -- 117, 1998. [5] D. Centola and M. Macy. Complex contagion and the weakness of long ties. ajs, 113(3):702 -- 734, 2007. [6] N. Chen. On the approximability of influence in social networks. SIAM Journal on Discrete Mathematics, 23(3):1400 -- 1415, 2009. [7] W. Chen, L. V. S. Lakshmanan, and C. Castillo. Information and influence propagation in social networks. Morgan & Claypool Publishers, 2013. [8] W. Chen, F. Li, T. Lin, and A. Rubinstein. Combining traditional marketing and viral marketing with amphibious influence maximization. In EC, pages 779 -- 796. ACM, 2015. [9] W. Chen, C. Wang, and Y. Wang. Scalable influence maximization for prevalent viral marketing in large-scale social networks. In KDD, pages 1029 -- 1038. ACM, 2010. [10] W. Chen, Y. Wang, and S. Yang. Efficient influence SIMPATH: An Efficient Algorithm for Influence Maximization under the Linear Threshold Model. In ICDM, pages 211 -- 220, 2011. [17] D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins. Information diffusion through blogspace. In WWW, pages 491 -- 501. ACM, 2004. [18] M. T. Hajiaghayi, K. Jain, K. Konwar, L. C. Lau, I. I. Mandoiu, A. Russell, A. Shvartsman, and V. V. Vazirani. The minimum k-colored subgraph problem in haplotyping and dna primer selection. In IWBRA, 2006. [19] J. He, S. Ji, R. Beyah, and Z. Cai. Minimum-sized influential node set selection for social networks under the independent cascade model. In MobiHoc, pages 93 -- 102. ACM, 2014. [20] D. Kempe, J. Kleinberg, and ´E. Tardos. Maximizing the spread of influence through a social network. In KDD, pages 137 -- 146. ACM, 2003. [21] D. Kempe, J. Kleinberg, and ´E. Tardos. Influential nodes in a diffusion model for social networks. In ICALP, pages 1127 -- 1138, 2005. [22] S. Khot. Ruling out ptas for graph min-bisection, dense k-subgraph, and bipartite clique. SIAM J. Comput, 36(4):1025 -- 1071, 2006. [23] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance. Cost-effective outbreak detection in networks. In KDD, pages 420 -- 429. ACM, 2007. [24] C. Long and R.-W. Wong. Minimizing seed set for viral marketing. In Data Mining (ICDM), 2011 IEEE 11th International Conference on, pages 427 -- 436. IEEE, 2011. [25] G. Nemhauser, L. Wolsey, and M. Fisher. An analysis of the approximations for maximizing submodular set functions. Mathematical Programming, 14:265 -- 294, 1978. [26] H. T. Nguyen, M. T. Thai, and T. N. Dinh. Stop-and-stare: Optimal sampling algorithms for viral marketing in billion-scale networks. In SIGMOD, pages 695 -- 710. ACM, 2016. [27] M. Richardson and P. Domingos. Mining (a) NetPHY τ = 0.1 (b) NetPHY τ = 0.3 (c) NetPHY τ = 0.5 (d) NetPHY τ = 0.7 Figure 7: Results of SM-CA on NetPHY knowledge-sharing sites for viral marketing. In KDD, pages 61 -- 70. ACM, 2002. [28] J. Tang, J. Sun, C. Wang, and Z. Yang. Social influence analysis in large-scale networks. In KDD, pages 807 -- 816. ACM, 2009. [29] Y. Tang, Y. Shi, and X. Xiao. Influence maximization in near-linear time: A martingale approach. In SIGMOD, pages 1539 -- 1554. ACM, 2015. [30] Y. Tang, X. Xiao, and Y. Shi. Influence maximization: near-optimal time complexity meets practical efficiency. In SIGMOD, pages 946 -- 957. ACM, 2014. [31] C. Wang, W. Chen, and Y. Wang. Scalable influence maximization for independent cascade model in large-scale social networks. Data Mining and Knowledge Discovery, 25(3):545 -- 576, 2012. [32] P. Zhang, W. Chen, X. Sun, Y. Wang, and J. Zhang. Minimizing seed set selection with probabilistic coverage guarantee in a social network. In KDD, pages 1306 -- 1315. ACM, 2014.
1611.01393	2	1611	2017-12-12T10:23:39	Hierarchical Overlapping Clustering of Network Data Using Cut Metrics	[ "cs.SI", "cs.IT", "cs.IT", "physics.soc-ph" ]	A novel method to obtain hierarchical and overlapping clusters from network data -i.e., a set of nodes endowed with pairwise dissimilarities- is presented. The introduced method is hierarchical in the sense that it outputs a nested collection of groupings of the node set depending on the resolution or degree of similarity desired, and it is overlapping since it allows nodes to belong to more than one group. Our construction is rooted on the facts that a hierarchical (non-overlapping) clustering of a network can be equivalently represented by a finite ultrametric space and that a convex combination of ultrametrics results in a cut metric. By applying a hierarchical (non-overlapping) clustering method to multiple dithered versions of a given network and then convexly combining the resulting ultrametrics, we obtain a cut metric associated to the network of interest. We then show how to extract a hierarchical overlapping clustering structure from the aforementioned cut metric. Furthermore, the so-called overlapping function is presented as a tool for gaining insights about the data by identifying meaningful resolutions of the obtained hierarchical structure. Additionally, we explore hierarchical overlapping quasi-clustering methods that preserve the asymmetry of the data contained in directed networks. Finally, the presented method is illustrated via synthetic and real-world classification problems including handwritten digit classification and authorship attribution of famous plays.	cs.SI	cs	IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 1 Hierarchical Overlapping Clustering of Network Data Using Cut Metrics Fernando Gama, Santiago Segarra, and Alejandro Ribeiro degrees that may be present. Hierarchical clustering solves this issue by providing a collection of partitions, indexed by a resolution parameter, that can be set by the user to determine the extent up to which nodes are considered similar or different [9]. In other words, each partition in the collection reflects a different degree of similarity between the nodes, ranging from considering all nodes different, to considering all nodes similar. Examples of hierarchical clustering methods are UPGMA [10], Ward's method [11], complete [12] or single linkage [13]. 7 1 0 2 c e D 2 1 ] I S . s c [ 2 v 3 9 3 1 0 . 1 1 6 1 : v i X r a Abstract-A novel method to obtain hierarchical and overlap- ping clusters from network data – i.e., a set of nodes endowed with pairwise dissimilarities – is presented. The introduced method is hierarchical in the sense that it outputs a nested collection of groupings of the node set depending on the resolution or degree of similarity desired, and it is overlapping since it allows nodes to belong to more than one group. Our construction is rooted on the facts that a hierarchical (non-overlapping) clustering of a network can be equivalently represented by a finite ultrametric space and that a convex combination of ultrametrics results in a cut metric. By applying a hierarchical (non-overlapping) clustering method to multiple dithered versions of a given network and then convexly combining the resulting ultrametrics, we obtain a cut metric associated to the network of interest. We then show how to extract a hierarchical overlap- ping clustering structure from the aforementioned cut metric. Furthermore, the so-called overlapping function is presented as a tool for gaining insights about the data by identifying meaningful resolutions of the obtained hierarchical structure. Ad- ditionally, we explore hierarchical overlapping quasi-clustering methods that preserve the asymmetry of the data contained in directed networks. Finally, the presented method is illustrated via synthetic and real-world classification problems including handwritten digit classification and authorship attribution of famous plays. Index Terms-Clustering, Network theory, Cut metrics, Hier- archical clustering, Covering, Dithering. I. INTRODUCTION C ONSIDER a dataset where each element can be rep- resented as a node in a network with a pairwise dis- similarity function. In this setting, the general objective of clustering is to group those nodes that are more similar to each other than to the rest, according to the relationship established by the dissimilarity function [1], [2]. Clustering and its generalizations are ubiquitous tools since they are used in a wide variety of fields such as psychology [3], social network analysis [4], political science [5], neuroscience [6], among many others [7], [8]. Traditional clustering methods provide only one partitioning of the node set in such a way that each data point belongs to one and only one block of the partition. An important limitation of these traditional clustering methods is that the dataset may present a complex data structure at several reso- lutions or levels of similarity, and outputting only one partition may not be adequate in portraying the different grouping Work in this paper is supported by NSF CCF 1217963. F. Gama and A. Ribeiro are with the Department of Electrical and Systems Engi- neering, University of Pennsylvania. S. Segarra is with the Institute for Data, Systems, and Society, Massachusetts Institute of Technology. Emails: [email protected], [email protected], [email protected]. However, hierarchical methods still have a major limitation, namely, that nodes are assigned to one and only one category or cluster. There are many situations in which this assignment does not constitute a reasonable approach. Specifically, some nodes might inherently have traits of more than one group and hence have similarities to multiple clusters of nodes that oth- erwise would be considered dissimilar [14], [15]. Two classes of methods have been proposed to overcome this issue. The first class corresponds to the so-called soft or fuzzy clustering methods that allow the allocation of a node to multiple clusters by assigning to each node a membership degree or probability of belonging to every cluster [16]. This can also be done in a hierarchical fashion [17], [18]. Nevertheless, these methods subscribe to the idea that nodes have more affinity to either one or another subset, as illustrated by their membership degree; or that the setting of the problem is not rich enough to determine to which one, thus assigning a probability of belonging to different subsets. Fundamentally, these methods do not contemplate the idea that a node can intrinsically be part in equal terms of more than one cluster. The second class of methods that attempt to solve this issue is the class of overlapping clustering methods, which perform a non-hierarchical deterministic assignment of nodes to more than one subset [19]. While overlapping methods enable nodes to belong to more than one cluster, they are still myopic – just like traditional clustering methods – to multiple affinity resolutions. they belong to only one group but In this paper, a hierarchical overlapping clustering method is proposed. Our method outputs a collection of groupings where each of them corresponds to different resolutions of similarity between the nodes (hierarchical), while allowing nodes in each grouping to deterministically belong to more than one cluster (overlapping). Essentially, the method is de- rived by generalizing the concepts of ultrametrics, equivalence relations, and dendrograms – typical of hierarchical (non- overlapping) clustering methods – to those of cut metrics [20], [21], tolerance relations [22], and nested collections of coverings. More precisely, we generate a family of networks IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 2 derived from dithering multiple times the dissimilarity function of a given network. We then proceed to apply a hierarchical (non-overlapping) clustering method to each of the dithered networks, resulting in a collection of ultrametrics. From this collection we obtain a cut metric related to the network of interest by means of a convex combination of the ultrametrics. Finally, based on the cut metric we construct a tolerance relation that connects the nodes that are at a distance lower than a pre-specified resolution, which we then use to derive a covering of the node set. By considering all possible resolution levels, we obtain a nested collection of coverings. Hierar- chical overlapping clustering methods have been previously developed in the context of community detection, but only for undirected, unweighted networks [23], [24]. In Section II, the concepts of partition, equivalence relation, and dendrogram are defined, and their relations to both tradi- tional and hierarchical (non-overlapping) clustering methods are established. These concepts are then respectively general- ized to those of covering, tolerance relation, and nested collec- tion of coverings. This generalization is achieved by dropping the requirements that prevent a node from belonging to more than one group. Moreover, a formal definition of a hierarchical overlapping clustering method is presented. Section III relates the introduced concepts with concrete finite metric construc- tions. More specifically, ultrametrics are defined and directly linked to hierarchical clustering methods (Theorem 1). We then introduce cut metrics and explain how they can be used to obtain a nested collection of coverings (Theorem 2). The proposed algorithm to obtain cut metrics from a given network (Algorithm 1), and hence nested collection of coverings, is detailed in Section IV. The algorithm is grounded on the fact that a convex combination of ultrametrics yields a cut metric (Proposition 1) and the concept of dithering [25], that can be used to obtain multiple noisy versions of the given network. Also, quasi-clustering methods that allow for a hierarchical structure and overlapping nodes are explored in Section V. These methods are particularly useful when there is a need to preserve in the grouping structure the asymmetric nature of directed networks. Section VI illustrates the implementation of the proposed method in synthetic experiments to show that the hierarchical overlapping clustering algorithm yields sensitive results that correspond to intuition. Finally, in Section VII, the proposed method is applied to the unsupervised classifica- tion problem of determining handwritten digits [26] and the authorship attribution problem of identifying plays that have been co-authored [27]. Concluding remarks in Section VIII close the paper. II. PRELIMINARIES Let N = (X, AX ) be a network defined by a finite set of nodes X and a dissimilarity function AX : X × X → R+ that measures how different two nodes are. This function satisfies that, for any two nodes x, x(cid:48) ∈ X, AX (x, x(cid:48)) ≥ 0 with AX (x, x(cid:48)) = 0 if and only if x = x(cid:48). AX is a dissimilarity function in the sense that the greater the value of AX, the more different two nodes are. Observe that the dissimilarity function need not satisfy the triangle inequality nor be symmetric, that is, we may have that AX (x, x(cid:48)) (cid:54)= AX (x(cid:48), x). This is particularly useful when modeling uneven levels of influence among nodes. We denote by N the set of all possible networks. Definition 1. A partition PX = {B1, . . . , Bm} is a collection i=1Bi = X and Bi ∩ Bj = ∅ for of subsets of X such that ∪m i (cid:54)= j, i, j = 1, . . . , m. Denote by P the set of all possible partitions. Definition 2. An equivalence relation ∼ is a binary set relation that, for any x, x(cid:48), x(cid:48)(cid:48) ∈ X, satisfies the properties of reflexivity (x ∼ x), symmetry (x ∼ x(cid:48) if and only if x(cid:48) ∼ x), and transitivity (if x ∼ x(cid:48) and x(cid:48) ∼ x(cid:48)(cid:48), then x ∼ x(cid:48)(cid:48)). Observe that partitions (Def. 1) uniquely define equivalence relations (Def. 2) as follows: x ∼ x(cid:48) if and only if x, x(cid:48) ∈ Bi for some Bi ∈ PX, i = 1, . . . , m. Equivalently, equivalence relations uniquely define partitions [28]. An equivalence re- lation obtained from a specific partition PX will be denoted by ∼PX . With these definitions in place, a (traditional) clustering method G can be defined as a structure-preserving map from the set of networks to the set of partitions, G : N → P. This map is structure preserving in the sense that the output partition is defined over the same node set X specific to the input network. That is, G(N ) = PX = {B1, . . . , Bm} for N = (X, AX ) with ∪m i=1Bi = X. Intuitively, a desirable clustering method G is one in which the nodes contained in each subset Bi are determined by the dissimilarity function AX so that similar nodes are grouped together. In many situations, having only one partition as the output of a clustering method may not be appropriate. It might be desirable to output several partitions that depend on a resolu- tion parameter that specifies the affinity required between two nodes to be deemed as similar. Definition 3. A dendrogram DX = {DX (δ), δ ≥ 0} is a collection of partitions, where δ is a resolution parameter and DX (δ) is a partition of the node set X, satisfying the following two properties: (i) DX (0) = {{x}, x ∈ X}, and there exists a δmax such (ii) it is a nested collection of partitions, that is, for δ ≤ δ(cid:48), that for all δ ≥ δmax, DX (δ) = {X}; if x ∼δ x(cid:48), then x ∼δ(cid:48) x(cid:48) for all x, x(cid:48) ∈ X; together with a technical condition for right-continuity [29]. Note that, in an effort to ease the exposition, x ∼δ x(cid:48) is being used as a shorthand notation for x ∼DX (δ) x(cid:48). Observe that property (i) encodes the fact that at resolution δ = 0 all nodes are considered to be different from each other, thus each node forms a singleton cluster, whereas for sufficiently large δ all nodes are considered to be similar and, hence, clustered in a single block. Property (ii) enforces the agglomeration to occur in a nested fashion in the sense that if two nodes belong to the same cluster at a specific resolution then they will remain co- clustered for any larger resolution. Denoting by D the set of all possible dendrograms, we define a hierarchical clustering method H as follows. IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 3 Figure 1. Dumbbell network with dissimilarities given by the Eu- clidean distance between points. Definition 4. A hierarchical clustering method H is a structure-preserving map from the set of networks to the set of dendrograms H : N → D. (1) A limitation of both traditional G and hierarchical H clustering methods is that each node is required to belong to one and only one cluster at any given resolution. This requirement stems from the nonintersecting nature of the blocks that form a partition. In many problems, however, some nodes may inherently have traits of more than group, making it reasonable for them to belong to more than one cluster. For instance, consider the dumbbell network in Fig. 1 where the dissimilarity between two nodes is given by their Euclidean distance. It is intuitive that the point clouds on each side should form separate clusters. However, it is unclear if, e.g., all of the handle (bridge) should be a separate cluster as it is not unreasonable to assign its borders to the respective point clouds. Throughout the paper, we develop a clustering method that enables both deterministic overlapping of nodes and hierarchi- cal outcomes. This is achieved by generalizing the concepts of equivalence relations, partitions, and dendrograms to those of tolerance relations, coverings, and nested collections of coverings, respectively; see Table I. A. Generalizing concepts: Coverings and tolerance relations The first step in order to allow a node to belong to more than one cluster is to drop the nonintersecting requirement of partitions. Definition 5. A covering QX = {C1, . . . , Cm} of the node set X is a collection of subsets such that ∪m i=1Ci = X, but not necessarily Ci ∩ Cj = ∅ for i (cid:54)= j. Denote by Q the set of all possible coverings. Note that partitions (Def. 1) are particular cases of cover- ings (Def. 5), i.e., P ⊂ Q. Furthermore, a non-hierarchical overlapping clustering method M is defined as a structure- preserving map from the set of networks to the set of coverings M : N → Q. Related to the concept of covering, we introduce the notion of a tolerance relation. Definition 6. A tolerance relation ↔ is a binary set relation that, for any x, x(cid:48) ∈ X, satisfies the properties of reflexivity (x ↔ x), and symmetry (x ↔ x(cid:48) if and only if x(cid:48) ↔ x), but is not necessarily transitive. Observe that a tolerance relation is a generalization of an equivalence relation (Def. 2) obtained by dropping transitivity. A tolerance relation defined on a node set X can be repre- sented by an unweighted and undirected graph where an edge Table I: Parallelism between Hierarchical clustering H and Hierarchical Overlapping clustering O. Method Metric Relation Grouping Hierarchy Construction Hierarchical H (Def. 4) Ultrametric uX (x, x(cid:48)) (Def. 9) Equivalence ∼ (Def. 2) Partition PX = {Bi} (Def. 1) Dendrogram DX (Def. 3) Theorem 1 Overlapping O (Def. 8) Cut Metric cX (x, x(cid:48)) (Def. 10) Tolerance ↔ (Def. 6) Covering QX = {Ci} (Def. 5) Nested Covering KX (Def. 7) Theorem 2 exists between nodes x and x(cid:48) if and only if x ↔ x(cid:48). There are two established ways of inducing a covering from a tolerance relation based on this graph [22]. The first methodology is referred to as covering by classes, wherein each cover consists of the neighbor set of each node, together with the node itself, and where all covers that are completely contained within a larger one are removed. The other established procedure is denominated covering by blocks, where each block of the covering is given by the maximal cliques of the graph. For the remainder of the paper, we adopt this latter procedure as the canonical way of inducing coverings from tolerance relations. Note that, unlike the case of equivalence relations and partitions, while any tolerance relation can induce a covering (by blocks) not all coverings of a node set X can be induced by a tolerance relation [22, Theorem 2]. Leveraging the connection between tolerance relations and coverings, we extend the notion of a dendrogram into the realm of overlapping clusters. Definition 7. A nested covering KX = {KX (δ), δ ≥ 0} is a collection of coverings, where δ is a resolution parameter and KX (δ) is a covering induced by a tolerance relation ↔δ, satisfying the following properties: (i) KX (0) = {{x}, x ∈ X} and there exists a δmax such (ii) For δ ≤ δ(cid:48), if x ↔δ x(cid:48), then x ↔δ(cid:48) x(cid:48) for all x, x(cid:48) ∈ X. Notice the resemblance between properties (i) and (ii) above and those in Def. 3. In this way, we extend the notion of nestedness from its intuitive definition for (non-overlapping) clusters to the setting of overlapping clusters. Denoting by K the set of all nested coverings, we define hierarchical overlapping clustering methods as follows. Definition 8. A hierarchical overlapping clustering method O is a structure-preserving map from the set of networks to the set of nested coverings that for all δ ≥ δmax, KX (δ) = {X}; O : N → K. (2) Given a network N = (X, AX ), a clustering method O outputs a structure O(N ) = KX which takes into account different resolution levels of similarity and, at each resolution, allows overlapping clusters. Building on metric concepts de- veloped in Section III, we present the practical construction of one such method O in Section IV. IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 4 S Sc x3 x1 x2 y3 y1 y2 2. A cut Figure semimetric = {x1, x2, x3, y1, y2, y3}. A cut space into S and Sc semimetric is δS(xi, xj) = δS(yi, yj) = 0 and δS(xi, yj) = 1 for all i, j = 1, 2, 3. is drawn. The resulting cut node partitions on that set X the the III. NESTED COVERINGS DERIVED FROM CUT METRICS Hierarchical clustering methods output dendrograms, which are nested collections of partitions. However, dendrograms are cumbersome to handle mathematically so their equivalence to ultrametric spaces proves to be useful [13]. Definition 9. An ultrametric is a function uX : X × X → R+ that, for any x, x(cid:48), x(cid:48)(cid:48) ∈ X satisfies (i) uX (x, x(cid:48)) ≥ 0 and uX (x, x(cid:48)) = 0 if and only if x = x(cid:48); (ii) uX (x, x(cid:48)) = uX (x(cid:48), x); (iii) uX (x, x(cid:48)(cid:48)) ≤ max{uX (x, x(cid:48)), uX (x(cid:48), x(cid:48)(cid:48))}. Ultrametrics can be interpreted as metrics that satisfy a more stringent version of the triangle inequality [Property (iii)] where the addition in the regular triangle inequality is replaced by a maximization. Dendrograms can be equivalently repre- sented as ultrametrics, as shown in the following theorem. Theorem 1. Given an ultrametric uX (x, x(cid:48)) defined over the node set X, define a collection of equivalence relations as follows uX (x, x(cid:48)) ≤ δ ⇐⇒ x ∼δ x(cid:48). The resulting collection of equivalence relations is such that it generates a dendrogram DX. Likewise, given a dendrogram DX that uniquely determines a collection of equivalence relations ∼δ defined on the node set X, an ultrametric uX can be obtained as uX (x, x(cid:48)) := min{δ : x ∼δ x(cid:48)}. Proof: See [29, Theorem 1]. Theorem 1 establishes a structure-preserving bijection between ultrametrics and dendrograms. Consequently, we may rein- terpret hierarchical clustering methods as structure-preserving maps from the set of networks to the set U of all possible ultrametrics, i.e., H : N → U (cf. Def. 4). With the objective of deriving a relation between nested coverings KX and a metric structure akin to that between dendrograms and ultrametrics, the notion of cut metrics is introduced. Let S be a subset of a given node set X and define a cut semimetric δS(x, x(cid:48)) as δS(x, x(cid:48)) = I{S ∩ {x, x(cid:48)} (cid:54)= ∅}I{Sc ∩ {x, x(cid:48)} (cid:54)= ∅} (5) where I{} is the indicator function. Essentially, the cut semi- metric partitions the node set into two blocks S and Sc and assigns a dissimilarity of 1 to nodes that are in different sides (3) (4) (cid:4) S1 λS1 Sc 1 y1 x3 x1 x2 y3 y2 Sc 2 S2 λS2 S3 λS3 Sc 3 Figure 3. Construction of a cut metric cX. Consider the node set X = {x1, x2, x3, y1, y2, y3}. The first cut separates {xi} from {yi}, i = 1, 2, 3 and is assigned a dissimilarity λS1. The cuts associated with Si for i = 2, 3 separate {xi, yi} from the rest of X and are assigned dissimilarity λSi. Finally, the cut metric is obtained as cX = λS1 δS1 + λS2 δS2 + λS3 δS3. of this partition, and 0 to nodes that fall in the same side; see Fig. 2. This cut semimetric can also be understood as being induced by a binary classifier which partitions the node set in two and assigns a unit dissimilarity to nodes belonging to different categories. If multiple cuts, or partitions, are considered, then a cut metric cX can be obtained as the combination of all the associated dissimilarities; see Fig. 3. Definition 10. A cut metric cX : X × X → R+ is a metric that can be written as a nonnegative linear combination of cut semimetrics cX (x, x(cid:48)) = λSδS(x, x(cid:48)) (6) (cid:88) S⊆X with λS ≥ 0 and where the sum ranges over all possible subsets of X. Note that λS(cid:48) = 0 for some S(cid:48) amounts to the associated cut not being considered in the construction of cX. It follows from Def. 10 that cX is a metric, thus, in particular it satisfies that for any x, x(cid:48) ∈ X, cX (x, x(cid:48)) = cX (x(cid:48), x) ≥ 0 (symmetry and nonnegativity) and cX (x, x(cid:48)) = 0 if and only if x = x(cid:48) (identity). We denote by V the set of cut metrics. In the same way that ultrametrics induce dendrograms (The- orem 1), cut metrics induce nested collections of coverings, as shown next. Theorem 2. Let X be a set of nodes and let cX be a cut metric defined on X. Let KX = {KX (δ), δ ≥ 0} be a collection of coverings. If, for each δ ≥ 0, the corresponding covering KX (δ) is obtained from the tolerance relation given by cX (x, x(cid:48)) ≤ δ =⇒ x ↔δ x(cid:48) (7) then KX is a nested collection of coverings as defined in Def. 7. Proof: Observe that (7) defines a valid tolerance relation since cX (x, x) = 0 and cX (x, x(cid:48)) = cX (x(cid:48), x) imply reflexivity and symmetry of ↔δ, respectively. Using the relation ↔δ in (7) to define a collection of coverings by blocks KX (δ) [22], it can be shown that KX satisfies properties (i) and (ii) in Def. 7. To be more specific, KX (0) = {{x}, x ∈ X} follows from the fact that cX (x, x(cid:48)) = 0 if and only if x = x(cid:48) and, since there exists a δmax such that cX (x, x(cid:48)) ≤ δmax for all x, x(cid:48) ∈ X, it follows that KX (δ) = {X} for all δ ≥ δmax. Finally, observe IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 5 that KX is nested since, for any δ < δ(cid:48), cX (x, x(cid:48)) ≤ δ < δ(cid:48) implies that x ↔δ x(cid:48) ⇒ x ↔δ(cid:48) x(cid:48). (cid:4) Notice that (7) is similar to the connection between ultramet- rics and equivalence relations in (3) but in a unidirectional fashion; see Remark 1. For 0 < δ < δmax the resulting tolerance relations define coverings in such a way that nodes might belong to more than one cluster. If there exists at least one such node, it is said that there is overlap. Consequently, we are interested in computing the number of overlapping nodes for each value of δ. Definition 11. Let KX = {KX (δ), δ ≥ 0} be a nested col- lection of coverings defined over node set X and where each covering KX (δ) is comprised of m(δ) blocks or covers, i.e. KX (δ) = {C1, . . . , Cm(δ)}. Then, the overlapping function fol : R+ → Z+ is given by n(cid:88) k=1 fol(δ) = I{xk ∈ Ci ∩ Cj, i (cid:54)= j, i, j = 1, . . . , m(δ)} . (8) The function fol(δ) counts the nodes that are overlapping at each value of δ, i.e. that belong to more than one block of the covering KX (δ). From property (i) in Def. 7, it follows that fol(0) = 0 and fol(δ) = 0 for all δ ≥ δmax. Moreover, we use the overlapping function to define the notion of clusterability as follows. Definition 12. Given a network N = (X, AX ), a cut metric cX, and an associated nested covering KX, we say that (X, cX ) is clusterable if there exists a δ such that fol(δ) = 0, KX (δ) (cid:54)= {{x}, x ∈ X}, and KX (δ) (cid:54)= {X}. In other words, a network is clusterable if for some resolution δ we obtain a valid partition (i.e., there is no overlap) that is different from those partitions obtained for extreme reso- lutions (δ = 0 where all nodes are clustered separately, and δ ≥ δmax where all nodes are clustered together). Even for non-clusterable networks, the overlapping function provides valuable information about the underlying grouping structure and may help to identify nodes that cannot be fully classified into one subset. Also, the overlapping function is useful in determining meaningful values of the resolution δ; see Sections VI and VII for additional details. To sum up, in the same way that in hierarchical (non- overlapping) clustering, ultrametrics are used to define equiva- lence relations that determine dendrograms, in the proposed hi- erarchical overlapping clustering method, cut metrics are used to define tolerance relations that determine nested coverings; see Table I. Remark 1. Recall that defining a partition on a node set is equivalent to defining an equivalence relation on the same set (cf. Defs. 1 and 2). This means that specifying any of them uniquely leads to the other. The same occurs for the definitions of dendrograms and ultrametrics; see Theorem 1. However, in the context of hierarchical overlapping clustering these bijections no longer hold. While it is true that given a cut metric, we can obtain a collection of tolerance relations that leads to a nested collection of coverings (see Theorem 2), we cannot, in general, uniquely determine a tolerance relation from a given covering [22, Theorem 2] or a cut metric from a given nested collection of coverings. Remark 2. In a strict sense, the proof of Theorem 2 does not require cX to be a cut metric. In fact, any nonnegative, sym- metric function satisfying the identity property would suffice to create a relationship between nodes as in (7). Constructing such a relation is known as the Rips complex [30]. However, since we ultimately want to group nodes that are more similar to each other than to the rest, we are looking for functions that actually reflect the relationship between nodes encoded in the dissimilarity function. The cut metric obtained as a convex combination of ultrametrics achieves this. Moreover, due to its construction (cf. Figs. 2 and 3), cut metrics are intrinsically related to the formation of coverings and partitions. Another recent alternative to use in (7) is projections on A-spaces as done in [31]. It is also worth pointing out that the process of generating a Rips complex as in (7) involves the computation of the maximal cliques of a graph which is a computationally intensive operation; however, there exists algorithms to carry this operation more efficiently by means of parallelization [32]. IV. HIERARCHICAL OVERLAPPING CLUSTERING ALGORITHM Given a network, our goal is to output a nested collection of coverings such that the grouping of nodes is done in accordance to the difference between them at varying levels of resolution. In virtue of Theorem 2, a cut metric can be used to induce a nested collection of coverings. Hence, the idea is to systematically construct a cut metric in such a way that it reflects the dissimilarity across nodes as given by the dissimilarity function AX. The following proposition states a key relation between ultrametrics and cut metrics. Proposition 1. A convex combination of ultrametrics yields a cut metric. Proof: Given (X, dX ) where dX is a metric, then dX is in particular a tree metric – there exists a tree graph in which it is possible to embed the distance on its edges [20, p. 147] – if and only if it satisfies the four-point condition dX (x, x(cid:48)) + dX (x(cid:48)(cid:48), x(cid:48)(cid:48)(cid:48)) ≤ max(cid:8)dX (x, x(cid:48)(cid:48)) + dX (x(cid:48), x(cid:48)(cid:48)(cid:48)), dX (x, x(cid:48)(cid:48)(cid:48)) + dX (x(cid:48), x(cid:48)(cid:48))(cid:9) (9) for all x, x(cid:48), x(cid:48)(cid:48), x(cid:48)(cid:48)(cid:48) ∈ X. Ultrametrics also satisfy the four- point condition so that they are particular cases of tree metrics [20, p. 311]. It can be shown that tree metrics, as well as convex combinations of tree metrics, are (cid:96)1-embeddable [20, Proposition 11.1.4, Fact 11.1.15]. Finally, since a metric is (cid:96)1- embeddable if and only if it is a cut metric [20, Proposition 4.4.2], a convex combination of ultrametrics yields a cut (cid:4) metric. From Proposition 1 it immediately follows that ultrametrics are particular cases of cut metrics. Intuitively, this feature is inherited from the facts that equivalence relations are particular cases of tolerance relations and that partitions are particular IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 6 Algorithm 1 Hierarchical overlapping clustering algorithm O. Input: J: no. of perturbations, N: network, H: hierarchical clustering method, perturbation(·) Output: KX: nested collection of coverings. for i = 1:J do 1: procedure O(J, N,H, perturbation) 2: 3: 4: 5: 6: 7: 8: end procedure N = perturbation(N); X =H( N ); u(i) end for cX = avg(u(1) KX = obtain coverings(cX); X , . . . , u(J) X ); cases of coverings; see Table I. Also, Proposition 1 plays a key role in the generation of cut metrics as it gives a systematic way to obtain cut metrics from ultrametrics, which are readily available from the application of hierarchical (non- overlapping) clustering methods H. The next step is to obtain multiple ultrametrics related to the dissimilarity function AX so that the resulting combination of these yields a cut metric that bears a relation with AX. In order to achieve this, the concept of dithering [25] proves to be useful, where we intentionally apply random noise to the dissimilarity function AX to obtain AX. Denoting by N = (X, AX ) the resulting perturbed network, and by uX = H( N ) the ultrametric obtained by hierarchically clustering N, the following corollary of Proposition 1 results. Corollary 1. The function cX obtained as cX = E [uX ] (10) is a cut metric, where the expectation is taken with respect to the probability distribution of the randomization introduced by dithering. In practice, several realizations of random perturbations are implemented, generating a whole family of closely related networks. Each one of these networks yields a different ul- trametric when a pre-specified hierarchical clustering method H is applied to them. When all these ultrametrics are averaged, a cut metric is obtained (cf. Proposition 1). Formally, given a network N = (X, AX ) denote by Ni = (X, A(i) X ) for i = 1, . . . , J the J networks resulting from independently per- turbing the dissimilarity function J times. Moreover, denoting by u(i) X the ultrametric output of hierarchically clustering Ni, we generate the cut metric cX given by (cf. Proposition 1) J(cid:88) i=1 cX (x, x(cid:48)) = 1 J X (x, x(cid:48)). u(i) (11) Finally, we use cX to obtain an associated nested collection of coverings KX (cf. Theorem 2). The algorithm is summarized in Algorithm 1. Remark 3. Proposition 1 also holds for non-negative linear combinations of ultrametrics, but only a convex combination guarantees that the distance scale described by the ultrametrics is preserved in the resulting cut metric. Therefore, if we want to interpret the scaling parameter δ in the same units as the ultrametric function, then we should proceed with a convex combination. This is useful for directly comparing how clusters are formed in hierarchical methods, both overlapping and non-overlapping. Remark 4. Observe that cut metrics conform a convex cone and hence, any non-negative combination of cut metrics is still a cut metric. This is not true in the case of ultrametrics: a non-negative linear combination of ultrametrics is not an ultrametric [29, Section VII-B]. In fact, it is cut metrics that form the conic hull of ultrametrics. Remark 5. The choice of the hierarchical non-overlapping clustering method H depends on the specific problem at hand and as such is an input is observed in the experiments (Sections VI and VII) that for clusterable datasets, Algorithm 1 outputs the same coverings as those obtained when using H directly. Furthermore, the number J of perturbations is not a fundamental parameter in the sense that it is only used to create a cut metric. For the experiments considered in this paper, we observed that for values of J ≥ 10 the obtained output covers do not depend on J. to the proposed algorithm. It V. QUASI-CUT METRICS The output structure of the hierarchical overlapping clus- tering method O is a nested collection of coverings K (cf. Def 8). This is a symmetric structure in the sense that if node x shares a block in a covering with node x(cid:48) then x(cid:48) must share a block with x. This feature is inherited from the fact that cut metrics are symmetric functions. When considering asymmetric (directed) networks, the symmetry in the output necessarily entails a loss of information, which might be an undesirable feature of a grouping algorithm in some applications [33]–[35]. Our objective is to provide a framework for developing a hierarchical overlapping quasi-clustering of the node set X. A quasi-cluster is a structure that consists of both a grouping of the node set and a collection of edges indicating the influence between different groups of nodes [36]. In extending this notion, we start by defining quasi-coverings, which are the basic units of the proposed method, and we also determine the restrictions required for a collection of quasi-coverings to be nested. Definition 13. The pair QX = (QX , FX ) is a quasi-covering if QX = {C1, . . . , Cm} is a covering of X and FX ⊆ QX × QX is a set (possibly empty) of ordered pairs representing directed edges between the elements of QX. A quasi-covering is a directed graph whose node set is given by the blocks of a covering. The edges represent the asym- metric influence that some blocks can exercise over others. The conditions needed for a collection of quasi-coverings to be nested are described next. Definition 14. The collection KX = { KX (δ), δ ≥ 0} KX (δ) = is a nested collection of quasi-coverings if IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 7 u(1) X 1 x1 x2 2 1 3 2 u(2) X 1 3 x3 x1 x2 3 1 2 2 cX 1 3 x3 x1 x2 1 2.5 2.5 2 3 x3 (a) Quasi-ultrametric u(1) X (b) Quasi-ultrametric u(2) X (c) Quasi-cut metric cX = 0.5(u(1) X + u(2) X ) KX x2 x2 x2 x2 x2 x3 x1 0 ≤ δ < 1 x3 x1 1 ≤ δ < 2 x3 x1 x1 2 ≤ δ < 2.5 x3 2.5 ≤ δ < 3 x1 x3 3 ≤ δ 0 1 2 2.5 3 δ (d) Nested collection of quasi-coverings KX Figure 4. Illustrative example of a hierarchical overlapping quasi-clustering method O. (a)-(b) Given quasi-ultrametrics. (c) Quasi-cut metric obtained by averaging the given quasi-ultrametrics. (d) Nested collection of quasi-coverings obtained from cX. For δ = 0 all nodes belong to separate covers and the edge set is empty. At δ = 1 nodes x1 and x3 start influencing x2. At δ = 2 node x3 also exercises influence over x1. The first two covers form at δ = 2.5 and they present an overlap, since x1 belongs to both covers. Additionally, node x3 still exercises influence over x2. Finally, for δ = 3 all nodes belong to the same single cover and the edge set is empty. (KX (δ), FX (δ)) is a quasi-covering for each resolution pa- rameter δ, KX = {KX (δ), δ ≥ 0} is a nested collection of coverings (cf. Def. 7), and FX = {FX (δ), δ ≥ 0} is a collection of edge sets satisfying (i) FX (0) = ∅ and there exists a δmax such that FX (δ) = ∅ for all δ ≥ δmax; (ii) For arbitrary resolutions δ < δ(cid:48) and covers Ci, Cj ∈ KX (δ) such that (Ci, Cj) ∈ FX (δ), there exist covers i, Cj ⊆ C(cid:48) i, C(cid:48) C(cid:48) j and either i ∩ C(cid:48) j. (C(cid:48) i, C(cid:48) j ∈ KX (δ(cid:48)) such that Ci ⊆ C(cid:48) j) ∈ FX (δ(cid:48)) or Ci ∩ Cj < C(cid:48) Denote by K the space of all nested collections of quasi- coverings. A nested collection of quasi-coverings reflects the different levels of similarity present in the network. More specifically, as δ grows, covers would typically start exercising influence one over the other, then some nodes start to overlap while the remaining ones still exercise influence, until eventually the covers merge into one; see Fig. 4d for an example. Require- ment (i) in Def. 14 enforces border conditions on the edge set FX that are consistent with those required for KX in Def. 7. More precisely, at resolution δ = 0, each node must belong to its own cluster and there should be no influence relations among these. By contrast, for a large enough resolution δmax every node must belong to a single cluster, hence, no influence relation is possible. Requirement (ii) enforces nestedness in the influence structure FX. Specifically, if an edge exists at resolution δ between blocks Ci and Cj then this edge must persist for larger resolutions δ(cid:48) or the blocks must become more similar, encoded by having a larger intersection. We consider a nested collection of quasi-coverings to be the desired output of hierarchical overlapping quasi-clustering methods. Formally, define such methods O as a structure- preserving map from the space of networks to the space of nested collections of quasi-coverings, O : N → K. In analogy to the development of Sections III and IV, in which we used hierarchical (non-overlapping) clustering methods to obtain ultrametrics that are then combined to obtain cut metrics, in what follows we propose to use existing hierar- chical (non-overlapping) quasi-clustering methods to obtain quasi-ultrametrics that can be combined into quasi-cut metrics (Def. 15). Quasi-ultrametrics are defined as non-negative func- tions uX : X × X → R+ such that uX (x, x(cid:48)) = 0 if and only if x = x(cid:48) and ux(x, x(cid:48)) ≤ max{uX (x, x(cid:48)(cid:48)), uX (x(cid:48)(cid:48), x(cid:48))} for all x, x(cid:48), x(cid:48)(cid:48) ∈ X but not necessarily symmetric uX (x, x(cid:48)) (cid:54)= uX (x(cid:48), x); see [36]. Denote by U the set of all possible quasi- ultrametrics. Definition 15. Denote as V the conic hull of the set of quasi- ultrametrics, V = cone{ U}. Then, given a node set X, an element cX ∈ V is a quasi-cut metric on the node set X. Notice that, by definition, a convex combination of quasi- ultrametrics outputs a quasi-cut metric. Moreover, nested col- lections of quasi-coverings can be constructed from quasi-cut metrics as shown in the following theorem. Theorem 3. Given a node set X and a quasi-cut metric cX ∈ V, we can obtain a nested collection of quasi-coverings KX = { KX (δ), δ ≥ 0}, with KX (δ) = (KX (δ), FX (δ)) a quasi- covering, as follows. The covering KX (δ) is obtained from IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 8 the tolerance relation ↔δ determined by max{cX (x, x(cid:48)), cX (x(cid:48), x)} ≤ δ ⇒ x ↔δ x(cid:48). (12) The edge set FX (δ) is constructed such that for Ci, Cj ∈ KX (δ), we have that (Ci, Cj) ∈ FX (δ) if cX (x, x(cid:48)) ≤ δ. (13) min x∈Ci\Ci∩Cj x(cid:48)∈Cj\Ci∩Cj Proof. Notice that max{cX (x, x(cid:48)), cX (x(cid:48), x)} is a nonnega- tive, symmetric function that satisfies the identity property. Hence, ↔δ in (12) is a valid tolerance relation and, from the combination of Theorem 2 and Remark 2, it follows that KX = {KX (δ), δ ≥ 0} is a nested collection of coverings as required by Def. 14. We now show that the collection of edge sets FX satisfies conditions (i) and (ii) in Def. 14. It follows from the definition of quasi-cut metrics that cX (x, x(cid:48)) = 0 if and only if x = x(cid:48) so that FX (0) = ∅. Also, since KX is a nested collection of coverings, it must be that for some δmax we have KX (δ) = {X} for all δ ≥ δmax. Consequently, having only one block, the edge set must be empty, i.e., FX (δ) = ∅ for all δ > δmax. To prove that FX satisfies condition (ii), let Ci, Cj ∈ KX (δ) and (Ci, Cj) ∈ FX (δ) at a given δ. Then, there is (at least) a pair of nodes xi ∈ Ci\Ci ∩ Cj and xj ∈ Cj\Ci ∩ Cj that satisfy (13). Focus now on an arbitrary resolution δ(cid:48) > δ. The fact that KX = {KX (δ), δ ≥ 0} is a nested collection of j ∈ KX (δ(cid:48)) coverings implies the existence of blocks C(cid:48) such that Ci ⊆ C(cid:48) j. This immediately implies i and xj ∈ C(cid:48) that xi ∈ C(cid:48) j. Consider then the following two alternatives: (1) if neither xi nor xj belong to C(cid:48) j, then (13) j) ∈ FX (δ(cid:48)); and (2) if either xi or xj belong to forces (C(cid:48) i, C(cid:48) j > Ci ∩ Cj. Given that i ∩ C(cid:48) C(cid:48) for any δ(cid:48) either (1) or (2) must be true, requirement (ii) in Def. 14 is satisfied, completing the proof. j, then we have that C(cid:48) i and Cj ⊆ C(cid:48) i ∩ C(cid:48) i∩C(cid:48) i, C(cid:48) In virtue of Theorem 3 we can readily obtain a nested collection of quasi-coverings from a quasi-cut metric. The quasi-cut metric can thus be constructed by performing a conic combination of quasi-ultrametrics. The algorithm proposed is analogous to the one developed in Section IV. Given a network (X, AX ), first dither J times the dissimilarity function AX. Then, obtain a quasi-ultrametric u(i) X by applying a hierarchical (non-overlapping) quasi-clustering method – see, e.g., [36] – to each of the dithered networks Ni = (X, A(i) X ) for i = 1, . . . , J. Finally, generate the quasi-cut metric cX by averaging the J quasi-ultrametrics. In Fig. 4 we show an illustrative example of a hierar- chical overlapping quasi-clustering method O. We are given two quasi-ultrametrics u(1) X – representing those obtained by dithering the same underlying network – de- picted in Figs. 4a and 4b, respectively. We then obtain a quasi-cut metric by averaging these two quasi-ultrametrics cX = 0.5(u(1) X ); see Fig. 4c. Observe that cX is not a quasi-ultrametric since cX (x2, x3) = 3 which is larger than max{cX (x2, x1), cX (x1, x3)} = 2.5. In Fig. 4d we portray the nested collection of quasi-coverings resulting from the quasi-cut metric cX as explained in Theorem 3. It is observed that for δ = 0 all nodes belong to separate covers and the X and u(2) X + u(2) edge set is empty; and that for δ = 3 all nodes belong to the same single cover and the edge set is also empty [cf. (i) in Def. 14]. We can also observe the progression of influence as δ grows larger. Nodes start influencing each other (δ ≥ 1), until covers are created (δ = 2.5) based on nodes that were already exercising influence on each other. Then, some nodes overlap (x1), while non-overlapping nodes still exercise influence (x3 onto x2). Finally, all covers and edges collapse into a single cover containing all nodes (δ ≥ 3). Remark 6. Quasi-coverings in Def. 13 act as generalizations of quasi-partitions introduced in [36]. Likewise, a nested collection of quasi-coverings is the generalization of a quasi- dendrogram. These natural extensions are obtained from the constructions in [36] by dropping the non-intersecting require- ment. VI. SYNTHETIC EXPERIMENTS The first three illustrations of Algorithm 1 are conducted on synthetic networks N = (X, AX ) where X is a set of points embedded in R2 and AX is the Euclidean distance between them. As a hierarchical clustering method H we apply single linkage [13] and average the clustering output of J = 100 different noisy realizations of N. The noisy N are obtained by perturbing the positions of the nodes in R2 with zero-mean gaussian noise with standard deviation σ obtained as 10−1 of the smallest distance between nodes unless otherwise stated. For comparison, we include the output of single linkage hi- erarchical clustering. We observe that, for clusterable datasets, the cut metric approach proposed in this work coincides with the ultrametric clustering (Sections VI-A and VI-B). Furthermore, for the dumbbell network (Section VI-C), the proposed method effectively overcomes the chaining effect. A. Two clouds Consider the simple network portrayed in Fig. 5a, where there are 121 nodes on each cloud. The average distance between adjacent nodes within each of the clouds is d1 = 1 whereas the distance between the centers of both clouds is d2 = 22. The overlapping function is found in Fig. 5e. As expected, this network is clusterable (see Definition 12), i.e., for δ = 1.11 a non-trivial (neither all nodes together nor all separated) partition KX (δ) = {C1, C2} is obtained; see Fig. 5a. The resulting dendrogram obtained from applying single linkage to this network is shown in Fig. 5h. By looking at resolution δ = 1.26 we obtain the same set of covers as for the cut metric algorithm; see Fig. 5a. Observe that the value of δ that generates this partition that identifies both clouds is in the order of d1. B. Multiple resolutions Consider now the network depicted in Fig. 5b which con- sists of five clouds of 81 points each. Within each of the four clouds on the left, the average distance between adjacent nodes is d1 = 1 whereas the average distance between neighboring nodes in different clouds is d2 = 2. The fifth cloud, on the right, has an average distance between adjacent nodes within IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 9 C1 C2 C1 C2 C5 C13 C1 C2 C1 C2 (a) Two clouds (covers for δ = 1.11) C3 C4 C85 (b) Multiple resolutions (covers for δ = 1.11) (c) Multiple resolutions (covers for δ = 2.07) (d) Dumbbell (covers for δ = 2.17) (e) Two clouds: overlapping function (f) Multiple resolutions: overlapping function (g) Dumbbell: overlapping function (h) Two clouds: dendrogram (i) Multiple resolutions: dendrogram (j) Dumbbell: dendrogram Figure 5. Synthetic experiments. (a)-(d) Location of nodes for the three synthetic networks considered: (a) two clouds, (b)-(c) multiple resolutions, and (d) dumbbell. Also shown are the coverings obtained at different values of δ. (e)-(g) Overlapping functions for each one of the mentioned networks. Shown in red are the values of δ for which either a zero – (e) and (f) – or a minimum of local minima – (g) – is attained. (h)-(j) Reduced dendrograms resulting from applying single linkage to each of the three synthetic networks, respectively. These dendrograms have been simplified to show only a few illustrative nodes for the sake of clarity (i.e. nodes that are grouped at resolution levels close to 0 are not shown, and only a representative node in each of these groups is shown). the cloud of d2 and the distance between the center of the right cloud and that of the other four clouds is d3 = 22. The resulting overlapping function is found in Fig. 5f. We have also applied single linkage to this network and obtained the dendrogram shown in Fig. 5i. The network is found to be clusterable. First, for δ = 1.11 we obtain the covering KX (δ) = {C1, C2, . . . , C85}, as illustrated in Fig. 5b. In this case, each subset C1, . . . , C4 contains one of the clouds on the left and each of the subsets C5, . . . , C85 contains a single point of the fifth cloud. Intuitively, since δ = 1.11 is close to d1 but smaller than d2, only nodes that are at distances around or below d1 are grouped together. The same result is obtained when looking at the output of the single linkage clustering for resolution δ = 1.7387. Additionally, the network is also clusterable for δ = 2.07 with the resulting covering KX (δ) = {C1, C2} portrayed in Fig. 5c. In this case, there are only two subsets: C1 contains the four clouds on the left, and C2 contains the right cloud. This is reasonable since δ = 2.07 is slightly greater than both the inter-cloud distance on the left and the inter-point distance on the right but smaller than the minimum distance between points in the left and right clouds. In the dendrogram, for resolution δ = 2.1198 we get the same partition. C. A solution to single linkage's chaining effect Consider the network formed by the set of 171 nodes in Fig. 5d where the distance between adjacent nodes is d = 1. For this simulation, the value of σ used is equal to the minimum distance between nodes. Unlike the previous examples considered, the expected clustering output is not unequivocal since it is not clear how to group the nodes along the line connecting both point clouds. More specifically, it is not clear whether these points should constitute a cluster in itself or if this putative group should have overlap with the ones corresponding to the point clouds at its extremes. As a matter of fact, if one applies (non-overlapping) single linkage clustering to the network of interest, the output dendrogram shown in Fig. 5j only consists of the two trivial partitions: every node in a different cluster for resolutions δ < 1 or all nodes clustered together for δ ≥ 1. This phenomenon is known as chaining effect [9] and is an undesirable concomitant of the δ0.70.80.911.11.2Number of nodes050100150200250δ0.511.522.5Number of nodes050100150200250300350δ0.811.21.41.61.822.2Number of nodes050100150200250300nodes16209210121882155597985650 1 4 61117 34823 22249 7/11.522.533.544.55 1 30 29 28 79 53 52 35 56 55 58 59 60 78 61 62 51 64 65 66 67 46 81 40 39 38 37 80 57nodes00.20.40.60.811.2/IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 10 definition of single linkage. Nevertheless, via the introduction of dithering, non-trivial coverings of the network of interest can be recovered, thus, overcoming the aforementioned chain- ing effect. The overlapping function is presented in Fig. 5g and, in accordance with the intuition explained, the network is found not to be clusterable, i.e., no zeros are attained for intermediate resolutions. In this case, we focus on the resolution δ leading to the minimum number of overlapping nodes, i.e., we look for a covering associated with the minimum of all local minima of the overlapping function. For δ = 2.3854, the resulting covering is depicted in Fig. 5d where there are only two subsets C1 and C2, one containing all the left cloud as well as the connecting line (except for the rightmost point in the line), while the other subset contains all the right cloud and the connecting line (except for the leftmost point). By admitting overlap between clusters and using the overlapping function to guide our analysis, useful insight on the structure of the network was recovered. VII. APPLICATIONS We consider two real-world classification problems: deter- mining handwritten digits from the MNIST image database (Section VII-A), and identifying the authors and co-authors of famous plays (Section VII-B). A. Handwritten digit classification i i The proposed hierarchical overlapping clustering algorithm is implemented to classify digits of the MNIST database [26]. The database consists of black and white images {Gi}i∈I of size 28 × 28 pixels. In order to construct a network from the images, we first reshape each image Gi into a column vector yi of size 784 and then apply a principal component analysis (PCA) transform [37] to each vector. For computing the PCA, 5, 000 training samples of each digit are used to estimate the mean µ and the covariance matrix C. The resulting transformed vectors are for i ∈ I. A total of n images are considered denoted yPCA for classification, and only the first 20 components of the associated transformed vectors yPCA are used while the rest are discarded. The reduced vectors xi ∈ R20 for i = 1, . . . , n constitute the node set X = {x1, . . . , xn} of our network. Each element of the (symmetric) dissimilarity function is computed as the Euclidean distance between vectors in the 20-PCA space, i.e., AX (xi, xj) = (cid:107)xi − xj(cid:107)2. In the application of Algorithm 1, J = 100 noise realizations are carried out for the dithering step in all simulations. The standard deviation of the noise σ = k minxi(cid:54)=xj AX (xi, xj) is computed as a factor k of the minimum distance between nodes. The hierarchical (non- overlapping) clustering method H used to obtain the ultra- metrics is Ward's clustering method [11]. We also include the dendrograms obtained from applying H directly to the network. It is observed that the proposed method based on cut metrics outputs the same clusters as the hierarchical (non-overlapping) method since the MNIST handwritten digit database is a clusterable set (i.e., each image is associated to only one digit). 1) Two-digit classification. Digits 1 and 7: First, the per- formance of the algorithm is analyzed for classification of just two digits: 1 and 7. The total number of images considered is n = 200 counting 100 images of each digit, selected at random from the test set of the MNIST database. Images of the digits 1 and 7 from this database are illustrated in Fig. 6a, where the top row shows typical images that are correctly classified when using the proposed method, and the bottom row shows numbers that are often confused. The overlapping function obtained is displayed in Fig. 6d for k = 0.01 in the computation of σ. For δ = 8656, the covering KX (δ) = {C1, C2} with C1 = {1(×100), 7} and C2 = {7(×99)} is obtained. Same result is obtained when applying the Ward's linkage method at resolution δ = 8234. In this case, the proposed algorithm makes one classification mistake, hence, the error rate is 0.5% (1 misclassified image out of 200). To help illustrate this example, a non-metric Mul- tidimensional Scaling (MDS) [38] is performed on the vectors in R20 to be able to depict them on R2. This representation is shown in Fig. 6g. Notice that points representing images of different digits are clearly separated, except for one instance of 7 that is projected on top of the points corresponding to digit 1. 2) Three-digit classification. Digits 6, 8 and 9: As done before, 100 images are chosen at random from the test dataset for each digit, resulting in n = 300 images. As can be observed from Fig. 6b that contains images of the numbers being classified, digits 8 and 9 are much harder to distinguish than digit 6. The overlapping function is portrayed in Fig. 6e for k = 0.001. The covering KX (δ) = {C1, C2, C3} for δ = 9247 yields an error rate of 2.67% since C1 = {6(×100)}, C2 = {8(×95), 9(×3)} and C3 = {8(×5), 9(×97)}. Appli- cation of Ward's linkage at resolution 8854 yields the same three clusters. The MDS representation is depicted in Fig. 6h, confirming that the major concern is separating 8 and 9 since these two digits occupy approximately half of the plane and 6 occupies the other half. 3) Four-digit classification. Digits 0, 1, 2 and 7: We again consider 100 images per digit totalizing n = 400 images randomly selected from the test dataset. Examples of these digits are in Fig. 6c and the plot of the overlap- ping function can be found in Fig. 6f for k = 0.005. For δ = 8981 the resulting covering is KX (δ) = {C1, C2, C3, C4} with C1 = {0(×100), 2(×3)}, C2 = {1(×99), 7(×2)}, C3 = {1, 2(×86), 7(×3)}, and C4 = {2(×11), 7(×95)}, resulting in an error rate of 5%. Same results are obtained after applying Ward's linkage at resolution δ = 8887. The MDS representation is shown in Fig. 6i where we see that 2 is as sparsely located as 0 but presents larger overlap with the clouds corresponding to 1 and 7. This constitutes the main source of error as can be seen by inspecting blocks C3 and C4. To wrap up this first application it is worth pointing out that, while both the proposed hierarchical overlapping clustering method using cut metrics as well as the hierarchical (non- overlapping) clustering method using ultrametrics yield the same coverings, in the former the overlapping function was the key tool used to determine which resolutions yield reasonable results. In other words, the zeros of the overlapping function IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 11 Digit 1 Digit 1 Digit 7 Digit 7 Digit 6 Digit 8 Digit 9 Digit 0 Digit 1 Digit 2 Digit 7 Digit 1 Digit 1 Digit 7 Digit 7 (a) Digits 1, 7: images Digit 6 Digit 8 Digit 9 (b) Digits 6, 8, 9: images Digit 0 Digit 1 Digit 2 Digit 7 (c) Digits 0, 1, 2, 7: images (d) Digits 1, 7: overlapping function (e) Digits 6, 8, 9: overlapping function (f) Digits 0, 1, 2, 7: overlapping function (g) Digits 1, 7: MDS (h) Digits 6, 8, 9: MDS (i) Digits 0, 1, 2, 7: MDS (j) Digits 1, 7: dendrogram (k) Digits 6, 8, 9: dendrogram (l) Digits 0, 1, 2, 7: dendrogram Figure 6. Classification of handwritten digits. (a)-(c) Samples of images of the digits for each one of the three classification problems: (a) digits 1 and 7, (b) digits 6, 8, and 9, (c) digits 0, 1, 2, and 7. The first row shows samples of numbers that are typically classified correctly by the proposed method whereas the second row presents numbers that are hard to classify. (d)-(f) Overlapping functions for each of the classification problems considered. (g)-(i) Multidimensional Scaling (MDS) representation in 2 dimensions of the points in the 20-PCA domain. (j)-(l) Simplified dendrograms resulting from applying Ward's linkage to each of the three classification problems. 010002000300040005000600070008000900010000δ0102030405060Number of nodesδ0200040006000800010000Number of nodes0510152025303540δ0200040006000800010000Number of nodes0102030405060708090100-1000-5000500100015002000x1MDS-1000-500050010001500x2MDS17-1500-1000-500050010001500x1MDS-1000-50005001000x2MDS689-1000010002000x1MDS-1000-500050010001500x2MDS012721231012171119221825 72024 91413 1 216 3 4 5 6 815nodes2000400060008000100001200014000/1216 1 3 714 220 41511 523 8 9171318 62210192128242629253027nodes4000600080001000012000/nodes16191713151827222023 4 912 8101429322428303133 111 5 7 2 6212526 3/#1040.40.60.811.21.41.61.8IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 12 (b) SF: covers for δ = 0.2005 (a) SF: nodes (c) SF: covers for δ = 0.5459 (d) SF: overlapping function (e) SF: dendrogram (f) SCJ: nodes (g) SCJ: covers for δ = 0.0336 (h) SCJ: overlapping function (i) SCJ: dendrogram Figure 7. Authorship attribution. (a)-(e) Classifying plays from Shakespeare and Fletcher (SF). (a) Location of plays with correct labeling. (b) Covers obtained for the proposed algorithm at resolution δ = 0.1968 and for Ward's linkage at resolution δ = 0.1885 showing no overlap. (c) Covers for δ = 0.5460 obtained by the proposed algorithm present overlap between Shakespeare and Fletcher. (d) Overlapping function showing the two resolutions of interest. (e) Dendrogram resulting from applying Ward's linkage in which label 1 represent Shakespeare's plays, label 2 are Fletcher's plays and 8 are the co-authored plays. (f)-(i) Classifying plays from Shakespeare, Chapman and Jonson (SCJ). (f) Location of plays with correct labeling. (g) Covers obtained for the proposed method at resolution δ = 0.0331 showing overlap between Chapman and Jonson. (h) Overlapping function highlighting the local minimum of interest. (i) Dendrogram obtained from applying UPGMA's linkage in which labels 1, 3 and 6 correspond to plays authored solely by Shakespeare, Jonson and Chapman, respectively and where label 7 is for the co-authored play by Chapman and Jonson. were used to select adequate resolution values to obtain the coverings for the hierarchical overlapping clustering method. B. Authorship Attribution We address the problem of authorship attribution [27], where our objective is to attribute a given play to its rightful author. The authors considered in this experiment are William Shakespeare, John Fletcher, Ben Jonson, and George Chap- man. We consider 33 plays written by Shakespeare, 21 by Fletcher, 17 by Jonson, 14 by Chapman, as well as 2 plays co-authored by Shakespeare and Fletcher and 1 co-authored by Jonson and Chapman. Following the procedure described in [27] based on word adjacency networks, for each play we are able to obtain a dissimilarity measure to the four authors of interest. Denote by yij the dissimilarity of play i to author j. The index j represents the initial of each author, j ∈ {S, F, C, J}. To be more specific, if yij is small, then play i follows a function-word structure that resembles the one typically used by author j; see [27] for details. In what follows, the node set X is comprised of the subset of plays correspond- ing to the authors under study. Our goal is to cluster together plays written by the same author, as well as identifying co- authored plays in the overlap between author-specific clusters. By comparing with hierarchical non-overlapping clustering, we observe that our method successfully identifies the co- authored plays as being part of overlapping clusters, while for clusterable resolutions, using cut metrics amounts to the same result as using ultrametrics. 1) Shakespeare and Fletcher: Denote by S the set of 33 plays written solely by Shakespeare, by F the set of 21 plays solely authored by Fletcher and by SF the set containing the two co-authored plays. Define by xi the difference of dissimilarities of play i to each author, xi = yiF − yiS, for all i ∈ S ∪ F ∪ SF. The node set is then comprised of these 56 points on R representing the plays under study, i.e., X = {xi}i∈S∪F∪SF. The location of these points is depicted in Fig. 7a. The dissimilarity function considered is the Euclidean distance AX (xk, x(cid:96)) = xk − x(cid:96) for every pair of plays k and l. In the application of Algorithm 1, we use Ward's linkage [11] as the method H, we set J = 100 dithering steps, and we fix k = 4 in the determination of the noise amplitude; see Section VII-A. The resulting overlapping function is displayed in Fig. 7d, while the dendrogram obtained from just applying Ward's linkage is found in Fig. 7e. Two coverings of interest, achieved for resolutions δ = 0.1968 and δ = 0.5460 are presented in Figs. 7b and 7c, respectively. Additionally, the clusters shown in Fig. 7b are obtained by applying Ward's linkage at resolution δ = 0.1185. In Fig. 7b it is observed that the data does not have any over- lap and that there are two resulting clusters C1 ={Shakespeare (x 33), Fletcher} and C2 ={Fletcher (x 20), Shakespeare and Fletcher (x 2)}. Thus, as a non-overlapping clustering method, it is seen that the proposed algorithm makes only one mistake but fails to identify the co-authored plays, assigning them to just one of the authors. By contrast, in Fig. 7c the coverings C1 ={Shakespeare (x 33), Fletcher (x 5), Shakespeare and -0.06-0.04-0.0200.020.040.060.080.1x1ShakespeareFletcherShakespeare and Fletcher-0.06-0.04-0.0200.020.040.060.080.1x1ShakespeareFletcher-0.06-0.04-0.0200.020.040.060.080.1x1ShakespeareFletcherOverlapδ00.10.20.30.40.50.6Number of nodes05101520nodes11111111111111111121111111111111112822282222222222222222/00.10.20.30.40.5-0.0200.020.040.06x1-0.03-0.02-0.0100.010.020.030.040.050.06x2ShakespeareJonsonChapmanChapman and Jonson-0.0200.020.040.06x1-0.03-0.02-0.0100.010.020.030.040.050.06x2ShakespeareJonsonChapmanOverlapδ00.010.020.030.040.050.060.07Number of nodes051015202530nodes11111111111111111111111111111111133333333733333333666366666666666/00.010.020.030.040.050.06IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 13 Cymbeline 0 . 1 8 0 . 2 0 0 6 . 2 1 . 80 6 2 . 0 0 . 0 1 . 9 2 7 x7 Jew of Malta 0 . 1 8 0 . 2 0 x1 4 0.2 9 0.1 5 0 . 2 2 0 0 . . 1 2 9 2 Edward III Othello x2 0 . 1 9 0 . 1 9 0.21 0.18 Coriolanus x3 0 . 2 0 . 2 0 . 1 9 0 . 1 0.19 3 6 0.21 0 . 2 3 0 . 2 0 0 . 2 2 0 . 3 0 0.22 0.24 x6 6 2 . 0 0 . 1 9 x4 6 0.2 0 . 2 9 7 0.1 0 . 1 9 2 0 . 2 7 0 . 1 4 0.2 9 0.1 0 . 2 0 0 . 2 0 x5 Tamburlaine C1 x2 C2 C1 x6 x2 C2 x6 x3 x1 x4 x5 x7 x3 x1 x4 C3 x5 x7 (c) Quasi-UM: δ = 0.2138 (b) Quasi-UM: δ = 0.2127 C2 C1 x2 x3 x1 x4 x5 C3 x7 x6 Edward II (a) Shakespeare and Marlowe: Directed Network (d) Quasi-CM: δ = 0.2572 (e) Overlapping Function Figure 8. Authorship attribution as a directed network. (a) Plays by Shakespeare in blue and Marlowe in red, and one (allegedly) co-authored play, in green. The weight of the directed edges is also shown. (b)-(c) Coverings obtained by applying directed single linkage (DSL) to the network. For resolution δ = 0.2127 the co-authored play forms a separate cluster, while for resolution δ = 0.2138 this play is incorporated to Shakespeare's plays. (d) Coverings obtained by applying the proposed hierarchical overlapping quasi-clustering method with DSL as the quasi-ultrametric. We observe that for δ = 0.2572 covers are generated such that the co-authored play correctly belongs to both Shakespeare's and Marlowe's clusters. (e) Overlapping function resulting from applying the hierarchical overlapping quasi-clustering algorithm. Fletcher (x 2)} and C2 ={Fletcher (x 20), Shakespeare and Fletcher (x 2)} do overlap. More specifically, one error is committed and 6 plays belong to the overlap, among which the two co-authored plays are contained. 2) Shakespeare, Chapman, and Jonson: In this example, plays by Shakespeare, Chapman and Jonson are classified, including one play co-authored by the last two authors. Denote by S, C, J , and CJ , the sets containing the plays solely authored by Shakespeare, Chapman, and Jonson, and the co- authored play, respectively. Mimicking the construction in the previous experiment, for a given play i let xi1 = yiC − yiS and xi2 = yiJ − yiS be the differences in similarities between Chapman and Shakespeare, and Jonson and Shakespeare, respectively. Define the vector xi = [xi1, xi2]T ∈ R2 for each play i to obtain the node set X = {xi}i∈S∪C∪J ∪CJ containing the 65 plays from all three authors. The position of these nodes in R2 is illustrated in Fig. 7f. In this case, the dissimilarity function is given by the Euclidean distance between the points, i.e., AX (xk, x(cid:96)) = (cid:107)xk − x(cid:96)(cid:107)2, for plays k and l. In running Algorithm 1 we select UPGMA [10] as the hierarchical method H, we set J = 100 dithering steps and the noise amplitude is chosen for k = 1. The resulting overlapping function is depicted in Fig. 7h and the dendrogram obtained from directly applying UPGMA linkage is shown in Fig. 7i. A covering obtained from the proposed method for δ = 0.0336 is displayed in Fig. 7g. This covering is comprised of the blocks C1 ={Shakespeare (x 33)}, C2 ={Jonson (x 16), Chapman and Jonson}, C3 ={Jonson (x 2), Chapman (x 14)} and C4 ={Jonson, Chapman, Chapman and Jonson}. Note that block C1 contains all of Shakespeare's plays and nothing else, thus making this classification perfect. Block C2 contains all Jonson plays but one and also the co-authored play, and C3 contains all Chapman plays and 2 Jonson plays, one of which overlaps with C2. If these were all the outputs, then the algorithm would be making one full mistake (Jonson play classified as Chapman's), the co-authored play would not be recognized (as it is classified as Jonson's) and attention would be drawn to a Jonson play that is being classified as both Chapman and Jonson. However, the algorithm outputs a fourth covering C4 that contains one play by Jonson ("Sejanus"), one play by Chapman ("May Day") and the co-authored play ("Eastward Ho"). Thus, by outputting this fourth cluster, now there are three overlapping plays: one by each author and the co-authored one. Therefore, the algorithm successfully recognizes the co-authored play and draws attention to two other plays that are hard to classify. By applying UPGMA hierarchical (non-overlapping) method we obtain at resolution δ = 0.0357 just three clusters DX (δ) = {C1, C2, C3} which are C1 = {Shakespeare (x 33)}, C2 = {Jonson (x 15), Chapman and Jonson (x 1)} and C3 = {Jonson (x2), Chapman (x14)}. This clustering incurs in an error rate of 4.6% and fails to identify the co-authored play. 3) Shakespeare and Marlowe: As a last example, we use plays by Shakespeare and Marlowe to illustrate the applica- tion of the hierarchical overlapping quasi-clustering method (Section V). In this case, the network is constructed using directed edges based on the relative entropy between plays, which is asymmetric in nature, instead of using the distance to an author profile; see [27]. The resulting network and the corresponding directed dissimilarities AX are shown in Fig. 8a. For applying the algorithm we use quasi-ultrametrics obtained from applying directed single linkage (DSL) [36] to dithered versions of the network. We carry out J = 100 perturbations with noise given by 0.5 times the smallest nonzero entry of the dissimilarity function AX. For the sake /00.050.10.150.20.25fol(/)00.511.522.533.544.55IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 14 of comparison, we also include results obtained from directly applying DSL to the original network. By applying DSL we observe two relevant quasi-partitions, one for resolution δ = 0.2127 (Fig. 8b) and one for δ = 0.2138 (Fig. 8c). In the first case, we have three clusters, one con- sisting of Shakespeare's plays, one containing all Marlowe's plays, and a separate third cover including only the (allegedly) co-authored play [39]. We observe that there is an edge be- tween Shakespeare's plays and the other two covers. This can be intuitively interpreted as an indication that Shakespeare's writing is rich enough to contain part of Marlowe's stylistic fingerprint, but not vice versa; see [27]. Same happens for the edge between the co-authored play and Marlowe's plays, since the co-authored play also contains Shakespeare's style. When increasing the resolution to δ = 0.2138 we note that the co- authored play is incorrectly included as a Shakespeare play. However, when applying the hierarchical overlapping quasi- clustering algorithm, for resolution δ = 0.2572 we obtain the quasi-coverings illustrated in Fig. 8d. We observe that we have three covers, one containing all Shakespeare's plays and the co-authored play, one containing all Marlowe's plays, and the last cover containing two Marlowe's plays and the co-authored play. This third cover indicates that the co-authored play can be classified both as Shakespeare's and as Marlowe's. The resulting overlapping function is shown in Fig. 8e. VIII. CONCLUSIONS The proposed hierarchical overlapping clustering method obtains a collection of coverings of the data in such a way that a nested structure exists among the coverings and nodes can simultaneously belong to more than one cluster. The essence of the method is drawn from the connection of cut metrics, tolerance relations, and nested collections of coverings, that are respective generalizations of ultrametrics, equivalence relations, and dendrograms, typical of hierarchical (non-overlapping) clustering methods. A systematic method for obtaining cut metrics by dithering the dissimilarity function of a network is also proposed. The overlapping function is introduced as a tool for gaining insight on the structure of the data. It is also helpful in identifying meaningful resolutions within the nested collection of coverings. Additionally, a hi- erarchical overlapping quasi-clustering method was proposed. The objective of such method is to design a grouping algorithm that accommodates for asymmetries in the data, inherent to directed networks. We devised an algorithm for constructing quasi-cut metrics and showed how to obtain the output of the quasi-clustering method from these. The hierarchical overlapping clustering method was applied to three synthetic networks to establish its performance in controlled and intuitive scenarios, as well as to illustrate how overlap can solve single linkage's chaining effect. Moreover, the method was applied to classification of handwritten digits obtained from the MNIST database showing satisfactory re- sults in clusterable datasets. Finally, the proposed algorithm was used to classify plays by author, and it was shown that it succeeds in recognizing co-authored plays. By comparing with pertinent hierarchical non-overlapping algorithms, it was ob- served that for clusterable resolutions, the proposed algorithm using cut metrics coincides with the use of ultrametrics. REFERENCES [1] A. K. Jain and R. C. Dubes, Algorithms for Clustering Data, ser. Prentice Hall Advanced Reference Series. Englewood Cliffs, NJ: Prentice Hall, 1988. [2] V. Estivill-Castro, "Why so many clustering algorithms: A position paper," SIGKDD Explor. Newsl., vol. 4, no. 1, pp. 65–75, June 2002. [3] R. B. Cattell, "The description of personality: Basic traits resolved into clusters," J. Abnormal and Social Psychology, vol. 38, no. 4, pp. 476– 506, Oct. 1943. [4] C. Lu, X. Hu, and J. Park, "Exploiting the social tagging network for web clustering," IEEE Trans. Syst., Man, Cybern. A, vol. 41, no. 5, pp. 840–852, Sep. 2011. [5] M. Paulus and L. Kristoufek, "Worldwide clustering of the corruption perception," Physica A: Statist. Mech. Appl., vol. 428, pp. 351–358, 2015. [6] A. Ozdemir, M. Bolanos, E. Bernat, and A. Selin, "Hierarchical spectral consensus clustering for group analysis of functional brain networks," IEEE Trans. Biomed. Eng., vol. 62, no. 9, pp. 2158–2169, Sep. 2015. [7] L. A. Dalton, M. E. Benalcazar, M. Brun, and E. R. Dougherty, "Analytic representation of Bayes labeling and Bayes clustering operators for random labeld point processes," IEEE Trans. Signal Process., vol. 63, no. 6, pp. 1605–1620, March 2015. [8] X. Zao and A. H. Sayed, "Distributed clustering and learning over networks," IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3285–3300, July 2015. [9] G. N. Lance and W. T. Williams, "A general theory of classificatory sorting strategies: 1. Hierarchical systems," Comput. J., vol. 9, no. 4, pp. 373–380, Apr. 1967. [10] R. R. Sokal and C. D. Michener, "A statistical method for evaluating systematic relationships," University Kansas Sci. Bulletin, vol. 38 (II), no. 22, pp. 1409–1438, March 1958. [11] J. H. Ward Jr., "Hierarchical grouping to optimize an objective function," J. Amer. Statist. Assoc., vol. 58, no. 301, pp. 236–244, March 1963. [12] D. Defays, "An efficient algorithm for a complete link method," Comput. J., vol. 20, no. 4, pp. 364–366, Apr. 1977. [13] G. E. Carlsson and F. M´emoli, "Characterization, stability and con- vergence of hierarchical clustering methods," J. Mach. Learning Res., vol. 11, pp. 1425–1470, Apr. 2010. [14] M. A. Youssef, A. Youssef, and M. F. Younis, "Overlapping multihop clustering for wireless sensor networks," IEEE Trans. Parallel Distrib. Syst., vol. 20, no. 12, pp. 1844–1856, Dec. 2009. [15] P. C. H. Ma and K. C. C. Chan, "A novel approach for discovering overlapping clusters in gene expression data," IEEE Trans. Biomed. Eng., vol. 56, no. 7, pp. 1803–1809, July 2009. [16] J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Al- gorithms, ser. Advanced Appl. Pattern Recognition. New York, NY: Plenum Press, 1981. [17] K. Yu, S. Yu, and V. Tresp, "Soft clustering on graphs," in Neural Inform. Process. Syst. (NIPS), Vancouver, BC, 5-8 Dec. 2005. [18] G. Bordogna and G. Pasi, "A quality driven hierarchical data divisive soft clustering for information retrieval," Knolwedge-Based Systems, vol. 26, pp. 9–19, Feb. 2012. [19] G. Cleuziou, "An extended version of the k-means method for overlap- ping clustering," in Int. Conf. Pattern Recognition 2008. Tampa, FL: Int. Assoc. Pattern Recognition, 8-11 Dec. 2008. [20] M. Deza and M. Laurent, Geometry of Cuts and Metrics. Paris, France: LIENS - Ecole Normale Sup´erieure, Nov. 1996. [21] J. Culbertson, D. P. Guralnik, and P. F. Stiller, "Injective metrizability and the duality theory of cubings," arXiv:1502.00126v1 [math.MG], 31 Jan. 2015. [Online]. Available: http://arxiv.org/abs/1502.00126 [22] W. Bartol, J. Mir´o, J. Pi´oro, and F. Rossell´o, "On the coverings by tolerance classes," Inform. Sci., vol. 166, no. 1-4, pp. 193–211, Oct. 2004. [23] A. Lancichinetti, S. Fortunato, and J. Kert´esz, "Detecting the overlap- ping and hierarchical community structure in complex networks," New Journal of Physics, vol. 11, pp. 033 015 (1–18), March 2009. [24] H. Shen, X. Cheng, K. Cai, and M.-B. Hu, "Detect overlapping and hierarchical community structure in networks," Physica A, vol. 388, no. 8, pp. 1706–1712, Apr. 2009. [25] L. Schuchman, "Dither signals and their effect on quantization noise," IEEE Trans. Commun. Technol., vol. 12, no. 4, pp. 162–165, Dec. 1964. IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS (ACCEPTED) 15 [26] Y. Le Cun, C. Cortes, and C. J. C. Burges, "The MNIST database of handwritten digits," Website, 18 Aug. 2015. [Online]. Available: http://yann.lecun.com/exdb/mnist/ [27] S. Segarra, M. Eisen, and A. Ribeiro, "Authorship attribution through function word adjacency networks," IEEE Trans. Signal Process., vol. 63, no. 20, pp. 5464–5478, Oct. 2015. [28] P. R. Halmos, Naive Set Theory, ser. University Series Undergraudate Math. New York, NY: Van Nostrand Reinhold Company, 1960. [29] G. E. Carlsson, F. M´emoli, A. Ribeiro, and S. Segarra, "Axiomatic construction of hierarchical clustering in asymmetric networks," arXiv:1301.7724v2 [cs.LG], 2 Sep. 2014. [Online]. Available: http: //arxiv.org/abs/1301.7724 [30] J.-C. Hausmann, "On the Vietoris-Rips complexes and a cohomology theory for metric spaces," in Prospects in Topology, ser. Ann. Math. Stud. Princeton, NJ: Princeton University Press, 1995, vol. 138, pp. 175–188. [31] J. Culbertson, D. P. Guralnik, and P. F. Stiller, "Functorial hierarchical clustering with overlaps," arXiv:1609.02513v1 [cs.LG], 8 September 2016. [Online]. Available: http://arxiv.org/abs/1609.02513 [32] C. Bron and J. Kerbosch, "Algorithm 457: finding all cliques of an undirected graph," Communications of the ACM, vol. 16, no. 9, pp. 575–577, Sep. 1973. [33] P. B. Slater, "A partial hierarchical regionalization of 3140 US counties on the basis of 1965-1970 intercounty migration," Environment and Planning A, vol. 16, no. 4, pp. 545–550, Apr. 1984. [34] W. Pentney and M. Meila, "Spectral clustering of biological sequence Pittsburgh, PA: Assoc. data," in 20th Nat. Conf. Artificial Intell. Advancement Artificial Intell., 9-13 July 2005. [35] Y. Zhao and G. Karypis, "Hierarchical clustering algorithms for docu- ment datasets," Data Mining and Knowledge Discovery, vol. 10, no. 2, pp. 141–168, March 2005. [36] G. Carlsson, F. M´emoli, A. Ribeiro, and S. Segarra, "Hierarchical quasi- clustering methods for asymmetric networks," in 31st Int. Conf. Mach. Learning. Beijing, China: J. Mach. Learning Res., 21-26 June 2014. [37] J. E. Jackson, A User's Guide to Principal Components. New York, NY: John Wiley and Sons, 1991. [38] J. B. Kruskal and M. Wish, Multidimensional Scaling, ser. Quantitative Appl. Social Sci. Iowa City, IA: Sage Publications, Inc., 1978. [39] T. Merriam, "Marlowe's hand in edward iii," Literary and linguistic computing, vol. 8, no. 2, pp. 59–72, 1993.
1507.03490	3	1507	2015-09-27T16:49:21	What are essential concepts about networks?	[ "cs.SI", "math.CO", "nlin.AO", "physics.data-an", "physics.soc-ph" ]	Networks have become increasingly relevant to everyday life as human society has become increasingly connected. Attaining a basic understanding of networks has thus become a necessary form of literacy for people (and for youths in particular). At the NetSci 2014 conference, we initiated a year-long process to develop an educational resource that concisely summarizes essential concepts about networks that can be used by anyone of school age or older. The process involved several brainstorming sessions on one key question: "What should every person living in the 21st century know about networks by the time he/she finishes secondary education?" Different sessions reached diverse participants, which included professional researchers in network science, educators, and high-school students. The generated ideas were connected by the students to construct a concept network. We examined community structure in the concept network to group ideas into a set of important themes, which we refined through discussion into seven essential concepts. The students played a major role in this development process by providing insights and perspectives that were often unrecognized by researchers and educators. The final result, "Network Literacy: Essential Concepts and Core Ideas", is now available as a booklet in several different languages from http://tinyurl.com/networkliteracy .	cs.SI	cs	What are essential concepts about networks? Hiroki Sayama∗ Center for Collective Dynamics of Complex Systems, Binghamton University, State University of New York, Binghamton, NY 13902-6000, USA Center for Complex Network Research, Northeastern University, Boston, MA 02115, USA ∗Corresponding author: [email protected] New York Hall of Science, Corona, NY 11368, USA Catherine Cramer Mason A. Porter Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute, University of Oxford, OX2 6GG, UK CABDyN Complexity Centre, University of Oxford, Oxford, OX1 1HP, UK Network Science Center, United States Military Academy at West Point, Lori Sheetz West Point, NY 10996, USA Stephen Uzzo New York Hall of Science, Corona, NY 11368, USA September 29, 2015 Abstract Networks have become increasingly relevant to everyday life as human society has become increasingly connected. Attaining a basic understanding of networks has thus become a necessary form of literacy for people (and for youths in particular). At the NetSci 2014 conference, we initiated a year-long process to develop an educational resource that concisely summarizes essential concepts about networks that can be used by anyone of school age or older. The process involved several brainstorming sessions on one key question: “What should every person living in the 21st century know about networks by the time he/she finishes secondary education?” Different sessions reached diverse participants, which included professional researchers in network science, educators, and high-school students. The generated ideas were connected by the students to construct a concept network. We examined community structure in the concept network to group ideas into a set of important themes, which we refined through discussion into seven essential concepts. The students played a major role in this development process by providing insights and perspectives that were often unrecognized by researchers and educators. The final result, “Network Literacy: Essential Concepts and Core Ideas”, is now available as a booklet in several different languages from http://tinyurl.com/networkliteracy. 1 Introduction Network science has matured over the past few decades, and its potential importance for improving under- standing of complex natural and human-made phenomena is now recognized in an increasingly diverse set of domains [16, 2]. Its potential beneficiaries include not only researchers, but also business practitioners, policymakers, educators, and virtually all lay people, as they need to understand and deal with complex real- world phenomena. To advance the use of networks as a lens on ourselves and on our world, there needs to be widespread and sophisticated knowledge about how the study of networks—the science of connectivity—can 1 be an important epistemological tool for all people. networks is a new kind of literacy that is important for everyone living in the 21st century. In other words, we assert that an understanding of We favor the term “Network Literacy” over “Network Science Literacy”, because we believe that the scope and impact of networks goes beyond what one would typically construe as “science”. The scope of Network Literacy is also rather different from what is currently being taught as “Digital Literacy” or “Internet Literacy” about the use of computers, the Internet, and other digital media [29, 6]. Network Literacy needs to include fundamental concepts about networks, including (1) how an overwhelming abundance of everyday phenomena can be viewed through the lens of connectivity, interactions, and interdependence; and (2) how one can analyze, understand, utilize, and improve their features. In response to such emerging societal needs, researchers in the network-science community have started educational outreach efforts to bring network science to secondary education [10, 22, 7]. There are also a few other dispersed outreach efforts that have considered, or even focused on, teaching ideas about networks (see, e.g., [4, 15, 12, 14, 27]). However, there do not yet exist educational materials about networks that are both systematically structured and easily accessible—such as grade-school-level textbooks, workbooks, curricular modules, and lesson plans (e.g., the educational materials [21, 20, 1] developed by the “System Dynamics” community [8, 25])—and which can be integrated readily into formal and informal educational programs. As a first step towards meeting this need, we undertook a year-long effort to develop a set of Essential Concepts for Network Literacy. We started by consulting people who had succeeded in similar educational initiatives. We were particu- larly inspired by the “Ocean Literacy” initiative [5, 23, 26], which was spearheaded by members of the US NSF Centers for Ocean Science Education Excellence (COSEE). The work on the Ocean Literacy initiative began in 2002 when a grassroots group of scientists and educators, frustrated by the lack of ocean-science content in national education standards, came together to discuss what they felt every person should know about the ocean by the time he/she graduates from high school.1 After discussing initial ideas for two years, the group organized a two-week online conference in 2004 that included more than 100 participants, a live keynote talk and virtual “rooms” organized according to scientific disciplines and age groups. Their next step was to hold a charrette, at which 30 people convened for two days and distilled the combined input into 7 principles (plus 44 fundamental concepts to more fully explain these principles). The final result was pub- lished in 2005 as a pamphlet called “Ocean Literacy: The Essential Principles and Fundamental Concepts of Ocean Sciences for Learners of All Ages”.2 To date, more than 30,000 brochures have been distributed in hard copy (in addition to digital dissemination). The success of Ocean Literacy was followed by similar processes in Climate Literacy, Atmospheric Literacy, Earth Science Literacy, Great Lakes Literacy, Energy Literacy, and more.3 Our goal is similar: we wish to provide a simple, concise, high-level guiding document that can facilitate the process of developing resources for disseminating Network Literacy to all people. To initiate this process, we posed one key question to ourselves and to the network-science community: What should every person living in the 21st century know about networks by the time he/she finishes secondary education? In the present paper, we describe the highly collaborative process by which we developed the first version of “Network Literacy: Essential Concepts and Core Ideas” (which is now available at http://tinyurl.com/ networkliteracy). Network-science researchers, educators, high-school students, and others collectively explored a wide variety of ideas and developed a network of concepts. We examined community structure [18] in the network of concepts to obtain several concept clusters, which were further discussed and refined to ultimately yield seven essential concepts about networks. We hope that the process and outcomes of this initiative will be useful for people who are interested in network science and in networks more generally, and we also hope that it will have positive effects on teaching, learning, communicating, and policy-making about networks. 1Their history is discussed at http://oceanliteracy.wp2.coexploration.org/ocean-literacy-network/foundations/ history/. 2See http://www.oceanliteracy.net/. 3See http://nagt.org/nagt/teaching_resources/literacies.html. 2 2 Methods 2.1 Brainstorming and Concept Generation Our starting point was to hold several brainstorming sessions to explore various concepts and ideas about networks that could be used for developing the list of essential concepts (see the top panel of Fig. 1). Each session followed the same basic procedure: 1. Have each participant introduce him/herself to the rest of the group. 2. Explain the objective of the session and then pose the following key question to the group: “What should every person living in the 21st century know about networks by the time he/she finishes sec- ondary education?” 3. Hand out sticky notes to the participants. 4. Let each participant write his/her ideas about what “network literacy” might mean on the sticky notes. Participants work on this phase as individuals without communication. This usually takes about 15 minutes. 5. Have participants place their sticky notes on a large wall or a blackboard, and then have them cluster the notes into several topical areas based on their similarity and relatedness. This phase usually takes 15 minutes or often longer, and it involves a lot of physical interaction and conversations among the participants (see Fig. 2). 6. Have a group discussion on the result of brainstorming to refine and distill a final set of essential concepts. This phase also takes about 15 minutes. 7. Break the group into subgroups and have each subgroup visualize one of the essential concepts that arise from the brainstorming. For educational purposes, we included this phase only at sessions with high-school students. Their visualization results were later compiled into a PowerPoint presentation and presented to the students to summarize and reinforce their learning experience. Each brainstorming session produced two primary sets of outcome data: (a) handwritten data on sticky notes; and (b) a final compiled list of essential concepts that the group produced (inspired by the sticky notes) through brainstorming. We used the former to qualitatively compare the outputs of different participant groups, and we used the latter to develop a final version of the essential concept list. We ran brainstorming sessions with the following four participant groups (see the bottom panel of Fig. 1): • Network-science researchers and educators who participated in NetSci 2014 (“professionals” or “pro”). The session was run at a NetSci 2014 pre-conference event at University of California at Berkeley on 1 June 2014. Unlike the other three sessions, this group also had a follow-up discussion during the next day at the NetSciEd3 satellite symposium4 to revise the list of essential concepts (and to get input from people who could not attend the previous day). • The 2014–2015 cohort of NetSci High [7] high-school students who just finished an initial training workshop on network science (“new students” or “new”). The session was run at the NetSci High Summer Workshop at Boston University on 22 July 2014. • The 2013–2014 cohort of NetSci High high-school students who had completed a year-long research experience in network science (“old students” or “old”). The session was run at the NetSci High Summer Workshop at Boston University on 23 July 2014. • Explainers at the New York Hall of Science (mostly high-school students, but also some university students) who did not have any learning or training experience about networks (“baseline group”). The session was run at the New York Hall of Science on 13 August 2014. 4See https://sites.google.com/a/nyscience.org/netscied3/. 3 Figure 1: Overview of the work flow of the process by which the “Network Literacy” booklet was devel- oped. (Top) Procedure of each initial brainstorming session. (Bottom) Process of concept generation and refinement. 4 Brainstorming session Writing ideas individually on sticky notes Organizing ideas by discussion (Fig. 2) Outcome data (a): Original ideas written on sticky notes Outcome data (b): List of essential concepts generated by the group NetSci researchers and educators (Pro) New students who started learning about networks Old students who finished year-long NetSci research Baseline students with no prior knowledge Concept generation and refinement Word counting, word-cloud visualization (Figs. 3, 4) Lists of concepts from three groups with NetSci knowledge (Tables 1–3) Concepts connected by participants Community detection (Figs. 5–7) Discussion & refinement Seven essential concepts (Table 5) Seven essential concepts & Core ideas Adding details Revision & Design Feedback from NetSci community Feedback from NetSci community Final booklet (Fig. 8) Brainstorming sessions by four different groups Figure 2: Initial brainstorming session by network-science researchers and educators at the NetSci 2014 pre-conference event held at University of California at Berkeley on 1 June 2014. No high-school students were present at this session. Participants posted their ideas of essential concepts about networks using sticky notes, and they then organized them into several concept clusters through discussion. The same format was used for the other brainstorming sessions with high-school students. (Photo credit: Hiroki Sayama) 5 2.2 Comparison Between Participant Groups To compile raw text data from each group, the original ideas written on the sticky notes by the participants were collected and typed up. We conducted simple content analysis of this data set based on word-frequency counting [28, 24]. Specifically, we broke the text data into words, and we then removed trivial words (e.g., articles and simple prepositions). We converted the remaining words into a canonical form by converting plural nouns to singular ones and by removing aspects and tenses from verbs. We also removed words that appeared only once in the data for each group. For initial visual inspection, we used Mathematica 10.2’s WordCloud function (in its default settings) to construct a word cloud for the words generated by each group. We then compared the word frequencies across the four groups to detect semantic differences between them. We generated word rankings based on their frequencies and visualized them in Mathematica 10.2 using a custom script. Because of the small number of samples, we did not conduct statistical tests. 2.3 Development of Final List of Essential Concepts We used the lists of essential concepts produced by the three participant groups with some experience in network science (“professionals”, “old students”, and “new students”) to help develop a final list of essential concepts about networks (see the bottom panel of Fig. 1). To facilitate this process, we presented the participants (researchers, educators, and high-school students) at the NetSci High Summer Workshop with a sheet of paper on which all of the concepts from the three groups were presented together, and we asked them to draw connections among those concepts. From the results of this activity, we produced a single multigraph, in which each essential concept produced by the three groups is represented by a node and in which an edge represents a connection that a participant made between a pair of concepts. We detected communities of concepts on this multigraph using Mathematica 10.2’s FindGraphCommunities function in its default settings (which uses a method based on maximizing the modularity objective function). This produced several distinct communities with clear topical themes, as well as a patchwork community (i.e., a community whose topical theme was not clear) that consisted of the remaining concepts. We applied the same community-detection technique to the patchwork community (extracted as an induced subgraph) repeatedly to attempt to discern more subtle concept clusters within it. This procedure was repeated twice (i.e., considering patchwork communities as networks and detecting communities in them), after which we were able to obtain a set of concept clusters that were thematically meaningful. Once we detected several meaningful concept clusters, we further discussed and refined them. We then used the resulting clusters to draft seven essential concepts for Network Literacy, and we announced this list (which we posted on Google Drive) to the network-science community via e-mail to collect feedback. We incorporated the community’s feedback and used it to help develop in-depth descriptions of each es- sential concept (called “core ideas”) to provide more details about each concept (including examples and implications). We then announced this version, now furnished with both essential concepts and core ideas, to the network-science community to receive additional feedback. We incorporated this feedback, which we also discussed and refined further among ourselves, and placed the final content into a booklet for public distribution. 3 Results 3.1 Comparison Between Participant Groups In Tables 1–4, we show the lists of essential concepts that were created by the four participant groups. There is a large difference between the concepts created by the baseline group (i.e., high-school or university students without any formal training with networks), whose concepts we show in Table 4, and the other groups. The concepts generated by the baseline group tended to be vague and general, and they were likely derived from everyday uses of the word “network” in English (e.g., social networking and computer networks). In contrast, the concepts that were created by the other three groups (see Tables 1–3), who had learning and training experiences about networks, included much more concrete notions about networks 6 Table 1: List of essential concepts that were created by network-science researchers and educators who par- ticipated in the NetSci 2014 pre-conference event and the NetSciEd3 satellite symposium (“professionals”). No. Concept 1. 2. 3. 4. 5. 6. 7. Networks describe how things connect & interact. The world around us can be represented as networks of interconnecting parts. Representing networks allows them to be used as a tool. Modeling systems as networks can help reveal and explain patterns and general principles. Network structure can influence behavior and vice versa. Networks help us understand similarities among systems found in our everyday life. Networks can be represented and studied in many different ways. Table 2: List of essential concepts that were created by the 2014–2015 cohort of NetSci High high-school students (“new students”) who just finished a short training workshop at the 2014 NetSci High Summer Workshop at Boston University. No. Concept 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. Networks connects various aspects of the world. Networks are connections and interactions. Networks are present in every aspect of life. Examples include economics/social/political sciences. Networks consist of nodes and links. Computers are often used to study networks. Networks can be used for making predictions. Visualization of networks helps understanding. Networks allows us to analyze human interactions. Network science should be incorporated into schools. Future of network science can include preinstalled programs and networks in media. that likely also encapsulated a better understanding of salient ideas. This demonstrates emphatically that education can and does have an impact on people’s understanding of networks. There are also some interesting differences between the concepts that were created by professionals and those that were created by students. The former (see Table 1) generally used abstract language and also had a clear focus on what could be done using network models (which they also described in abstract ways). In contrast, the concepts that were created by students (see Tables 2 and 3) consisted predominantly of definitions of networks and concrete examples, together with research methodologies that they had learned (e.g., visualization and use of computer software). These differences convey a clear gap between professionals and lay people regarding what is perceived as most essential to the study of networks, and they suggest that it is crucial to actively involve non-professionals (in addition to practicing network scientists) in the development of a guiding booklet of essential concepts about networks. The Powerpoint presentations of concept visualizations that the NetSci High students constructed (see our description above) are available online.5 Our aforementioned observations are supported further by word-cloud visualizations (see Fig. 3) and word-ranking visualizations (see Fig. 4), which we constructed using the methods described in Section 2. The word “network” was the most frequently-used word in all four groups, and we thus removed it from the analysis. The other most common words give identifiable characterizations for each of the groups. For example, the baseline group was the only one to include career-related words (such as “job”, “work”, and “opportunity”), which came from the participants’ recognition of the importance of “networking” for their future career development. Additionally, the total number of words produced by the baseline group was the smallest, suggesting that they were not aware of the broad applicability of network-related concepts to a wide 5Presentation by old students: http://www.bu.edu/networks/files/2012/08/Networks-by-NetSciHigh2013-2014-student-team.pptx Presentation by new students: http://www.bu.edu/networks/files/2012/08/Networks-by-NetSciHigh2014-2015-student-team.pptx 7 Table 3: List of essential concepts that were created by the 2013–2014 cohort of NetSci High high-school students (“old students”) who had completed a year-long research experience. No. Concept 1. 2. 3. 4. 5. 6. Networks are everywhere. A connection between more than one person can form a link. People should understand basic terminology in a network such as nodes, edges and hubs. Visualizations are the key to understanding the network. Networks can be applied to various fields, whether through entertainment, disease spread, finding quality furniture, etc. Life is hard... then you die. Table 4: List of essential concepts that were created by high-school students (explainers at the New York Hall of Science) who had not had any learning or training experience about networks (i.e., the “baseline group”). No. Concept 1. 2. 3. People and connections Act of network Technology + networking variety of phenomena. The words from the old students included more technical terms (e.g., “node”, “link”, and “edge”) than any other group, indicating that they had become familiar with those technical words from their year of research experience in network science. The new students produced the most diverse set of words (which is apparent in both word clouds and word rankings), which nicely conveys their intellectual excitement immediately after learning about networks. Finally, the words produced by professionals often tended to be about high-level, abstract concepts. Interestingly, the word “people” was not used as often by the professionals as by other groups. 3.2 Development of Final List of Essential Concepts Detecting communities in our final concept multigraph yielded several distinct concept clusters. These include basic terminology, definitions of networks, ubiquity of networks, and social applications (see Fig. 5). Detecting communities in networks consisting of the remaining patchwork communities revealed concept clusters about visualization, computer-based representation, education, and pattern identification (see Figs. 6 and 7). Three of those clusters (basic terminology, visualization, and education) arose only in the student groups. This again illustrates the limitation of brainstorming without active student participation. The algorithmically-detected concept clusters served as the basis for additional discussion and refinement of the essential concepts. Through iterative discussion sessions among the authors, we decided (to reduce potential redundancy) to include social applications in the cluster about the ubiquity of networks and to combine basic terminology with definitions of networks. We also decided not to include education explicitly as a separate concept, because the importance of education was the main motivation of the Network Literacy initiative. This yielded the following five major concepts: 1. Ubiquity of networks, including social applications 2. Definitions of networks and basic terminology 3. Pattern identification 4. Visualization 5. Computer-based representation Through further discussion about the above five concepts, we recognized that some of the contemporary (and actively-researched) concepts about network science were missing. Specifically, this includes the power 8 Figure 3: Word clouds generated (using Mathematica 10.2) to visualize the frequencies of words used by each of the four participant groups (excluding “network”, which was the most frequently-used word). We scaled the word sizes logarithmically based on their number of appearances. 9 peopleconnectcitizengraduateschoolhighinternetjobcreatefutureworkopportunitycommunicateimportantcellphoneavailabledifferentpersonbenefitthingrightmeanwantthinknewformunittalkallBaselineGroupconnectpeopletransportationsciencerelationshipcommunicateunderstandbetweendifferentthinggroupsystemmakeeverythinginteractsocialrepresentcomputerexampleprogrameveryonestudytogetherincludeanythinghelpanalyzevisualizedatasubjectsocietydiseasehumanallownodefacebookideabiologythroughoutspreadpatternpredictworldanotherobjectimportantorganizemanybasicschoolinformationlinkconclusionthinktwitterformdegreepythontrendshowlifegraphfoodfieldfashioncreatemediabettervisualfuturefriendshipneedeverydaycomposewebconceptalldeterminegeneralexplaincountryenginebehaviorestablishessentialefficientdirectedeverycertainedgetieclustereasiereventfriendentirecomedrawfacefindNewStudentspeopleunderstandnodeconnectanalyzedifferentvisualizeedgefieldscienceeverywherelinkmanybetweeninformationsocialbasiccommonterminologydataimportantfindconceptfriendprovideideashowpersongandhimeasurethinkgraphcertainstudyclusterapplytypelifeknowledgelookepidemiccentralityeverydaypropertygeneralpatternqualityelementspreadfurniturefamilywebsiteinteractanotherhairballallhumaneverymediacreatehelpwhenordermakeknowmosttermnewkindtreehubseevsOldStudentsconnectsomethingthingunderstandimportantinformationphysicssystemrelationshipwhylookchangethinkvisualizeelementdifferentscienceentitysocialworldmeanbasicmathideaconsequencecomponenteverythingnaturallyproblemmultipleconceptallidentifyliteracypeoplestructureinteractrangepointaroundeffectpartgivenodeneedlimittermlinkbadskillProfessionalsFigure 4: Visualization of words from the raw text data written on sticky notes during each of brainstorming the sessions (generated by Mathematica 10.2 using a custom script). In each case, we rank the words (again excluding the word “network”) from highest frequency at the top to lowest frequency at the bot- tom. The background color intensity (which is not scaled linearly) represents the relative frequencies of the words. When the same word appears in multiple groups, we use a curve to connect the word in the different groups. 10 peopleconnectcitizenschooljobhighworkinternetgraduatefuturecreatewantunitthinkthingtalkrightpersonopportunitynewmeanimportantformdifferentcommunicatecellphonebenefitavailableallnodefieldlinkpeopleconnectunderstandedgedifferentvisualizescienceanalyzesocialmanyfinddatabasicideabetweenthinkshowlifeinformationfriendeverywherecommontypeterminologystudyprovidepersonmeasurelookimportantgraphgandhiconceptclustercertainapplyallwhenwebsitevstreetermspreadseequalitypropertypatternordernewmostmediamakeknowledgeknowkindinteracthumanhubhelphairballgeneralfurniturefamilyeverydayeveryepidemicelementcreatecentralityanotherconnectsciencepeoplemakethingbetweensocialdifferentdatagroupunderstandinteracthelptransportationsystemstudyrelationshipincludeideacommunicatenodelinklifeallowworldvisualizeprogramhumanexamplediseasetogetherthinksubjectsocietyrepresentformeverythingeveryonecomputerbasicanythinganalyzetwittertrendspreadshowschoolpredictpatternobjectmanyfoodfieldfacebookbiologyanotherallwebvisualtiethroughoutpythonorganizeneedmediainformationimportantgraphfuturefashiondegreecreateconclusionbetteryoungwhywherewealthvariousundirectedtooltodaysupportstemsiteseriesseparatesenseseesearchrequirerelatedrealputportionpopularpiecepartparadoxoftennewmillionlearnknowledgekidinvolveinterrelatedinternetinstallinfluencehubgephigeneralfriendshipfriendfindfaceexplaineverydayeveryeventestablishessentialentireengineefficientedgeeasierdrawdirecteddeterminecountryconceptcomposecomeclustercertainbehaviorconnectthingwhyunderstandsomethinglookthinksystemphysicsinformationimportantworldvisualizesocialsciencerelationshipmeanmathideaentityelementdifferentchangebasicalltermstructureskillrangeproblempointpeoplepartnodeneednaturallymultipleliteracylinklimitinteractidentifygiveeverythingeffectconsequenceconceptcomponentbadaroundcommunicatecommunicateelementelementformformfuturefutureinternetinternetlooklookmeanmeanschoolschooltermtermthingthingBaselineOldStudentsNewStudentsProfessionals” s l a n o i s s e f o r p “ d n a , ) ” N “ ( ” s t n e d u t s w e n “ , ) ” O “ ( ” s t n e d u t s d l o “ : s p u o r g e e r h t y b e d a m s t p e c n o c g n i t c e n n o c y b d e t a r e n e g k r o w t e n t p e c n o C : 5 e r u g i F c i s a B “ t n e s e r p e r y e h T . n o i t c n u f s e i t i n u m m o C h p a r G d n i F s ’ 2 . 0 1 a c i t a m e h t a M g n i s u d e t c e t e d e w t a h t s e i t i n u m m o c e h t e t a c i d n i s d u o l c y a r g e h T . ) ” P “ ( , ) o r p + w e n + d l o , t h g i r m o t t o b ( ” s k r o w t e n f o y t i u q i b U “ , ) o r p + w e n + d l o , t h g i r p o t ( ” s k r o w t e n f o s n o i t i n fi e D “ , ) w e n + d l o , t f e l p o t ( ” y g o l o n m r e t i . 7 d n a 6 . s g i F n i ” r e h t O “ d e l e b a l y t i n u m m o c e h t n o i t i t r a p e W . ” r e h t O “ d n a , ) o r p + w e n + d l o , t f e l m o t t o b ( ” s n o i t a c i l p p a l a i c o S “ 11 Basic terminology Definitions of networks Social applications Ubiquity of networks Other O3 People should understand basic terminology in a network such as nodes, edges and hubs. N4 Networks consist of nodes and links. N8 Networks allow us to analyze human interactions. O5 Networks can be applied to various fields. N3 Examples include economics/ social/political sciences. P5 Network structure can influence behavior and vice versa. P1 Networks describe how things connect and interact. O2 A connection between more than one person can form a link. N1 Networks are connections and interactions. O1 Networks are everywhere. P6 Networks help us understand similarities among systems found in our everyday life. N2 Networks are present in every aspect of life. N11 Networks connects various aspects of the world. P2 The world around us can be represented as networks of interconnecting parts. n o i t i t r a p r e h t r u f e W . ) w e n + d l o ( ” n o i t a z i l a u s i V “ t u o b a y t i n u m m o c r e h t o n a d n fi e W . 5 . g i F n i y t i n u m m o c ” r e h t O “ e h t f o n o i t i s o p m o c e D : 6 e r u g i F . 7 . g i F n i y t i n u m m o c ” r e h t O “ g n i n i a m e r e h t 12 Other Visualization N7 Visualization of networks helps understanding. O4 Visualizations are the key to understanding the network. w e n , p o t ( ” n o i t a t n e s e r p e r d e s a b - r e t u p m o C “ : s e i t i n u m m o c e e r h t d n fi e W . 6 . g i F n i y t i n u m m o c ” r e h t O “ g n i n i a m e r e h t f o n w o d k a e r b l a n F i : 7 e r u g i F . ) d l o + o r p + w e n , t f e l m o t t o b ( i ” g n d n fi n r e t t a P “ d n a , ) w e n , t h g i r m o t t o b ( ” n o i t a c u d E “ , ) o r p + 13 P4 Modeling systems as networks can help reveal and explain patterns and general principles. Computer-based representation Education N5 Computers are often used to study networks. P7 Networks can be represented and studied in many different ways. N10 Future of network science can include preinstalled programs and networks in media. N9 Network science should be incorporated into schools. Pattern finding N6 Networks can be used for making predictions. P3 Representing networks allows them to be used as a tool. O6 Life is hard... then you die. Table 5: Final version of the Network Literacy: Essential Concepts. No. Concept 1. 2. 3. 4. 5. 6. 7. Networks are everywhere. Networks describe how things connect and interact. Networks can help reveal patterns. Visualizations can help provide an understanding of networks. Today’s computer technology allows you to study real-world networks. Networks help you to compare a wide variety of systems. The structure of a network can influence its state and vice versa. of networks for interdisciplinary comparison of systems and the interactions between structures and temporal dynamics of networks. These concepts were present in the list of concepts produced by the professionals, but they were misunderstood and misrepresented in the concept network when the students connected concepts together. This is possibly because these concepts are among the most difficult ones about networks for non-professionals to grasp. We decided to add these concepts to the final list because of their importance. This yielded a list of seven essential concepts. Incorporating feedback from the network-science commu- nity then resulted in the final version of the seven essential concepts about networks (see Table 5). Naturally, each of the seven concepts subsumes several more detailed ideas. We thus developed in-depth descriptions of each concept as a list of “core ideas”. We also circulated these core ideas to the network- science community and incorporated their feedback. After numerous iterations of online and offline edits and reviews of the contents, the current version of “Network Literacy: Essential Concepts and Core Ideas” was finalized on 4 December 2014, approximately six months after the initial pre-conference event at NetSci 2014. We include the text of the final version in the Supplemental Material. The final text of “Network Literacy: Essential Concepts and Core Ideas” was developed into a profes- sionally designed, printable booklet for use by teachers, students, and other learners (see the top panel of Fig. 8). It was first published online for free download at http://tinyurl.com/networkliteracy on 12 March 2015. Hard copies of the booklet were also disseminated at CompleNet 2015 on 25 March 2015 and at NetSci 2015 on 1 June 2015. Several very positive responses followed these initial disseminations. An important development, facilitated through social media and personal contacts, has been the translation of the booklet into non-English languages by volunteers. As of 22 September 2015, seven translated versions (Dutch, German, Italian, Japanese, Korean, Persian, and Spanish) are available (see the bottom panel of Fig. 8), and several other translations are in progress. 4 Conclusions In this paper, we reported our recent initiative of developing an educational resource that concisely sum- marizes essential concepts about networks in an easily accessible format. The result, “Network Literacy: Essential Concepts and Core Ideas”, is the product of a collaborative effort that involved a large number of network-science researchers, educators, school teachers, and students. Since its initial release, we have received many very positive responses from across the globe. We hope that this initiative will help inspire the network-science community to make a large societal impact by launching more educational outreach activities that spread the concepts and ideas about networks into various formal and informal learning settings—with a particular emphasis on the next generation, who will live their whole lives in a networked world. An important lesson from this process is that the deep involvement of students is crucial. The outcomes of the brainstorming sessions revealed that network-science researchers and educators tended to consider abstract concepts as “essential” but placed less emphasis on specific examples or research methods (e.g., visualization and computational tools). However, the students who were exposed to network science through the NetSci High program frequently mentioned that concrete examples of networks and the specific research methods that they learned were among the most essential components of their experience. We believe that such a down-to-earth, hands-on viewpoint is extremely important when communicating the value of network science (i.e., “the study of connectivity”) to the public. Although theoretical abstraction of systems is crucial for the modeling and analysis of networks, the general public needs—and wants!—to know what networks are 14 Figure 8: “Network Literacy: Essential Concepts and Core Ideas” booklets. Top: Original (English) version. Bottom: Non-English versions translated by volunteers. (The Japanese version was produced by Yoshi Fujiwara, Toshihiro Tanizawa, and Hiroki Sayama; the Dutch version was produced by Paul van der Cingel; the Italian version was produced by by Paolo Tieri; the Korean version was produced by Sang Hoon Lee and Mi Jin Lee; the Spanish version was produced by Rosa Benito; the German version was produced by Andreas Joseph and Florian Klimm; and the Persian version was produced by Taha Yasseri.) The graphical design of the original English version was created by Eri Yamamoto; the design of the Korean version was created by Mi Jin Lee; the design of the Japanese version was created by Hiroki Sayama; the design of the Persian version was created by Taha Yasseri based on the English and Japanese versions; and the other four versions adopted the design template of the Japanese one. All of these versions are available at http://tinyurl.com/networkliteracy. Translations into other languages are in progress. 15 and how to use them. We are confident that our final booklet will play a large role in successfully addressing this issue, and the booklet has benefited greatly from the significant contributions made by high-school students. The above successes notwithstanding, our “Network Literacy” initiative is limited in some important respects. First, the booklet that we produced introduces concepts and ideas only. It does not provide lesson plans (for teachers) or further study guides (for learners). Developing lesson plans in particular will require substantially more effort and resources, as it will need to include details of class instructions, which needs to sufficiently match local school curricular structures, contents, and instruction methods (and these are rather heterogeneous) to facilitate adoption in classrooms. Second, the current list of essential concepts may not accurately reflect the scientific importance of various aspects of network science. For example, a major portion of recent discoveries in the study of networks have focused on the dynamical nature of networks and/or on various structural intricacies (e.g., synchronization, diffusion and contagion [19, 17], adaptive networks [9], temporal networks [11], and multiplex and other multilayer networks [13, 3]), but this topical area was represented only briefly in concept 7 (and it was not present in the students’ input). How to accessibly convey the richness of complex network dynamics to the general public remains an important problem to address. We are currently seeking additional resources to make further progress on these frontiers of network science and education, and we invite the entire network-science community to join this challenging yet highly meaningful and societally important endeavor. Acknowledgements The Network Literacy initiative would not have been possible without the participation, support, and en- couragement of the US Army Research Office, Albert-L´aszl´o Barab´asi, Raissa D’Souza, Sarah Schoedinger, H. Eugene Stanley, Craig Strang, the Network Science Society, NetSci High students and teachers, the University of California at Berkeley, and all of the members of the network-science community who have contributed to and supported this effort. Specifically, the following individuals have made important con- tributions to the contents of Network Literacy: Alvar Agusti, Chris Arney, Robert F. Chen, Arthur Hjorth, Khaldoun Khashanah, Yasamin Khorramzadeh, Erik Laby, Toshi Tanizawa, Paolo Tieri, Brooke Foucault Welles, Robin Wilkins, and Eri Yamamoto. Translations of “Network Literacy: Essential Concepts and Core Ideas” were made possible thanks to the following volunteers: Paul van der Cingel (Dutch), Yoshi Fujiwara (Japanese), Toshihiro Tanizawa (Japanese), Paolo Tieri (Italian), Mi Jin Lee (Korean), Sang Hoon Lee (Korean), Rosa Benito (Spanish), Andreas Joseph (German), Florian Klimm (German), Taha Yasseri (Persian), Ralucca Gera (Romanian, in progress), and Peter Polner (Hungarian, in progress). The authors also thank Katherine Culp, Amanda Jaksha and Matty Lau for their valuable feedback on Fig. 1. The material in this paper is based on work supported in part by the US National Science Foundation under Grant No. 1027752 and 1139478. References [1] Jennifer Anderson and Anne LaVigne. Systems in Motion: Exploring Complexity through an Interdis- ciplinary Lens. Creative Learning Exchange, 2014. [2] Albert-L´aszl´o Barab´asi. Linked: How Everything Is Connected to Everything Else and What It Means for Business, Science, and Everyday Life (new edition). Basic Books, 2014. [3] Stefano Boccaletti, G Bianconi, R Criado, Charo I Del Genio, J G´omez-Gardenes, M Romance, I Sendina-Nadal, Z Wang, and M Zanin. The structure and dynamics of multilayer networks. Physics Reports, 544(1):1–122, 2014. [4] Chris J Budd. Chris Budd’s Home Page. http://people.bath.ac.uk/mascjb/, Accessed on 9 July 2015. 16 [5] Francesca Cava, Sarah Schoedinger, Craig Strang, and Peter Tuddenham. Science Content and Stan- dards for Ocean Literacy: A Report on Ocean Literacy. College of Exploration Berkeley, 2005. [6] Council of Europe. Internet Literacy Handbook. http://www.coe.int/t/dghl/StandardSetting/ InternetLiteracy/hbk_en.asp, Accessed on 9 July 2015. [7] Catherine Cramer, Lori Sheetz, Hiroki Sayama, Paul Trunfio, H Eugene Stanley, and Stephen Uzzo. NetSci High: Bringing network science research to high schools. In Complex Networks VI: Proceedings of the Sixth Workshop on Complex Networks (CompleNet 2015), volume 597 of Studies in Computational Intelligence, pages 209–218. Springer, 2015. [8] Jay Wright Forrester. Industrial dynamics. Journal of the Operational Research Society, 48(10):1037– 1041, 1997. [9] Thilo Gross and Hiroki Sayama. Adaptive Networks. Springer, 2009. [10] Heather A Harrington, Mariano Beguerisse-D´ıaz, M Rombach, Laura M Keating, and Mason A Porter. Commentary: Teach network science to teenagers. Network Science, 1(02):226–247, 2013. [11] Petter Holme and Jari Saramaki. Temporal networks. Physics Reports, 519(3):97–125, 2012. [12] G Jenkinson, J Goutsias, D Jenkinson, and S Frennesson. Teach Engineering: Curricular Unit “It’s a Connected World: The Beauty of Network Science”. Accessed on 9 July 2015. [13] Mikko Kivela, Alex Arenas, Marc Barthelemy, James P Gleeson, Yamir Moreno, and Mason A Porter. Multilayer networks. Journal of Complex Networks, 2(3):203–271, 2014. [14] Espen Knoop et al. Tangible Networks. http://www.tangiblenetworks.net/, Accessed on 9 July 2015. [15] E Meeks and M Krishnan. An Interactive Introduction to Network Analysis and Representation. http: //dhs.stanford.edu/dh/networks/, Accessed on 9 July 2015. [16] Mark E. J. Newman. Networks: An Introduction. Oxford University Press, 2010. [17] Romualdo Pastor-Satorras, Claudio Castellano, Piet Van Mieghem, and Alessandro Vespignani. Epi- demic processes in complex networks. Reviews of Modern Physics, 87:925–979, 2015. [18] M. A. Porter, J.-P. Onnela, and P. J. Mucha. Communities in networks. Not. Am. Math. Soc., 56(9):1082–1097, 1164–1166, 2009. [19] Mason A. Porter and James P. Gleeson. Dynamical systems on networks: A tutorial. Frontiers in Applied Dynamical Systems: Reviews and Tutorials, in press (arXiv:1403.7663v2), 2015. [20] Jeff Potash and John Heinbokel. Dollars and Sense — Stay in the Black: Saving and Spending. Creative Learning Exchange, 2011. [21] Rob Quaden, Alan Ticotsky, and Debra Lyneis. The Shape of Change. Creative Learning Exchange, 2005. [22] Angel S´anchez and Cristina Brandle. More network science for teenagers. arXiv preprint arXiv:1403.3618, 2014. [23] Sarah Schoedinger, Francesca Cava, and Beth Jewell. The need for ocean literacy in the classroom: Part i. The Science Teacher, page 44, 2006. [24] Steve Stemler. An overview of content analysis. Practical Assessment, Research & Evaluation, 7(17):137– 146, 2001. [25] John D Sterman. Business Dynamics: Systems Thinking and Modeling for a Complex World. Irwin/McGraw-Hill, 2000. 17 [26] Craig Strang, A DeCharon, and S Schoedinger. Can you be science literate without being ocean literate? Current: The Journal of Marine Education, 23(1):7–9, 2007. [27] Virginia Tech Virginia Bioinformatics Institute. Virus Tracker. https://virustracker.vbi.vt.edu/, Accessed on 9 July 2015. [28] Robert Philip Weber. Basic Content Analysis. Sage, 1990. [29] Wikipedia. Digital literacy. https://en.wikipedia.org/wiki/Digital_literacy, Accessed on 9 July 2015. 18 Supplemental Material: Final Text of Network Literacy: Essential Concepts and Core Ideas 1: Networks are everywhere. • The concept of networks is broad and general, and it describes how things are connected to each other. Networks are present in every aspect of life. • There are networks that form the technical infrastructure of our society—e.g., communication systems, semantic systems, the Internet, electrical grids, the water supply, etc. • There are networks of people—e.g., families and friends, e-mail/text exchanges, Facebook/Twitter/Instagram, professional groups, etc. • There are economic networks—e.g., networks of products, financial transactions, corporate partner- ships, international trades, etc. • There are biological and ecological networks—e.g., food webs, gene/protein interactions, neuronal networks, pathways of disease spreading, etc. • There are cultural networks—e.g., language/literature/art connected by their similarities, historical events linked by causal chains, religions connected by their shared roots, people connected to events, etc. • Networks can exist at various spatial and/or temporal scales. 2: Networks describe how things connect and interact. • There is a subfield of mathematics that applies to networks. It is called graph theory. Many networks can be represented mathematically as graphs. • Connections are called links, edges, or ties. The entities that are connected to each other are called nodes, vertices, or actors. • Connections can be undirected (symmetric) or directed (asymmetric). They can also indicate ties of different strengths, and can indicate either positive or negative relationships. • The number of connections of a node is called the degree of that node. • Many networks have more than one type of connection—e.g., offline friendships and Facebook connec- tions, different modes of transportation, etc. • A sequence of edges that leads from one node, through other nodes, to another node is called a path. • A group of nodes within which a path exists from any one entity to any other entity is called a connected component. Some networks have multiple connected components that are isolated from each other. • Some networks are studied using mathematical structures that are more complicated than graphs. 3: Networks can help reveal patterns. • You can represent something as a network by describing its parts and how they are connected to each other. Such a network representation is a very powerful way to study a system’s properties. • The properties in a network that you can study include: – how the degrees are distributed across nodes – which parts or connections are most important – strengths and/or weaknesses of the network 19 – if there is any sub-structure or hierarchy – how many steps, on average, are needed to move from one node to another in the network • In some networks, you can find a small number of nodes that have much larger degrees than others. They are often called hubs. • In some networks, you can find a group of nodes that are better connected to each other than chance would dictate. They are sometimes called clusters or communities. Some of them can occupy a central, or core, part of a network. • Using these findings, you can sometimes infer how a network was formed and/or make predictions about dynamical processes on the network or about its future structure. 4: Visualizations can help provide an understanding of networks. • Networks can be visualized in many different ways. • You can draw a diagram of a network by connecting nodes to each other using edges. • There are a variety of tools available for visualizing networks. • Visualization of a network often helps to understand it and communicate ideas about connectivity in an intuitive, non-technical way. • Creative information design plays a very important role in making an effective visualization. • It is important to be careful when interpreting and evaluating visualizations, because they typically do not tell the whole story about networks. 5: Today’s computer technology allows you to study real-world networks. • Computer technology has dramatically enhanced the ability to study networks, and this is especially important for large ones with rich structure. • There are many free software tools available for network visualization and analysis. • Using personal computers, everyone (not just scientists) can construct, visualize, and analyze networks. • Through the Internet, everyone has access to many interesting network data sets. • Computers allow you to simulate hypothetical or virtual networks, as well as to simulate dynamical processes on both real and hypothetical networks. • Learning computer literacy skills opens the door to myriad possibilities for a career. These include scientist, data analyst, software engineer, educator, web developer, media creator, and many others. 6: Networks help you to compare a wide variety of systems. • Various kinds of systems, once represented as networks, can be compared to examine their similarities and differences. • Certain network properties commonly appear in many seemingly unrelated systems. This implies that there exist some general principles about connectivity that apply to multiple domains. • Other network properties are different in different systems. These properties can help to classify networks in different families and to gain insight into why they are different. • Science is typically conducted in separate areas of research called disciplines. Networks can help to cross disciplinary boundaries to achieve a holistic and more complete understanding of the world. • Networks can assist in the transfer of knowledge across different areas of study. 20 7: The structure of a network can influence its state and vice versa. • Network structure indicates how parts are connected in a network. • Network state indicates the properties of a network’s nodes and edges. • Network structure and state can each change over time. • The time scales on which network structure and state co-evolve can be either similar or different. • Network structure can influence changes of network state. Examples include the spread of diseases, behaviors, or memes in a social network, and traffic patterns on the road network in a city. • Network state can influence changes of network structure. Examples include the creation of new “following” edges in social media and the construction of new roads to address traffic jams. 21
1110.5057	1	1110	2011-10-23T15:48:25	Patterns of Emotional Blogging and Emergence of Communities: Agent-Based Model on Bipartite Networks	[ "cs.SI", "cs.HC", "physics.soc-ph" ]	Background: We study mechanisms underlying the collective emotional behavior of Bloggers by using the agent-based modeling and the parameters inferred from the related empirical data. Methodology/Principal Findings: A bipartite network of emotional agents and posts evolves through the addition of agents and their actions on posts. The emotion state of an agent,quantified by the arousal and the valence, fluctuates in time due to events on the connected posts, and in the moments of agent's action it is transferred to a selected post. We claim that the indirect communication of the emotion in the model rules, combined with the action-delay time and the circadian rhythm extracted from the empirical data, can explain the genesis of emotional bursts by users on popular Blogs and similar Web portals. The model also identifies the parameters and how they influence the course of the dynamics. Conclusions: The collective behavior is here recognized by the emergence of communities on the network and the fractal time-series of their emotional comments, powered by the negative emotion (critique). The evolving agents communities leave characteristic patterns of the activity in the phase space of the arousal--valence variables, where each segment represents a common emotion described in psychology.	cs.SI	cs	Patterns of Emotional Blogging and Emergence of Communities: Agent-Based Model on Bipartite Networks Department of theoretical physics; Jozef Stefan Institute; Box 3000 SI-1001 Ljubljana Slovenia Marija Mitrovi´c and Bosiljka Tadi´c1 Abstract: Background: We study mechanisms underlying the collective emotional behavior of Bloggers by using the agent- based modeling and the parameters inferred from the related empirical data. Methodology/Principal Findings: A bipartite network of emotional agents and posts evolves through the addition of agents and their actions on posts. The emotion state of an agent, quantified by the arousal and the valence, fluctuates in time due to events on the connected posts, and in the moments of agent's action it is transferred to a selected post. We claim that the indirect communication of the emotion in the model rules, combined with the action-delay time and the circadian rhythm extracted from the empirical data, can explain the genesis of emotional bursts by users on popular Blogs and similar Web portals. The model also identifies the parameters and how they influence the course of the dynamics. Conclusions: The collective behavior is here recognized by the emergence of communities on the network and the fractal time-series of their emotional comments, powered by the negative emotion (critique). The evolving agents communities leave characteristic patterns of the activity in the phase space of the arousal–valence variables, where each segment represents a common emotion described in psychology. 1 1 0 2 t c O 3 2 ] I S . s c [ 1 v 7 5 0 5 . 0 1 1 1 : v i X r a 1e-mail: [email protected] Contents 1 Introduction 2 Materials and Methods 2.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Dynamics: Model of Emotional Bloggers on Evolving Bipartite Network . . . . . . . . . . . . . 2.2.1 Emotional states of individual agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Agents interaction and network evolution: Rules & Implementation . . . . . . . . . . . . 2.2.3 Control parameters: Definitions and inference from the empirical data . . . . . . . . . . . 3 Results 3.1 Time-series of the emotional comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Emergent network and communities of the emotional agents . . . . . . . . . . . . . . . . . . . . 4 Discussion 2 3 3 4 5 6 7 8 8 9 11 4.1 Patterns of agents behavior in time and space of the emotion variables . . . . . . . . . . . . . . . 11 4.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 5 Figure Legends and Supporting Information 14 1 1 Introduction The Internet experience in recent years has revolutionized the mechanisms that an individual can exploit to par- ticipate in global social dynamics. Consequently, new techno-social phenomena emerge on the Web [1, 2, 3, 4], boosting an intensive multidisciplinary research. In technology research, for example, new generation of services are developing in the direction to integrate human capabilities in a service-oriented manner [5]. Behavior of the users in the virtual world has impact on real-life events, which becomes a concern of both social sciences and every day's practice. On the other hand, the data collected from massive use of the Web provide the basis to study human behavior "experimentally" at unprecedented scale. For instance, from the high-resolution data stored at various Web portals (social networks, Blogs, forums, chat-rooms, computer games, etc) information related to user preferences, patterns of behavior, attitudes, and emotions can be inferred for each individual user and user communities gathered around certain popular subjects [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]. Physics of complex systems and, in particular, the statistical physics of social dynamics, are focused on the dynamical processes in which human collective behaviors emerge from large number of individual actions [4, 6, 7, 9]. Combining the concepts of statistical physics with the machine-learning methods for the emotion detection in texts of messages [13, 17], we have recently performed analysis of large datasets from bbcblog.com and digg.com and determined quantitative measures of the collective behaviors in which the emotions are involved [14, 15]. Complementary to our work in Refs. [14, 15] where the empirical data are analyzed to extract various complex- systems properties, the present work is a theoretical study of the processes, underlying the emergence of the collective emotional behavior of Blog users, within the framework of agent-based modeling. The quantitative analysis of users collective behavior in the empirical data from diggs.com and bbcblog.com in Refs. [14, 15] has been enabled by mapping the high-resolution data onto bipartite networks of users and posts, as two natural partitions. The idea of bipartite networks makes the "firm ground" also in the present theoretical model, where the agents interact indirectly over the posts. We also make use of several other features, observed in various empirical data, that are relevant for designing the dynamic rules of the theoretical model: • Universality of user's behavior related with the action-delay and the circadian cycles [9, 6, 10, 7]; • User communities occurring in the cyberspace are reminiscent to the ones in real life, however, different time scales and grouping mechanisms might be involved [10, 14, 18, 15]; • Quantitative measures of emotions have been introduced in psychology research [19]. In particular, based on Russell's multidimensional model of affect [20], each known emotion can be represented by a set of nu- merical values in the corresponding multidimensional space. Two fundamental components of emotion, to which we refer in this work, are the arousal, related to reactivity to a stimulation, and the valence, measur- ing intrinsic attractiveness or aversiveness to a stimulation. These components of emotion can be measured in laboratory based on the related psychophysiological and neurological activity [21, 22]. Moreover, a systematic association has been recognized [23] between individual emotional characteristics and word use. The arousal and valence components of an emotion can be retrieved from a written text by suitable machine-learning methods, which are being developed for a specific type of data [22, 24, 25]. Systematic analysis of the patterns of user behaviors and the emotion contents in the texts of comments in the empirical data from popular Blogs [14], discussion-driven Diggs [15], and Forums [26], suggests that negative emotions (critique) drive the activity on these Web portals. However, the mechanisms working behind this global picture have not been well understood. In order to elucidate the role of emotions in the blogging interactions, and to point out potential parameters and levels where the process can be controlled, we devise an agent-based model. The agents are spreading their emotions in a bipartite network environment. The agent's properties, the rules and the parameters of the model are closely related with the empirical data from Blogs and Diggs. Agent-based modeling [27, 28, 4], where different properties of agents influence their actions, provides suitable theoretical framework for numerical simulations of social phenomena. Recently a model for product-review with the emotional agents in a mean-field environment has been introduced [29], with the agents emotional states described by two state variables (ai,vi). These variables correspond to the psychological values of the arousal and the valence, respectively, in view of the Russell's two dimensional circumples model [20, 30, 19]. 2 Building on the Ref. [29], here we study an agent-based-model to explore the emergence of user communities on Blogs, where the emotional contents are communicated indirectly via comments that they leave on the posts. For this purpose, our emotional agents are situated on a weighted bipartite network consisting of users (agents) and posts, where by definition no direct link between the nodes of the same partition is allowed. The weighted links between the nodes of different partitions represent the number of comments of an agent to the linked post. Motivated by the realistic situation on Blogs, in our model the network itself evolves over time due to the arrival of new users and the addition of new posts, and due to user's actions on previous posts of their preference. The emotional state, measured by the arousal and the valence variables, which are attached to each agent-node, is influenced by time-varying fields on the posts surrounding that agent on the evolving network. As in real-life, the elevated arousal may induce an action of the emotional agent on a post, according to the rules introduced below in section 2.2. In the moment of action on a post, the agent's current emotion arousal and valence components are transferred to the comment that the agent leaves on that post, where it can be experienced by other agents. Thus the fields themselves evolve with the network evolution and differ for each agent, depending on its position on the network. In order to have realistic dynamics, we design the rules of actions that are motivated by systematic observations of the activity patterns in the empirical data at Blogs and Diggs, as described in section 2.1 and in the supportive information. Moreover, the set of parameters that control the dynamics of our model are inferred from the empirical data of popular discussion-driven Diggs, as explained below. 2 Materials and Methods 2.1 Datasets In this work we use large datasets2 related to popular posts, from which we (i) study the temporal patterns of events, that motivate the dynamic rules of our model, and (ii) extract realistic values of certain control parameters of the model. As it will be clear below, the present analysis aims for different features of these empirical datasets, yielding several new results in comparison with the ones presented in our previous study [14, 15]. The datasets that we use are collected from bbcblog.com and diggs.com [14, 15] and have high temporal resolution, information about identity of each user and of each post, as a unique ID, and precise relationship between users and comments-on-posts, as well as full text of all posts and comments. The subsets of data related with popular posts as posts with more than 100 comments are selected together with all users linked to them, as good candidates for the analysis of collective behavior of users [14, 15]. In addition, diggs.com data contain information about comment-on-comment. Thus, we select the subset of popular posts, termed discussion-driven Diggs (ddDiggs), on which more than 50% of comments represent reply to the comments of other users. This data consists of NP = 3984 discussion-driven Digg stories, on which NC = 917708 comments are written by NU = 82201 users [15]. In addition, texts of posts and comments are classified by machine-learning methods with the emotion classifier designed in Ref. [13] and trained at Blog-type of texts. Using the emotion classifier the texts are designated as carrying either positive or negative emotion valence, or otherwise are neutral [14]. From the discussion-driven Diggs dataset here we analyze the temporal patterns of activity related to both users and posts. Parts of these patterns are shown in Figs. 1a,b. Each user (post) occurring in that dataset is given a unique index, plotted along vertical axis, sorted by the time of first appearance in the dataset. For each user index, points along time axis indicate the times when an activity of that user occurred to anyone of the posts. Analogously, the points on the posts pattern indicate the times when an activity occurred at that post by anyone of the users. In the post-activity pattern, shown in Fig. 1b, dense points in a narrow time window following the post appearance time indicate an intensive activity at that post. This might be related with certain exposure of the posts to users during that time period. The width of the exposure time window, T0, will be recognized as a relevant parameter in the dynamics. Whereas, different type of the dynamics beyond the exposure window is manifested in systematically reduced activity until eventually the post ceases to be active (expires). The situation is entirely different when looked from the point of view of the users. The patterns of activity of every user over time is shown in Fig. 1a. The user indexes are ordered by the time of their first appearance in 2Data accessible on http://www.cyberemotions.eu/data.html under terms and conditions of the CYBEREMOTIONS project. 3 the dataset, hence the top boundary of the plot indicates the appearance of new users, relative to the beginning of the dataset. The profile of the top boundary shows that new users arrive in "waves". Moreover, the arrivals of new users boost the activity of previous users, which is manifested in the increased density of points in depth of the plot below each "wave". This feature of the dynamics is utilized for designing the model rules in section 2.2. Further quantitative measures of the temporal patterns in Figs. 1a,b are given in relation to the model parameters in section 2.2.3. The number of newly arrived users (with respect to the beginning of the dataset) in suitably defined time bin can be readily extracted from the dataset. The time series of new users p(t) per time bin (in units of tbin = 5 minutes), inferred from the ddDiggs dataset is shown by the red line in Figure 2. It exhibits characteristic daily cycles superimposed on the fractal fluctuations with long-range correlations. Note that these cycles are related with the occurrence of "waves" in the activity patterns in Fig 1a. The signal has the power-spectrum of the type S(n ) ∼ 1/n f , with f ≈ 1.5, shown in top panel in Fig. 2. The time series of the number of all active users per time bin, extracted from the same dataset is shown in the same Figure 2 by the green line. It has a similar fractal structure and the power-spectrum of 1/n -type. The power spectrum is correlated over the range of frequencies, which correspond to times larger than 2 hours in the time domain. Further analysis of this dataset, which is relevant for this work, is given in respect to extracting the control parameters of the dynamics in section 2.2.3. Detailed analysis of the emotion contents of the comments and the related time series of the emotional comments can be found in Ref. [15]. In the supporting information, Figure S1 shows the time-series of the number of emotional comments with positive (negative) valence from the same dataset which is analyzed here, and the excess of the negative emotions. 2.2 Dynamics: Model of Emotional Bloggers on Evolving Bipartite Network In the spirit of agent-based modeling, the agents (representing users on Blogs) are given certain properties that may affect their actions, i.e., the dynamic rules of the model. Conversely, these agent's properties are changed due to dynamic interactions between them, which imply further changed actions, and so on. To describe emotional actions of users on Blogs, we adapt the agents whose emotion states are described by two variables-arousal and valence, first introduced in Ref. [29]. Apart from the emotion variables, the agents in our model have additional properties indicated below in Eq. (1), and are subjected to the dynamically active networked environment, which affects their actions. Moreover, thy are designed for multiple actions in course of the dynamics, thus contributing to the emergence of collective emotional behavior. As mentioned above, here the emotional agents are adapted to interact indirectly via posts on a bipartite network of the agents and posts. Thus the essential elements of our model are: • 2-dimensional local maps, describing the emotion variables arousal and valence {ai(t),vi(t)} of each agent; • Interaction environment, represented by an evolving bipartite network of agents and posts, through which the agents emotion is spreading; • Driving noise, applied to systematically perturb the system boosting its internal dynamics. In our model the system is driven by adding new users, according to the time-series p(t) of new users, which is inferred from the empirical data of ddDiggs, as explained above. No other type of driving is considered in this work. It should be stressed that the bipartite network type, with the users (agents) and the posts as two partitions and the weighted links between them representing the number of comments of the user to the post, is necessary in order to take into account the fundamental nature of the dynamics on Blogs: users, as the nodes of the same partition, are never connected directly, but only through the posts. Technically, both types of nodes appear as the objects at the same level. Therefore, we actually have two types of "agents", user and post agent, with different properties indicated here: U[g;vi(t),ai(t);Listsi...,D t] ; P[tP; < vp(t) >, < ap(t) >;Listsp...] . (1) The dynamical variables arousal and valence ai(t) and vi(t) are properties of the user nodes, which vary in time as explained in detail below. In analogy to real Blogs, at the moments of actions the emotion (arousal and valence) 4 of the user are transferred to the post (precisely to the comment that the user puts on the post), and thus contribute to the current overall emotional content on that post, < vp(t) >, < ap(t) >. Both user and post objects have individual Lists of connections on the evolving bipartite network, which are updated through the user actions, as explained below in section 2.2.2. Additional properties that may affect the dynamics in our model are the post life-time, tP, and the user action-delay, D t, as well as the user probability for posting a new post, g. These properties can be inferred from the considered dataset, as done below in section 2.2.3. Beside the dynamic rules, which are motivated by the blogging processes, other gross features of our model are different from previous work on the emotional agents in Ref. [29]. The relevant differences are due to the role of networked environment and the endogenous driving, without any external noise. Moreover, at the level of individual agents, apart form the two emotion variables which are given by the same type of the nonlinear maps Eqs. (2-3) as in Ref. [29], the agents in our model have additional properties which affect their actions. These are the inclination towards posting new posts and the action delay, measured by the quantities g and D t, respectively, as well as the lists of connections (posts) on the network, which are unique for each agent. Hence the environmental fields in Eqs. (2-3) are practically different for each agent on the network. In this way the network environment induces (and keeps track of) the heterogeneity among the agents in a natural way, through the lists of posts to which they were connected in the course of their actions. Having these general remarks, we introduce details of the model in the remaining part of this section. We first explain the dynamic rules of the model and define all the parameters that control the dynamics. 2.2.1 Emotional states of individual agents Following Ref. [29], we assume that the individual emotional state (arousal and valence) of each agent can be described by two nonlinear equations, which are subject of the environmental fields. For our system on bipartite networks, the arousal and the valence are associated with each user-node(!) and their values, kept in the intervals ai(t) ∈ [0,1] and vi(t) ∈ [−1,1], are updated according to the following nonlinear maps: m f (t)](d1 + d2(ai(t)− ai(t)2))(1− ai(t)) i (t) + qha (2) if D ti < 1 otherwise and (1− g a)ai(t) ai(t + 1) =(cid:26) (1− g a)ai(t) + [ha vi(t + 1) =(cid:26) (1− g v)vi(t) + [hv (1− g v)vi(t) i (t) + qhv m f (t)](t)(c1 + c2(vi(t)− vi(t)3))(1−vi) if D ti < 1 otherwise (3) i (t) and hv m f (t) and hv i (t), and the mean fields ha m f (t). Note that the local fields ha where i = 1,2···NU (t) indicates the index of user node and t-the time bin. The coefficients d1,d2 and c1,c2 characterize the maps themselves, while the network environment effects appear through two types of fileds: the local fields ha i (t) vary not only in time but also from user to user, depending on their connections on the network, and due to evolution of the network itself (see details below). Whereas, the mean fields ha m f (t) may act on a larger number of users, while also fluctuating in time. In our model they steam from a currently active posts and, thus, may be seen by all users who are attached to these posts. The mean fields indicate how the overall activity moves through posts, as a kind of "atmosphere" at the Blogsite. The contribution from the mean fields in our model is taken with a fraction 0 ≤ q ≤ 1, which is varied as a free parameter in Eqs. (2) and (3), and it is added to the contributions from the local fields. A user-node receives the stochastic inputs from the network in certain instants of time, when the events occur in its network surrounding, and reacts to them with a delay. The delay time D t of each user is counted continuously. When the delay time is smaller than the computational time bin D t < 1 tbin, the user is prompted for the update of its arousal and valence according to the full expressions indicated by top lines in the Eqs. (2) and (3). Otherwise, the arousal and valence values are only relaxing, with the rates g a = g v ≡ g . m f (t) and hv i (t) and hv For understanding the dynamics at local level, let us consider for a moment the Eqs. (2) and (3) as two-dimensional nonlinear maps of an isolated node subjected to a given constant field. In Fig. 3 we show the situation for several values of the field pre-factors, while keeping all other parameters fixed to the values which are used later in the simulations. Depending on the parameters, the maps can reach different fixed points. In particular, the arousal map always leads to an attractive fixed point, the position of which depends on the strength of the field-larger arousal is reached when the field is stronger, cf. Fig. 3a. In the case of valence, two fixed points can be reached, one at the positive valence, when the fields are positive (upper branch), and the other one in the area of negative 5 valence, which is attractive when the fields are negative (lower branches in the Figure 3b). In general, when the nonlinear maps are coupled on a network, the network environment affects each individual map through a feedback loop, causing synchronization [31, 32] or other self-organization effects [33] among the nodes. In our case the network affects the fields, which thus fluctuate at every time step and depending on the node's particular position on that network. The dynamics of the fields can thus be visualized in the local map as jumping of the trajectories vi(t),ai(t) from one branch of the map to another branch, and consequently, being attracted to another area of the phase space. As will be explained in detail below in section 2.2.2, the nonlinear mapping takes part only when the agent is prompted to act. Meanwhile, the maps are only relaxing towards the origin. 2.2.2 Agents interaction and network evolution: Rules & Implementation m f (t) in Eq. (2), which affect user i arousal at step t + 1, are determined from the posts in i (t) and ha The fields ha the currently active part of the network, C (t,t − 1), along the links of that user. Specifically, p∈C (t,t−1) aC p (t) p∈C (t,t−1) nC p (t) p∈C (t,t−1) AipaC p∈C (t,t−1) AipnC p (t)(1 + vi(t)vC p (t)(1 + vi(t)vC p (t)) p (t)) ha m f (t) = ha i (t) = ; , (4) p (t) and vC p (t) are the total arousal and the average valence of the post p calculated from the comments where aC in two preceding time steps, while nC p (t) is the number of all comments posted on it during that time period. Aip represents the matrix elements of the network, i.e., Aip > 0 if user i is connected with the active post p, while Aip = 0 if there is no link between them at the time when the fields are computed. Note that such links may appear later as the system evolves! In Eq. (4) the individual arousal fields ha i (t) is modified by (dis)similarity in user's p (t). actual valence, vi(t), and the valence of recent comments on the post, vC Regarding the valence fields in Eq. (3), we take into account contributions from the positive and the negative comments separately, while the neutral comments do not contribute to valence field. Depending on the current emotional state of the agent, positive and negative fields can lead to different effects [29], in particular, positive (negative) state will be influenced more with negative (positive) field, and vice versa. Here we assume that both components influence user valence, but with different strength according to the following expression: hv i (t) = 1− 0.4ri(t) 1.4 p∈C (t,t−1) AipN+ p (t) p∈C (t,t−1) AipNemo p (t) − 1 + 0.4ri(t) 1.4 p∈C (t,t−1) AipN−p (t) p∈C (t,t−1) AipNemo (t) p , (5) where the valence polarity of the user i is given by ri(t) = vi(t) , and N±p (t) is the number of positive/negative vi(t) comments written on post p in the period [t − 1,t]. The normalization factor Nemo (t) = N+ p (t) + N−p (t). The mean-field contributions to the valence steam from the entire set of currently active posts C (t,t − 1), and are independent on how users are linked to them: (t) is defined as Nemo p p However, the mean-field effects are perceived individually by each user, depending on the polarity ri(t) of user's current valence. The rules of agents interactions on the network are formulated in view of user behavior on real Blogs and Diggs and the observations from the quantitative analysis of the related empirical data. In particular, the dynamic rules are motivated by the temporal patterns in Figs. 1a,b and the time-series in Fig. 2, indicating how the number of active users arises in response to the arrival of new users. Moreover, additional features of the dynamics on ddDiggs, shown in supporting information, Figure S1 and Movie S2, suggest the dynamics with the dominance of negative emotions and with user's focus systematically shifting towards different posts. In the implementation of the model rules, we also make use of some general features of human dynamics, i.e., the occurrence of circadian cycles and delayed action to the events, mentioned in the Introduction, and assume that the arousal drives an action, as commonly accepted in the psychological literature. The rules are implemented in the C++ code as follows. The system is initialized with typically 10 Users who are connected to 10 Posts, to start the lists of the exposed and the active posts and the prompted and the active users. Then at each time step: 6 hv i,m f (t) = 1− 0.4ri(t) 1.4 p∈C (t,t−1) N+ p (t) p∈C (t,t−1) Nemo p (t) − 1 + 0.4ri(t) 1.4 p∈C (t,t−1) N−p (t) p∈C (t,t−1) Nemo (t) p . (6) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) • The system is driven by adding p(t) new users (Note the correspondence of one simulation step with one tbin = 5 min of real time); Their arousal and valence are given as uniform random values from ai ∈ [0,1], vi ∈ [−1, +1], then updated with the actual mean-field terms. By the first appearance each user is given a probability g ∈ P(g) to start a new post. The new users are then moved to the active user list; • The emotional states for all present users are relaxed with the rate g , according to the second row in the Eqs. (2) and (3); • The network area C (t,t − 1) of the active posts is identified as post on which an activity occurred in two preceding time steps; then the lists of active users is updated from the users linked to these posts, as follows: – Users linked to the active posts are considered as exposed to the posted material and decide when they will act on it, i.e., they are given new delay-time from the distribution P(D t); All users whose current delay time D t < 1tbin are prompted for update the emotional states according to the first rows in Eqs. (2) and (3), with their actual network fields computed from the Eqs. (4-6). An updated user is moved to the active user list with the probability a0ai(t) proportional to its current arousal, else it gets a new delay time D t); t ∈ P(D • Every active user: 0.5(1+vC (cid:229) p[0.5(1+vC p (t)vi(t))+Nc p(t) p (t)vi(t))+Nc p(t)] , depending on the number of comments on it Nc – adds a new post with the probability g or otherwise comment to one of the exposed posts, which are not older than T0 steps; Users are linked to posts preferentially with the probability pp(t) = p(t) and the valence similarity; – and with probability m comments a post which is older than T0 steps. The post is selected prefer- entially according to the negativity of the charge of all comments on it, with (properly normalized) probabilities p j,old(t) ∼ 0.5 +Q j(t), if the charge is negative, else p j,old(t) ∼ 0.5; Lifetimes of the posts are systematically monitored (already expired posts are not considered); – Current values of the valence and arousal of the user are transferred to the posted comment or the new post; User is given a new delay-time D t ∈ P(D t); New posts are given life-time tP ∈ P(tP). • Delay-time D t for all other users is decreased by one. Time-step closes with updating the lists of the exposed and the active posts, and the lists of the exposed and the prompted users. 2.2.3 Control parameters: Definitions and inference from the empirical data According to the above dynamic rules of the model, one can identify the parameters which control the dynamics at different levels. In particular, we use the following parameters, distributions or time-series which characterize, respectively: • the local maps: c1 = d1 = 1, c2 = 2.0, d2 = 0.5, g = 0.05; • the properties of posts and users: tP ∈ P(tP), D • the driving: {p(t)}, q = 0.4, a0 = 0.5. t ∈ P(D t), T0 = 2days, m (T0) = 0.05, g ∈ P(g); As stated above, the time-series of the number of new users per time bin, {p(t)}, is shown in Fig. 2. Several other parameters and distributions can be also inferred from the high-resolution data of Blogs and Diggs. Specifically, in Fig. 4 we show the distributions of the lifetime of posts, P(tP), time delay of users actions, P(D t), and the fraction of new posts per user, P(g), as well as the functional dependence of the probability m (T0) that a user looks for a post older than the exposed posts window T0. These quantities are inferred from the ddDiggs dataset. The value T0 = 576 time bins (corresponding to 2 days of real time) can be approximately estimated from the posts activity pattern, cf. Fig. 1b. The numerical values of the remaining parameters can not be extracted from this kind of empirical data. Hence they are considered as free parameters, that can be varied within theoretical limits. The values quoted above are used for the simulations in this work. 7 tp tkp>t0p+T0 tp (cid:229) NP p=1(cid:16) 1 Here we describe a general methodology how such parameters are determined from the empirical data of high resolution. The delay-time distribution P(D t) in Fig. 4c, is directly related with the user activity pattern, cf. Fig. 1a: for a given user (fixed index along y-axis) the delay time D t is defined as the distance between two subsequent points along the time axis. The distribution is then averaged over all users in the dataset. Similarly, the distribution of the life-time of posts P(tP) in Fig. 4d, is related to the pattern of posts activity as the distance between the first and the last point on the time axis for a given post, cf. Fig. 1b. The parameter T0 is roughly estimated as the width of the time window, during which new posts were 'exposed' (dense points area in Fig. 1b). When T0 is fixed, then the probability that a user finds a post which is older than T0 can be extracted from the data as the fraction of points beyond the dense area in the posts activity pattern until the post expires, cf. Fig. 1b. Then we have m (T0) = 1 1(cid:17), where NP is the number of posts, tp is the expire time of the post p, NP while tkp and t0p indicate the moments of the activity at the post p and its creation time, respectively. Averaged over all posts in the dataset, gives the parameter m (T0), plotted in Fig. 4b against T0. In the case of user properties, looking at the activity list of a given user, we can determine the fraction g of new posts that the user posted out of all posts on which the user were active in the entire dataset. The values appear to vary over time and users, the distribution P(g) averaged over time and all users in the dataset is shown in Fig. 4a. Strictly speaking, the values of the control parameters will depend on the empirical dataset considered. Specifically, the parameters as the life- time of posts, tP, and users inclination to posting new posts, g, or to looking towards old material m (T0), strongly depend on the dataset. Note also that they might have hidden inter-dependences in view of the nonlinear process underlying the original dataset. For instance, if on a certain Blogsite users are more inclined towards posting new material, which would yield increased probabilities of large g, then the life-time of posts may decline, resulting in a steeper distribution. Therefore, it is important to derive these parameters from the same dataset in order to ensure their mutual consistency. Note, however, that certain universal features apply, in particular, in the power- law dependences of the delay-time [9, 6] distributions P(D t), and the circadian cycles [7] in the time series {p(t)}. Although our model works for a wide range of parameter values, here we keep the parameters extracted from the empirical data of the popular discussion-driven Diggs in order to enable a comparison of the results to largest possible extent. In contrast, the relaxation rate of the arousal and the valence and the parameters d1,d2,c1,c2 of the maps in Eqs. (2-3) can not be extracted from this type of empirical data. The values shown above are chosen such that the fixed points of the maps do not fall to corner areas for typical values of the environmental fields occurring in our simulations, cf. Figs. 3a,b. 3 Results 3.1 Time-series of the emotional comments As mentioned above, we drive the system by adding p(t) new users at each time step and letting them to boost the activity of the system, according to the model rules introduced in sec 2.2.2. We sample different quantities in analogy to those that we can define and compute from the empirical data, see section 2.1 and Refs. [14, 15]. Compared with the empirical data, the advantage of the agent-based model is that we can keep track of the fluctuations in the valence and the arousal of each agent ("user") at all time steps. Typically, the arousal and the valence of an agent, who is linked through the posts to other agents, experiences stochastic inputs from the active environment, as shown in Fig. 5. Between such events the emotion arousal and valence decay with the rate g towards zero values. It should be stressed that at each agent (user-node) different patterns of the activity are expected. They depend not only on the current network structure surrounding the agent, but also on the fact that at given time the activity might be transferred to another part of the network, i.e., due to the aging of posts and the preferences of other agents towards particular types of posts. Two illustrative examples shown in Fig. 5 are from the same simulation run, but for two agents who are located at different areas of the network. The actions of individual agents contribute to the overall activity that can be monitored at each post and at the whole (evolving) network, as well as at the network parts, for instance the topological communities, that can be identified when the network is large enough. In the simulations we monitor the fluctuations of the number of active posts, Nap(t), the number of different agents that are active at these posts, Nau(t), and the number of comments 8 (cid:229) that these agents posted at each time step, Nc(t). Furthermore, we distinguish between the comments that carry positive (negative) valence, N±(t), and the overall charge of these emotional comments, which is defined as the difference between the number of comments with positive and negative valence, Q(t) ≡ N+(t) − N−(t). The temporal fluctuations of these quantities are shown in Fig. 6a,b, where only the initial part of the time-series are shown, corresponding to four weeks of real time. Notice, the circadian cycles of the driving signal are reflected to the time-series of the number of active agents and the number of their comments. The power spectra S(n ) of these time series are shown in the upper panels in Figs. 6c,d. A characteristic peak cor- responding to the daily cycles of the time-series is visible. In addition, long-range correlations with S(n ) ∼ 1/n f occur in most of these time series (except for the charge fluctuations!) for the range of frequencies, indicated by the slopes of the straight lines in both Figures. The simulated time-series can be compared with the ones observed in the empirical data, for instance Fig. 2 and Figure S1 and with similar data analyzed in Refs. [14, 15]. The fractality of these time series, leading to the power spectrum of the type 1/n f , as well as the dominance of the negative charge suggest that our model captures the basic features of the blogging dynamics. Specifically, in response to the same driving signal, which has the power spectrum with the exponent f = 1.5, the simulated blogging process builds the long-range correlations yielding the time-series with smaller exponents f = 1.33 and f = 1, in the number of active agents and the number of comments, respectively, and increased range of corre- lations, qualitatively similar to the popular Diggs. Further comparison between the simulated and the empirical data can be considered at the level of the emergent network topology, studied below in section 3.2. These time series have comprised the agent's activity at the whole (evolving) network. In sec. 4 we analyze the time series which are recorded at the level of each emerging agent-community separately. 3.2 Emergent network and communities of the emotional agents The networks that emerge through the activity of our emotional agents on posts can be studied in full analogy with the bipartite networks mapped from the empirical data of the same structure, which are discussed in section 2.1 and in Refs. [10, 14, 34]. (For other interesting examples and review of complex monopartite networks, see [35] and references therein). In our simulations the network evolves due to the addition of nodes of both partitions, as well as the evolution of links. The lists of user-posts connections are updated at every time step. With the prevailing negative comments, as demonstrated above in the time-series, cf. Fig. 6, arrivals of the negative comments at posts generate an environment that, in view of the linking rules (large negative charge preference), strongly affects the network evolution (see also discussion in sec. 4). In this way some posts with a large number of negative comments and thus large topological strength may appear. A part of the emergent network obtained in our simulations is shown in Fig. 7, where three such hubs-popular posts are visible together with the users linked to them. Some topology measures of the emergent bipartite network-the degree distributions and the assortativity mea- sures, are shown in Fig. ??a,b. For comparison, the corresponding topology measures of the network obtained from the empirical dataset of discussion-driven popular Diggs are also computed and shown in supporting infor- mation, Fig. S3. As expected, the degree distributions for each of the partitions-agent(user)-nodes and post- nodes appear to be different (some other examples of bipartite networks representing the empirical data of various techno-social interactions have been analyzed in [34]). Specifically, the broad distributions are dominated by different type of cut-offs. They can be approximated by the following mathematical expressions, motivated by fitting the corresponding empirical data, cf. Fig. S3: P(qu) = Cuq−t u e−qu/Xu0; P(qp) = Cp(cid:20)1− (1− q)(cid:18) qp Xp0(cid:19)(cid:21)1/1−q ; (7) for the agent(user)-node and the post-node distributions, respectively. The agent-degree distribution P(qU ) ex- hibits a short power-law region and a large exponential cut-off, whereas the distribution related to the post-nodes P(qP) has a dominant cut-off at smaller degree followed by a power-law tail, compatible with the q-exponential form with q > 1. The fitted values of the parameters are shown in the Figure legends in Fig. ??a. In principle, they depend on the simulation parameters of the agent-based model. However, the expressions appear to be stable with respect to the simulation time (size of the network). The results are shown for two simulation runs with 16384 and 25000 time steps, resulting in the networks with NP = 13504 + NU = 64852, and NP = 22757 + NU = 107933 9 nodes, respectively. The same mathematical expressions in eq. (7) apply (with different parameters) to the cor- responding distributions derived from the empirical data, cf. Fig. S3a. Hence, in these measures the topology of the bipartite network emerging in the emotional blogging of our model shares qualitative similarity with the one from the popular posts in real data. To certain extent, similar conclusions apply in the case of the mixing patterns, shown in Fig. ??b and in Fig. S3b. Namely, the posts making the network neighborhood of an agent-node (user- node in the empirical data) exhibit no assortativity measure, which is indicated by the line of zero slope before a cut-off at large user-node degree. In the case of post-nodes, the empirical data indicate slight disassortativity (decrease) just before the cut-off, Fig. S3b, a feature that seems not be properly captured by our model in the present parameter range. Systematic study of the network topologies, emerging when the parameters of blogging dynamics within the agent-based model are varied in a wide range away from their empirical values, is left for a separate work [36]. Here we focus on the mesoscopic structure of the emergent network obtained for the current set of parameters, which are listed above. In particular, we are interested in communities of the emotional agents, that may potentially occur on the respective monopartite projection of the network. Communities, as topological subgraphs with stronger connections between the nodes inside the community com- pared to the rest of the network, can be accurately identified using different methods [37, 38, 39, 40]. The bipartite networks representing the dynamics on Blogs and Diggs exhibit abundance of different communities [10, 14, 15]. A systematic analysis of such bipartite networks, representing various online interactions, can be found in [34]. The monopartite user- or post-projections of these networks appear to be highly clustered and weighted networks [10, 14, 15, 34], which limits applications of classical methods of community analysis [40]. Therefore, we use the methods based on the eigenvalues spectral analysis of the network [41, 39] and the maximum likelihood method adapted for multi-graphs [42]. The networks emerging through the actions of the emotional agents in our simulations have similar features. We perform the spectral analysis of the normalized Laplacian operator [43, 39] which is related to the weighted user-projection network, whose matrix elements CW i j represent the common number of posts per pair of users, including the multiplicity of user–post connections, which is indicated by the superscript. It is constructed from the symmetric matrix of commons as where li is the strength of node i, defined as the sum of wights of its links. As discussed in detail in Refs. [43, 39], the spectrum of the Laplacian (8) is limited in the range l i ∈ [0,2]. When the communities exists, the lowest non-zero eigenvalues of the Laplacian (8) appear separated from the rest of the spectrum and the corresponding eigenvectors are localized on the network subgraphs (communities). The localization of the eigenvectors is visu- alized as a characteristic branched structure of the scatter-plot in the space of these eigenvectors. This property of the eigenvectors is then utilized to identify the nodes of the network that belong to each community (detailed discussion and various examples studied by this methods can be found in [39, 44, 10, 15]). It should be stressed that the network grown through the emotional actions of of the agents in our model has specific properties which may be reflected in the community structure. Namely, the bipartite network is already weighted, which shifts the distribution of weights in the monopartite projection away from pure topology of commons [34]. In addition, the network evolves in such way that the center of the activity is shifting to ever new groups of (exposed) posts. Here we analyze the structure of the network after 4032 time steps (two weeks) of the evolution. The network projected onto user (agents) partition contains NU = 4572 users, only users with the degree larger than 5 are considered as relevant for the community formation. The eigenvalue spectrum of the Laplacian operator Eq. 8 with the CW i j matrix related with this user-projected weighted network, is then computed. The results for the eigenvalues shown in the ranking order are given in Fig. 8a. The scatter-plot of three eigenvectors belonging to the three lowest eigenvalues is shown in Fig. 8b. The spectrum, as well as the scatter-plot in Fig. 8b, indicate that five agent-communities can be differentiated. These are denoted by Gk, with k = 1,2,···5 corresponding to top-to-bottom branches in Fig. 8b. In the following we first identify the nodes representing the agents in each of these communities. Then we analyze how the communities actually evolved on that network and discuss the fluctuations of the emotional states of each agent in the communities through the evolution time. The communities on user-projected networks are our main concern, in view of the collective behavior of the emotional agents. Formally, the same methodology can be applied to the post-projected network as well as the 10 Li j = d i j − CW i j p(lil j) , (8) weighted bipartite network directly. Studies of the empirical data [10, 14] suggest that the communities of posts often appear in relation to their subjects, age, and sometimes authorship. The features is prominent in the case of Blogs of "normal" popularity. Whereas, in the communities on popular posts the subjects are often mixed, leaving potentially different driving force for user's intensive activity on that posts. In our model the post have no defined subjects, and the agents are driven by the emotion content alone. Nevertheless, the results of the spectral analysis of the post-projected network reveals that several communities of the posts can be identified. The network projection is described in supporting information and the scatter-plot of the respective eigenvectors is shown in Fig. S4. Four communities can be differentiated. Looking at the node's identity in these branches, we find that the "age" of the posts prevails as a grouping principle. Other attributes of posts that we have in the model, as "authorship", "popularity" and "charge", seem to be mixed in all present communities. 4 Discussion 4.1 Patterns of agents behavior in time and space of the emotion variables Having identified the agents in each of the communities, we can track of their group activity and the emotion fluctuations over time from our simulation data. The time-series of the number of comments of all agents in a given community Gk are shown in Fig. 9a and the emotional charge of the valence of these comments-in the Fig. 9b. Note that a fraction of comments with the valence values close to zero in the range between (−0.01, +0.01) are considered as neutral, and do not contribute to the charge. The profile of the time series indicates that all communities started to grow at early stages of the network evolution. However, two of them, G1 and G5, ceased to grow and reduced the activity relatively quickly after their appearance. Looking at the fluctuations of charge of the emotional comments in these two communities, we find that it is well balanced, fluctuating around zero at early times, and eventually leveling up to zero. Whereas, in the other two medium-size communities, G2,G4, the activity is slowly decreasing, while the largest central community, G3, shows constantly large activity. Comparing the activity (number of comments) with the fluctuations in the charge of the emotional comments, we can see that in these three communities the excess negative charge settles after some time, breaking the initial balance in the charge fluctuations. In this way our model reveals the correlations between the prolonged activity and the size of a community (i.e., number of different agents), on one side, with the occurrence of the negative charge of the related comments, on the other, a feature also observed in the empirical data on Blogs and Diggs [14, 15]. In view of the preference towards the posts with negative charge, a comment regarding the breaking of the charge balance and its consequences to the network topology is in order. Note that, according to the model rules, cf. sec. 2.2, the probability of a post to receive first negative/positive comment depends on the valence of the agent who is active on that post and its similarity with the average valence of the currently active posts. Contrary to the naive preference towards node's degree, which is known to lead to a scale-free degree distribution in growing monopartite networks (for theoretical derivation and the conditions when power-law distributions occur, see Ref. [45]), in our model agents preference is driven by another quality of a post node-its emotional content. Hence, in this process no scale-free distributions of the posts degree is expected, as also shown above. More importantly, the negative charge appears to have limited fluctuations. The time-series of the (negative) charge and of the number of comments remains stationary over large periods of time, before the activity ceases, as shown in Figs. 9 for the communities, and in Fig. 6b for the entire network. Another interesting feature of these communities can be observed by visualizing the patterns of their activity in the phase space of the emotion variables. In order to match the emotion measures accepted in the psychology literature, i.e., according to the 2-dimensional Russell's model [20, 19], we use the circumplex map suggested in Ref. [46]. The values of the arousal and the valence variables are thus mapped onto a surface enclosed by a circle, where, according to Refs. [19, 20], the emotions commonly known, as for instance, "afraid", "astonished", "bored", "depressed", etc., can be represented by different points (or segments) of the surface. In particular, the values of the arousal and valence are mapped as follows [46]: a′ = a1 √1 + z2 ,v′ = v √1 + z2 , 11 (9) where z ≡ min(a1,v) and a1 ∈ [−1, +1] is obtained from our arousal variable by first mapping a1 ≡ 2a− 1. Computing the transformed values of the arousal and the valence for each agent in a given community at all time steps when an action of that agent is recorded in our simulations, we obtain the color-plots shown in Figs. 10. Specifically, the color map indicates how often a particular state on the circumplex was occupied in four of the above communities, normalized with the all actions in that community. As the Figure 10 shows, the communities leave different patterns in the space of emotions. For instance, the community G1, that have balanced charge fluctuations, appears to cover a larger variety of the emotional states, leading to the pattern on the top left figure. Whereas, when a large community is formed, it may induce large negative fields which keep the agents in the negative valence area of the circumplex map. The situation corresponding to the community G3 is shown in bottom left plot in Fig. 10. Majority of the comments is this case are centered in the area of the arousal and the valence where the negative emotional states known as "worried", "apathetic", and "suspicious", "impatient", "annoyed" etc, are found on the circumplex map (see Refs. [19, 46] for coordinates of some other well known emotional states covered by these patterns). Plots on the right-hand side of Fig. 10 correspond to the communities G2 and G4, in which charge fluctuations are moderately negative, as discussed above. From the Figures 10 we can also observe that the well defined lower bound for the agent's arousal emerges in a self-organized manner inside each community, although no sharp threshold exists in the model rules. Moreover, the arousal drives the valence when the agents are active. This is clearly displayed in the case of communities with a balanced charge, as our community G1, Fig. 10 top left. Similar pattern of the arousal–valence was found in the laboratory experiments [47], where the values of the arousal and valence are inferred from skin conduction, heart beat and facial expression measurements on users reading a selection of posted texts. The characteristic patterns in Figs. 10 emanating from the emotional blogging of our agents suggest the processes with anomalous diffusion, in which certain parts of the phase space are more often visited than the others. They reflect the self-organized dynamics of the agent's emotion variables and the network topology. Formally, these patterns are between two extreme situations: synchronous behavior, focusing at a lower-dimensional areas, and random diffusion, spreading evenly over the entire space. The most visited areas of the phase space are in the vicinity of the attractors of the nonlinear maps. The positions of these attractors for each agent map move, depending on the actual values of the fields acting on it. The fields themselves fluctuate over time for each map, being tunned by the the agent's emotion variables and the local topology of the network (community) where the agent is situated. 4.2 Conclusions In this work we have adapted the idea of the emotional agents with two-dimensional emotion states following Ref. [29], and implemented them onto a networked environment with a bipartite network of agents (users) and posts, to model the dynamics on Blogs and similar Web portals, where the emotions are communicated indirectly via posts. In addition to the emotions, measured by the dynamical variables of the arousal and the valence of each agent, appearance of our agents on the network follows circadian cycles and and the action-delay from a distribution, which is, like some other parameters of the dynamics, inferred from the empirical data. Novel and the most important features of our agent-based model are: • the agent's emotion is spreading via indirect contacts through their actions on the posts on a bipartite network, and • the network itself evolves through the addition of the agents and their actions on the posts. • We also present a systematic methodology to extract the relevant parameters of the model from the empirical datasets of high temporal resolution. Apart from the inference of the control parameters from a given empirical dataset, the structure of the model allows experimenting on the dynamics and potential extensions at three different levels: (a) changing the properties of each individual agent; (b) modifying the linking rules, i.e., posts exposure, agent's preferences, role of the 12 emotions, etc.; (c) introducing different and/or additional driving of the system, and varying the balance between the influences of the local and the global events. In this work the focus was on the following aspects of the dynamics: • the emergence of the collective states due to agent's emotional communications; • the patterns of their emotional behaviors in time and in phase space of emotion variables; • the role of underlying contacts and the structure of emergent network. For this purpose, in the simulations we consider no external input to drive the system except for the agents arrivals according to the time signal {p(t)}, inferred from the empirical data of the discussion-driven popular Diggs. We have demonstrated that the communities of agents emerge through the emotional commenting of posts, and can be identified as topological subgraphs on the weighted user-projected network. Most of the activity of the (user) agent occurs inside the community where the agent belongs. Moreover, the growth of the community is self- amplified and prolonged with the excess negative charge of the related comments. Another quantitative measure of the collective behavior is found in building the correlations in the streams of events-1/n f -type of the power spectrum for the time series is found for the emotional comments, in the response to the driving signal, which has weaker correlations. Moreover, this also applies for the time-series of comments with positive/negative emotion valence. These features are in qualitative agreement with those found in the empirical data from the discussion- driven Diggs and Blogs, analysed by the same mathematical approaches [34, 15], cf. also Fig. 2. Hence we can conclude that the dynamic rules of our agent-based model reveal the key mechanisms behind the collective emotional behaviors, which are observed on the popular posts in real Blogs and Diggs. To examine the role of the driving signal in building the collective response, we perform comparative simulations in which the features of the p(t) signal are completely ignored. Instead, we drive the system by adding a constant number p = 6 of agents per time step. The simulated time-series are shown in supporting information, Fig. S5a and its power-spectra in Fig. S5b. Apart from higher average activity and the absence of daily cycles, the fractality of the time-series is preserved. The power-spectrum with the exponent f = 1.25 is observed for the number of comments, although in a smaller range. The negative charge sets-in after certain time period and fluctuates in a stationary manner. These results suggest that the occurrence of long-range temporal correlations is inherent to the stochastic process of our model, which might be enhanced, but not imported, by the profile of the driving signal. It is also interesting to point that the network topology grown in this process exhibits the degree distributions of the same type as Eq. (7) and three communities of the emotional agents, as shown in Fig. S5c and d. Compared with the empirical data, where the emotion is extracted only from the text on posts, in the agent-based model we can follow the fluctuations of the emotion of each agent over time. In this respect our model can interpolate between the psychology experiments at individual users, on one side, and the global emotional states, that can be recognized at larger scale [8], on the other. We have shown that the activities within each community of the agents leave a characteristic pattern in the space of emotions. Specifically, "normal" blogging leads to balanced emotions in a segment of the circumplex map, driven by large arousal (i.e., between the lines marking the limits of "high power/control" and "obstructive" emotions in Ref. [19]). This is in agreement with the laboratory measurements of the emotional states obtained on individual users [47, 13]. However, their collective behaviors may be entirely different. Our simulations suggests how the large communities can get caught into excessive negative emotions (critique), which prolongs the activity and the number of agents involved. The model provides the underlying mechanisms and the parameters through which such collective behavior may be influenced. In a broader view of the science of affective computing, our work aims to quantitative accounting of the dynamics of emotions. With the rules and parameters motivated by the discussions on popular Diggs, our agent-based model describes the dynamic environment with indirect communications among users, where the critique prevails over positive emotions. Keeping track of the relations between simulated events in terms of an evolving bipartite network is another aspect of this work, which makes the basis for the quantitative analysis of agents collective behaviors. Furthermore, the network with its local structure is essential for the discussion type dynamics that we study. By taking away the network of posts, the agents would be literally disconnected and their dynamics greatly simplified, consisting of a single appearance-and-relaxation event per agent. The role of the network structure can be altered by increasing the mean-field term (i.e., its strength q and the selection of posts which 13 contribute to it), or by adding an external noise to Eqs. (2-3) which acts directly to each agent, as in Ref. [29]. Qualitative agreement of the simulation results with the ones of the empirical dataset indicates how the emotional communications prevail on popular posts, leading to the bursting events and the emergence of communities. Whereas, the quantitative differences, in particular in the user heterogeneity shown in Fig. S3 compared to the the degree distribution of our emotional agents in Fig. ??, suggest to what extent features other than emotion may drive users behavior. Potentially different outcomes can be predicted by the model when parameters are changed or extracted from another dataset. On the other hand, when the posts of "normal" popularity are considered, we expect that subjects of the posts may play a role, affecting the way that certain emotions are communicated. Such situations can be studied by appropriate modifications of the dynamic rules within our model and inference of the related parameters. In conclusion, compared to extracting user's emotional behavior from texts of Blogs and Diggs by quantitative analysis of the empirical datasets, the agent-based modeling has the advantage in that the emotional processes are studied at each agent (representing user) directly. Moreover, variety of social experiments can be devised and performed on the agents, thus avoiding the ethical issues related with the real users. Despite its mathematical complexity, our agent-based model obviously represents simplified reality both in respect to the dimensionality of emotions and the agent properties. For instance, in the science of emotion it is known that social aspects of the emotions such as "guilt", "pride", "shame" are different compared to "anger", "depression", "joy", and others. The psychological theory also suggests that personality profile determines how certain emotion is expressed. The theoretical modeling along the lines of our agent-based model will benefit from current developments in the sci- ence of affective computing, which aims for quantitative measures behind the psychology theory and for extracting different dimensions of emotions [22]. The agent properties can also include more realistic personal differences and influence of the real-world processes on users decisions and rhythm of their stepping into the virtual world. In the present model such real-world processes are implicitly taken into account through the parameters-the driving t), which are involved in generating the respective empirical data of Diggs. This is a step beyond the first approximation, which we hope is opening the way for further research towards realistic modeling of the emotion dynamics on the Web. signal {p(t)} and the delay-time distribution P(D 5 Figure Legends and Supporting Information Figure 1:Patterns of activity differ for user and post nodes. Example of temporal patterns of (a) user actions and (b) activity at posts, obtained from the original dataset of discussion-driven Diggs (ddDiggs). Indexes are ordered by the user (post) first appearance in the dataset, while time is given in minutes. Figure 2:Time-series with circadian cycles and fractal features. (bottom) Time-series of the number of new users (red-dark) and the number of all active users (green-pale) per time bin of 5min derived from the ddDiggs dataset; (top) Power spectra of these time series as indicated (shifted vertically for better vision). Daily and weekly cycles can be easily noticed on both plots. Figure 3: Two-dimensional maps of the emotion variables of an isolated user-node. Maps for arousal (left) and valence (right) for four different values of the fields are shown. The fixed line X(t + 1) = X(t) is indicated. Figure 4:Parameters of the model as inferred from the empirical data of ddDiggs. (a) Distribution of g-the fraction of new posts per user, relative to all posts on which that user was active, averaged over all users in the dataset. (b) Probability m that a user looks at a post which is older than the specified time window T0 time bins, averaged over all users and plotted against T0. (c) Distribution of the time-delay D t between two consecutive user actions, averaged over all users in the dataset. (d) Distribution of the life-time of posts tP, averaged over all posts in the dataset (log-binned data). In Figs. (c) and (d) the time axis is given in the number of time bins, each time bin corresponds to tb = 5 minutes of real time. Figure 5:Valence and arousal of each agent linked on the network fluctuate in time. Two examples of the valence and the arousal are shown against time for two agents (users) located in different areas of the bipartite network, resulting in different activity patterns: a very active agent (left) and a sporadically active agent (right). 14 Figure 6: Time series simulated in the model of interacting emotional agents on bipartite network. The number of all comments per time step (cyan) and the number of comments with positive (red) and negative (black) valence per time step (a), and the double-logarithmic plot of their power spectra (c). The number of active posts (indigo) and the number of active agents (magenta) per time step, and the charge of all emotional comments (blue), are shown in panel (b), and their corresponding power spectra, panel (d). Straight lines indicate slopes f ={1, 1.33, 0}. For clear display, the power-spectra are logaritmically binned. Figure 7: Emergent bipartite network of the emotional agents and the posts exhibits strong inhomogeneity. Shown is a part of the network structure in the vicinity of three popular posts. Figure 8: Some topology measures of the bipartite network emerging in the dynamics of emotional agents. (a) The degree distributions of the user-partition ((cid:13),+) and the post-partition ((cid:3),×). Fitting lines explained in the Legend. (b) Assortativity measures: The average degree of the posts linked to the user node of a given degree versus user degree ((cid:13),+), and the average user degree linked to the post of a given degree, plotted against post degree ((cid:3),×). Empty symbols are for the simulation time 16384 steps, while the crosses indicate the respective results from runs with 25000 time steps. Figure 9: Spectral analysis of the emergent network reveals community structure. For the agent-projected network (a) the eigenvalue spectrum, and (b) the scatter-plot of three eigenvectors belonging to the lowest nonzero eigenvalues, indicate five communities Gk, k = 1,2···5, related to five branches in the scatter-plot, from top to bottom. Note that each point in the scatter-plot represents a unique node on the network. Figure 10: Active communities grow in correlation with the excess of negative charge. Time series of the number of comments by the agents belonging to a given community Gk (a), and the charge of these comments (b), computed for each of five communities identified in Fig. 8. Figure 11: Activity patterns of the communities projected onto 2-dimensional space of emotions. Circum- plex map of the emotional states of the agents belonging to four communities identified on the emergent network: G1–top left, G2–top right, G3–bottom left, G4–bottom right. Color map indicates occupancy of a given state, normalized relative to the number of comments in each community. Figure S1: Prevalence of negative comments on popular Diggs. Shown are the time-series and their power-spectra of the number of comments N±(t) carrying positive emotion (red) and negative emotion (black), and the time- series of "charge" of the emotional comments, Q(t) = N+(t)− N−(t) (cyan). The Figure shows that comments are correlated in time, leading to 1/n -type of power spectra and that the negative comments prevail. Video S2: User interests shift daily towards new posts. The movie shows the evolution of activity of one com- munity on the weighted bipartite network of discussion driven Digg stories during 90 days. In Ref. [15], we have identified such communities in the network constructed from the subset of the discussion-driven Diggs and keeping only very active users (with more than 100 comments). The network is then analyzed by the eigenvalue spectral methods[39] and three communities are identified [15]. For the movie, we selected nodes belonging to one of these communities, g2, which consists of 236 users and posts, and then determined weighted bipartite subnetworks of it for each consecutive day. Thus, the weight of the link corresponds to the number of comments written by the user on the post on a given day, while the color of the link indicates overall emotional contents of these comments (black–negative, white–neutral and red–positive). The networks are then visualized using the program Pajek. Every frame corresponds to the window of one day in real time. The frames are combined using Avidemux program package. The movie shows that on each day different posts were in the focus. Figure S3: Some topology measures of the bipartite network derived from empirical data on Diggs. For com- parison of the network topology, the dataset of the popular discussion-driven Diggs, described in section 2.1, is mapped onto a weighted bipartite network, with the weights of links representing the number of comments of a user to a post. The degree distribution of user-nodes and post-nodes of that network are computed and shown in Fig. S3(a), while the assortativity measures-the average degree of the post nodes linked to the user of a given degree, and vice versa, the average degree of the user nodes linked to the post of a given degree, are shown in Fig. S3(b). For completeness, shown are also the respective quantities computed from whole dataset of all Diggs (including the posts with normal popularity), indicated in the Legend. Figure S4: Community structure of post-projected network obtained from the emotional agents dynamics. Shown 15 is the scatter-plot of the eigenvectors belonging to three lowest eigenvalues of the post-projected weighted bipartite network, emerging in the simulations of the emotional agents dynamics after 4000 time steps. The network of size is reduced to NP = 1156 posts by taking the posts with strength ℓP > 50, as relevant for the community formation. Figure S5: When the bloggers never sleep. Simulations results obtained for the system of the emotional agents is driven by adding a constant number of users p = 6 per time step: Time-series of the number of comments and charge (a) and their power-spectra (b). Degree distributions of user-nodes and post-nodes, P(qU ) and P(qP), are shown in panel (c).Scatter-plot of the eigenvectors related with the user-projected network (NU = 4418 users with strength ℓU > 10 considered), indicating the community structure (d). Funding The research leading to these results has received funding from the European Community's Seventh Framework Programme FP7-ICT-2008-3 under grant agreement no 231323 (CYBEREMOTIONS). B.T. thanks support from the national program P1-0044 (Slovenia). Author Contributions Conceived and designed the experiments: BT. Performed the experiments: MM. Con- tributed analysis tools: MM. Analyzed the data: MM, BT. Wrote the paper: BT. References [1] J. Kleinberg. The Convergence of Social and technological Networks. Communications of the ACM, 51:66– 72, 2008. [2] J. Giles. Social science lines up its biggest challenges. Nature, 470:18–19, 2011. [3] A. Cho. Ourselves and Our Interactions: The Ultimate Physics Problem? Science, 325:406–408, 2009. [4] C. Castellano, S. Fortunato, and V. Loreto. Statistical physics of social dynamics. Reviews of Modern Physics, 81(2):591–646, 2009. [5] F. Skopik, D. Schall, H. Psaier, M. Treiber, and S. Dustdar. Towards social crowd environments using service-oriented architectures. Information Technology, 3:108–116, 2011. [6] A. V´azquez, J. G. Oliveira, Z. Dezso, K.-I. Goh, I. Kondor, and A.-L. Barab´asi. Modeling bursts and heavy tails in human dynamics. Physical Review E, 73(3):036127, 2006. [7] R. D. Malmgren, D. B. Stouffer, A. S. L. O. Campanharo, and L. A. Amaral. On Universality in Human Correspondence Activity. Science, 325(5948):1696–1700, 2009. [8] P.S. Dodds and C.M. Danforth. Measuring the Happiness of Large-Scale Written Expression: Songs, Blogs, and Presidents. Journal of Happiness Studies, 11(4):441–456, 2009. [9] R. Crane, F. Schweitzer, and D. Sornette. Power law signature of media exposure in human response waiting time distributions. Physical Review E, 81(5):056101, 2010. [10] M. Mitrovi´c and B. Tadi´c. Bloggers Behavior and Emergent Communities in Blog Space. European Physical Journal B, 73(2):293–301, 2010. [11] M. Szell, R. Lambiotte, and S. Thurner. Multirelational organization of large-scale social networks. Pro- ceedings of the National Academy of Sciences USA, 107(31):13636–13641, 2010. [12] S. Gonzalez-Bailon, R.E. Banchs, and A. Kaltenbrunner. Emotional Reactions and the Pulse of Public Opinion: Measuring the Impact of Political Events on the Sentiment of Online Discussions. in Proceedings of CoRR, 2010. e-print arXiv:1009.4019. [13] M. Thelwall, K. Buckley, G. Paltoglou, D. Cai, and A. Kappas. Sentiment strength detection in short informal text. J. Am. Soc. Inf. Sci. Technol., 61:2544–2558, 2010. 16 [14] M. Mitrovi´c, G. Paltoglou, and B. Tadi´c. Networks and emotion-driven user communities at popular blogs. European Physical Journal B, 77:597–609, 2010. [15] M. Mitrovi´c, G. Paltoglou, and B. Tadi´c. Quantitative analysis of bloggers' collective behavior powered by emotions. Journal of Statistical Mechanics: Theory and Experiment, 2011(02):P02005+, February 2011. [16] P.S. Dodds, K.D. Harris, I.M. Koloumann, C.A. Bliss, and C.M. Danforth. Temporal patterns of happiness and information in a global social network: Hedonometric and Twitter. arXiv:1101.5120v3, 2011. [17] G. Paltoglou, S. Gobron, M. Skowron, M. Thelwall, and D. Thalmann. Sentiment analysis of informal textual communication in cyberspace. In Proc. ENGAGE 2010 (Springer LNCS State-of-the-Art Survey), pages 13–23. Heidelberg: Springer, 2010. [18] A. Lancichinetti, M. Kivela, J. Saramki, and S. Fortunato. Characterizing the community structure of com- plex networks. PLoS ONE, 5(8):e11976, 2010. [19] K. Scherer. What are emotions? And how can they be measured? Social Science Information, 44(4):695– 729, 2005. [20] J. A. Russell. A circumplex model of affect. Journal of Personality and Social Psychology, 39:1161–1178, 1980. [21] J. A. Coan, and John J. B. Allen, Editors, The Handbook of Emotion Elicitation and Assessment Oxford University Press Series in Affective Science, 2007. [22] R.A. Calvo and S. D'Mello. Affect Detection: An Interdisciplinary Review of Models, Methods, and Their Applications, Affective Computing, IEEE Transactions on, 1:18–37, 2010. [23] T. Yarkoni. Personality in 100.000 words: A large-scale analysis of personality and word use among blog- gers, Journal of research in personality, 44:363–373, 2010. [24] M. A. Nicolaou, H. Gunes, and M. Pantic, Continuous Prediction of Spontaneous Affect from Multiple Cues and Modalities in Valence-Arousal Space, IEEE Transactions on Affective Computing, 2:92–105, 2011. [25] G. Paltoglou, M. Theunis, A. Kappas, and M. Thelwall, Prediction of valence and arousal in forum discus- sions, Journal of IEEE Transactions on Affective Computing, in review, 2011. [26] A. Chmiel, P. Sobkowicz, J. Sienkiewicz, G. Paltoglou, K. Buckley, M. Thelwall, and J. Holyst. Negative emotions boost user activity at BBC forum. Physica A, 390:2936–2944, 2011. [27] F. Schweitzer. Browning Agents and Active Particles: Collective Dynamics in the Natural and Social Sci- ences. Springer-Verlag Berlin Heidelberg, 2007. [28] G. N. Gilbert. Agent-based models. Sage Publications, Los Angeles, 2008. [29] F. Schweitzer and D. Garcia. An agent-based model of collective emotions in online communities. European Physical Journal B, 77(4):533–545, 2010. [30] L. Feldman Barrett and J. A. Russell. Independence and bipolarity in the structure of current affect. Journal of Personality and Social Psychology, 74(4):967–984, 1998. [31] A. Arenas, A. Diaz-Guilera, J. Kurths, Y. Moreno, and C.S. Zhou. Synchronization in complex networks. Physics Reports, 469:93–153, 2008. [32] L. Buzna, S. Lozano, and A. D´ıaz-Guilera. Synchronization in symmetric bipolar population networks. Physical Review E, 80:066120, 2009. [33] Z. Levnaji´c, , and B. Tadi´c. Self-organization in trees and motifs of two-dimensional chaotic maps with time delay. Journal of Statistical Physics: Theory and Experiment, P03003, 2008. 17 [34] M. Mitrovi´c and B. Tadi´c. Emergence and structure of cybercommunities. in Handbook of Optimization in Complex Networks: Communication and Social Networks (vol 2), part 2 "Online Social Networks", Eds. M.T. Thai and P. Pardalos. Springer, Berlin, 2011. [35] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang. Complex networks: Structure and dynamics. Physics Reports, 424:175–308, 2006. [36] M. Mitrovi´c and B. Tadi´c. Topology of bipartite networks grown by blogging of the emotional agents. in preparation, 2011. [37] M. Rosvall and C. T. Bergstrom. Maps of random walks on complex networks reveal community structure. Proceedings of the National Academy of Sciences of the United States of America, 105(4):1118–1123, 2008. [38] T. S. Evans and R. Lambiotte. Line Graphs, Link Partitions and Overlapping Communities. Physical Review E, 80(1):016105+, 2009. [39] M. Mitrovi´c and B. Tadi´c. Spectral and dynamical properties in classes of sparse networks with mesoscopic inhomogeneities. Physical Review E, 80(2):026123–+, 2009. [40] S. Fortunato. Community detection in graphs. Physics Reports, 486(3-5):75–174, 2010. [41] L. Donetti and M. A. Munoz. Detecting network communities: a new systematic and efficient algorithm. Journal of Statistical Mechanics: Theory and Experiment, P10012, 2004. [42] M. Mitrovi´c and B. Tadi´c. Search of Weighted Subgraphs on Complex Networks with Maximum Likelihood Methods. Lecture Notes in Computer Science, 5102:551–558, 2008. [43] S. N. Dorogovtsev, A. V. Goltsev, J. F.F. Mendes, and A. N. Samukhin. Spectra of complex networks. Physical Review E, 68(4):046109–+, 2003. [44] J. Gruji´c, M. Mitrovi´c, and B. Tadi´c. Mixing patterns and communities on bipartite graphs on web-based In Digital Signal Processing 2009: 16th International Conference on, IEEEXplore, social interactions. arnumber=5201238, 2009. [45] S. N. Dorogovtsev, J. F.F. Mendes, and A. N. Samukhin. Structure of Growing Networks with Preferential Linking. Physical Review Letters, 85:4633–4666, 2000. [46] J. Ahn, S. Gobron, Q. Silvestre, and D. Thalmann. Asymmetrical Facial Expressions based on an Advanced Interpretation of Two-dimensional Russell's Emotional Model. In Proceedings of ENGAGE, 2010. [47] M. M. Bradley, M. Codispoti, B. N. Cuthbert, and P. J. Lang. Emotion and motivation I: defensive and appetitive reactions in picture processing. Emotion, 1(3):276–298, 2001. 18 Figure 1 Figure 2 19 Figure 3 Figure 4 20 Figure 5 Figure 6 21 Figure 7 22 Figure 8 23 Figure 9 Figure 10 24 Figure 11: S1 25 Figure 12: S3 26 Figure 13: S4 27 Figure 14: S5 28
1301.2561	1	1301	2013-01-11T17:59:20	Modeling complex systems with adaptive networks	[ "cs.SI", "nlin.AO", "physics.soc-ph" ]	Adaptive networks are a novel class of dynamical networks whose topologies and states coevolve. Many real-world complex systems can be modeled as adaptive networks, including social networks, transportation networks, neural networks and biological networks. In this paper, we introduce fundamental concepts and unique properties of adaptive networks through a brief, non-comprehensive review of recent literature on mathematical/computational modeling and analysis of such networks. We also report our recent work on several applications of computational adaptive network modeling and analysis to real-world problems, including temporal development of search and rescue operational networks, automated rule discovery from empirical network evolution data, and cultural integration in corporate merger.	cs.SI	cs	Preprint Hiroki Sayama, Irene Pestov, Jeffrey Schmidt, Benjamin James Bush, Chun Wong, Junichi Yamanoi, and Thilo Gross, Modeling complex systems with adaptive networks, Computers and Mathematics with Applications, 2013, http://dx.doi.org/10.1016/j.camwa.2012.12.005. Disclaimer: One of the authors carried out this research on behalf of the Government of Canada, and as such the copyright for that part of the paper belongs to the Crown, that is to the Canadian Government. No provision exists for the transfer of any such Crown copyright. Non-exclusive permission is granted to use, reproduce and communicate the copyright article in any way you wish, as long as its source is acknowledged. 3 1 0 2 n a J 1 1 ] I S . s c [ 1 v 1 6 5 2 . 1 0 3 1 : v i X r a Modeling Complex Systems with Adaptive Networks Hiroki Sayamaa,∗, Irene Pestovb, Jeffrey Schmidta, Benjamin James Busha, Chun Wonga, Junichi Yamanoic, Thilo Grossd aCollective Dynamics of Complex Systems Research Group, Binghamton University, State University of New York, Binghamton, NY, USA bDefence Research & Development Canada, Centre for Operational Research and Analysis, Ottawa, Ontario cFaculty of Policy Studies, Chuo University, Tokyo, Japan dDepartment of Engineering Mathematics, University of Bristol, Bristol, UK Abstract Adaptive networks are a novel class of dynamical networks whose topologies and states coevolve. Many real-world complex systems can be modeled as adaptive networks, including social networks, transportation networks, neural net- works and biological networks. In this paper, we introduce fundamental concepts and unique properties of adaptive networks through a brief, non-comprehensive review of recent literature on mathematical/computational modeling and analysis of such networks. We also report our recent work on several applications of computational adaptive network modeling and analysis to real-world problems, including temporal development of search and rescue opera- tional networks, automated rule discovery from empirical network evolution data, and cultural integration in corporate merger. Keywords: adaptive networks, complex systems, complex networks, state-topology coevolution, dynamics, generative network automata 1. Introduction The rapidly growing research on complex networks has presented a new approach to complex systems modeling and analysis [1–5]. It addresses the self-organization of complex network structure and its implications for system behavior, which holds significant cross-disciplinary relevance to many fields of natural and social sciences, particularly in today’s highly networked society. Interestingly, complex network research has historically addressed either “dynamics on networks” or “dynamics of networks” almost separately, without much consideration given to both at the same time. In the former, “dynamics on networks” approach, the focus is on the state transition of nodes on a network with a fixed topology and the trajectories of the system states in a well-defined phase space [6–12]. This is a natural extension of traditional dynamical systems research to a high-dimensional phase space with non-trivial interaction between state variables. On the other hand, in the latter, “dynamics of networks” approach, the focus is on the topological transformation of a network and its effects on statistical properties of the entire network [13–19], where a number of key concepts and techniques utilized are borrowed from statistical physics and social network analysis. When looking into real-world complex networks, however, one can find many instances of networks whose states and topologies “coevolve”, i.e., they interact with each other and keep changing, often over the same time scales, due to the system’s own dynamics (Table 1). In these “adaptive networks”, state transition of each component and topological transformation of networks are deeply coupled with each other, producing emergent behavior that would not be seen in other forms of networks. Modeling and predicting state-topology coevolution is now becoming well recognized as one of the most significant challenges in complex network research [1, 5, 20, 21]. ∗Corresponding author Email address: [email protected] (Hiroki Sayama) Preprint submitted to Computers & Mathematics with Applications May 10, 2014 Table 1: Real-world examples of adaptive networks whose states and topologies interact with each other and coevolve. Network Nodes Links Examples of node states Gene/protein activities Population, intraspecific diversities Pathologic states Cell adhesions, intercellular communica- tions Ecological relationships (predation, symbiosis, etc.) Physical contacts Examples of node addition or removal Cell division, cell death Examples of topological changes Cell migration Speciation, invasion, extinction Death, quarantine Entry to or withdrawal from community Changes in ecological relationships via adaptation Reduction of physical contacts Establishment or renouncement of relationships Organism Cells Ecological community Species Epidemiological network Individuals Social network Individuals Social relationships, conversations, collaborations Socio-cultural states, political opinions, wealth In this paper, we introduce fundamental concepts and unique properties of adaptive networks through a brief, non- comprehensive review of recent literature on mathematical/computational modeling and analysis of such networks. We also report our recent work on several applications of computational adaptive network modeling and analysis to real-world problems, including temporal development of search and rescue operational networks, automated rule discovery from empirical network evolution data, and cultural integration in corporate merger. The rest of the paper is structured as follows. In the next section, some of the recent literature is reviewed briefly to illustrate the increasing attention to the field of adaptive networks. In Section 3, we introduce Generative Network Automata (GNA), a theoretical framework for modeling adaptive network that we have proposed. In Sections 4–5, we present the aforementioned three examples of applications of adaptive network modeling to study the dynamics of complex systems. The final section summarizes and concludes the paper. 2. Growing Literature on Adaptive Networks Over the past decade, several mathematical/computational models of state-topology coevolution in adaptive net- works have been developed and studied on various subjects, ranging from physical, biological to social and engineered systems. In this section, we introduce a small number of samples taken from the recent literature, categorizing them into five major subjects of interest in the field. 2.1. Self-Organized Criticality in Adaptive Neural Systems The present interest in adaptive networks was triggered by a paper published by Bornholdt and Rohlf in 2000 [22]. They built on an observation by Christensen et al. [23] who investigated the dynamics of a simple dynamical model for extremal optimization on complex networks. In the penultimate paragraph of their paper, Christensen et al. remarked that letting the structure of the network coevolve with the dynamics on the network, leads to a peculiar self-organization such that topological properties of the network approached a critical point, where the dynamics on the network changed qualitatively. Inspired by Christensen et al., Bornholdt and Rohlf proposed a different model in which the self-organization to the critical state could be understood in greater detail. Their investigation showed that the dynamical processes taking place on the network effectively explored the network topology and thereby made topological information available in every network node. This information then fed into the local topological evolution and steered the dynamics toward the critical state. Thus a global self-organization is possible through the interplay of two local processes. Importantly, 2 the paper of Bornholdt and Rohlf demonstrated that this is not only the case in rare, specifically engineered examples, but should be expected under fairly general conditions. Self-organized criticality is interesting because it can be argued that every information processing system should be in a critical state. It was therefore suspected that also the brain should reside in a critical state [24]. The mechanism of Bornholdt and Rohlf provided a plausible mechanism, explaining how criticality in the brain could be achieved. This “criticality hypothesis” [25] was subsequently supported by further models [26, 27] and laboratory experiments [28, 29]. Today there is growing evidence that self-organized criticality is a central process for brain functionality. Adap- tive networks models remain a major tool for understanding this process. In the neural context, understanding self- organized criticality in adaptive networks is thus paving the way to new diagnostic tools and a deeper understanding of neural disorders [30]. Furthermore, understanding self-organized criticality in biological neural networks is thought to hold the key to the artificial systems that can self-organize to a state where they can process information. This may be enable the use of future nano-scale electronic components that are too small to arrange precisely using photolitog- raphy, and thus have to use adaptive principles to self-tune to a functional state after quasi-random assembly. 2.2. Epidemics on Adaptive Networks While adaptive self-organized criticality requires dynamics on different time scales, other dynamical phenomena in adaptive networks occur when topology and node states evolve simultaneously. The resulting interplay has been investigated in detail in a class of epidemiological models where the agents rewire their social contacts in response to the epidemic state of other agents. The first adaptive-network-based epidemic model was the adaptive Susceptible-Infected-Susceptible (SIS) model studied by Gross et al. [31]. By a so-called moment closure approximation, the authors were able to compute transition points in the model analytically. The main value of this work was to provide detailed analytical insights into the emergence of system-level phenomena from the node-level coevolution. Today, the model remains a benchmark for the performance of analytical approximations to adaptive networks [32–35]. Furthermore, it triggered a large body subsequent investigations into the effect of social responses to epidemics on disease propagation and vaccination strategies [36–39]. Adaptive networks have produced significant implications for real-world epidemiological practice, as they capture more realistic dynamics of social networks where people tend to alter social behaviors according to epidemiological states of their neighbors [40]. For example, Epstein et al. [41] and Funk et al. [42] considered a spatial context for the influence of human behavior in the outbreak of epidemics (although the former did not explicitly use network models). Also, Shaw and Schwartz [43] recently showed that vaccine control of a disease is significantly more effective in adaptive social networks than in static ones, because of individuals’ adaptive behavioral responses to the vaccine application. 2.3. Adaptive Opinion Formation and Collective Behavior Another active direction in adaptive networks research focuses on models of collective opinion formation. These models describe the diffusion of competing opinions through a networked population, where agents can modify their contacts depending on the opinions held by their neighbors. Two similar pioneering models in this direction were published by Holme and Newman [44] and Zanette and Gil [45] in 2006. A central question in opinion formation is whether the coevolutionary dynamics will eventually lead to consensus or to a fragmentation splitting the population into two disconnected camps. The transition points between these two long-term outcomes is known as the fragmentation transition. The simplest and best-understood model exhibiting the fragmentation transition is the adaptive voter model [46]. A detailed understanding of the fragmentation transition in this model was gained through the work of Vazquez et al. [47] and the independent parallel study of Kimura and Hayakawa [48]. Although the adaptive voter model is similar to the adaptive SIS model, analytical tools that perform well for the SIS model yield poor results for the voter model [48]. Nevertheless, the transition point can be computed analytically, using a different approach [49, 50]. It was sometimes criticized that mathematical models of opinion formation fall short of the complexity of opinion formation processes in the real world, and that hence no connection to real-world observations and experiments can be 3 made. However, Centola et al. [51] studied agent-based adaptive network models of more realistic cultural drift and dissemination processes, finding similar dynamics including the fragmentation transition. Centola also experimentally examined how social network structures interact with human behaviors [52, 53]. More recently, the works of Huepe et al. [54] and Couzin et al. [55] showed that voter-like models can be used to understand the dynamics of decision making in the collective motion of swarms of locusts [54] and schools of fish [55]. Their studies demonstrated that analytically tractable adaptive network models could predict the result of laboratory experiments. 2.4. Social Games on Adaptive Networks Besides opinion formation, also other types of social dynamics have been investigated on adaptive networks. In particular, many adaptive extensions of classical game theoretical models have been proposed. Three early works that appeared already in 2000 are a study of the minority game on adaptive networks by Paczuski, Bassler, and Corral [56], an exploration of various coordination and cooperation games by Skyrms and Pemantle [57], and a study of the Prisoner’s dilemma by Zimmermann et al. [58]. Another influential work is a paper by by Bornholdt and Ebel, which remains unfinished but is available as a preprint [59]. These papers above triggered a large number of subsequent work that explored how coevolutionary dynamics affects the evolution of cooperation in adaptive networks. Notable examples are the work of Pacheco et al. [60] and van Segbroeck et al. [61] who demonstrated clearly that coevolution can lead to increased levels of cooperation; Poncela et al. [62], who showed that coevolutionary dynamics can facilitate cooperation not only by building up beneficial structures, but through the dynamics of growth itself; and Zschaler et al. [63], who identified an unconventional dynamical mechanism leading to full cooperation. Research in adaptive networks also gave rise to a different class of games, where agents do not aim to optimize some abstract payoff, but struggle for an advantageous position in the network. The earliest example of these adap- tive network formation games is perhaps the paper of Bala and Goyal [64] which was published in 2001. Another early paper is Holme and Ghoshal’s model [65], where the nodes tried to maximize their centralities by adaptively changing their links based on locally available information, without paying too much costs (i.e., maintaining too many connections). The resulting time evolution of the network was highly nontrivial, involving a cascade of strategic and topological changes, leading to a network state that was close to the transition between well-connected and fragmented states. A recent work by Do et al. [66] presents an analytical investigation of network formation and cooperation on an adaptive weighted network. 2.5. Organizational Dynamics as Adaptive Networks Applications of adaptive networks do not stop at abstract social models like those reviewed above. One of the latest application areas of adaptive networks is the computational modeling of complex organizational behavior, including the evolution of organizational networks, information/knowledge/culture sharing and trust formation within a group or corporation. Studies on organizational network structures actually have several decades of history (including the well-known structural holes argument by Burt [67]), but computational simulation studies of organizational adaptive networks have begun only recently, e.g., the work by Buskens and Van de Rijt on the simulation of social network evolution by actors striving for structural holes [68]. More recent computational models of organizational adaptive networks are hybrids of dynamical networks and agent-based models, where mechanisms of the coevolution of network topologies and node states can be a lot more complex and detailed than other more abstract mathematical models. Such models are therefore hard to study ana- lytically, yet systematic computational simulations provide equally powerful tools of investigation. Adaptive network models are still quite novel in management and organizational sciences, and thus the relevant literature has just begun to develop. For example, Dionne et al. [69] developed an agent-based model of team development dynamics, where agents (nodes) exchange their knowledge through social ties and then update their self-confidence and trust to other team members dynamically. In their model, the self-confidence (node state) and trust (link weight) were represented not by a simple scalar number, but by a complex function defined over a continuous knowledge domain. Computational simulations illustrated the nontrivial effects of team network topology and other parameters on the overall team per- formance after the team development process. Another computational model addressing organizational dynamics at a larger scale was developed by Lin and Desouza [70] on the coevolution of informal organizational network and individual behavior. In their model, a node 4 state includes behavioral patterns and knowledge an individual has, and the knowledge is transferred through informal social links that are changed adaptively. Computational simulations showed that knowledgeable individuals do not necessarily gain many connections in the network, and that when high knowledge diversity exists in the organization, the network tends to evolve into one with small characteristic path lengths. Our most recent work on cultural integration in corporate merger [71] also models organizational dynamics as adaptive networks, which will be discussed in more detail in Section 6. Note that the literature introduced in this section is not meant to be a comprehensive review of adaptive network research. More extensive information about the literature and other resources can be found online [72]. 3. Generative Network Automata In this and the following sections, we present some of our recent work on computational modeling of adaptive networks and its applications to complex systems. To provide a useful modeling framework for adaptive network dynamics, we have proposed to use graph rewriting systems [73, 74] as a means of uniform representation of state-topology coevolution. This framework, called Gener- ative Network Automata (GNA), is among the first to systematically integrate graph rewritings in the representation and computation of complex network dynamics that involve both state transition and topological transformation. 3.1. Definitions A working definition of GNA is a network made of dynamical nodes and directed links between them. Undirected links can also be represented by a pair of directed links symmetrically placed between nodes. Each node takes one of the (finitely or infinitely many) possible states defined by a node state set S . The links describe referential relationships between the nodes, specifying how the nodes affect each other in state transition and topological transformation. Each link may also take one of the possible states in a link state set S (cid:48). A configuration of GNA at a specific time t is a combination of states and topologies of the network, which is formally given by the following: • Vt: A finite set of nodes of the network at time t. While usually assumed as time-invariant in conventional dynamical systems theory, this set can dynamically change in the GNA framework due to additions and removals of nodes. • Ct : Vt → S : A map from the node set to the node state set S . This describes the global state assignment on the network at time t. If local states are scalar numbers, this can be represented as a simple vector with its size potentially varying over time. • Lt : Vt → {Vt × S (cid:48)}∗: A map from the node set to a list of destinations of outgoing links and the states of these links, where S (cid:48) is a link state set. This represents the global topology of the network at time t, which is also potentially varying over time. States and topologies of GNA are updated through repetitive graph rewriting events, each of which consists of the following three steps: 1. Extraction of part of the GNA (subGNA) that will be subject to change. 2. Production of a new subGNA that will replace the subGNA selected above. 3. Embedding of the new subGNA into the rest of the whole GNA. The temporal dynamics of GNA can therefore be formally defined by the following triplet (cid:104)E, R, I(cid:105): • E: An extraction mechanism that determines which part of the GNA is selected for the updating. It is defined as a function that takes the whole GNA configuration and returns a specific subGNA in it to be replaced. It may be deterministic or stochastic. • R: A replacement mechanism that produces a new subGNA from the subGNA selected by E and also specifies the correspondence of nodes between the old and new subGNAs. It is defined as a function that takes a subGNA configuration and returns a pair of a new subGNA configuration and a mapping between nodes in the old subGNA and nodes in the new subGNA. It may be deterministic or stochastic. 5 Figure 1: GNA rewriting process. (a) The extraction mechanism E selects part of the GNA. (b) The replacement mechanism R produces a new subGNA as a replacement of the old subGNA and also specifies the correspondence of nodes between old and new subGNAs (dashed line). This process may involve both state transition of nodes and transformation of local topologies. The “bridge” links that used to exist between the old subGNA and the rest of the GNA remain unconnected and open. (c) The new subGNA produced by R is embedded into the rest of the GNA according to the node correspondence also specified by R. In this particular example, the top gray node in the old subGNA has no corresponding node in the new subGNA, so the bridge links that were connected to that node will be removed. (d) The updated configuration after this rewriting event. • I: An initial configuration of GNA. The above E, R, I are sufficient to uniquely define specific GNA models. The entire picture of a rewriting event is illustrated in Figure 1, which visually shows how these mechanisms work together. This rewriting process, in general, may not be applied synchronously to all nodes or subGNAs in a network, because simultaneous modifications of local network topologies at more than one places may cause conflicting results that are inconsistent with each other. This limitation will not apply, though, when there is no possibility of topological conflicts, e.g., when the rewriting rules are all context-free, or when GNA is used to simulate conventional dynamical networks that involve only local state changes but no topological changes. 3.2. Uniqueness and Generality of GNA The function of the extraction and replacement mechanisms (E and R) may be defined as either deterministic or stochastic, as opposed to typical deterministic graph grammatical systems [75]. A stochastic representation of GNA dynamics will be particularly useful when applied to the modeling of real-world complex network data, in which a considerable amount of random fluctuations and observation errors are inevitable. Also, the GNA framework is unique in that the mechanism of subGNA extraction is explicitly described in the formalism as an algorithm E, not implicitly assumed outside the replacement rules like what other graph rewriting systems typically adopt (e.g., [76]). Such algorithmic specification allows more flexibility in representing diverse network evolution and less computational complexity in implementing their simulations, significantly broadening the areas of application. For example, the preferential attachment mechanism widely used in network science to construct scale-free networks is hard to describe with pure graph grammars but can be easily written in algorithmic form in GNA. The GNA framework is highly general and flexible so that many existing dynamical network models can be rep- resented and simulated within this framework. For example, if R always conserves local network topologies and modifies states of nodes only, then the resulting GNA is a conventional dynamical network model, including cellular automata, artificial neural networks, and random Boolean networks (Figure 2 (a), (b)). A straightforward application of GNA typically comes with asynchronous updating schemes, as introduced above. Since asynchronous automata 6 Agent Class Sensor Router Actor Database Controller Table 2: Summary of Canadian Arctic SAR agent classes. Functions senses, detects, and passes gathered information distributes the flow of information and enables communication links executes a response action as tasked stores and provides access to information coordinates a response and tasks other agents networks can emulate any synchronous automata networks [77], the GNA framework covers the whole class of dynam- ics that can be produced by conventional dynamical network models. Moreover, as mentioned earlier, synchronous updating schemes could also be implemented in GNA for this particular class of models because they involve only state changes on each localized node but no topological transformation. On the other hand, many network growth models developed in network science can also be represented as GNA if appropriate assumptions are implemented in the subGNA extraction mechanism E and if the replacement mechanism R causes no change in local states of nodes (Figure 2 (c)). We also conducted extensive computational experiments of simple binary-state GNA, which revealed several dis- tinct types of the GNA dynamics, illustrating the richness and subtleness in the dynamics of this modeling framework [74]. 4. Application I: Dynamics of Operational Networks In this section, we consider an application of adaptive network models to socio-technical systems – a special type of complex systems that comprise social and technological components or a combination of both types in one entity [78, 79]. The services provided by socio-technical systems can be categorized into two main functions: (1) detection of a significant event, and (2) execution of an appropriate response action. The second function involves a creation of a new network between system components that will be called upon to execute a response. This new network that dynamically develops on the nodes of the existing network is termed as an operational network [80]. In what follows, an adaptive network-based model of the operational network will be illustrated on an example of the Canadian Arctic Search and Rescue (SAR) system. The case log of a real incident in the Arctic will be used to develop and analyze a sample operational network. 4.1. SARnet – an Adaptive Network Model of the Canadian Arctic SAR System The Canadian Arctic Search and Rescue (SAR) system comprises a large number of highly specialized SAR assets that are trained or designed to provide a comprehensive range of SAR services. These are Canadian Coast Guard (CCG) officers, teams of SAR technicians, Joint Rescue Coordination Centres (JRCCs), aircraft and ships equipped with the crew, and various information and communication systems. A detailed description of the Canadian Arctic SAR system is given in [81]. SARnet is a network model of the Canadian Arctic SAR system that comprises multiple networks with embedded heterogeneous agents, where agents are SAR assets. Unlike agents of other typical social network models, heterogeneous agents of the SAR system cannot easily be re-trained and replace other agents. The agent specialization results in a distinctive pattern of network dynamics, as we elaborate below. SARnet distinguishes between five classes of agents according to their specialization (sensor, router, actor, database, and controller; see Table 2), six environmental realms in which these agents operate (maritime, land, air, space, cyber, and cognitive), and four SAR operational domains according to traditional subdivision of SAR services (Air, Mar- itime, Ground, and Joint SAR). In addition, SARnet represents such agent properties as skill sets, access to resources, home organizations, and technical specifications. The SARnet agent is represented by a string of data of dimension N, i.e., where σi is a binary, categorical or continuous variable that represents a property of the agent. σ = [σ1, σ2, . . . , σN], (1) 7 Figure 2: Various dynamical network models simulated using GNA. These examples were represented in the same format of (cid:104)E, R, I(cid:105) (see text) and simulated using the same simulator package implemented in Mathematica. (a) Simulation of asynchronous 2-D binary cellular automata with von Neumann neighborhoods and local majority rules. Space size: 100×100. (b) Simulation of an asynchronous random Boolean network with N = 30 and K = 2. Time flows from left to right. Nodes of random Boolean networks are non-homogeneous, i.e., they obey different state-transition rules. Here each node’s own state-transition rule is embedded as part of its state, and the replacement mechanism R refers to that information when calculating the next state of a node. (c) Simulation of a network growth model with the Barab´asi-Albert preferential attachment scheme [15]. Time flows from left to right. Each new node is attached to the network with one link. The extraction mechanism E is implemented so that it determines the place of attachment preferentially based on the node degrees, which causes the formation of a scale-free network in the long run. 8 We say that agents belong to the same heterotype if they are identical in the first several key positions of string σ. The distribution of the numbers of agent heterotypes can be used to measure the agent heterogeneity. If K is the number of heterotypes and Xk is a fraction of agents of heterotype k (k = 1, . . . , K), then the network entropy can be defined as follows: (2) In Eq. 2, the network entropy S is normalized by its maximum value S max = ln K. As follows from Eq. 2, S ∈ [0, 1]. The minimum value S = 0 corresponds to a network composed of one heterotype. The maximum value S = 1 corresponds to a network composed of agents evenly distributed between all K heterotypes (i.e. Xk = 1/K for k = 1, . . . , K). As the network entropy approaches 1, the agent distribution between heterotypes becomes uniform. Xk ln Xk. k=1 On a day-to-day basis, SAR assets are connected in a standby network, which represents the standby posture of the system [81]. The operational network dynamically develops on the nodes of the standby network in response to a particular SAR incident. It links SAR assets, which are called upon to provide specified SAR services. A responsible JRCC initiates a SAR response by appointing one of the controller agents, as the Search Master. The Search Master is responsible for the SAR operation in question until closure of the case. The sequence of services, which will be provided after a distress alert is received, follows prescribed protocols and procedures, which serve as a blueprint for tasking SAR assets based on their specialization and availability. The nature and size of the incident (e.g., location of the crash site and number of people on board) also determine the choice of SAR assets being called upon. The dynamics of the operational network differs from that of the standby network, as the architecture of the former evolves at the time scale of minutes or hours instead of months or even years, as in the latter case. We developed the SARnet simulation software, called OpNetSim, for automated generation of operational net- works, which is described in [82]. OpNetSim has its theoretical basis on GNA [73, 74]. The simulation code was developed in Python, and NetworkX [83] was used for network representation and analysis. The network dynamics are described as a set of possible rewriting events. A rewriting event is defined as an establishment of a new link between two agents, possibly involving changes of their states. Each possible event is specified by the following eight properties: 1. Conditions: (Optional) Logical expression(s) that indicate when this event can be executable. 2. Source: Agent from which the new link departs. 3. Destination: Agent to which the new link points. 4. Link type: Type of the event (i.e., interaction between the two agents). The following three types are allowed: K(cid:88) S = − 1 ln K • “Request”: The source agent requests the destination agent for specific information. • “Flow”: The source agent sends specific information to the destination agent. • “Task”: The source agent commands the destination agent to do a particular task. 5. Knowledge required: (Optional) List of internal variables the source agent needs to have in order for the event to occur. 6. Knowledge transferred: (Optional) List of internal variables whose values are requested or shared between the two agents during the “Request” or “Flow” event. 7. Duration: Amount of time the event takes. 8. Duration variation: Amount of stochastic variation for the duration. OpNetSim reads the set of possible rewriting events given in the above format. The algorithm of simulation of this network proceeds in the following steps: 1. Select all the events that are currently executable (i.e., all conditions are met and the source agent has all the knowledge required). 2. Make the selected events active and set a duration time (with stochastic variation added according to the duration variation property of the event) to their respective internal time counters. 3. Decrease the time counters of all of the active events by a unit time. 9 Figure 3: Graphical user interface of OpNetSim. 4. If the counter of any of those active events hits zero, establish a new directed link from the source agent to the destination agent in the network. Also, depending on the type of the event, update the internal variables of both agents. Then deactivate the event. 5. Repeat the process above until no more executable events exist. The operational network will emerge as the simulation progresses and more agents are connected by information exchange and task allocation. OpNetSim implements interactive graphical user interface (GUI) by which the user can operate and inspect the simulation status (Figure 3). 4.2. The December 2008 SAR Incident in the Arctic OpNetSim was used to simulate the operational network of a real SAR incident in the Arctic that occurred in December 2008. On 7 December 2008 a small two-engine Cessna plane with two people on board crash-landed in the Arctic approximately 120 nautical miles (nm) from Iqaluit, Nunavut. Three mayday calls were intercepted by a commercial aircraft and by a Royal Canadian Air Force (RCAF) aircraft, and then relayed to JRCC Halifax. (JRCC Halifax, located in Halifax NS, is a JRCC responsible for that sector of the Arctic.) The Canadian SAR system mounted a response to the incident, which involved three RCAF SAR squadrons, Canadian Coast Guard (CCG) resources (including database resources and marine communication systems), regional units of the Royal Canadian Mounted Police (RCMP) and Civil Air Search and Rescue Association (CASARA), local police, air-ground-air communication systems, and private-sector assets. One of the CCG officers on duty was appointed as the Search Master to coordinate the SAR operation. In less than 18 hours from the time of the first mayday call, the two survivors were rescued (with mild frostbites, otherwise in good condition). The concept of the operational network was used to represent and analyze the operational architecture of the SAR response to this incident, and to identify factors contributing to a successful outcome. The agent heterogeneity was identified as the main driving mechanism for the development of the operational net- work. Such agent attributes as agent class, realm, and SAR domain influenced the formation of network architecture. The agents’ skill sets and access to resources as well as the crash-site information also plays a role in shaping the operational network. The architecture of the resulting network evolved at the time scale of minutes or hours instead of months or even years, as in the case of the standby network. Figure 4 shows the snapshots of the actual operational network one, three and 18 hours after the response initiation, respectively. The number of agent heterotypes was increasing in the course of the SAR response, meaning that the network heterogeneity was also increasing. At the same time, the distribution of agents between heterotypes became less 10 Figure 4: The development of the operational network drawn based on the real SAR incident in the Arctic in December 2008: (a) 1 hour, (b) 3 hours, and (c) 18 hours after the response initiation. Controllers are shown by red, actors by blue, sensors by gray, routers by green, and databases by yellow circles. The Search Master is shown by an enlarged red circle. The ORA network analysis and visualization software [84] was used to visualize the networks. balanced, as follows from a decline in normalized entropy after 6 hours of the response. Table 3 summarizes the development of the operational network after 1, 3, and 18 hours. Table 3: Network development after 1, 3, and 18 hours. 1 hour 28 41 3 hours 40 64 4 1 1.37 6 1 1.48 18 hours 79 140 12 1 1.96 Node count: Link count: Link weight: Max. Min. Average Network composition: Actors: Controllers: Databases: Routers: Sensors: Number of heterotypes: Network entropy (normalized): 7 (25%) 10 (36%) 4 (14%) 5 (18%) 2 (7%) 14 0.92520 10 (25%) 15 (37.5%) 4 (10%) 7 (17.5%) 4 (10%) 17 0.89210 32 (40.5%) 21 (26.5%) 7 (9%) 12 (15%) 7 (9%) 23 0.86274 According to our analysis, all major players were quickly identified and added to the network at early stages of its development. By the end of the first hour of the response, the operational network included 28 agents and 41 links, i.e. 30% of the final network. After the first three hours, 40 agents and 64 links, i.e. nearly 50% of the final net, were in place. By the time when the survivors were rescued (i.e. 18 hours after the response initiation), more than 80% of the operational net had developed. The Search Master (which was an isolate in the standby network) quickly became the most influential entity of the operational network, coming first in all node-level measures, including standard Social Network Analysis measures of degree centrality and extended measures of cognitive demand and shared situation awareness (see [84] for measure definitions). 11 Figure 5: The development of the operational network simulated using OpNetSim. Time flows from left to right. The gray nodes and links represent the standby network, while colored nodes are the activated ones that form the dynamically changing operational network. Visualization is done using Python with NetworkX. Node color schemes are the same as in Figure 4. After 18 hours, the Search Master’s sphere of influence5 encompassed 67% of the entire operational network. For comparison, the sphere of influence of the next most influential node contained about 14% of the network. The Search Master had direct interactions with 44 out of 78 other agents. (In total, 79 agents were included in the operational network.) This value was almost an order of magnitude higher than that of the second-ranked agent. Average com- munication speed between any two nodes within the Search Master’s sphere of influence (or 67% of the network) was close to 0.5, and the average speed with which the Search Master interacted with 67% of all agents was 1.0 – the maximum value for this measure. High centralization of the operational net was identified as a contributing factor to the operational effectiveness. However, it can also be viewed as a vulnerability factor. According to our analysis results, the removal of the Search Master will lead to maximum network fragmentation when almost 80% of SAR assets will become disconnected. Moreover, there was no other entity in the network capable of assuming the leadership role in the SAR response in question. A detailed summary of network analysis results can be found in [81]. We examined the actual log of inter-agent communications during this SAR incident, and manually reconstructed the rewriting rules that drove the operational network formation. OpNetSim was then used to simulate the temporal development of the operational network under several hypothetical scenarios. Figure 5 shows snapshots of the simu- lated operational network produced by OpNetSim. Since the simulation algorithm involves stochasticity, the topology of the simulated network does not exactly match the actual one, but the general trend of increasing agent heterogeneity and concentration on the Search Master node were correctly represented in this model. 5. Application II: Automated Rule Discovery from Empirical Network Evolution Data In the previous section, we developed an adaptive network model based on our knowledge and understanding about local dynamics of node and link interactions. In the meantime, it has remained an open question how one could derive dynamical rules of an adaptive network model directly from a given empirical data of network evolution. In this section, we describe an algorithm that automatically discovers a set of dynamical rules that best captures state transition and topological transformation expressed in the empirical data [85]. Network evolution is formulated using the GNA framework and the subnetwork extraction and replacement phases are analyzed separately. Within the scope of this paper, we will simplify the problem by requiring the data to satisfy the following: 1. A given data set is a series of configurations of labeled directed or undirected networks in which labels (states) and topologies coevolve over discrete time steps (Figure 6(a)). 2. The data set contains information about the correspondence of nodes between every pair of two successive time points (Figure 6(a)). 3. States are discrete, finite, and assigned only to nodes, not to links. 5A Sphere of influence of a node is a sub-network of radius 1 that includes all nodes to whom that node has direct connections plus connections between those nodes. 12 4. Changes that take place between successive time points are reasonably small so that they can be identified as one small network rewriting event per each time step. 5. The extraction mechanism E and the replacement mechanism R are memoryless, i.e., they produce outputs solely based on inputs given to them. We note that the GNA framework has a significant advantage for the algorithm design. It formulates the network evolution using two separate phases, i.e., the extraction of subGNA (performed by E) and its replacement (performed by R). Therefore, the estimation and construction of models of E and R can be conducted independently and concur- rently using separate training data sets, which will make the algorithm simple and tractable. 5.1. Proposed Algorithm A general procedure of the proposed algorithm is as follows (Figure 6): 1. Preprocess the original network evolution data using data-dependent heuristics, if necessary, so that they meet all the aforementioned requirements. 2. Detect the difference between each pair of configurations at two successive time points (Gt, Gt+1) and represent it as a rewriting event st ⇒ rt (Figure 6(b)), where st is a subGNA to be replaced, rt is another subGNA that replaces st, and “⇒” denotes correspondence from nodes in st to nodes in rt. The difference between two configurations (Gt = (cid:104)Vt, Ct, Lt(cid:105), Gt+1 = (cid:104)Vt+1, Ct+1, Lt+1(cid:105)) will be detected in the following way: (a) Let A be a set of nodes in Gt which disappeared in Gt+1 (A = {xx ∈ Vt ∧ x (cid:60) Vt+1}). (b) Let B be a set of nodes in Gt+1 which did not exist in Gt (B = {xx ∈ Vt+1 ∧ x (cid:60) Vt}). (c) Add to A and B all the nodes whose states or neighbors changed between Gt and Gt+1 (D = {xx ∈ Vt ∧ x ∈ Vt+1 ∧ (Ct(x) (cid:44) Ct+1(x) ∨ Lt(x) (cid:44) Lt+1(x))}, A = A ∪ D, B = B ∪ D). At this point, A and B contain the nodes that experienced some changes (enclosed by solid lines in (d) Add to A and B all the nodes which have a link to any of the nodes in A (D(cid:48) = {xx ∈ Vt ∧ x ∈ Vt+1∧ Lt(x)∩ Figure 6(b)). A (cid:44) {∅}}, A = A ∪ D(cid:48), B = B ∪ D(cid:48)). The above step includes in A and B additional nodes that may have influenced the rewriting event (enclosed by dashed lines in Figure 6(b)). (e) Let st and rt be subgraphs of Gt and Gt+1 induced by nodes in A and B, respectively. Then the detected rewriting event is represented as st ⇒ rt , where “⇒” is the set of all the node correspondences between st and rt present in the original data. 3. Construct a model of the extraction mechanism E by using (Gt, st) as training data, where Gt is the input given to E and st the output that E should produce (Figure 6(c),(e)). This step is the most challenging part in this algorithm development effort. The task to be achieved in this step is to identify an unknown mechanism that chooses a subset of a given set of nodes. Exact identification of an unknown computational mechanism is theoretically not possible in general. Here, we will assume several predefined candidate mechanisms (e.g., random selection, preferential selection based on node degrees, motif-based selection, etc.) and calculate the likelihood for each extraction result given in the training data to occur with each candidate mechanism. This calculation will be conducted and multiplied sequentially over the whole training data to evaluate how likely the given training data could result from each of the candidate mechanisms. If a mechanism includes parameters, they will be optimized to attain the maximal probability. Then the mechanism with the highest likelihood will be returned as the estimated mechanism of E. 4. Construct a model of the replacement mechanism R by using (st, st ⇒ rt) as training data, where st is the input given to R and st ⇒ rt the output R should produce (Figure 6(d),(f)). In this step, the task can be achieved in a much simpler manner than in step (3) (though technically it still remains identification of an unknown mechanism). This is because a single rewriting event typically involves just a few nodes so the number of possible inputs given to the replacement mechanism R is virtually finite in contrast to the number of possible inputs to E that is virtually infinite. Therefore we use straightforward pattern matching methods to construct a model of R from the data. Specifically, the algorithm will construct R as a simple procedure that searches for a rewriting event in the training data whose left hand side matches the given 13 Figure 6: Overview of the proposed algorithm for automatic discovery of GNA rewriting rules. (a) Original network evolution data starting with the initial configuration I. (b) Detection of rewriting events at every time step. (c) Training data for the extraction mechanism E. (d) Training data for the replacement mechanism R. (e, f) Construction of models of E and R based on the training data. (g) Final GNA model. 14 input. If there is only one such event found, the event itself will be the output of R. If multiple events are found, the output will be determined either deterministically (e.g., event with greatest frequency) or stochastically (e.g., random selection with weights set proportional to event frequencies). Or, if no event is found, either identity (“input ⇒ input”; no change) will be returned or seek similar events will be sought using partial graph matching schemes. 5. Construct a complete GNA model by combining the results of the above steps (3) and (4) together with the initial configuration I (Figure 6(g)). 5.2. Software Implementation We have designed the details of the algorithm described above and implemented them in Python with NetworkX and GraphML. The software, called PyGNA [86, 87], is designed to automatically discover a set of dynamical rules that best captures both state transition and topological transformation in the data of spatio-temporal evolution of a complex network. PyGNA is still at its alpha stage, but is publicly available from SourceForge.net1. We conducted preliminary experiments applying PyGNA to data generated by abstract adaptive network models, in order to test if it could correctly identify the actual network generation mechanisms used to reproduce the input data. The following four abstract network models were used as inputs to PyGNA: (a) Barabasi-Albert network, grown using the standard degree-based preferential attachment method [15]. (b) “Degree-state” network, grown by degree-based preferential attachment whose mechanism is influenced by the randomly determined state of the newcomer node. The state of the target node could also be altered by the attachment. (c) “State-based” network, grown by repeated random edge addition between a node that has a particular state (shown in red in Figure 7) and any other randomly selected node. New isolated nodes are also continuously introduced to the network, with a randomly selected state. (d) “Forest fire” network, generated by the method proposed in [88]. Figure 7 shows typical results that visually compare the original input networks and the networks reconstructed by PyGNA. For the Barabasi-Albert (a), degree-state (b) and state-based (c) networks, both input and reconstructed networks have visually similar structures. PyGNA also correctly identified that the growth of those networks was determined by degrees (for (a)), degrees and states (for (b)), and states (for (c)). For the forest fire network (d), however, PyGNA failed to capture the unique topological characteristics of the original input network, because of the complexity in the original network generation method. We also quantified the accuracy of the reconstructed network models by measuring the distance of probability distributions of extracted subgraphs between original and simulated networks. Specifically, for the original input data and the reconstructed network simulation results, we counted how many times each of the different kinds of subgraphs was selected for graph rewriting events, and then computed the Bhattacharyya distance [89] between the two distributions, defined as (3) (cid:112) (cid:88) s∈S DB = − ln p(s)q(s), where s is the unique subgraph, S the set of all extracted subgraphs, and p(s) and q(s) the probability distributions of subgraphs extracted for rewriting in the input network and in the reconstructed network, respectively. DB = 0 means the two distributions were exactly the same, while higher value of DB means they are far apart. The results are summarized in Figure 8. For the Barabasi-Albert network (a), the low DB value indicates that the simulated network is indeed very close to the original input network. The DB value was a little higher for the degree- state (b) and state-based (c) networks, but the overall trends of the extracted subgraph distributions were generally in agreement between the input and simulated networks. For the forest fire network, however, the extraction mechanism selected by PyGNA was over-choosing certain subgraphs and was unable to generate many subgraphs seen in the input data, resulting in the apparent topological difference seen in Figure 7. The DB value for this case is therefore substantially larger than the other three cases. 1http://sourceforge.net/projects/gnaframework/ 15 Figure 7: Results of experiments to reconstruct dynamical network models from artificially generated network data. Sizes of nodes represent their relative degrees. t represents the number of iterations. (a) Barabasi-Albert network. (b) Degree-state network. (c) State-based network. (d) Forest fire network. See text for details. 16 (a) (b) (c) (d) Figure 8: Accuracy of reconstructed network models measured using the Bhattacharyya distance (DB) of probability distributions of extracted subgraphs between original and simulated networks. Horizontal axes represent different kinds of extracted subgraphs. (a) Barabasi-Albert network. (b) Degree-state network. (c) State-based network. (d) Forest fire network. See text for details. 17 (a) (b) (c) (d) These preliminary results tell us that the current algorithm in PyGNA is effective for certain types of networks while still limited for the analysis of others, especially those that involve pure randomness and/or mesoscopic topo- logical structures such as motifs. We are currently revising and expanding our algorithm by addressing these issues in order to improve the performance of PyGNA. 6. Application III: Cultural Integration in Corporate Merger The final example we present is a computational model of cultural integration taking place on a dynamically changing adaptive social network when two firms merge into one. This example is more complex than the previous two, firstly because the model has continuous link weights that adaptively change due to node state dynamics, but more importantly because the node states are far more complex than in the previous two models, in order to represent complex sociocultural aspects of agents. In this sense, this example can be better understood as a hybrid of agent-based models and adaptive network models. It is recognized that cultural integration, or sharing a common corporate culture, is crucial for the success of cor- porate mergers. However, previous studies have been limited to firm-level analyses only, while cultural adoption and diffusion in a merged firm actually occurs among individuals. We thus explored, using the computational model, how cultural integration emerges from the patterns of dynamic social interactions among individuals [71]. Our computer simulation model is an agent-based model operating on a dynamic network structure, where individuals (nodes) ex- change elements of a corporate culture with others who are connected to it through social ties (links). In this model, we set two merging firms, A and B, each consisting of 50 individuals. Our goal is to find initial network structures that promote or impede post-merger cultural integration. Although the number of individuals in the firms is by far smaller than that of publicly traded firms, we found that this parameter has a negligible impact on the simulation results when network density is kept at the same level. 6.1. Representations of Corporate Cultures We represent a corporate culture as a vector in a multi-dimensional continuous cultural space. The cultural space is composed of several cultural dimensions; each dimension represents an element of a corporate culture. We set 10 cultural dimensions for the cultural space; this number is founded on previous empirical studies of corporate culture. For example, O’Reilly et al. [90], who investigated eight large U.S. public accounting firms, found eight dimensions of organizational cultures: innovation, attention to detail, outcome orientation, aggressiveness, supportiveness, em- phasis on rewards, team orientation, and decisiveness. Likewise, Chatterjee et al. [91] measured cultural distance perceived by the top management teams of acquired firms across seven dimensions of organizational cultures: innova- tion and action orientation, risk-taking, lateral integration, top management contact, autonomy and decision making, performance orientation, and reward orientation. Therefore, setting 10 dimensions as elements of corporate cultures would be a more conservative approach. In our model, we characterize the distance between two cultures by the Euclidean distance between two vectors in the cultural space. The average cultural difference between the two merging firms is characterized as the average cultural distance between two individuals—one in Firm A and the other in Firm B. If the value of this measurement is large, the corporate culture that individuals perceive in Firm A is, on average, far different from that in Firm B. We initialized the individual cultural vectors as follows: First, two cultural “center” vectors were created for the two merging firms, and these center vectors were separated by 3.0 (in an arbitrary unit) in the cultural space. Then individual cultural vectors were created for individuals in each firm by adding a small random number drawn from a normal distribution with a mean of 0 and a standard deviation of 0.1 (in the same unit used above) to each component of the cultural center vector of that firm. This setting creates an initial condition where the average between-firm cultural difference is approximately seven times larger than the average within-firm cultural difference. 6.2. Adaptive Changes of Cultural States and Tie Strengths Individuals in our model are connected to each other through directed social ties. A tie going from one individual to another works as a conduit that can transmit, from the origin node to the destination node, information and knowledge that include the elements of their corporate cultures. Each tie has a weight associated with it, called tie strength in the social network literature [92, 93]. The range of possible tie strength values is bounded between 0 and 1. Corporate 18 Figure 9: Adaptive changes of tie strength caused by cultural acceptance or rejection (from [71]). cultures diffuse among individuals through their ties. The algorithm for simulating the dynamics of cultural diffusion, and subsequent social network changes, is as follows. One iteration in a simulation consists of simulations of individual actions for all individuals in a sequential order (therefore there are always 100 individual actions simulated in each iteration). When it is its turn to take an action, an individual first selects an information source. For 99% of the time, the individual chooses the information source from its local in-neighbors, that is, the nodes from which directed ties are coming to the individual. The probability for a neighbor to be selected as the information source is proportional to the strength of the tie that connects the neighbor to the individual; this represents that individuals tend to listen more often to others whom they trust more or with whom they have stronger connections. Otherwise (with a 1% chance), the individual chooses as the information source any individual in the connected component in which the individual belongs. If there is no existing tie from the randomly selected source to the individual, a new tie with a very weak strength (0.01) will be created between them. This represents an informal, incidental communication, like a “water-cooler” conversation within an organization. Once the information source is selected, the individual receives the source’s cultural vector and then measures the distance between the received cultural vector and its own cultural vector. With a probability that decreases monotoni- cally with increasing cultural distance, the individual accepts the received culture. The probability of acceptance, PA, is mathematically represented as (cid:33)d/dc (cid:32)1 2 PA(d) = , (4) where d is the distance between the two cultural vectors and dc is the characteristic cultural distance at which PA becomes 50%. We used dc = 0.5 for our simulations. If the individual accepts the received cultural vector, it adopts the mean of the two vectors (i.e., the sum of the two vectors divided by 2) as its new cultural vector, and the strength of the tie from the source to the individual is increased by the following formula: S new = logistic(logit(S current) + 1) (5) Here S current and S new are the current and updated tie strengths, respectively (this formula guarantees that the tie strength is always constrained between 0 and 1). On the other hand, if the individual rejects the received cultural vector, its own vector will not change, and the tie strength is decreased by the following formula: S new = logistic(logit(S current) − 1) (6) The mechanism of the update of tie strength caused by cultural acceptance or rejection is illustrated in Figure 9. If the tie strength falls below 0.01, the tie is considered insignificant and is removed from the social network. 19 6.3. Initial Network Structures We set the network structures within and between merging firms so that there are substantially more within-firm ties than between-firm ties at the beginning of each simulation. The number of ties within each merging firm is 490. Since the number of individuals in each firm is 50, the network density of the firm is 490/(50*49) = 0.2. The number of ties from one merging firm to the other (that is, A→B or B→A) is 50 for each direction. All tie strengths of those connections are initialized using random numbers drawn from a uniform distribution between 0 and 1. Pw(i) ∼ (i/n)w In our computational experiments, we set two experimental parameters that control topological characteristics of the initial social network among individuals. One is what we call the within-firm concentration, denoted by variable w. This parameter determines the probability for each individual to be selected as an information source of a within-firm tie. It is mathematically defined as (7) where i is the ID of the individual within a firm, n the firm size (n = 50 in our simulations), and Pw(i) the probability for individual i to be selected as an information source when within-firm ties are initially created. When w = 0, within-firm ties are uniformly distributed within the firm so that the organizational structure of the firm is “flat”. For larger w, the within-firm information sources are more concentrated on a small number of individuals with greater IDs, which represent a highly centralized organizational structure of the firm, such as that with a one-man CEO. In our model, we used w = 1, 3, 5, 10, 20, and 30. (i = 1, 2, . . . , n), The other experimental parameter is what we call the between-firm concentration, denoted by variable b. This parameter determines the probability for each individual to be selected as either an origin or a destination of a between- firm tie. It is mathematically defined as Pb(i) ∼ cb i (i = 1, 2, . . . , n), (8) where i and n are the same as in the previous formula, ci the within-firm closeness centrality of individual i, and Pb(i) the probability for individual i to be selected as a connecting person, either as origin or destination, when between-firm ties are created, which is done only after all the within-firm ties have been created. When b = 0, between-firm ties randomly connect individuals across firms, regardless of their social positions. For larger b, the between-firm ties are more concentrated on a small number of individuals with higher centralities that represent the formation of top-level (only) inter-firm communication channels. In our model, we used b = 0.1, 0.5, 1, 3, and 5. Figure 10 illustrates images of within-firm and between-firm concentrations. Note that the above two parameters affect only the initial social network structure. As cultural integration pro- gresses, the network topologies will change dynamically in our simulations. 6.4. Outcome Measurements and Results As a primary dependent variable of our computational experiments, we measure the average cultural distance between individuals who used to belong to different pre-merger firms and who still remain in the largest connected component of the social network. If the average cultural distance decreases from its initial value, cultural integration proceeds among individuals in the merged firm. Likewise, we use three measures of the consequences of cultural integration: turnover, interpersonal conflict, and organizational communication ineffectiveness. All the measures should influence overall firm performance. Turnover is measured by the number of individuals in the simulations who do not stay in the largest connected component of the social network. In our model, if an individual terminates all ties with his neighbors, he is considered to have left the merged firm. Interpersonal conflict is calculated as the cultural distance across a social tie between two individuals, multiplied by their tie strength. This quantity is summed up for all the tied pairs of individuals within the largest connected component. Since tie strength can be considered to represent communication frequency [92], individuals who are strongly tied to neighbors with different perceptions of corporate culture would often encounter greater communication conflict in the workplace. Lastly, organizational communication ineffectiveness is calculated by the cultural distance across a social tie be- tween two individuals, multiplied by the edge betweenness of the tie between them. This quantity is, again, summed up for all the tied pairs of individuals within the largest connected component. Edge betweenness is defined as the 20 Within-firm concentration Between-firm concentration Figure 10: Illustrations of within-firm and between-firm concentrations. The directions of ties are omitted in this figure for clarity (from [71]). number of geodesics (shortest paths) going through an edge [93]. If a tie with high edge betweenness is filled with cultural conflict, most communication between individuals in a firm would be conflicted. As a result, information and knowledge transfer in the firm would be delayed or impeded. We implemented the simulation model and analysis tools by using Python with NetworkX. The program codes of the model are available from the authors upon request. We set 200 time steps in one simulation. We ran 50 simulations for each experimental condition and conducted statistical analysis of the generated simulation results. Figure 11 plots the results showing the effects of within-firm and between-firm concentrations. The highest level of cultural integration is achieved when social ties are more centralized within each merging firm and the social ties between the merging firms are less concentrated on central individuals. Additionally, the results show that within-firm and between-firm network structures significantly affect individual turnover, interpersonal conflict and organizational communication ineffectiveness, and that these three outcome measurements do not vary in tandem. The most turnovers were observed when within-firm concentration was high while between-firm concentration was low, which is the same condition as that promoted cultural integration. Interpersonal conflict was highest when within-firm concentration was low, without much interaction with between-firm concentration. Organizational communication ineffectiveness was highest when both within- and between-firm concentrations were high. For more detailed discussion of the results, see [71]. Note that those findings were all the outcomes of the adaptive changes of social ties in our model. Results would be different if the social network structure was fixed like in other, more typical opinion spreading network models. 7. Conclusions As briefly reviewed above, the co-evolution of network states and topologies is an emerging research topic that has great potential and applicability to many real-world complex systems. It combines two separate dynamics, i.e., dynamic state changes on a network and topological transformations of a network, into a single picture that will allow 21 Figure 11: Cultural distance and organizational dysfunctions by within-firm and between-firm concentrations obtained through simulations of our adaptive network model of corporate merger (from [71]). one to better understand and represent the nature of evolving complex systems, possibly leading to new properties that were not discovered before. The application areas of adaptive networks are now expanding to various disciplines, not only social sciences and operations research (as demonstrated by a few examples in this paper) but also biology, ecology and physical sciences. The key challenges in adaptive network research include (1) how to generate meaningful dynamical models from large-scale temporal network data, and (2) how to mathematically analyze the dynamics of adaptive networks in which the time scales of state changes and topological transformations are inseparable. We hope that the work reviewed in this paper helps indicate the future directions of this exciting field of study. This material is based upon work supported by the US National Science Foundation under Grant No. 1027752. The development of OpNetSim was supported by the Canadian Government contract W7714-125419. References [1] R. Albert, A.-L. Barab´asi, Statistical mechanics of complex networks, Rev. Mod. Phys. 74 (2002) 47-97. [2] M.E.J. Newman, The structure and function of complex networks, SIAM Review 45 (2003) 167-256. [3] M. Newman, A.-L. Barab´asi, D.J. Watts (Eds.), The Structure and Dynamics of Networks, Princeton University Press, Princeton, NJ, 2006. [4] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D.-U. Hwang, Complex networks: structure and dynamics, Physics Reports 424 (2006) [5] S.N. Dorogovtsev, A.V. Goltsev, J.F.F. Mendes, Critical phenomena in complex networks, Rev. Mod. Phys. 80 (2008) 1275-1335. [6] R. Pastor-Satorras, A. Vespignani, Epidemic spreading in scale-free networks, Phys. Rev. Lett. 86 (2001) 3200-3203. [7] A. Pikovsky, M. Rosenblum, J. Kurths, Synchronization: A Universal Concept in Nonlinear Sciences, Cambridge University Press, Cam- 175-308. bridge, UK, 2003. [8] V. Sood, S. Redner, Voter model on heterogeneous graphs, Phys. Rev. Lett. 94 (2005) 178701. [9] H. Zhou, R. Lipowsky, Dynamic pattern evolution on scale-free networks, PNAS 102 (2005) 10052-10057. 22 [10] M. Tomassini, Generalized automata networks, in: Cellular Automata: Proc. 7th Intl. Conf. Cellular Automata for Research and Industry (ACRI 2006), Springer, Berlin/Heidelberg, 2006, pp. 14-28. [11] S.B. Otto, B.C. Rall, U. Brose, Allometric degree distributions facilitate food-web stability, Nature 450 (2007) 1226-1229. [12] H. Nakao, A.S. Mikhailov, Turing patterns in network-organized activator-inhibitor systems, Nature Physics 6 (2010) 544-550. [13] D.J. de Solla Price, Networks of scientific papers, Science 149 (1965) 510-515. [14] D.J. Watts, S.H. Strogatz, Collective dynamics of ‘small-world’ networks, Nature 393 (1998) 440-442. [15] A.-L. Barab´asi, R. Albert, Emergence of scaling in random networks, Science 286 (1999) 509-512. [16] B. Shargel, H. Sayama, I.R. Epstein, Y. Bar-Yam, Optimization of robustness and connectivity in complex networks, Phys. Rev. Lett. 90 [17] L. da Fontoura Costa, Reinforcing the resilience of complex networks, Phys. Rev. E 69 (2004) 066127. [18] P.L. Krapivsky, S. Redner, Network growth by copying, Phys. Rev. E 71 (2005) 036118. [19] A. Beygelzimer, G.M. Grinstein, R. Linsker, I. Rish, Improving network robustness by edge modification, Physica A 357 (2005) 593-612. [20] T. Gross, B. Blasius, Adaptive coevolutionary networks: a review, J. R. Soc. Interface 5 (2008) 259-271. [21] T. Gross, H. Sayama (Eds.), Adaptive Networks: Theory, Models and Applications, Springer, Berlin/Heidelberg, 2009. [22] S. Bornholdt, T. Rohlf, Topological evolution of dynamical networks: global criticality from local dynamics, Phys. Rev. Lett. 84 (2000) [23] K. Christensen, R. Donangelo, B. Koiller, K. Sneppen, Evolution of random networks, Phys. Rev. Lett. 81 (1998) 2380-2383. [24] D.R. Chialvo, P. Bak, Learning from mistakes, Neuroscience 90 (1999) 1137-1148. [25] B.A. Pearlmutter, C.J. Houghton, Tuning for criticality: a new hypothesis for sleep, Neural Comput. 21 (2009) 1622-1641. [26] A. Levina, J.M. Herrmann, T. Geisel, Phase transitions towards criticality in a neural system with adaptive interactions, Phys. Rev. Lett. 102 (2003) 068701. 6114-6117. (2009) 118110. 5 (2009) e1000314. 2 (2010) 155-164. Interface 7 (2010) 1247-1256. explorations, PLOS ONE 3(12) (2008) e3955. [27] C. Meisel, T. Gross, Adaptive self-organization in a realistic neural network model, Phys. Rev. E 80 (2009) 061917. [28] J. Beggs, D. Plenz, Neuronal avalanches in neocortical circuits, J. Neurosci. 23 (2003) 11167-11177. [29] M. Kitzbichler, M. Smith, S. Christensen, E. Bullmore, Broadband criticality of human brain network synchronization, PLoS Comput. Biol. [30] C. Meisel, A. Storch, S. Hallmeyer-Elgner, E. Bullmore, T. Gross, Failure of adaptive self-organized criticality during epileptic seizure attacks, PLoS Comput. Biol. 84 (2012) e1002312. [31] T. Gross, C. Dommar D’Lima, B. Blasius, Epidemic dynamics on an adaptive network, Phys. Rev. Lett. 96 (2006) 208701. [32] T. Gross, I.G. Kevrekidis, Robust oscillations in SIS epidemics on adaptive networks: coarse graining by automated moment closure, Euro- [33] V. Marceau, P.-A. Noel, L. H´ebert-Dufresne, A. Allard, L.J. Dube, Adaptive networks: coevolution of disease and topology, Phys. Rev. E 82 [34] B. Guerra, J. Gomez-Gardenes, Annealed and mean-field formulations of disease dynamics on static and adaptive networks, Phys. Rev. E 82 [35] S. Wieland, T. Aquino, A. Nunes, The structure of coevolving infection networks, Europhys. Lett. 97 (2012) 18003. [36] L. Wang, J.-R. Yan, J.-G. Zhang, Z.-R. Liu, Controlling disease spread on networks with feedback mechanism, Chinese Phys. 16 (2007) phys. Lett. 82 (2008) 38004. (2010) 036116. (2010) 035101. 2498. [37] L.B. Shaw, I.B. Schwartz, Fluctuating epidemics on adaptive networks, Phys. Rev. E 77 (2008) 066101. [38] D.H. Zanette, S. Risau-Gusm´an, Infection spreading in a population with evolving contacts, J. Biol. Phys. 34 (2008) 135-148. [39] E. Volz, S.D.W. Frost, R. Rothenberg, L.A. Meyers, Epidemiological bridging by injection drug use drives an early HIV epidemic, Epidemics [40] S. Funk, M. Salath´e, V. A. A. Jansen, Modelling the influence of human behaviour on the spread of infectious diseases: a review, J. R. Soc. [41] J. M. Epstein, J. Parker, D. Cummings, R. A. Hammond, Coupled contagion dynamics of fear and disease: mathematical and computational [42] S. Funk, E. Gilad, C. Watkins, V. A. A. Jansen, The spread of awareness and its impact on epidemic outbreaks, PNAS 106 (2009) 6872-6877. [43] L.B. Shaw, I.B. Schwartz, Enhanced vaccine control of epidemics in adaptive networks, Phys. Rev. E 81 (2010) 046120. [44] P. Holme, M.E.J. Newman, Nonequilibrium phase transition in the coevolution of networks and opinions, Phys. Rev. E 74 (2006) 056108. [45] D.H. Zanette, S. Gil, Opinion spreading and agent segregation on evolving networks, Physica D 224 (2006) 156-165. [46] B. Kozma, A. Barrat, Consensus formation on adaptive networks, Phys. Rev. E 77 (2008) 016102. [47] F. Vazquez, V.M. Egu´ıluz, M. San Miguel, Generic absorbing transition in coevolution dynamics, Phys. Rev. Lett. 100 (2008) 108702. [48] D. Kimura, Y. Hayakawa, Coevolutionary networks with homophily and heterophily, Phys. Rev. E 78 (2008) 016103. [49] G.A. Bohme, T. Gross, Analytical calculation of fragmentation transitions in adaptive networks, Phys. Rev. E 83 (2011) 035101. [50] G. Zschaler, G.A. Bohme, M. Seissinger, C. Huepe, T. Gross, Early fragmentation in the adaptive voter model on directed networks, Phys. Rev. E 85 (2012) 046107. Conflict Resolution 51 (2007) 905-929. [51] D. Centola, J. C. Gonzalez-Avella, V. M. Eguiluz, M. San Miguel, Homophily, cultural drift, and the co-evolution of cultural groups, J. [52] D. Centola, The spread of behavior in an online social network experiment, Science 329 (2010) 1194-1197. [53] D. Centola, An experimental study of homophily in the adoption of health behavior, Science 334 (2011) 1269-1272. [54] C. Huepe, G. Zschaler, A.-L. Do, T. Gross, Adaptive-network models of swarm dynamics, New J. Phys. 13 (2011) 073022. [55] I.D. Couzin, C.C. Ioannou, G. Demirel, T. Gross, C.J. Torney, A. Hartnett, L. Conradt, S.A. Levin, N.E. Leonard, Uninformed individuals promote democratic consensus in animal groups, Science 334 (2011) 1578-1580. [56] M. Paczuski, K.E. Bassler, A. Corral, Self-organized network of competing Boolean agents, Phys. Rev. Lett. 84 (2000) 3185-3188. [57] B. Skyrms, R. Pemantle, A dynamic model of social network formation, PNAS 97 (2000) 9340-9346. [58] M.G. Zimmermann, V.M. Egu´ıluz, M. San Miguel, A. Spadaro, Cooperation in an adaptive network, Adv. Complex Syst. 3 (2000) 283-297. 23 [59] H. Ebel, S. Bornholdt, Evolutionary games and the emergence of complex networks, arXiv:cond-mat/0211666 (2002). Accessed on [60] J.M. Pacheco, A. Traulsen, M.A. Nowak, Coevolution of strategy and structure in complex networks with dynamical linking, Phys. Rev. Lett. [61] S. van Segbroeck, F.C. Santos, T. Lenaerts, J.M. Pacheco, Selection pressure transforms the nature of social dilemmas in adaptive networks, 8/30/2012. 97 (2006) 258103. New J. Phys. 13 (2011) 013007. [62] J. Poncela, J. Gomez-Gardenes, A. Traulsen, Y. Moreno, Evolutionary game dynamics in a growing structured population, New J. Phys. 11 [63] G. Zschaler, A. Traulsen, T. Gross, A homoclinic route to full cooperation in adaptive networks and its failiure, New J. Phys. 12 (2010) [64] V. Bala, S. Goyal, Conformism and diversity under social learning, Econ. Theory 17 (2001) 101-120. [65] P. Holme, G. Ghoshal, Dynamics of networking agents competing for high centrality and low degree, Phys. Rev. Lett. 96 (2006) 098701. [66] A.-L. Do, L. Rudolf, T. Gross, Patterns of cooperation: fairness and coordination in networks of interacting agents, New J. Phys. 12 (2010) (2009) 083031. 093015. 063023. [67] R. S. Burt, Structural Holes: The Social Structure of Competition, Harvard University Press, 1995. [68] V. Buskens, A. Van de Rijt, Dynamics of networks if everyone strives for structural holes, Amer. J. Sociol. 114 (2008) 371-407. [69] S.D., Dionne, H. Sayama, C. Hao, B.J. Bush, The role of leadership in shared mental model convergence and team performance improvement: an agent-based computational model, Leadership Quart. 21 (2010) 1035-1049. [70] Y. Lin, K.C. Desouza, Co-evolution of organizational network and individual behavior: an agent-based model of interpersonal knowledge transfer, in: ICIS 2010 Proc. (2010) paper 153. Available online at http://aisel.aisnet.org/icis2010 submissions/153. Accessed on 8/30/2012. [71] J. Yamanoi, H. Sayama, Post-merger cultural integration from a social network perspective: A computational modeling approach, Computa- tional and Mathematical Organization Theory, in press. [72] T. Gross, Adaptive networks website, http://adaptive-networks. wikidot.com/. Accessed on 8/30/2012. [73] H. Sayama, Generative network automata: a generalized framework for modeling complex dynamical systems with autonomously varying topologies, in: Proc. IEEE Symp. Artif. Life, 2007, pp. 214-221. [74] H. Sayama, C. Laramee, Generative network automata: a generalized framework for modeling adaptive network dynamics using graph rewritings, in: T. Gross, H. Sayama (Eds.), Adaptive Networks: Theory, Models and Applications, Springer, Berlin/Heidelberg, 2009, pp. 311-332. [75] G. Rozenberg (Ed.) Handbook of Graph Grammars and Computing by Graph Transformation, Volume 1: Foundations, World Scientific, Singapore, 1997. 719-739. [76] W. Kurth, O. Kniemeyer, G.H. Buck-Sorlin, Relational growth grammars – a graph rewriting approach to dynamical systems with a dynamical structure, in: Proc. 2004 Intl. Workshop on Unconventional Programming Paradigms, Springer, Berlin/Heidelberg, 2005, pp. 56-72. [77] C.L. Nehaniv, Asynchronous automata networks can emulate any synchronous automata network, Intl. J. Algebra & Comput. 14 (2004) [78] I. Pestov, Defence and security applications of network technologies, in: Proc. Workshop on Link Analysis, Counterterrorism and Security, SIAM Int. Conf. on Data Mining, 2009, pp. 1-7. Defence Applications (CISDA 2009), 2009, 8 pp. Defence Applications (CISDA 2011), 2011, 8 pp. TM 2012-067, Ottawa, Canada, 2012. Visualization Methods (MSV 2012), 2012, 7 pp. [79] I. Pestov, S. Verga, Dynamical networks as a tool for system analysis and exploration, in: Proc. IEEE Symp. Comput. Intell. in Security and [80] I. Pestov, M. Pilat, Modelling search and rescue systems with dynamical networks, in: Proc. IEEE Symp. Comput. Intell. in Security and [81] I. Pestov, M. Pilat, A net-enabled approach to modelling joint interagency systems: the Arctic search and rescue network, DRDC CORA [82] I. Pestov, H. Sayama, C. Wong, Modeling discrete distributed heterogeneous systems, in: Proc. Int. Conf. on Modeling, Simulation and [83] A.A. Hagberg, D.A. Schult, P.J. Swart, Exploring network structure, dynamics, and function using NetworkX, in: G. Varoquaux, T. Vaught, J. Millman (Eds.), Proc. 7th Python in Science Conf., 2008, pp. 11-15. [84] K.M. Carley, D. Columbus, M. DeReno, J. Reminga, I.C. Moon, ORA User’s Guide 2008, The CASOS Technical Report CMU-ISR-08-125, Carnegie Mellon University, Pittsburgh, PA, 2008. [85] H. Sayama, An algorithm for automatically discovering dynamical rules of adaptive network evolution from empirical data, in: J. Suzuki, T. Nakano (Eds.), Proc. 5th Intl. ICST Conf. on Bio-Inspired Models of Network, Information, and Computing Systems (BIONETICS 2010), Springer, Berlin/Heidelberg, 2012, pp. 497-504. [86] J. Schmidt, B.J. Bush, H. Sayama, A Python implementation of generative network automata, in: Unifying Themes in Complex Systems Volume VIII: Proc. Eighth Intl. Conf. on Complex Systems (ICCS 2011), NECSI Knowledge Press, Cambridge, MA, 2011, pp. 439-440. [87] J. Schmidt, PyGNA: Designing and Evaluating Algorithms for Automated Discovery of Adaptive Network Models Based on Generative Network Automata, Master’s Thesis, Binghamton University, State University of New York, December 2012. [88] J. Leskovec, J. Kleinberg, C. Faloutsos, Graphs over time: densification laws, shrinking diameters and possible explanations, in: Proc. Eleventh ACM SIGKDD Intl. Conf. on Knowledge Discovery in Data Mining, ACM, New York, 2005, pp. 177-187. [89] A. Bhattacharyya, On a measure of divergence between two statistical populations defined by their probability distributions, Bull. Calcutta [90] C.A. O’Reilly, J. Chatman, D.F. Caldwell, People and organizational culture: a profile comparison approach to assessing person-organization [91] L. Chatterjee, M. Lubatkin, D. Schweiger, Y. Weber. Cultural differences and shareholder value in related mergers: linking equity and human Math. Soc. 35 (1943) 99-109. fit, Acad. Management J. 34 (1991) 487-516. capital, Strategic Management J. 13 (1992) 319-334. [92] M. Granovetter, The strength of weak ties, Amer. J. Sociol. 78 (1973) 1360-1380. [93] S. Wasserman, K. Faust, Social Network Analysis: Methods and Applications, Cambridge University Press, Cambridge, UK, 1994. 24
1312.7742	1	1312	2013-12-27T12:12:32	Information Spreading on Almost Torus Networks	[ "cs.SI", "physics.soc-ph" ]	Epidemic modeling has been extensively used in the last years in the field of telecommunications and computer networks. We consider the popular Susceptible-Infected-Susceptible spreading model as the metric for information spreading. In this work, we analyze information spreading on a particular class of networks denoted almost torus networks and over the lattice which can be considered as the limit when the torus length goes to infinity. Almost torus networks consist on the torus network topology where some nodes or edges have been removed. We find explicit expressions for the characteristic polynomial of these graphs and tight lower bounds for its computation. These expressions allow us to estimate their spectral radius and thus how the information spreads on these networks.	cs.SI	cs	Information Spreading on Almost Torus Networks Antonia Maria Masucci ∗1 and Alonso Silva †2 1ETIS/ENSEA - Universit´e de Cergy Pontoise - CNRS, 6, avenue du Ponceau, 95014 2Alcatel-Lucent Bell Labs France, Centre de Villarceaux, Route de Villejust, 91620 Nozay, Cergy-Pontoise, France France Abstract Epidemic modeling has been extensively used in the last years in the field of telecommunications and com- puter networks. We consider the popular Susceptible- Infected-Susceptible spreading model as the metric for information spreading. In this work, we analyze information spreading on a particular class of net- works denoted almost torus networks and over the lattice which can be considered as the limit when the torus length goes to infinity. Almost torus networks consist on the torus network topology where some nodes or edges have been removed. We find explicit expressions for the characteristic polynomial of these graphs and tight lower bounds for its computation. These expressions allow us to estimate their spectral radius and thus how the information spreads on these networks. 1 Introduction There exists an abundant literature on epidemic mod- eling and in particular on epidemics on networks (see e.g., the books [1, 2, 3] and references therein). In the last decades, epidemic modeling has been extensively used in the field of telecommunications and computer networks. For example, epidemic models have been applied in order to analyze the spread of computer ∗Email: [email protected] †Email: [email protected] To whom corre- spondence should be addressed. viruses and worms [4], they have also been applied on epidemic routing for delay tolerant networks [5], etc. The structure of networks plays a critical role in the spread of a viral message and the authors of [6, 7, 8] have identified some conditions for successful perfor- mance of viral marketing. In particular, the authors of [7] have shown that the spectral radius of the graph determines the epidemic lifetime and its coverage. More recently on [9], the authors have studied the intertwined propagation of two competing "memes" (or viruses, rumors, etc.) in a composite network (individual agents are represented across two planes e.g. Facebook and Twitter). The authors have re- alized that the meme persistence δ and the meme strength β on each plane, and the spectral radius of the graph, completely determine which of the two competing "memes" prevail. The topologies of focus in this work, torus net- work topologies, are commonly used in the produc- tion of high-end computing systems [10]. A num- ber of supercomputers on the TOP500 list use three- dimensional torus networks [11]. For instance, IBM's Blue Gene L [12, 13] and Blue Gene P [14]; Cray's XT and XT3 [15] systems use three-dimensional torus networks for node communication. IBM Blue Gene/Q uses a five-dimensional torus network [16]. Fujitsu's K computer and the PRIMEHPC FX10 use a proprietary six-dimensional torus interconnect called Tofu [17]. Torus networks are used because of a combination of their linear per node cost scaling and their competitive overall performance. 1 In this work, we analyze information spreading on a particular class of networks denoted almost torus networks where we assume the popular Susceptible- Infected-Susceptible model as model of information spreading. Almost torus networks consist on the torus network topology where some nodes or edges have been removed. This situation can model the failure of some computer nodes or connections be- tween the computer nodes. As we will see, in those graphs, the study of the spectral radius of the graph is determinant in order to analyze the information spreading. We find explicit expressions for the char- acteristic polynomial of these graphs and very tight lower bounds for its computation. These expressions allow us to estimate their spectral radius and thus how the information spreads on these networks. The outline of the work is as follows. In Section 2, we recall some preliminary notions of graph theory. In Section 3, we present the Susceptible-Infected- Susceptible model of information spreading. Then, in Section 4, we analyze the information spreading in the almost torus network where one node has been removed and in Section 5 we extend our results to the cases when a set of nodes has been removed and when an edge has been removed from the torus network. In Section 6, we provide lower bounds for the two- dimensional torus network. In Section 7, we present numerical results that validate our analysis. Finally, in Section 8 we conclude. 2 Preliminaries Let G = (V,E) denote an undirected graph with no self-loops. We denote by V = V(G) = {v1, . . . , vn} the set of nodes and by E = E(G) ⊆ V × V the set of undirected edges of G. If {vi, vj} ∈ E(G) we call nodes vi and vj adjacent (or neighbors), which we denote by vi ∼ vj. We define the set of neighbors of node v as Nv = {w ∈ V : {v, w} ∈ E}. The degree of a node v, denoted by degv, corresponds to the num- ber of neighbors of v, i.e. the cardinality of the set Nv. We define a walk of length k from v0 to vk to be an ordered sequence of nodes (v0, v1, . . . , vk) such that vi ∼ vi+1 for i = 0, 1, . . . , k − 1. If v0 = vk, then the walk is closed. Graphs can be algebraically represented via matri- ces. The adjacency matrix of an undirected graph G, denoted by A = A(G), is an n × n symmetric matrix defined entry-wise as Aij = (cid:26) 1 0 if vi and vj are adjacent, otherwise. We recall the well-known result that for k ∈ N, Ak ij is the number of paths of length k connecting the i- th and j-th vertices (proof by induction). Since A0 is the identity matrix, we thus accept the existence of walks of length zero. We use I to denote the identity matrix, where its order is determined by the context. We define the Laplacian matrix L for graphs with- out loops or multiple edges, as follows: degvi −1 0 Lij =   if vi = vj, if vi and vj are adjacent, otherwise. We notice that the Laplacian of a graph can be writ- ten as L = D − A where D is a diagonal matrix whose diagonal entries correspond to the degree of each node and A is the adjacency matrix. The spectral radius of a graph G, denoted ρ(A), is the size of the largest eigenvalue (in absolute value) of the adjacency matrix of the graph, i.e. ρ(A) = maxi(λi). Since A is a symmetric matrix with non- negative entries, all its eigenvalues are real. The char- acteristic polynomial of G, denoted φ(G, x) is defined as det(xI − A), that corresponds to the character- istic polynomial of the adjacency matrix A. The walk generating function1 W (G, x) is defined to be (I − xA)−1. The ij-entry of W (G, x) will be written as Wij (G, x). If S is a subset of V(G) then G \ S is the subgraph of G induced by the vertices not in S. We normally write G \ i instead of G \ {i} and G \ ij instead of G \ {i, j}. 1This can be viewed indifferently as a matrix with rational functions as entries, or as a formal power series in x over the ring of all polynomials in the matrix A. 2 3 Model of Information Spreading We use the popular Susceptible-Infected-Susceptible (SIS) model of viral spreading [18] as the metric for information spreading. We remark that our results can be easily extended to the Susceptible-Infected1- Infected2-Susceptible (SI1I2S) model [9] of spreading over composite networks. We consider that each node can be in two possible states: susceptible (of being infected) or infected. We denote these two states as S and I, respectively. The result presented on this section was first found in [19] and [20], through mean-field approximations of the Markov chain evolution of the 2n possible states. We believe this alternative proof to be simpler and we presented here for completeness. Consider a population of n nodes interconnected via an undirected graph G = (V,E). Time is slotted. In each time slot, infected nodes attempt to contami- nate their susceptible neighbors, where each infection attempt is successful with a probability β, indepen- dent of other infection attempts. The parameter β is called the virus birth rate (or meme strength as denoted in [9]). Each infected node recover in time slot t with probability δ. The parameter δ is called virus curing rate (or meme persistence). We notice that this parameter captures the meme persistence in an inverse way, i.e. a high δ means low persistence. This means a very contagious virus will be modeled with a low δ value. We define pi,t as the probability that node i is in- fected at time t. We define ζi,t+1 as the probability that node i will not receive infections from its neigh- bors in the next time-slot, which is given by ζi,t+1 = Yj∈Ni (1 − βpj,t), (1) where Ni denotes the set of neighbors of node i. This expression can be interpreted as the probability that none of the infection attempts is successful. The probability for a node i of not being infected at time t + 1 depends on the conditioning event if the node i was or was not infected at time t: 1. For the first case, if the node i was infected at time t, then the probability for not being infected at time t + 1 is equal to δζi,t+1, 2. For the second case, if the node i was not infected at time t, then the probability for not being in- fected at time t + 1 is equal to ζi,t+1. From 1) and 2), the probability for a node i of not being infected at time t + 1 is equal to 1 − pi,t+1 = ζi,t+1(1 − pi,t) + δζi,t+1pi,t. Replacing from eq. (1), we obtain that (1−βpj,t)+δpi,t Yj∈Ni 1−pi,t+1 = (1−pi,t) Yj∈Ni (1−βpj,t). (2) We focus on the criterion based on the asymptotic stability of the disease-free equilibrium p∗ i (t) = 0 for all i. For doing this, we will make use of the following theorem. 102 101 s e d o n d e t c e f n i f o r e b m u N 100 100 delta/beta=2 delta/beta=6 101 Timestep 102 103 Figure 1: Number of infected nodes vs time. Theorem 1 (Hirsch and Smale, 1974 [21]). The sys- tem given by pt+1 = g(pt) is asymptotically stable at an equilibrium point p∗, if the eigenvalues of the Ja- cobian J = ∇g(p) are less than 1 in absolute value, where Jk,l = [∇g(p∗)]k,l = ∂pk,t+1 ∂pl,t . ∗ (cid:12)(cid:12)(cid:12)(cid:12)pt=p 3 4 3 8 1 6 5 1 0 2 7 1 2 1 1 1 6 1 5 9 1 4 1 3 4 3 8 1 6 7 1 1 5 2 1 2 5 1 1 6 1 3 1 4 9 1 0 Figure 2: A two-dimensional grid of length 4. Figure 3: A two-dimensional torus of length 4. We rewrite eq. (2) as follows: pi,t+1 = 1−(1−pi,t) Yj∈Ni (1−βpj,t)−δpi,t Yj∈Ni (1−βpj,t). information spreading dies out, however for the first scenario the virus or information spreading may con- tinue as time increases. Then Jk,l =   if k = l, 1 − δ β if l is neighbor of k, 0 otherwise, 4 (3) Information Spreading over the Torus with one removed node ρ(A) < . (4) δ β where we recall that the evaluating point is p∗ = 0. Eq. (3) can be written in a more compact way as J = (1 − δ)I + βA. Using Theorem 1, we obtain that for an asymptotic stability of the disease-free equi- librium we need to impose that the eigenvalues of (1− δ)I + βA are in absolute value smaller than 1, or equivalently, In Fig. 1, we present two scenarios where we con- sider a total population of 900 nodes and an initial seeding of 20 nodes with the same probability of in- fection β1 = β2 = 0.1 but with different probabilities of recovery, δ1 = 0.2 for the solid curve and δ2 = 0.6 for the dashed curve. We consider a graph with spec- tral radius ρ(A) = 4. In the second scenario, as pre- dicted by the analysis and equation (4), the virus or 4 In a regular grid topology, each node in the network is connected with at most two neighbors along one or more directions (see e.g. Fig. 2). If the network is one-dimensional and we connect the first and last nodes, then the resulting topology consists on a chain of nodes connected in a circular loop, which is known as ring. In general, when an n-dimensional grid net- work is connected circularly in more than one dimen- sion, the resulting network topology is a torus (see e.g. Fig. 3). In this work, we consider torus networks with the same number of nodes in every direction. From Fig. 2, by connecting each first node to the last node in each direction we obtain Fig. 3. For exam- ple, if we connect node 1 with node 4 in the horizon- tal direction and node 1 with node 13 in the vertical direction, we obtain the neighbors of node 1 on the torus (nodes 2, 4, 5, 13) as shown in Fig. 3. From the previous section, we obtained that the spectral radius is an important quantity to study if our interest is the spreading of information (or virus spreading) through the network. The follow- ing proposition give us a relationship between the spectral radius and the degrees of the nodes. Lemma 1. [22] Let degmin denote the minimum de- gree of G, let deg be the average degree of G, and let degmax be the maximum degree of G. For every graph G, max{deg,pdegmax} ≤ ρ(A) ≤ degmax. We observe that for a k-regular graph, its average degree and its maximum degree are equal to k, and thus k corresponds to its spectral radius. This means that for a d-dimensional torus, 2d corresponds to its spectral radius. The previous lemma seems to prove that there is no interest in studying the spectral radius over the torus since it is a well-known quantity. However, if the torus network is modified (removing some of the nodes or edges to the graph) then the propagation of information will change (see Fig. 4 and Fig. 5). In the following analysis, we will give an explicit closed-form expression of these changes on the spreading. First of all, we recall a well-known result in lin- ear algebra. Cramer's rule [23] states that a system of n linear equations with n unknowns, represented in matrix multiplication form M x = y, where the n × n matrix M has a nonzero determinant and the vector x = (x1, . . . , xn)T is the column vector of the unknown variables, has a unique solution, whose in- dividual values for the unknowns are given by xi = det(Mi) det(M ) , ∀i ∈ {1, . . . , n}, Lemma 2. [24] For any graph G we have x−1Wii(G, x−1) = φ(G \ i, x)/φ(G, x). Proof. We have that W (G, x) = (I − xA)−1 = x−1(x−1I − A)−1. The entries of (x−1I − A)−1 are given by Cramer's rule (by noting that (x−1I − A)−1 corresponds to the matrix M such that (x−1I − A)M = I). The i-th diagonal entry of (x−1I − A)−1 is given by the i-th principal diagonal minor of (x−1I − A) divided by det(x−1I − A) = φ(G, x−1). We note that the i-th principal diagonal minor of (x−1I − A) is φ(G \ i, x−1), and so the lemma follows immediately from this. We are interested on finding the spectral radius of the almost regular torus, in order to analyze the information spreading on this topology. In the case of the removal of one node, this is equivalent to find φ(G \ i, x). From Lemma 2, we need to know the characteristic polynomial of the regular torus φ(G, x) and the diagonal entries of the walking generating function of the regular torus Wii(G, x−1) since φ(G \ i, x) = x−1Wii(G, x−1)φ(G, x). (5) In the next proposition, we give an explicit ex- pression for the characteristic polynomial of the two- dimensional torus network of length m. Proposition 1. The characteristic polynomial of the two-dimensional torus network, denoted T 2 m, is given by φ(T 2 m, x) = Y1≤i,j≤m (x − 2 cos(2πi/m) − 2 cos(2πj/m)) . where Mi is the matrix formed by replacing the i-th column of M by the column vector y. We use Cramer's rule in the next lemma, in order to establish the connection between the characteris- tic polynomials and walk generating functions. This connection will allow us to compute the spectral ra- dius of the modified torus when we remove a node. Proof. Consider Rm the ring graph which has edge set {(u, u + 1) : 1 ≤ u < m} ∪ {(m, 1)}. The Lapla- cian of Rm has eigenvectors [25]: xk(u) = sin(2πku/m) and yk(u) = cos(2πku/m). (6) for k ≤ m/2. Both of these eigenvectors have eigen- value λk = 2 − 2 cos(2πk/m). We notice that x0 5 7 6 5 4 3 2 1 0.125 0.12 0.115 0.11 0.105 0.1 0.095 0.09 0.085 1 2 3 4 5 6 7 Figure 5: Spectral radius reduction by removing two nodes from a torus network of length 7. One of the re- moved nodes is the central node. We remove another node and we compute the spectral radius reduction and we assign the value of this reduction at the re- moved node's position. a k-regular graph of n nodes we have det(L − λI) = det(D − A − λI) = det(−[A − (k − λ)I]) = (−1)ndet(A − (k − λ)I), which means that λ is an eigenvalue of L if and only if k − λ is an eigenvalue of A. In our case, the two-dimensional torus network is a 4-regular graph, and thus the eigenvalues of the two-dimensional torus network of length m are given by 2 cos(2πi/m) + 2 cos(2πj/m) for all 1 ≤ i, j ≤ m. Thus we conclude that the characteristic polynomial is equal to φ(T 2 (x − 2 cos(2πi/m) − 2 cos(2πj/m)) . m, x) = Y1≤i,j≤m We notice from the proof that the eigenvectors of the Laplacian coincide with the eigenvectors of the adjacency matrix and are given by the Kronecker product of the eigenvectors given by (6). In the following proposition, we find an explicit expression for the diagonal entries of the walk gener- ating function. 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 i s u d a r l a r t c e p s e h t f o n o i t c u d e R 0 0 5 10 15 20 Torus length 25 30 35 40 Figure 4: Spectral radius reduction by removing one node vs torus length. should be ignored and y0 is the all 1s vector. If m is even, then also xm/2 should be ignored. We recall that the product of two graphs G = (V,E) and H = (W,F ), denoted by G × H, corresponds to the graph with vertex set V × W and edge set ((v1, w1), (v2, w2)) where (v1, v2) ∈ E and (w1, w2) ∈ F . If G has Laplacian eigenvalues λ1, . . . , λm and eigenvectors p1, . . . , pm; and H has Laplacian eigenvalues µ1, . . . , µm and eigenvectors q1, . . . , qm; then for each 1 ≤ i ≤ m and 1 ≤ j ≤ m, G×H has an eigenvector z of eigenvalue λi + µj given by z(v, w) = pi(v)qj (w). In our case, the two-dimensional torus can be writ- ten as the product of two rings, G = H = Rm, and then λi = µi = 2 − 2 cos(2πi/m) ∀1 ≤ i ≤ m. Thus the eigenvalues of the Laplacian of the two- dimensional torus network of length m are given by λi+µj = 4−2 cos(2πi/m)−2 cos(2πj/m) ∀1 ≤ i, j ≤ m. We recall that the eigenvalues of a matrix M are the solutions λ to the equation det(M − λI) = 0. For 6 Proposition 2. The diagonal entries of the walk generating function of the two-dimensional torus of length m are: Wii(T 2 m, x) = xℓ Xℓ≥0 m2 X1≤i,j≤m φ(T 2 m, x−1) = Wii(T 2 x−ℓ Xℓ≥0 m2 X1≤i,j≤m 4ℓ (cos(π(i + j)/m))ℓ (cos(π(i − j)/m))ℓ . m, x) = Y1≤i,j≤m φ(T 2 (7) node has been removed is given by: m \ i, x) = x−1Wii(T 2 m, x−1)φ(T 2 m, x), where 4ℓ (cos(π(i + j)/m))ℓ (cos(π(i − j)/m))ℓ , (x − 2 cos(2πi/m) − 2 cos(2πj/m)) . Proof. We recall from Section 2 that Aℓ ii corresponds to the closed walks of length ℓ. Since each node is indistinguishable on the torus network, we know that tr(Aℓ) = n×# of closed walks of length ℓ, where n = m2 is the total number of nodes. But we also have that n tr(Aℓ) = (λi(A))ℓ, Xk=1 = X1≤i,j≤m = X1≤i,j≤m (2 cos(2πi/m) + 2 cos(2πj/m))ℓ, (4 cos(π(i + j)/m) cos(π(i − j)/m))ℓ. Thus Aℓ ii = 1 m2 X1≤i,j≤m (4 cos(π(i+j)/m) cos(π(i−j)/m))ℓ. (8) Let us notice that Wii(G, x) = [(I − xA)−1]ii =  Xn≥0 (xA)n ii = Xℓ≥0 From eq. (8) and eq. (9) we conclude eq. (7). We observe that all the previous calculations do not depend on the particular removed node i. This means that the spreading of information over the modified torus, if we remove one node, is not affected by the position of the removed node. The importance is that one and only one node is removed. In the next sec- tion, we will see that this is very different from the case of the removal of two nodes. The previous results can also be derived for the d- dimensional torus network. Since the proofs are sim- ilar to the proofs of Proposition 1 and Proposition 2 we do not include them here. Theorem 3. The characteristic polynomial of the d-dimensional torus network of length m where one node has been removed is given by: φ(T d m \ i, x) = x−1Wii(T d m, x−1)φ(T d m, x), where the diagonal entries of the walk generating function are xℓAℓ ii. m, x−1) = Wii(T d x−ℓ md X1≤i1,i2...,id≤m Xℓ≥0 (9) (2 cos(2πi1/m) + . . . + 2 cos(2πid/m))ℓ , the and polynomial dimensional torus network is given by characteristic of the d- From Proposition 1 and Proposition 2, we obtain the following theorem. Theorem 2. The characteristic polynomial of the two-dimensional torus network of length m where one m, x) = φ(T d Y1≤i1,...,id≤m 7 (x − 2 cos(2πi1/m) − . . . − 2 cos(2πid/m)) . 5 Information Spreading over the Torus with a set of re- moved nodes Following a similar approach to Lemma 2, Godsil [24] is able to prove the following theorem. Theorem 4. [Nodes removal][24] Let S be a subset of s nodes from the graph G. Then x−sdet WS,S (G, x−1) = φ(G \ S, x)/φ(G, x), where WS,S(G, x) denotes the submatrix of W (G, x) with rows and columns indexed by the elements of S. We observe that if S consists of two nodes i and j then This implies that the only unknowns to compute the removal of a set of nodes over the torus, in par- ticular for a set of two nodes, are given by Wij , which can be represented as the number of walks between node i and node j. The following theorem provides us a way to compute the values Wij . Theorem 5. [26] For any two vertices i and j in the graph G and any non-negative integer ℓ, we have that the number of walks of length ℓ, denoted Wij (G, ℓ), is given by Wij (G, ℓ) = Xθ uθ(i)T θℓuθ(j), where the sum is over all eigenvalues θ of G and uθ(i) denotes the i-th row of Uθ where Uθ is the ma- trix whose columns form an orthonormal basis for the eigenspace belonging to θ. detWS,S (G, x) = Wii(G, x)Wjj (G, x)−Wij (G, x)Wji(G, x). (10) and since G is an undirected graph, Wij (G, x) = Wji(G, x). For the case of the removal of two nodes, from The- orem 4 and eq. (10), we obtain the following corollary. From the previous theorem, we conclude that the walk generating function can be written as Wij (G, x−1) = Xℓ≥0 x−ℓXθ uθ(i)T θℓuθ(j). Corollary 1. [24] For any graph G we have: x−1Wij (G, x−1) = pφ(G \ i, x)φ(G \ j, x) − φ(G, x)φ(G \ ij, x)/φ(G, x). (11) The next corollary is extremely important since it guarantees that independently of the number of nodes we remove from the torus graph, we can re- strict our study of the characteristic polynomials to the case of the removal of two nodes. Corollary 2. [24] If C is a subset of V(G) then φ(G\C, x) is determined by the polynomials φ(G\S, x) where S ranges over all subsets of C with at most two vertices. From Corollary 1, we obtain that φ(G \ ij, x) = φ(G \ i, x)φ(G \ j, x) φ(G, x) (12) − φ(G, x)(cid:0)x−1Wij (G, x−1)(cid:1)2 . 8 In order to compute the walk generating function of the two-dimensional torus, we recall that the ring of length m has eigenvectors: xk(u) = sin(2πku/m) and yk(u) = cos(2πku/m). for k ≤ m/2. Both of these eigenvectors have eigen- value λk = 2 − 2 cos(2πk/m). Here x0 should be ig- nored and y0 is the all 1s vector. If m is even, then also xm/2 should be ignored. We denote the matrix of eigenvectors of the ring of length m as V . Then the two-dimensional torus has matrix of eigenvectors Z = V ⊗ V where ⊗ denotes the Kronecker prod- uct. Since the matrix is symmetric the eigenvectors form an orthogonal basis and we may normalize its eigenvectors to obtain an orthonormal basis. We call this orthonormal basis Z. We consider the matrix of eigenvalues of the two-dimensional torus, denoted by Λ, which are given by λi + µj = 2 cos(2πi/m) + 2 cos(2πj/m) ∀1 ≤ i, j ≤ m. From here we obtain the number of closed walks of the two-dimensional torus between node i and node j. Following a similar approach to the previous case of the removal of one node, Godsil [24] is able to prove the following theorem which deals with the removal of one edge of the graph. Theorem 6 (Edge removal). [24] Let e = {i, j} be an edge in G. Then φ(G, x) = φ(G \ e, x) − φ(G \ ij, x) (13) − 2pφ(G \ i, x)φ(G \ j, x) − φ(G, x)φ(G \ ij, x). We observe that all the previous terms in eq. (13) except φ(G \ e, x) are known. This implies that the characteristic polynomial of the two-dimensional torus with one edge removed is completely deter- mined by the previous expressions for the torus with one node removed and two nodes removed. 6 Lower bounds for the two- dimensional torus networks In this section, we will compare the previous expres- sions with the case of the infinite lattice. We notice that this scenario can be seen as the case where the length of the torus goes to infinity. Proposition 3. The diagonal entries of the walk generating function over the lattice L2 are given by Wii(L2, x) = Xℓ≥0 ℓ (cid:19)2 x2ℓ(cid:18)2ℓ . (14) Proof. The number of closed walks of length 2ℓ over the two-dimensional lattice2 L2 is equal to (see e.g. EIS A002894): ℓ Xi=0 (2ℓ)! i!i! (ℓ − i)! (ℓ − i)! ℓ (cid:19)2 = (cid:18)2ℓ . 2 We notice that to consider even length is not restrictive since every closed walk over the lattice network has to have even length. 9 In the following, we give a lower bound to the walk generating function of the two-dimensional torus with one removed node. Proposition 4. For the two-dimensional torus, we have x−1Wii(T 2 m, x−1) ≥ 1 x 4π e4 Xℓ≥0 1 ℓ (cid:18) 4 x(cid:19)2ℓ . Proof. The Stirling approximation formula tell us that e(cid:19)ℓ √2πℓ(cid:18) ℓ e(cid:19)ℓ ≤ ℓ! ≤ e√ℓ(cid:18) ℓ . Then applying it to the binomial coefficient we get the lower bound (cid:18)2ℓ ℓ (cid:19) = (2ℓ)! ℓ!ℓ! ≥ √2π2ℓ(cid:0) 2ℓ e (cid:1)2ℓ e(cid:1)ℓ = e(cid:1)ℓe √ℓ(cid:0) ℓ e√ℓ(cid:0) ℓ √4π e2 1 √ℓ 4ℓ. (15) We know that the number of walks over the torus network is greater than the number of walks over the lattice network since every walk over the lattice net- work can be mapped into a walk over the torus net- work. We notice that the contrary is not true in general since for example following m + 1 steps in the same direction will be a closed walk over the torus of length m but not a closed walk over the lattice. Therefore, from eq. (14) and eq. (15), we have that x−1Wii(T 2 m, x−1) ≥ x−1Wii(L2, x−1) = 4π = x−1Xℓ≥0 e4 Xℓ≥0 ℓ (cid:19)2 x−2ℓ(cid:18)2ℓ 1 x ℓ (cid:18) 4 x(cid:19)2ℓ ≥ 1 . Proposition 4 allow us to obtain also a lower bound for the characteristic polynomial of the two- dimensional torus with one removed node. The lower bound is given by φ(T 2 φ(T 2 m \ i, x) ≥ φ(T 2 m, x) = Y1≤i,j≤m m, x) 1 x 4π e4 Xℓ≥0 1 ℓ (cid:18) 4 x(cid:19)2ℓ where (x − 2 cos(2πi/m) − 2 cos(2πj/m)) . 1020 1015 1010 l s k a w d e s o c f l o r e b m u N 105 100 0 Lattice Torus m=5 5 Closed walks length 10 15 e c n e r e f f i d t n e c r e P 100 90 80 70 60 50 40 30 20 10 0 0 m=5 m=10 m=15 5 Closed walks length 10 15 Figure 6: Number of closed walks vs closed walks length. Figure 7: Percent difference on the number of closed walks between the torus and the lattice vs closed walks length. The case of the number of walks between any two nodes over the two-dimensional torus of length m is more involved. We define the following func- tion which allow us to map each node on the two- dimensional torus of length m to a node on the lat- tice Z × Z. The mapping function h : {1, . . . , m2} → Z × Z is as follows: h(u) = (xu, yu) where xu ≡ u − 1 (mod m) and yu = j u mk . Proposition 5. The number of walks of length ℓ over the two-dimensional lattice from the origin (0, 0) to (a, b) (ℓ + a + b is even) is equal to of closed walks of length ℓ with i steps going up is . Summing up for equal to all possible values of nU , the number of paths is equal to eq. (16). 2 −i)!( ℓ+b−a i!(i−b)!( ℓ+a+b 2 −i)! ℓ! From Proposition 5 and the mapping function h, we obtain a lower bound on the number of walks of length ℓ on the two-dimensional torus T 2 m between two different nodes i and j by mapping i and j to h(i) = (xi, yi) and h(j) = (xj , yj) and computing the number of walks from (0, 0) to (xj − xi, yj − yi). We observe that the previous lower bound is tight only when ℓ + a + b is even. b+(ℓ−b)/2 Xi=b ℓ! i!(i − b)!(cid:0) ℓ+a+b 2 − i(cid:1)!(cid:0) ℓ+b−a 2 − i(cid:1)! 7 Numerical Simulations . (16) Proof. We consider nU the number of steps you go up, nD the number of steps you go down, nL the number of steps you go left, nR the number of steps you go right. The number of walks of length ℓ from (0, 0) to (a, b) is equal to the number of ways you can choose ℓ steps between up, down, left, and right, such that nU = nD + b, nR = nL + a, and nU + nD + If we fix nU = i then nD = i − b, nL + nR = ℓ. nL = ( ℓ+b−a 2 − i) and the number 2 − i), nR = ( ℓ+a+b 10 In Fig. 6, we compute the number of closed walks over the torus of length m = 5 and over the lattice vs the length of the closed walks ℓ. We notice that for small values of ℓ to approximate the number of closed walks over the torus is more precise than for large values of ℓ. This fact is compensated since in the characteristic polynomial the important terms are the terms of smaller order. In Fig. 7, we compute the difference on percent error of this approximation vs the length of the closed walks ℓ. In Fig. 8, we compute the number of closed walks over the torus and over the lattice by varying the l s k a w d e s o c f l o r e b m u N 1018 1016 1014 1012 1010 108 106 104 2 Lattice l=5 Torus l=5 Lattice l=10 Torus l=10 Lattice l=15 Torus l=15 e c n e r e f f i d t n e c r e P 100 90 80 70 60 50 40 30 20 10 0 0 l=5 l=10 l=15 2 4 6 8 10 Torus length 12 14 16 18 20 4 6 8 10 12 Torus length 14 16 18 20 Figure 8: Number of closed walks vs torus length. torus length m and keeping constant the length of the closed walks ℓ. We notice that for small values of the torus length m the approximation is not tight, however it becomes tighter relatively fast. In Fig. 9, we compute the difference on percent error of this approximation vs the torus length m. 8 Conclusions In this work, we have analyzed information spreading on almost torus networks assuming the Susceptible- Infected-Susceptible spreading model as a metric of information spreading. Almost torus networks con- sist on the torus network topology where some nodes or edges have been removed. We have provided an- alytical expressions for the characteristic polynomial of these graphs and we have provided as well tight lower bounds for its computation. Using these ex- pressions we are able to estimate their spectral ra- dius and thus to know how the information spreads on these networks. Simulations results that validated our analysis are presented. Acknowledgments The work presented in this paper has been partially carried out at LINCS (http://www.lincs.fr). 11 Figure 9: Percent difference on the number of closed walks between the torus and the lattice vs torus length. References [1] F. Brauer, P. van den Driessche, and J. Wu, eds., Mathematical epidemiology, vol. 1945 of Lecture Notes in Mathematics. Berlin: Springer-Verlag, 2008. Mathematical Biosciences Subseries. [2] D. J. Daley and J. Gani, Epidemic Modeling: An Introduction. Cambridge, UK: Cambridge University Press, 1999. [3] M. Draief and L. Massouli, Epidemics and ru- mours in complex networks. Cambridge Univer- sity Press, 2010. [4] J. Kephart and S. White, "Directed-graph epi- demiological models of computer viruses," in Re- search in Security and Privacy, 1991. Proceed- ings., 1991 IEEE Computer Society Symposium on, pp. 343 -- 359, may 1991. [5] X. Zhang, G. Neglia, J. Kurose, and D. Towsley, "Performance modeling of epidemic routing," Computer Networks (Amsterdam, Netherlands: 1999), vol. 51, pp. 2867 -- 2891, July 2007. [6] M. Bampo, M. T. Ewing, D. R. Mather, D. Stew- art, and M. Wallace, "The effects of the social structure of digital networks on viral marketing performance.," vol. 19, pp. 273 -- 290, 2008. [7] A. J. Ganesh, L. Massouli´e, and D. F. Towsley, "The effect of network topology on the spread of epidemics," in INFOCOM, pp. 1455 -- 1466, IEEE, 2005. [8] Leskovec, Adamic, and Huberman, "The dy- namics of viral marketing," in CECOMM: ACM Conference on Electronic Commerce, 2006. [9] X. Wei, N. Valler, B. A. Prakash, I. Neamtiu, M. Faloutsos, and C. Faloutsos, "Competing memes propagation on networks: a case study of composite networks," Computer Communica- tion Review, vol. 42, no. 5, pp. 5 -- 12, 2012. [10] W. Tang, Z. Lan, N. Desai, D. Buettner, and Y. Yu, "Reducing fragmentation on torus- connected supercomputers," in IPDPS, pp. 828 -- 839, IEEE, 2011. [11] TOP 500 Supercomputing website, "http://www.top500.org," [12] A. Gara, M. A. Blumrich, D. Chen, G. L.-T. Chiu, P. Coteus, M. E. Giampapa, R. A. Haring, P. Heidelberger, D. Hoenicke, G. V. Kopcsay, T. A. Liebsch, M. Ohmacht, B. D. Steinmacher- Burow, T. Takken, and P. Vranas, "Overview of the Blue Gene/L system architecture," IBM Journal of Research and Development, vol. 49, pp. 195 -- 212, march 2005. [13] N. R. Adiga, M. A. Blumrich, D. Chen, P. Co- teus, A. Gara, M. E. Giampapa, P. Heidelberger, S. Singh, B. D. Steinmacher-Burow, T. Takken, M. Tsao, and P. Vranas, "Blue Gene/L torus interconnection network," IBM Journal of Re- search and Development, vol. 49, pp. 265 -- 276, mar./may 2005. [14] IBM Blue Gene team, "Overview of the IBM Blue Gene/P project," IBM Journal of Research and Development, vol. 52, pp. 199 -- 220, 2008. jan. [15] J. Brooks and G. Kirschner, "Cray XT3 and cray XT series of supercomputers," in Encyclo- pedia of Parallel Computing (D. A. Padua, ed.), pp. 457 -- 470, Springer, 2011. [16] D. Chen, N. Eisley, P. Heidelberger, R. Senger, Y. Sugawara, S. Kumar, V. Salapura, D. Sat- terfield, B. Steinmacher-Burow, and J. Parker, "The ibm blue gene/q interconnection network and message unit," in High Performance Com- puting, Networking, Storage and Analysis (SC), 2011 International Conference for, pp. 1 -- 10, nov. 2011. [17] Y. Ajima, S. Sumimoto, and T. Shimizu, "Tofu: A 6D mesh/torus interconnect for exascale com- puters," IEEE Computer, vol. 42, pp. 36 -- 40, Nov. 2009. Fujitsu. [18] H. W. Hethcote, "The mathematics of infectious diseases," SIAM Review, vol. 42, pp. 599 -- 653, Dec. 2000. [19] Y. Wang, D. Chakrabarti, C. Wang, and C. Faloutsos, "Epidemic spreading in real net- works: An eigenvalue viewpoint," in SRDS, pp. 25 -- 34, IEEE Computer Society, 2003. [20] D. Chakrabarti, Y. Wang, C. Wang, J. Leskovec, and C. Faloutsos, "Epidemic thresholds in real networks," ACM Trans. Inf. Syst. Secur, vol. 10, no. 4, 2008. [21] M. W. Hirsch and S. Smale, Differential equa- tions, dynamical systems, and linear algebra. New York: Academic Press, 1974. [22] L. Lov´asz, "Eigenvalues of graphs," 12 [23] G. Cramer, "Introduction `a l'analyse des lignes courbes alg´ebriques (in french)," Geneva: Euro- peana, pp. 656 -- 659, 1750. [24] C. D. Godsil, "Walk generating functions, christoffel-darboux identities and the adjacency matrix of a graph," Combinatorics, Probability and Computing, vol. 1, no. 01, pp. 13 -- 25, 1992. [25] D. Spielman, "Spectral graph theory and its ap- plications," [26] A. Chan and C. D. Godsil, "Symmetry and eigenvectors," in Graph symmetry (Montreal, PQ, 1996), vol. 497 of NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., pp. 75 -- 106, Dordrecht: Kluwer Acad. Publ., 1997. 13
1602.03780	6	1602	2017-03-01T07:33:37	Interplay between Social Influence and Network Centrality: A Comparative Study on Shapley Centrality and Single-Node-Influence Centrality	[ "cs.SI", "physics.soc-ph" ]	We study network centrality based on dynamic influence propagation models in social networks. To illustrate our integrated mathematical-algorithmic approach for understanding the fundamental interplay between dynamic influence processes and static network structures, we focus on two basic centrality measures: (a) Single Node Influence (SNI) centrality, which measures each node's significance by its influence spread; and (b) Shapley Centrality, which uses the Shapley value of the influence spread function --- formulated based on a fundamental cooperative-game-theoretical concept --- to measure the significance of nodes. We present a comprehensive comparative study of these two centrality measures. Mathematically, we present axiomatic characterizations, which precisely capture the essence of these two centrality measures and their fundamental differences. Algorithmically, we provide scalable algorithms for approximating them for a large family of social-influence instances. Empirically, we demonstrate their similarity and differences in a number of real-world social networks, as well as the efficiency of our scalable algorithms. Our results shed light on their applicability: SNI centrality is suitable for assessing individual influence in isolation while Shapley centrality assesses individuals' performance in group influence settings.	cs.SI	cs	Interplay between Social Influence and Network Centrality: A Comparative Study on Shapley Centrality and Single-Node-Influence Centrality Wei Chen Microsoft Research Beijing, China [email protected] Shang-Hua Teng University of Southern California Los Angeles, CA, U.S.A. [email protected] 7 1 0 2 r a M 1 ] I S . s c [ 6 v 0 8 7 3 0 . 2 0 6 1 : v i X r a ABSTRACT We study network centrality based on dynamic influence propagation models in social networks. To illustrate our in- tegrated mathematical-algorithmic approach for understand- ing the fundamental interplay between dynamic influence processes and static network structures, we focus on two basic centrality measures: (a) Single Node Influence (SNI) centrality, which measures each node's significance by its influence spread;1 and (b) Shapley Centrality, which uses the Shapley value of the influence spread function - formu- lated based on a fundamental cooperative-game-theoretical concept - to measure the significance of nodes. We present a comprehensive comparative study of these two centrality measures. Mathematically, we present axiomatic character- izations, which precisely capture the essence of these two centrality measures and their fundamental differences. Al- gorithmically, we provide scalable algorithms for approxi- mating them for a large family of social-influence instances. Empirically, we demonstrate their similarity and differences in a number of real-world social networks, as well as the effi- ciency of our scalable algorithms. Our results shed light on their applicability: SNI centrality is suitable for assessing individual influence in isolation while Shapley centrality as- sesses individuals' performance in group influence settings. Keywords Social network; social influence; influence diffusion model; interplay between network and influence model; network centrality; Shapley values; scalable algorithms 1. INTRODUCTION Network science is a fast growing discipline that uses mathematical graph structures to represent real-world net- works - such as the Web, Internet, social networks, bio- logical networks, and power grids - in order to study fun- damental network properties. However, network phenom- ena are far more complex than what can be captured only by nodes and edges, making it essential to formulate net- work concepts by incorporating network facets beyond graph structures [36]. For example, network centrality is a key concept in network analysis. The centrality of nodes, usu- ally measured by a real-valued function, reflects their signif- icance, importance, or crucialness within the given network. Numerous centrality measures have been proposed, based on 1The influence spread of a group is the expected number of nodes this group can activate as the initial active set. degree, closeness, betweenness and eigenvector (i.e., PageR- ank) (cf. [23]). However, most of these centrality measures focus only on the static topological structures of the net- works, while real-world network data include much richer interaction dynamics beyond static topology. Influence propagation is a wonderful example of interac- tion dynamics in social networks. As envisioned by Domin- gos and Richardson [28, 14], and beautifully formulated by Kempe, Kleinberg, and Tardos [18], social influence propa- gation can be viewed as a stochastic dynamic process over an underlying static graph: After a group of nodes becomes active, these seed nodes propagate their influence through the graph structure. Even when the static graph structure of a social network is fixed, dynamic phenomena such as the spread of ideas, epidemics, and technological innova- tions can follow different processes. Thus, network central- ity, which aims to measure nodes' importance in social in- fluence, should be based not only on static graph structure, but also on the dynamic influence propagation process. In this paper, we address the basic question of how to formulate network centrality measures that reflect dynamic influence propagation. We will focus on the study of the interplay between social influence and network centrality. A social influence instance specifies a directed graph G = (V, E) and an influence model PI (see Section 2). For each S ⊆ V , PI defines a stochastic influence process on G with S as the initial active set, which activates a random set I(S) ⊇ S with probability PI(S, I(S)). Then, σ(S) = E[I(S)] is the influence spread of S. The question above can be restated as: Given a social-influence instance (V, E, PI), how should we define the centrality of nodes in V ? A natural centrality measure for each node v ∈ V is its influence spread σ({v}). However, this measure - referred to as the single node influence (SNI) centrality - completely ignores the influence profile of groups of nodes and a node's role in such group influence. Thus, other more sensible cen- trality measures accounting for group influence may better capture nodes' roles in social influence. As a concrete formu- lation of group-influence analyses, we apply Shapley value [31] - a fundamental concept from cooperative game theory - to define a new centrality measure for social influence. Cooperative game theory is a mathematical theory study- ing people's performance and behavior in coalitions (cf. [21]). Mathematically, an n-person coalitional game is de- fined by a characteristic function τ : 2V → R, where V = [n], and τ (S) is the utility of the coalition S [31]. In this (τ ) of v ∈ V is v's expected game, the Shapley value φShapley v marginal contribution in a random order. More precisely: φShapley v (τ ) = Eπ[τ (Sπ,v ∪ {v}) − τ (Sπ,v)], (1) where Sπ,v denotes the set of players preceding v in a ran- dom permutation π of V : The Shapley value enjoys an ax- iomatic characterization (see Section 2), and is widely con- sidered to be the fairest measure of a player's power in a cooperative game. Utilizing the above framework, we view influence spread σ(·) as a characteristic function, and define the Shapley cen- trality of an influence instance as the Shapley value of σ. In this paper, we present a comprehensive comparative study of SNI and Shapley centralities. In the age of Big Data, networks are massive. Thus, an effective solution concept in network science should be both mathematically meaningful and algorithmically efficient. In our study, we will address both the conceptual and algorithmic questions. Conceptually, influence-based centrality can be viewed as a dimensional reduction from the high dimensional influence model PI to a low dimensional centrality measure. Dimen- sional reduction of data is a challenging task, because in- evitably some information is lost. As highlighted by Ar- row's celebrated impossibility theorem on voting [3], for var- ious (desirable) properties, conforming dimensional reduc- tion scheme may not even exist. Thus, it is fundamental to characterize what each centrality measure captures. So, "what do Shapley and SNI centralities capture? what are their basic differences?" Axiomatization is an instru- mental approach for such characterization. In Section 3, we present our axiomatic characterizations. We present five ax- ioms for Shapley centrality, and prove that it is the unique centrality measure satisfying these axioms. We do the same for the SNI centrality with three axioms. Using our ax- iomatic characterizations, we then provide a detailed com- parison of Shapley and SNI centralities. Our characteriza- tions show that (a) SNI centrality focuses on individual in- fluence and would not be appropriate for models concerning group influence, such as threshold-based models. (b) Shap- ley centrality focuses on individuals' "irreplaceable power" in group influence settings, but may not be interpreted well if one prefer to focus on individual influence in isolation. The computation of influence-based centralities is also a challenging problem: Exact computation of influence spread in the basic independent cascade and linear-threshold models has been shown to be #P-complete [37, 12]. Shapley cen- trality computation seems to be more challenging since its definition as in Eq. (1) involves n! permutations, and existing Shapley value computation in several simple network games have quadratic or cubic time complexity [19]. Facing these challenges, in Section 4, we present provably-good scalable algorithms for approximating both Shapley and SNI central- ities of a large family of social influence instances. Surpris- ingly, both algorithms share the same algorithm structure, which extends techniques from the recent algorithmic break- throughs in influence maximization [10, 34, 33]. We further conduct empirical evaluation of Shapley and SNI centrali- ties in a number of real-world networks. Our experiments - see Section 5 - show that our algorithms can scale up to networks with tens of millions of nodes and edges, and these two centralities are similar in several cases but also have noticeable differences. These combined mathematical/algorithmic/empirical analyses together present (a) a systematic case study of the interplay between influence dynamics and network centrality based on Shapley and SNI centralities; (b) axiomatic characterizations for two basic centralities that precisely capture their similarities and differences; and (c) new scalable algorithms for influence models. We believe that the dual axiomatic-and-algorithmic characterization provides a comparative framework for evaluating other influence-based network concepts in the future. For presentation clarity, we move the technical proofs into the appendix, which also contains additional technical materials for (algorithmic and axiomatic) generalization to weighted influence models. 1.1 Related Work Network centrality has been extensively studied (see [23] and the references therein for a comprehensive introduction). Most classical centralities, based on degree, closeness, be- tweenness, eigenvector, are defined on static graphs. But some also have dynamic interpretations based on random- walks or network flows [8]. Eigenvector centrality [6] and its closely related Katz-[17] and Alpha-centrality [7] can be viewed as some forms of influence measures, since their dynamic processes are non-conservative [15], meaning that items could be replicated and propagated, similar to diffu- sion of ideas, opinions, etc. PageRank [11, 25] and other random-walk related centralities correspond to conservative processes, and thus may not be suitable for propagation dy- namics. Percolation centrality [27] also addresses diffusion process, but its definition only involves static percolation. None of above maps specific propagation models to network centrality. Ghosh et al. [16] maps a linear dynamic process characterized by parameterized Laplacian to centrality but the social influence models we consider in this paper are be- yond such linear dynamic framework. Michalak et al. use Shapley value as network centrality [19], but they only con- sider five basic network games based on local sphere of influ- ence, and their algorithms run in (least) quadratic time. To the best of our knowledge, our study is the first to explicitly map general social network influence propagation models to network centrality. Influence propagation has been extensively studied, but most focusing on influence maximization tasks [18, 37, 12], which aims to efficiently select a set of nodes with the largest influence spread. The solution is not a centrality measure and the seeds in the solution may not be the high centrality nodes. Borgatti [9] provides clear conceptual discussions on the difference between centralities and such key player set identification problems. Algorithmically, our construction extends the idea of reverse reachable sets, recently intro- duced in [10, 34, 33] for scalable influence maximization. In terms of axiomatic characterizations of network central- ity, Sabidussi is the first who provides a set of axioms that a centrality measure should satisfy [29]. A number of other studies since then either provide other axioms that a central- ity measure should satisfy (e.g. [24, 5, 30]) or a set of axioms that uniquely define a centrality measure (e.g. [2] on PageR- ank without the damping factor). All of these axiomatic characterizations focus on static graph structures, while our axiomatization focuses on the interplay between dynamic in- fluence processes and static graph structures, and thus our study fundamentally differs from all the above characteri- zations. While we are heavily influenced by the axiomatic characterization of the Shapley value [31], we are also in- spired by social choice theory [3], and particularly by [26] on measures of intellectual influence and [2] on PageRank. 2. INFLUENCE AND CENTRALITY In this section, we review the basic concepts about social influence models and Shapley value, and define the Shapley and single node influence centrality measures. 2.1 Social Influence Models A network-influence instance is usually specified by a triple I = (V, E, PI), where a directed graph G = (V, E) represents the structure of a social network, and PI defines the influence model [18]. As an example, consider the classi- cal discrete-time independent cascade (IC) model, in which each directed edge (u, v) ∈ E has an influence probability pu,v ∈ [0, 1]. At time 0, nodes in a given seed set S are activated while other nodes are inactive. At time t ≥ 1, for any node u activated at time t − 1, it has one chance to activate each of its inactive out-neighbor v with an inde- pendent probability pu,v. When there is no more activation, the stochastic process ends with a random set I(S) of nodes activated during the process. The influence spread of S is σ(S) = E[I(S)], the expected number of nodes influenced by S. Throughout the paper, we use boldface symbols to represent random variables. Algorithmically, we will focus on the (random) triggering In this model [18], which has IC model as a special case. model, each v ∈ V has a random triggering set T (v), drawn from a distribution defined by the influence model over the power set of all in-neighbors of v. At time t = 0, triggering sets {T (v)}v∈V are drawn independently, and the seed set S is activated. At t ≥ 1, if v is not active, it becomes activated if some u ∈ T (v) is activated at time t − 1. The influence spread of S is σ(S) = E[I(S)], where I(S) denotes the random set activated by S. IC is the triggering model that: For each directed edge (u, v) ∈ E, add u to T (v) with an independent probability of pu,v. The triggering model can be equivalently viewed under the live-edge graph model: (1) Draw independent random triggering sets {T (v)}v∈V ; (2) form a live-edge graph L = (V,{(u, v) : u ∈ T (v)}), where (u, v), u ∈ T (v) is referred as a live edge. For any subgraph L of G and S ⊆ V , let Γ(L, S) be the set of nodes in L (cid:80) reachable from set S. Then set of active nodes with seed set S is Γ(L, S), and influence spread σ(S) = EL[Γ(L, S)] = L Pr(L = L) · Γ(L, S). We say a set function f (·) is monotone if f (S) ≤ f (T ) whenever S ⊆ T , and submodular if f (S ∪ {v}) − f (S) ≥ f (T ∪ {v}) − f (T ) whenever S ⊆ T and v (cid:54)∈ T . As shown in [18], in any triggering model, σ(·) is monotone and submodular, because Γ(L, S) is monotone and submodular for each graph L. More generally, we define an influence instance as a triple I = (V, E, PI), where G = (V, E) represents the underlying network, and PI : 2V × 2V → R defines the probability that in the influence process, any seed set S ⊆ V activates exactly nodes in any target set T ⊆ V and no other nodes: If II(S) denotes the random set activated by seed set S, then Pr(II(S) = T ) = PI(S, T ). This probability profile is commonly defined by a succinct influence model, such as the triggering model, which interacts with network G. We also require that: (a) PI(∅,∅) = 1, PI(∅, T ) = 0, ∀T (cid:54)= ∅, and (b) if S (cid:54)⊆ T then PI(S, T ) = 0, i.e., S always activates itself (S ⊆ II(S)). Such model is also referred to as the progressive influence model. The influence spread of S is: σI(S) = E[II(S)] = PI(S, T ) · T. (cid:88) T⊆V,S⊆T 2.2 Coalitional Games and Shapley Values An n-person coalitional game over V = [n] is specified by a characteristic function τ : 2V → R, where for any coalition S ⊆ V , τ (S) denotes the cooperative utility of S. In cooper- ative game theory, a ranking function φ is a mapping from a characteristic function τ to a vector in Rn. A fundamental solution concept of cooperative game theory is the ranking function given by the Shapley value [31]: Let Π be the set of all permutations of V . For any v ∈ V and π ∈ Π, let Sπ,v denote the set of nodes in V preceding v in permutation π. Then, ∀v ∈ V : 1 n! φShapley (τ ) = v (τ (Sπ,v ∪ {v}) − τ (Sπ,v)) S!(n − S − 1)! (cid:88) (cid:88) π∈Π (τ (S ∪ {v}) − τ (S)) . = S⊆V \{v} n! We use π ∼ Π to denote that π is a random permutation uniformly drawn from Π. Then: φShapley v (τ ) = Eπ∼Π[τ (Sπ,v ∪ {v}) − τ (Sπ,v)]. (2) The Shapley value of v measures v's marginal contribution over the set preceding v in a random permutation. Shapley [31] proved a remarkable representation theo- (cid:80) rem: The Shapley value is the unique ranking function that satisfies all the following four conditions: (1) Efficiency: v∈V φv(τ ) = τ (V ). (2) Symmetry: For any u, v ∈ V , if τ (S∪{u}) = τ (S∪{v}), ∀S ⊆ V \{u, v}, then φu(τ ) = φv(τ ). (3) Linearity: For any two characteristic functions τ and ω, for any α, β > 0, φ(ατ + βω) = αφ(τ ) + βφ(ω). (4) Null Player: For any v ∈ V , if τ (S∪{v})−τ (S) = 0, ∀S ⊆ V \{v}, then φv(τ ) = 0. Efficiency states that the total utility is fully distributed. Symmetry states that two players' ranking val- ues should be the same if they have the identical marginal utility profile. Linearity states that the ranking values of the weighted sum of two coalitional games is the same as the weighted sum of their ranking values. Null Player states that a player's ranking value should be zero if the player has zero marginal utility to every subset. 2.3 Shapley and SNI Centrality The influence-based centrality measure aims at assigning a value for every node under every influence instance: Definition 1 (Centrality Measure). An (influence-based) centrality measure ψ is a mapping from an influence instance I = (V, E, PI) to a real vector (ψv(I))v∈V ∈ RV . The single node influence (SNI) centrality, denoted by (I), assigns the influence spread of node v as v's cen- v ψSNI trality measure: ψSNI (I) = σI({v}). v The Shapley centrality, denoted by ψShapley(I), is the Shapley value of the influence spread function σI: ψShapley(I) = φShapley(σI). As a subtle point, note that φShapley maps from a 2V dimensional τ to a V -dimensional vector, while, formally, ψShapley maps from PI - whose di- mensions is close to 22V - to a V -dimensional vector. stance I = (V, E, PI),(cid:80) Axiom 2 (Normalization). For every influence in- v∈V ψv(I) = V . influence. The next axiom characterizes the centrality of a type In instance I = of extreme nodes in social (V, E, PI), we say v ∈ V is a sink node if ∀S, T ⊆ V \ {v}, PI(S ∪ {v}, T ∪ {v}) = PI(S, T ) + PI(S, T ∪ {v}). In the extreme case when S = T = ∅, PI({v},{v}) = 1, i.e., v can only influence itself. When v joins another S to form a seed set, the influence to a target T ∪{v} can always be achieved by S alone (except perhaps the influence to v itself). In the triggering model, a sink node is (indeed) a node without outgoing edges, matching the name "sink". Because a sink node v has no influence on other nodes, we can "remove" it and obtain a projection of the influence model on the network without v: Let I \{v} = (V \{v}, E \ {v}, PI\{v}) denote the projected instance over V \{v}, where E\{v} = {(i, j) ∈ E : v (cid:54)∈ {i, j}} and PI\{v} is the influence model such that for all S, T ⊆ V \ {v}: PI\{v}(S, T ) = PI(S, T ) + PI(S, T ∪ {v}). Intuitively, since sink node v is removed, the previously dis- tributed influence from S to T and T ∪ {v} is merged into the influence from S to T in the projected instance. For the triggering model, influence projection is simply remov- ing the sink node v and its incident incoming edges without changing the triggering set distribution of any other nodes. Axiom 3 below considers the simple case when the influ- ence instance has two sink nodes u, v ∈ V . In such a case, u and v have no influence to each other, and they influence no one else. Thus, their centrality should be fully determined by V \ {u, v}: Removing one sink node - say v - should not affect the centrality measure of another sink node u. Axiom 3 (Independence of Sink Nodes). For any influence instance I = (V, E, PI), for any pair of sink nodes u, v ∈ V in I, it should be the case: ψu(I) = ψu(I \ {v}). The next axiom considers Bayesian social influence through a given network: Given a graph G = (V, E), and (cid:80)r r influence instances on G: I η = (V, E, PIη ) with η ∈ [r]. Let λ = (λ1, λ2, . . . , λr) be a prior distribution on [r], i.e. η=1 λη = 1, and λη ≥ 0, ∀η ∈ [r]. The Bayesian influ- ence instance IB({Iη},λ) has the following influence process for a seed set S ⊆ V : (1) Draw a random index η ∈ [r] according to distribution λ (denoted as η ∼ λ). (2) Apply the influence process of I η with seed set S to obtain the activated set T . Equivalently, we have for all S, T ⊆ V , In the triggering model, we can view each live-edge graph and the determin- istic diffusion on it via reachability as an influence instance, and the diffusion of the triggering model is by the Bayesian (or convex) combination of these live-edge instances. The next axiom reflects the linearity-of-expectation principle: PIB({Iη},λ) (S, T ) = (cid:80)r η=1 ληPIη (S, T ). Axiom 4 (Bayesian Influence). For any network G = (V, E) and Bayesian social-influence model IB({Iη},λ): r(cid:88) ψv(IB({Iη},λ)) = Eη∼λ [ψv(I η)] = λη · ψv(I η),∀v ∈ V. η=1 The above axiom essentially says that the centrality of a Bayesian instance before realizing the actual model I η is the same as the expected centrality after realizing I η. Figure 1: Example on Shapley and SNI centrality. 3 (1 − p) · 1.5, To help understand these definitions, Figure 1 provides a simple example of a 3-node graph in the IC model with influ- ence probabilities shown on the edges. The associated table shows the result for Shapley and SNI centralities. While SNI is straightforward in this case, the Shapley centrality calculation already looks complex. For example, for node u, its second term in the Shapley computation, 1 accounts for the case where u is ordered in the second place (with probability 1/3), in which case only when the first- place node (either v or w) does not activate u (with prob- ability 1 − p), it could have marginal influence of 1 in acti- vating itself, and 0.5 in activating the remaining node. Sim- ilarly, the third term for the Shapley computation for node v accounts for the case where v is ordered second and w is ordered first (with probability 1/6), in which case if w does not activate u (with probability 1 − p), v's marginal influ- ence spread is 1 for itself and p for activating u; while if w activates u (with probability p), only when u does not ac- tivate v (with probability 0.5), v has marginal influence of 1 for itself. The readers can verify the rest. Based on the result, we find that for interval p ∈ (1/2, 2/3), Shapley and SNI centralities do not align in ranking: Shapley places v, w higher than u while SNI puts u higher than v, w. This sim- ple example already illustrates that (a) computing Shapley centrality could be a nontrivial task; and (b) the relationship between Shapley and SNI centralities could be complicated. Addressing both the computation and characterization ques- tions are the subject of the remaining sections. 3. AXIOMATIC CHARACTERIZATION In this section, we present two sets of axioms uniquely characterizing Shapley and SNI centralities, respectively, based on which we analyze their similarities and differences. 3.1 Axioms for Shapley Centrality Our set of axioms for characterizing the Shapley centrality is adapted from the classical Shapley's axioms [31]. The first axiom states that labels on the nodes should have no effect on centrality measures. This ubiquitous axiom is similar to the isomorphic axiom in some other centrality characterizations, e.g. [29]. Axiom 1 (Anonymity). For any influence instance I = (V, E, PI), and permutation π ∈ Π, ψv(I) = ψπ(v)(π(I)), ∀v ∈ V . In Axiom 1, π(I) = (π(V ), π(E), π(PI)) denotes the isomor- phic instance: (1) ∀u, v ∈ V , (π(u), π(v)) ∈ π(E) iff (u, v) ∈ E, and (2) ∀S, T ⊆ V , PI(S, T ) = Pπ(I)(π(S), π(T )). The second axiom states that the centrality measure di- vides the total share of influence V . In other words, the average centrality is normalized to 1. The last axiom characterizes the centrality of a family of simple social-influence instances. For any ∅ ⊂ R ⊆ U ⊆ V , a critical set instance IR,U = (V, E, PIR,U ) is such that: (1) The network G = (V, E) contains a complete directed bipartite sub-graph from R to U \ R, together with isolated nodes V \ U . (2) For all S ⊇ R, PIR,U (S, U ∪ S) = 1, and (3) For all S (cid:54)⊇ R, PIR,U (S, S) = 1. In IR,U , R is called the critical set, and U is called the target set. In other words, a seed set containing R activates all nodes in U , but missing any node in R the seed set only activates itself. We use IR,v to denote the special case of U = R ∪ {v} and V = U . That is, only if all nodes in R work together they can activate v. (Bargaining with Critical Sets). In any R R+1 , i.e. the centrality of v is Axiom 5 critical set instance IR,v, ψv(IR,v) = R R+1 . Qualitatively, Axiom 5 together with Normalization and Anonymity axioms implies that the relative importance of v comparing to a node in the critial set R increases when R increases, which is reasonable because when the critical set R grows, individuals in R becomes weaker and v becomes relatively stronger. This axiom can be interpreted through Nash's solution [22] to the bargaining game between a player representing the critical set R and the sink node v. Let r = R. Player R can influence all nodes by itself, achieving utility r + 1, while player v can only influence itself, with utility 1. The threat point of this bargaining game is (r, 0), which reflects the credits that each player agrees that the other player should at least receive: Player v agrees that player R's contribution is at least r, while player R thinks that player v may not have any contribution because R can activate everyone. The slack in this threat point is ∆ = r + 1 − (r + 0) = 1. However, in this case, player R is actually a coalition of r nodes, and these r nodes have to cooperate in order to influence all r +1 nodes - missing any node in R will not influence v. The need to cooperative in order to bargain with player v weakens player R. The ratio of v's bargaining weight to that of R is thus 1 to 1/r. Nash's bargaining solution [22] provides a fair division of this slack between the two players: (x1, x2) ∈ argmax x1≥r,x2≥0,x1+x2=r+1 (x1 − r)1/r · x2. The unique solution is (x1, x2) = (r + 1 v should receive a credit of r+1 , r r+1 , as stated in Axiom 5. r r+1 ). Thus, node Our first axiomatic representation theorem can now be stated as the following: Theorem 1. (Axiomatic of Shapley Centrality) The Shapley centrality ψShapley is the unique centrality measure that satisfies Axioms 1-5. Moreover, every axiom in this set is independent of others. Characterization The soundness of this representation theorem - that the Shapley centrality satisfies all axioms - is relatively sim- ple. However, because of the intrinsic complexity in influ- ence models, the uniqueness proof is in fact complex. We give a high-level proof sketch here and the full proof is in Appendix A.1. We follow Myerson's proof strategy [21] of Shapley's theorem. The probabilistic profile PI of influence instance I = (V, E, PI) is viewed as a vector in a large space RM , where M is the number of independent dimensions in PI. Bayesian Influence Axiom enforces that any conform- ing centrality measure is an affine mapping from RM to Rn. We then prove that the critical set instances IR,U form a full-rank basis of the linear space RM . Finally, we prove that any axiom-conforming centrality measure over critical set instances (and the additional null instance in which ev- ery node is a sink node) must be unique. The uniqueness of the critical set instances and the null instance, the linear independence of critical set instances in RM , plus the affine mapping from RM to Rn, together imply that the centrality measure of every influence instance is uniquely determined. Our overall proof is more complex and - to a certain degree - more subtle than Myerson's proof, because our axiomatic framework is based on the influence model in a much larger dimensional space compared to the subset utility functions. Finally, for independence, we need to show that for each ax- iom, we can construct an alternative centrality measure if the axiom is removed. Except for Axiom 5, the construc- tions and the proofs for other axioms are nontrivial, and they shed lights on how related centrality measures could be formed when some conditions are relaxed. 3.2 Axioms for SNI Centrality We first examine which of Axioms 1-5 are satisfied by SNI centrality. It is easy to verify that Anonymity and Bayesian Influence Axioms hold for SNI centrality. For the Indepen- dence of Sink Node Axiom (Axiom 3), since every sink node can only influence itself, its SNI centrality is 1. Thus, Ax- iom 3 is satisfied by SNI because of a stronger reason. For the Normalization Axiom (Axiom 2), the sum of sin- gle node influence is typically more than the total number of nodes (e.g., when the influence spread is submodular), and thus Axiom 2 does not hold for SNI centrality. The Bar- gaining with Critical Sets Axiom (Axiom 5) does not hold either, since node v in IR,v is a sink node and thus its SNI centrality is 1. We now present our axiomatic characterization of SNI centrality, which will retain Bayesian Influence Axiom 4, strengthen Independence of Sink Node Axiom 3, and rechar- acterize the centrality of a node in a critical set: Axiom 6 (Uniform Sink Nodes). Every sink node has centrality 1. Axiom 7 (Critical Nodes). In any critical set in- stance IR,U , the centrality of a node w ∈ R is 1 if R > 1, and is U if R = 1. These three axioms are sufficient to uniquely characterize SNI centrality, as they also imply Anonymity Axiom: Theorem 2. (Axiomatic Characterization of SNI Centrality) The SNI centrality ψSNI is the unique cen- trality measure that satisfies Axioms 4, 6, and 7. Moreover, each of these axioms is independent of the others. Theorems 1 and 2 establish the following appealing prop- erty: Even though all our axioms are on probabilistic pro- files PI of influence instances, the unique centrality measure satisfying these axioms is in fact fully determined by the in- fluence spread profile σI. We find this amazing because the distribution profile PI has much higher dimensionality than its influence-spread profile σI. 3.3 Shapley Centrality versus SNI Centrality We now provide a comparative analysis between Shapley and SNI centralities based on their definitions, axiomatic characterizations, and various other properties they satisfy. Comparison by definition. The definition of SNI cen- trality is more straightforward as it uses individual node's influence spread as the centrality measure. Shapley cen- trality is more sophisticatedly formulated, involving groups' influence spreads. SNI centrality disregards the influence profile of groups. Thus, it may limit its usage in more com- plex situations where group influences should be considered. Meanwhile, Shapley centrality considers group influence in a particular way involving marginal influence of a node on a given group randomly ordered before the node. Thus, Shapley centrality is more suitable for assessing marginal influence of a node in a group setting. Comparison by axiomatic characterization. Both SNI and Shapley centralities satisfy Anonymity, Independence of Sink Nodes, and Bayesian Influence axioms, which seem to be natural axioms for desirable social-influence centrality measures. Their unique axioms characterize exactly their differences. The first difference is on the Normalization Ax- iom, satisfied by Shapley but not SNI centrality. This in- dicates that Shapley centrality aims at dividing the total share of possible influence spread V among all nodes, but SNI centrality does not enforce such share division among nodes. If we artificially normalize the SNI centrality values of all nodes to satisfy the Normalization Axiom, the normal- ized SNI centrality would not satisfy the Bayesian Influence Axiom. (In fact, it is not easy to find a new characteriza- tion for the normalized SNI centrality similar to Theorem 2.) We will see shortly that the Normalization Axiom would also cause a drastic difference between the two centrality mea- sures for the symmetric IC influence model. The second difference is on their treatment of sink nodes, exemplified by sink nodes in the critical set instances. For SNI centrality, sink nodes are always treated with the same centrality of 1 (Axiom 6). But the Shapley centrality of a sink node may be affected by other nodes that influence the sink. In particular, for the critical set instance IR,v, v has centrality R/(R+1), which increases with R. As discussed earlier, larger R indicates v is getting stronger comparing to nodes in R. In this aspect, Shapley centrality assignment is sensible. Overall, when considering v's centrality, SNI cen- trality disregards other nodes' influence to v while Shapley centrality considers other nodes' influence to v. 1 The third difference is their treatment of critical nodes in the critical set instances. For SNI centrality, in the critical set instance IR,v, Axiom 7 obliviously assigns the same value 1 for nodes u ∈ R whenever R > 1, effectively equalizing the centrality of node u ∈ R with v. In contrast, Shapley centrality would assign u ∈ R a value of 1 + R(R+1) , de- creasing with R but is always larger than v's centrality of R R+1 . Thus Shapley centrality assigns more sensible values in this case, because u ∈ R as part of a coalition should have larger centrality than v, who has no influence power at all. We believe this shows the limitation of the SNI centrality - it only considers individual influence and disregards group influence. Since the critical set instances reflect the thresh- old behavior in influence propagation - a node would be influenced only after the number of its influenced neighbors reach certain threshold - this suggests that SNI centrality could be problematic in threshold-based influence models. Finally, we Comparison by additional properties. compare additional properties they satisfy. First, it is straightforward to verify that both centrality measures sat- isfy the Independence of Irrelevant Alternatives (IIA)) prop- erty: If an instance I = (V, E, PI) is the union of two dis- joint and independent influence instances, I1 = (V1, E1, PI1 ) and I2 = (V2, E2, PI2 ), then for k ∈ {1, 2} and any v ∈ Vk: ψv(I) = ψv(Ik). The IIA property together with the Normalization Axiom leads to a clear difference between SNI and Shapley central- ity. Consider an example of two undirected and connected graphs G1 with 10 nodes and G2 with 3 nodes, and the IC model on them with edge probability 1. Both SNI and Shapley centralities assign same values to nodes within each graph, but due to normalization, Shapley assigns 1 to all nodes, while SNI assigns 10 to nodes in G1 and 3 to nodes in G2. The IIA property ensures that the centrality does not change when we put G1 and G2 together. That is, SNI considers nodes in G1 more important while Shapley consid- ers them the same. While SNI centrality makes sense from individual influence point of view, the view of Shapley cen- trality is that a node in G1 is easily replaceable by any of the other 9 nodes in G1 but a node in G2 is only replaceable by two other nodes in G2. Shapley centrality uses marginal influence in randomly ordered groups to determine that the "replaceability factor" cancels out individual influence and assigns same centrality to all nodes. The above example generalizes to the symmetric IC model where pu,v = pv,u, ∀u, v ∈ V : Every node has Shapley cen- trality of 1 in such models. The technical reason is that such models have an equivalent undirected live-edge graph representation, containing a number of connected compo- nents just like the above example. The Shapley symme- try in the symmetric IC model may sound counter-intuitive, since it appears to be independent of network structures or edge probability values. But we believe what it unveils is that symmetric IC model might be an unrealistic model in practice - it is hard to imagine that between every pair of individuals the influence strength is symmetric. For ex- ample, in a star graph, when we perceive that the node in the center has higher centrality, it is not just because of its center position, but also because that it typically exerts higher influence to its neighbors than the reverse direction. This exactly reflects our original motivation that mere po- sitions in a static network may not be an important factor in determining the node centrality, and what important is the effect of individual nodes participating in the dynamic influence process. From the above discussions, we clearly see that (a) SNI centrality focuses on individual influence in isolation, while (b) Shapley centrality focuses on marginal influence in group influence settings, and measures the irreplaceability of the nodes in some sense. 4. SCALABLE ALGORITHMS In this section, we first give a sampling-based algorithm for approximating the Shapley centrality ψShapley(I) of any influence instance in the triggering model. We then give a slight adaptation to approximate SNI centrality. In both cases, we characterize the performance of our algorithms and prove that they are scalable for a large family of social- influence instances. In next section, we empirically show that these algorithms are efficient for real-world networks. 4.1 Algorithm for Shapley Centrality In this subsection, we use ψ as a shorthand for ψShapley. Let n = V and m = E. To precisely state our result, we make the following general computational assumption, as in [34, 33]: Assumption 1. The time to draw a random triggering set T (v) is proportional to the in-degree of v. The key combinatorial structures that we use are the fol- lowing random sets generated by the reversed diffusion pro- cess of the triggering model. A (random) reverse reachable (RR) set R is generated as follows: (0) Initially, R = ∅. (1) Select a node v ∼ V uniformly at random (called the root of R), and add v to R. (2) Repeat the following process until every node in R has a triggering set: For every u ∈ R not yet having a triggering set, draw its random triggering set T (u), and add T (u) to R. Suppose v ∼ V is selected in Step (1). The reversed diffusion process uses v as the seed, and follows the incoming edges instead of the outgoing edges to iteratively "influence" triggering sets. Equivalently, an RR set R is the set of nodes in a random live-edge graph L that can reach node v. Lemma 1 The following key lemma elegantly connects RR sets with Shapley centrality. We will defer its intuitive explanation to the end of this section. Let π be a random permutation on V . Let I{E} be the indicator function for event E. (Shapley Centrality Identity). Let R be a random RR set. Then, ∀u ∈ V , u's Shapley centrality is ψu = n · ER[I{u ∈ R}/R]. This lemma is instrumental to our scalable algorithm. It guarantees that we can use random RR sets to build unbi- ased estimators of Shapley centrality. Our algorithm ASV- RR (standing for "Approximate Shapley Value by RR Set") is presented in Algorithm 1. It takes ε, (cid:96), and k as input parameters, representing the relative error, the confidence of the error, and the number of nodes with top Shapley val- ues that achieve the error bound, respectively. Their exact meaning will be made clear in Theorem 3. ASV-RR follows the structure of the IMM algorithm of [33] but with some key differences. In Phase 1, Algorithm 1 estimates the number of RR sets needed for the Shapley es- timator. For a given parameter k, we first estimate a lower bound LB of the k-th largest Shapley centrality ψ(k). Fol- lowing a similar structure as the sampling method in IMM [33], the search of the lower bound is carried out in at most (cid:98)log2 n(cid:99)− 1 iterations, each of which halves the lower bound target x = n/2i and obtains the number of RR sets θi needed in this iteration (line 6). The key difference is that we do not need to store the RR sets and compute a max cover. Instead, for every RR set R, we only update the estimate est u of each node u ∈ R with an additional 1/R (line 9), which is based on Lemma 1. In each iteration, we select the k-th largest estimate (line 11) and plug it into the condition in line 12. Once the condition holds, we calculate the lower bound LB in line 13 and break the loop. Next we use this LB to obtain the number of RR sets θ needed in Phase 2 (line 17). In Phase 2, we first reset the estimates (line 19), then generate θ RR sets and again updating est u with 1/R increment for each u ∈ R (line 22). Finally, these estimates are transformed into the Shapley estimation in line 24. Unlike IMM, we do not reuse the RR sets generated in Phase 1, because it would make the RR sets dependent and the resulting Shapley centrality estimates biased. Moreover, our entire algorithm does not need to store any RR sets, and Input: Network: G = (V, E); Parameters: random trigger- ing set distribution {T (v)}v∈V , ε > 0, (cid:96) > 0, k ∈ [n] Output: ψv, ∀v ∈ V : estimated centrality measure 1: {Phase 1. Estimate the number of RR sets needed } √ 2 · ε; θ0 = 0 2: LB = 1; ε(cid:48) = 3: est v = 0 for every v ∈ V 4: for i = 1 to (cid:98)log2 n(cid:99) − 1 do 5: (cid:108) n·(((cid:96)+1) ln n+ln log2 n+ln 2)·(2+ 2 x = n/2i (cid:109) 3 ε(cid:48)) generate a random RR set R for every u ∈ R, est u = est u + 1/R end for est (k) = the k-th largest value in {est v}v∈V if n · est (k)/θi ≥ (1 + ε(cid:48)) · x then θi = ε(cid:48)2·x for j = 1 to θi − θi−1 do LB = n · est (k)/(θi · (1 + ε(cid:48))) break 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: end if 16: end for 17: θ = 18: {Phase 2. Estimate Shapley value} 19: est v = 0 for every v ∈ V 20: for j = 1 to θ do 21: 22: 23: end for 24: for every v ∈ V , ψv = n · est v/θ 25: return ψv, v ∈ V (cid:108) n(((cid:96)+1) ln n+ln 4)(2+ 2 ε2·LB (cid:109) 3 ε) generate a random RR set R for every u ∈ R, est u = est u + 1/R Algorithm 1: ASV-RR(G, T , ε, (cid:96), k) thus ASV-RR does not have the memory bottleneck encoun- tered by IMM when dealing with large networks. The follow- ing theorem summarizes the performance of Algorithm 1, where ψ and ψ(k) are Shapley centrality and k-th largest Shapley centrality value, respectively. Theorem 3. For any > 0, (cid:96) > 0, and k ∈ [n], Algo- rithm ASV-RR returns an estimated Shapley value ψv that satisfies (a) unbiasedness: E[ ψv] = ψv,∀v ∈ V ; (b) absolute ψv = n in every run; and (c) robust- ness: under the condition that ψ(k) ≥ 1, with probability at least 1 − 1 n(cid:96) : normalization: (cid:80) (cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). (3) Under Assumption 1 and the condition (cid:96) ≥ (log2 k − log2 log2 n)/ log2 n, the expected running time of ASV-RR is O((cid:96)(m + n) log n·E[σ(v)]/(ψ(k)ε2)), where E[σ(v)] is the ex- pected influence spread of a random node v drawn from V with probability proportional to the in-degree of v. v∈V Eq. (3) above shows that for the top k Shapley values, ASV-RR guarantees the multiplicative error of ε relative to node's own Shapley value (with high probability), and for the rest Shapley value, the error is relative to the k-th largest Shapley value ψ(k). This is reasonable since typically we only concern nodes with top Shapley values. For time com- plexity, the condition (cid:96) ≥ (log2 k−log2 log2 n)/ log2 n always hold if k ≤ log2 n or (cid:96) ≥ 1. When fixing ε as a constant, the running time depends almost linearly on the graph size (m + n) multiplied by a ratio E[σ(v)]/ψ(k). This ratio is upper bounded by the ratio between the largest single node influence and the k-th largest Shapley value. When these two quantities are about the same order, we have a near- linear time, i.e., scalable [35], algorithm. Our experiments show that in most datasets tested the ratio E[σ(v)]/ψ(k) is indeed less than 1. Moreover, if we could relax the robust- ness requirement in Eq. (3) to allow the error of ψv − ψv to be relative to the largest single node influence, then we could indeed slightly modify the algorithm to obtain a near- linear-time algorithm without the ratio E[σ(v)]/ψ(k) in the time complexity (see Appendix C.5). The accuracy of ASV-RR is based on Lemma 1 while the time complexity analysis follows a similar structure as in [33]. The proofs of Lemma 1 and Theorem 3 are presented in Appendix C. Here, we give a high-level explanation. In the triggering model, as for influence maximization [10, 34, 33], a random RR set R can be equivalently obtained by first generating a random live-edge graph L, and then con- structing R as the set of nodes that can reach a random v ∼ V in L. The fundamental equation associated with this live-edge graph process is: σ(S) = Pr L (L = L) Pr v (v ∈ Γ(L, S)) · n. (4) (cid:88) L Our Lemma 1 is the result of the following crucial observa- tions: First, the Shapley centrality ψu of node u ∈ V can be equivalently formulated as the expected Shapley centrality of u over all live-edge graphs and random choices of root v, from Eq. (4). The chief advantage of this formulation is that it localizes the contribution of marginal influences: On a fixed live-graph L and root v ∈ V , we only need to compute the marginal influence of u in terms of activating v to obtain the Shapley contribution of the pair. We do not need to compute the marginal influences of u for activating other nodes. Lemma 1 then follows from our second cru- cial observation. When R is the fixed set that can reach v in L, the marginal influence of u activating v in a random order is 1 if and only if the following two conditions hold concurrently: (a) u is in R - so u has chance to activate v, and (b) u is ordered before any other node in R - so u can activate v before other nodes in R do so. In addition, in a random permutation π ∼ Π over V , the probability that u ∈ R is ordered first in R is exactly 1/R. This explains the contribution of I{u ∈ R}/R in Lemma 1, which is also precisely what the updates in lines 9 and 22 of Algorithm 1 do. The above two observations together establish Lemma 1, which is the basis for the unbiased estimator of u's Shap- ley centrality. Then, by a careful probabilistic analysis, we can bound the number of random RR sets needed to achieve approximation accuracy stated in Theorem 3 and establish the scalability for Algorithm ASV-RR. 4.2 Algorithm for SNI Centrality u Algorithm 1 relies on the key fact given in Lemma 1 about = n · ER[I{u ∈ R}/R]. A the Shapley centrality: ψShapley u = σ({u}) = similar fact holds for the SNI centrality: ψSNI n· ER[I{u ∈ R}] [10, 34, 33]. Therefore, it is not difficult to verify that we only need to replace est u = est u + 1/R in lines 9 and 22 with est u = est u + 1 to obtain an approx- imation algorithm for SNI centrality. Let ASNI-RR denote the algorithm adapted from ASV-RR with the above change, and let ψv below denote SNI centrality ψSNI and ψ(k) denote the k-th largest SNI value. v Table 1: Datasets used in the experiments. "# Edges" refers to the number of undirected edges for the first three datasets and the number of directed edges for the last dataset. # Nodes 679 29,357 654,628 4,847,571 Dataset Data mining (DM) Flixster (FX) DBLP (DB) LiveJournal (LJ) # Edges Weight Setting 1687 WC, PR, LN 212,614 1,990,159 68,993,773 LN WC, PR WC Theorem 4. For any > 0, (cid:96) > 0, and k ∈ {1, 2, . . . , n}, Algorithm ASNI-RR returns an estimated SNI centrality ψv that satisfies (a) unbiasedness: E[ ψv] = ψv,∀v ∈ V ; and (b) robustness: with probability at least 1 − 1 n(cid:96) : (cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). (5) Under Assumption 1 and the condition (cid:96) ≥ (log2 k − log2 log2 n)/ log2 n, the expected running time of ASNI-RR is O((cid:96)(m + n) log n · E[σ(v)]/(ψ(k)ε2)), where E[σ(v)] is the same as defined in Theorem 1. Together with Algorithm ASV-RR and Theorem 3, we see that although Shapley and SNI centrality are quite different conceptually, surprisingly they share the same RR-set based scalable computation structure. Comparing Theorem 4 with Theorem 3, we can see that computing SNI centrality should be faster for small k since the k-th largest SNI value is usu- ally larger than the k-th largest Shapley value. 5. EXPERIMENTS We conduct experiments on a number of real-world social networks to compare their Shapley and SNI centrality, and test the efficiency of our algorithms ASV-RR and ASNI-RR. 5.1 Experiment Setup The network datasets we used are summarized in Table 1. The first dataset is a relatively small one used as a case study. It is a collaboration network in the field of Data Mining (DM), extracted from the ArnetMiner archive (ar- netminer.org) [32]: each node is an author and two authors are connected if they have coauthored a paper. The mapping from node ids to author names is available, allowing us to gain some intuitive observations of the centrality measure. We use three large networks to demonstrate the effective- ness of the Shapley and SNI centrality and the scalability of our algorithms. Flixster (FX) [4] is a directed network ex- tracted from movie rating site flixster.com. The nodes are users and a directed edge from u to v means that v has rated some movie(s) that u rated earlier. Both network and the influence probability profile are obtained from the authors of [4], which shows how to learn topic-aware influence prob- abilities. We use influence probabilities on topic 1 in their provided data as an example. DBLP (DB) is another aca- demic collaboration network extracted from online archive DBLP (dblp.uni-trier.de) and used for influence studies in [37]. Finally, LiveJournal (LJ) is the largest network we tested with. It is a directed network of bloggers, obtained from Stanford's SNAP project [1], and it was also used in [34, 33]. Table 2: Top 10 authors from DM dataset, ranked by Shapley, SNI, and degree centrality. We use the independent cascade (IC) model in our exper- iments. The schemes for generating influence-probability profiles are also shown in Table 1, where WC, PR, and LN stand for weighted cascade, PageRank-based, and learned from real data, respectively. WC is a scheme of [18], which assigns pu,v = 1/dv to edge (u, v) ∈ E, where dv is the in-degree of node v. PR uses the nodes' PageRanks [11] in- stead of in-degrees: We first compute the PageRank score r(v) for every node v ∈ V in the unweighted network, us- ing 0.15 as the restart parameter. Note that in our influ- ence network, edge (u, v) means u has influence to v; then when computing PageRank, we should reverse the edge di- rection to (v, u) so that v gives its PageRank vote to u, in order to be consistent on influence direction. Then, for each original edge (u, v) ∈ E, PR assigns an edge probability of r(u)/(r(u) + r(v)) · n/(2mU ), where mU is the number of undirected edges in the graph. The assignment achieves the effect that a higher PageRank node has larger influence to a lower PageRank nodes than the reverse direction (when both directions exist). The scaling factor n/(2mU ) is to normalize the total edge probabilities to Θ(n), which is similar to the setting of WC. PR defines a PageRank-based asymmetric IC model. LN applies to DM and FX datasets, where we obtain learned influence probability profiles from the authors of the original studies. For the DM dataset, the influence proba- bilities on edges are learned by the topic affinity algorithm TAP proposed in [32]; for FX, the influence probabilities are learned using maximum likelihood from the action trace data of user rating events. We implement all algorithms in Visual C++, compiled in Visual Studio 2013, and run our tests on a server computer with 2.4GHz Intel(R) Xeon(R) E5530 CPU, 2 processors (16 cores), 48G memory, and Windows Server 2008 R2 (64 bits). 5.2 Experiment Results Case Study on DM. We set ε = 0.01, (cid:96) = 1, and k = 50 for both ASV-RR and and ASNI-RR algorithms. For the three influence profiles: WC, PR, and LN, Table 2 lists the top 10 nodes in both Shapley and SNI ranking together with their numerical values. The names appeared in all ranking results are well-known data mining researchers in the field, at the time of the data collection 2009, but the ranking details have some difference. We compare the Shapley ranking versus SNI ranking un- der the same probability profiles. In general, the two top-10 ranking results align quite well with each other, showing that in these influence instances, high individual influence usually translates into high marginal influence. Some notice- able exception also exists. For example, Christos Faloutsos is ranked No.3 in the DM-PR Shapley centrality, but he is not in Top-10 based on DM-PR individual influence ranking. Conceptually, this would mean that, in the DM-PR model, Professor Faloutsos has better Shapley ranking because he has more unique and marginal impact comparing to his in- dividual influence. In terms of the numerical values, SNI values are larger than the Shapley values, which is expected due to the normalization factor in Shapley centrality. We next compare Shapley and SNI centrality with the structure-based degree centrality. The results show that the Shapley and SNI rankings in DM-WC and DM-PR are sim- ilar to the degree centrality ranking, which is reasonable be- cause DM-WC and DM-PR are all heavily derived from node degrees. However, DM-LN differs from degree ranking a lot, since it is derived from topic modeling, not node degrees. This implies that when the influence model parameters are learned from real-world data, it may contain further infor- mation such that its influence-based Shapley or SNI ranking may differ from structure-based ranking significantly. When comparing the numerical values of the same cen- trality measure but across different influence models, we see that Shapley values of top researchers in DM-LN are much higher than Shapley values of top researchers under DM-WC or DM-PR, which suggests that influence models learned from topic profiles differentiating nodes more than the synthetic WC or PR methods. The above results differentiating DM-LN from DM-WC and DM-PR clearly demonstrate the interplay between so- cial influence and network centrality: Different influence processes can lead to different centrality rankings, but when they share some aspects of common "ground-truth" influ- ence, their induced rankings are more closely correlated. Tuning Parameter ε We now investigate the impact of our ASV-RR/ASNI-RR pa- rameters, to be applied to our tests on large datasets. Pa- rameter (cid:96) is a simple parameter controlling the probability, 1 − 1 n(cid:96) , that the accuracy guarantee holds. We set it to 1, which is the same as in [34, 33]. For parameter ε, a smaller value improves accuracy at the cost of higher running time. Thus, we want to set ε at a proper level to balance accuracy and efficiency. ing the average value. The results on all datasets in general show that both Shapley and SNI centrality performs reason- ably well for the influence maximization task, but in some cases IMM is still noticeably better. This is because IMM is specially designed for the influence maximization task while Shapley and SNI are two centrality measures related to in- fluence but not specialized for the influence maximization task. For the FX-LN dataset, Shapley top rankers performs noticeably better than SNI top rankers (average 8.3% im- provement). This is perhaps due to that Shapley centrality accounts for more marginal influence, which is closer to what is needed for influence maximization. This is also the test where they both significantly outperform the baseline De- gree heuristic, again indicating that influence learned from the real-world data may contain significantly more informa- tion than the graph structure, in which case degree centrality is not a good index for node importance. The behavior of DBLP-PR needs a bit more attention. For ASNI-RR (as well as IMM and Degree), the first seed selected already generates influence spread of 95K, but subsequent seeds only have very small marginal contribution to the in- fluence spread. On the contrary, the first seed selected by ASV-RR only has influence spread of 77K, and the spread reaches the level of ASNI-RR at the fourth seed. Looking more closely, the first seed selected by ASV-RR has Shap- ley centrality of 10.3 but its influence spread of 77K is only ranked at around 68K on SNI ranking, while the first seed of ASNI-RR has Shapley centrality of 3.15, with Shapley rank- ing beyond 2100. This shows that when a large portion of nodes have high individual but overlapping influence (due to the emergence of the giant component in live-edge graphs), they all become more or less replaceable, and thus Shapley ranking, which focuses on marginal influence in a random order, would differs from SNI ranking significantly. Finally, we evaluate the scalability of ASV-RR and ASNI- RR, and use IMM as a reference point, even though IMM is designed for a different task. We use the same setting of ε = 0.5, (cid:96) = 1, and k = 50. Table 3 reports the running time of the three algorithms on four large influence instances. For FX-LN, DB-WC, and LJ-WC, the general trend is that IMM is the fastest, followed by ASNI-RR, and then ASV-RR. This is expected, because the theoretical running time of IMM is Θ((k + (cid:96))(m + n) log n · E[σ(v)]/(OP Tk · ε2)), where OP Tk is maximum influence spread with k seeds. Thus comparing to the running time results in Theorems 3 and 4, typically OP Tk is much larger than the k-th largest SNI centrality, which in turn is much larger than the k-th largest Shapley centrality, which leads to the observed running time result. Nevertheless, both ASNI-RR and ASV-RR could be considered efficient in these cases and they can scale to large graphs with tens of millions of nodes and edges. DB-PR again is an out-lier, with ASNI-RR faster than IMM, and ASV-RR being too slow and inefficient. This is because a large portion of nodes have large individual but overlapping influence, so that OP T50 = 95.9K is almost the same as the 50-th largest SNI value (94.2K), in which case the (k + (cid:96)) factor in the running time of IMM dominates and makes IMM slower than ASNI-RR. As for ASV-RR, due to the severe overlapping influence, the 50-th largest Shapley value (5.10) is much smaller than the 50-th largest SNI value or OP T50, resulting in much slower running time for ASV-RR. In summary, our experimental results on small and large datasets demonstrate that (a) Shapley and SNI centrality (a) Shapley computation (b) SNI computation Figure 2: Relative error of centrality computation when ε setting increases. We test different ε values from 0.1 to 2, on both DM and FX datasets, for both algorithms. To evaluate the accuracy, we use the results from ε∗ = 0.01 as the benchmark: For v ∈ V , suppose s∗ v and sv are the Shapley values computed for ε∗ = 0.01 and a larger ε value, respectively. Then, we compute sv − s∗ v/s∗ v and use it as the relative error at v. Since the top rankers' relative errors are more important, we take top 50 nodes from the two ranking results (using ε∗ and ε respectively), and compute the average relative error over the union of these two sets of top 50 nodes. Accordingly, we set parameter k = 50. We also apply the same relative error computation to SNI centrality. Figure 2 reports our results on the three DM options and the FX dataset, for both Shapley and SNI computations. We can see clearly that when ε ≤ 0.5, the relative errors of all datasets are within 0.05. In general, the actual relative error is below one tenth of ε in most cases, except for DM-PR dataset with ε ≥ 1. Hence, for the tests on large datasets, we use ε = 0.5 to provide reasonable accuracy for top values. Comparing to ε = 0.01, this reduces the running time 2500 fold, because the running time is proportional to 1/ε2. Results on Large Networks. We conduct experiments to evaluate both the effectiveness and the efficiency of our ASV-RR algorithm on large networks. For large networks, it is no longer easy to inspect rankings manually, especially when these datasets lack user profiles. For the effective- ness, we assess the effectiveness of Shapley and SNI cen- trality rankings through the lens of influence maximization. In particular, we use top rankers of Shapley/SNI centrality as seeds and measure their effectiveness for influence maxi- mization. We compare the quality and performance of our algorithm with the state-of-the-art scalable algorithm IMM proposed in [33] for influence maximization. Note that the IMM algorithm is based on the RR set approach. For IMM, we set its parameters as ε = 0.5, (cid:96) = 1, and k = 50, match- ing the parameter settings we use for ASV-RR/ASNI-RR. We also choose a baseline algorithm Degree, which is based on degree centrality to select top degree nodes as seeds for in- fluence maximization. We run ASV-RR, ASNI-RR, IMM, and Degree on four influ- ence instances: (1) the Flixster network with learned prob- ability, (2) the DBLP network with WC parameters, (3) the DBLP network with PR parameters, and (4) the Live- Journal network with WC parameters. Figure 3 shows the results of these four tests whose objectives are to identify 50 influential seed nodes. The influence spread in each case is obtained by running 10K Monte Carlo simulations and tak- (a) Flixster-LN (b) DBLP-WC (c) DBLP-PR (d) LiveJournal-WC Figure 3: Influence maximization test on IMM, ASV-RR, and Degree. Table 3: Running time (in seconds). Algorithm FX-LN DB-WC DB-PR LJ-WC ASV-RR 8295.57 ASNI-RR 267.50 IMM 54.88 594752 28.42 336.63 24.83 1.36 0.62 838.27 61.41 18.08 behaves similarly in these networks, but with noticeable differences; (b) for the influence maximization task, they perform close to the specially designed IM algorithm, with Shapley centrality noticeably better than SNI in some case; and (c) both can scale to large graphs with tens of millions of edges, with ASNI-RR having better scalability. except that ASV-RR would not be efficient for graphs with a huge gap be- tween individual influence and marginal influence. Finally, we remark that ASV-RR and ASNI-RR do not need to store RR sets, which eliminates a memory bottleneck that could be encountered by IMM on large datasets. 6. CONCLUSION AND FUTURE WORK Through an integrated mathematical, algorithmic, and empirical study of Shapley and SNI centralities in the con- text of network influence, we have shown that (a) both en- joy concise axiomatic characterizations, which precisely cap- ture their similarity and differences; (b) both centrality mea- sures can be efficiently approximated with guarantees under the same algorithmic structure, for a large class of influ- ence models; and (c) Shapley centrality focuses on nodes' marginal influence and their irreplaceability in group influ- ence settings, while SNI centrality focuses on individual in- fluence in isolation, and is not suitable in assessing nodes' ability in group influence setting, such as threshold-based models. There are several directions to extend this work and fur- ther explore the interplay between social influence and net- work centrality. One important direction is to formulate centrality measures that combine the advantages of Shapley and SNI centralities, by viewing Shapley and SNI centrali- ties as two extremes in a centrality spectrum, one focusing on individual influence while the other focusing on marginal influence in groups of all sizes. Then, would there be some intermediate centrality measure that provides a better bal- ance? Another direction is to incorporate other classical centralities into influence-based centralities. For example, SNI centrality may be viewed as a generalized version of de- gree centrality, because when we restrict the influence model to deterministic activation of only immediate neighbors, SNI centrality essentially becomes degree centrality. What about the general forms of closeness, betweenness, PageRank in the influence model? Algorithmically, efficient algorithms for other influence models such as general threshold models [18] is also interesting. In summary, this paper lays a foundation for the further development of the axiomatic and algorith- mic theory for influence-based network centralities, which we hope will provide us with deeper insights into network structures and influence dynamics. Acknowledgment We thank Tian Lin for sharing his IMM implementation code and helping on data preparation. We thank David Kempe and four anonymous reviewers for their valuable feedback. Wei Chen is partially supported by the National Natural Science Foundation of China (Grant No. 61433014). Shang-Hua Teng is supported in part by a Simons Investiga- tor Award from the Simons Foundation and by NSF grant CCF-1111270. 7. REFERENCES [1] Stanford network analysis project. https://snap.stanford.edu/data/. [2] A. Altman and M. Tennenholtz. Ranking systems: The pagerank axioms. In ACM, EC '05, pages 1–8, 2005. [3] K. J. Arrow. Social Choice and Individual Values. Wiley, New York, 2nd edition, 1963. [4] N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influence propagation models. In ICDM, 2012. [5] P. Boldi and S. Vigna. Axioms for centrality. Internet Mathematics, 10:222–262, 2014. [6] P. Bonacich. Factoring and weighting approaches to status scores and clique identification,. Journal of Mathematical Sociology, 2:113–120, 1972. [7] P. Bonacich. Power and centrality: A family of measures. American Journal of Sociology, 92(5):1170–1182, 1987. [8] S. P. Borgatti. Centrality and network flow. Social Networks, 27(1):55–71, 2005. [9] S. P. Borgatti. Identifying sets of key players in a social network. Computational and Mathematical Organizational Theory, 12:21–34, 2006. [10] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier. Maximizing social influence in nearly optimal time. In ACM-SIAM, SODA '14, pages 946–957, 2014. Games, Volume II, pages 307–317. Princeton University Press, 1953. [32] J. Tang, J. Sun, C. Wang, and Z. Yang. Social influence analysis in large-scale networks. In KDD, 2009. [33] Y. Tang, Y. Shi, and X. Xiao. Influence maximization in near-linear time: a martingale approach. In SIGMOD, pages 1539–1554, 2015. [34] Y. Tang, X. Xiao, and Y. Shi. Influence maximization: near-optimal time complexity meets practical efficiency. In SIGMOD, pages 75–86, 2014. [35] S.-H. Teng. Scalable algorithms for data and network analysis. Foundations and Trends in Theoretical Computer Science, 12(1-2):1–261, 2016. [36] S.-H. Teng. Network essence: Pagerank completion and centrality-conforming markov chains. In J. N. Martin Loebl and R. Thomas, editors, A Journey through Discrete Mathematics. A Tribute to Jir´ı Matousek. Springer Berlin / Heidelberg, 2017. [37] C. Wang, W. Chen, and Y. Wang. Scalable influence maximization for independent cascade model in large-scale social networks. DMKD, 25(3):545–576, 2012. [11] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. Computer Networks, 30(1-7):107–117, 1998. [12] W. Chen, Y. Yuan, and L. Zhang. Scalable influence maximization in social networks under the linear threshold model. In IEEE, ICDM '10, pages 88–97, 2010. [13] F. R. K. Chung and L. Lu. Concentration inequalities and martingale inequalities: A survey. Internet Mathematics, 3(1):79–127, 2006. [14] P. Domingos and M. Richardson. Mining the network value of customers. In ACM, KDD '01, pages 57–66, 2001. [15] R. Ghosh and K. Lerman. Rethinking centrality: The role of dynamical processes in social network analysis. Discrete and Continuous Dynamical Systems Series B, pages 1355–1372, 2014. [16] R. Ghosh, S.-H. Teng, K. Lerman, and X. Yan. The interplay between dynamics and networks: centrality, communities, and cheeger inequality. In ACM, KDD '14, pages 1406–1415, 2014. [17] L. Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43, March 1953. [18] D. Kempe, J. M. Kleinberg, and ´E. Tardos. Maximizing the spread of influence through a social network. In KDD, pages 137–146, 2003. [19] T. P. Michalak, K. V. Aadithya, P. L. Szczepanski, B. Ravindran, and N. R. Jennings. Efficient computation of the shapley value for game-theoretic network centrality. J. Artif. Int. Res., 46(1):607–650, Jan. 2013. [20] M. Mitzenmacher and E. Upfal. Probability and Computing. Cambridge University Press, 2005. [21] R. B. Myerson. Game Theory : Analysis of Conflict. Harvard University Press, 1997. [22] J. Nash. The bargaining problem. Econometrica, 18(2):155–162, April 1950. [23] M. Newman. Networks: An Introduction. Oxford University Press, Inc., New York, NY, USA, 2010. [24] U. Nieminen. On the centrality in a directed graph. Social Science Research, 2(4):371–378, 1973. [25] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. In Proceedings of the 7th International World Wide Web Conference, pages 161–172, 1998. [26] I. Palacios-Huerta and O. Volij. The measurement of intellectual influence. Econometrica, 72:963–977, 2004. [27] M. Piraveenan, M. Prokopenko, and L. Hossain. Percolation centrality: Quantifying graph-theoretic impact of nodes during percolation in networks. PLoS ONE, 8(1), 2013. [28] M. Richardson and P. Domingos. Mining knowledge-sharing sites for viral marketing. In ACM, KDD '02, pages 61–70, 2002. [29] G. Sabidussi. The centrality index of a graph. Psychometrika, 31(4):581–603, 1966. [30] D. Schoch and U. Brandes. Re-conceptualizing centrality in social networks. European Journal of Applied Mathematics, 27:971–985, 2016. [31] L. S. Shapley. A value for n-person games. In H. Kuhn and A. Tucker, editors, Contributions to the Theory of APPENDIX A. PROOFS ON AXIOMATIC CHARAC- TERIZATION A.1 Proof of Theorem 1 We use A to denote the set of Axioms 1-5. Analysis of Sink Nodes We first prove that the involvement of sink nodes in the in- fluence process is what we have expected: (1) The marginal contribution of a sink node v is equal to the probability that v is not influenced by the seed set. (2) For any other node u ∈ V , u's activation probability is the same whether or not v is in the seed set. Lemma 2. Suppose v is a sink node in I = (V, E, PI). Then, (a) for any S ⊆ V \ {v}: σI(S ∪ {v}) − σI(S) = Pr(v (cid:54)∈ II(S)). (b) for any u (cid:54)= v and any S ⊆ V \ {u, v}: Pr(u (cid:54)∈ II(S ∪ {v})) = Pr(u (cid:54)∈ II(S)). Proof. For (a), by the definitions of σI and sink nodes: σI(S ∪ {v}) = PI(S ∪ {v}, T ) · T Lemma 4. If v is a sink in I, then for any S ⊆ V \ {v}: σI\{v}(S) = σI(S) − Pr(v ∈ II(S)). Proof. By the definition of influence projection: (PI(S, T ) + PI(S, T ∪ {v})) · T σI\{v}(S) T⊇S,T⊆V \{v} PI\{v}(S, T ) · T (cid:88) (cid:88) (cid:88) PI(S, T ) · T − (cid:88) PI(S, T ) · T + T⊇S∪{v} T⊇S,T⊆V \{v} T⊇S,T⊆V \{v} (cid:88) T⊇S = = = = = σI(S) − Pr(v ∈ II(S)). (cid:88) T⊇S∪{v} PI(S, T ) PI(S, T ) · (T − 1) Lemma 5. For any two sink nodes u and v in I: σI\{v}(S ∪ {u}) − σI\{v}(S) = σI(S ∪ {u}) − σI(S). Proof. By Lemmas 4 and 2 (b), we have σI\{v}(S ∪ {u}) − σI\{v}(S) = σI(S ∪ {u}) − Pr(v ∈ II(S ∪ {u})) − (σI(S) − Pr(v ∈ II(S))) = σI(S ∪ {u}) − Pr(v ∈ II(S)) − (σI(S) − Pr(v ∈ II(S))) = σI(S ∪ {u}) − σI(S). (PI(S, T \ {v}) + PI(S, T )) · T (cid:88) PI(S, T )(T + 1) + PI(S, T ) · T Soundness T⊇S∪{v} T⊇S,T⊆V \{v} PI(S, T ) · T + (cid:88) T⊇S,T⊆V \{v} PI(S, T ) 5. Lemma 6. The Shapley centrality satisfies all Axioms 1- (cid:88) (cid:88) (cid:88) T⊇S∪{v} T⊇S∪{v} = = = (cid:88) T⊇S = σI(S) + Pr(v (cid:54)∈ II(S)). For (b), (cid:88) (cid:88) (cid:88) Pr(u (cid:54)∈ II(S ∪ {v})) = T⊇S∪{v},T⊆V \{u} = = T⊇S∪{v},T⊆V \{u} T⊇S,T⊆V \{u} PI(S ∪ {v}, T ) (PI(S, T \ {v}) + PI(S, T )) PI(S, T ) = Pr(u (cid:54)∈ II(S)). Lemma 2 immediately implies that for any two sink nodes u and v, u's marginal contribution to any S ⊆ V \ {u, v} is the same as its marginal contribution to S ∪ {v}: Lemma 3 (Independence between Sink Nodes). If u and v are two sink nodes in I, then for any S ⊆ V \{u, v}, σI(S ∪ {v, u}) − σI(S ∪ {v}) = σI(S ∪ {u}) − σI(S). Proof. By Lemma 2 (a) and (b), both sides are equal to Pr(u (cid:54)∈ II(S)). The next two lemmas connect the influence spreads in the original and projected instances. Proof. Axioms 1, 2, and 4 are trivially satisfied by ψShapley, or are direct implications from the original Shapley axiom set. Next, we show that ψShapley satisfies Axiom 3, the Axiom of Independence of Sink Nodes. Let u and v be two sink nodes. Let π be a random permutation on V . Let π(cid:48) be the random permutation on V \{v} derived from π by removing v from the random order. Let {u ≺π v} be the event that u is ordered before v in the permutation π. Then we have u (I) = Eπ[σI(Sπ,u ∪ {u}) − σI(Sπ,u)] ψShapley = Pr(u ≺π v)Eπ[σI(Sπ,u ∪ {u}) − σI(Sπ,u) u ≺π v]+ Pr(v ≺π u)Eπ[σI(Sπ,u ∪ {u}) − σI(Sπ,u) v ≺π u] = Pr(u ≺π v)Eπ(cid:48) [σI(Sπ(cid:48),u ∪ {u}) − σI(Sπ(cid:48),u)]+ Pr(v ≺π u)Eπ[σI(Sπ,u ∪ {u}) − σI(Sπ,u) v ≺π u] = Pr(u ≺π v)Eπ(cid:48) [σI(Sπ(cid:48),u ∪ {u}) − σI(Sπ(cid:48),u)]+ Pr(v ≺π u)· Eπ[σI(Sπ,u \ {v} ∪ {u}) − σI(Sπ,u \ {v}) v ≺π u] (6) = Pr(u ≺π v)Eπ(cid:48) [σI(Sπ(cid:48),u ∪ {u}) − σI(Sπ(cid:48),u)]+ Pr(v ≺π u)Eπ(cid:48) [σI(Sπ(cid:48),u ∪ {u}) − σI(Sπ(cid:48),u)] = Eπ(cid:48) [σI(Sπ(cid:48),u ∪ {u}) − σI(Sπ(cid:48),u)] = Eπ(cid:48) [σI\{v}(Sπ(cid:48),u ∪ {u}) − σI\{v}(Sπ(cid:48),u)] = ψShapley (I \ {v}). (7) u Eq.(6) above uses Lemma 3, while Eq.(7) uses Lemma 5. Finally, we show that ψShapley satisfies Axiom 5, the Criti- cal Set Axiom. By the definition of the critical set instance, we know that if influence instance I has critical set R, then σI(S) = V if S ⊇ R, and σI(S) = S if S (cid:54)⊇ R. Then for v (cid:54)∈ R, for any S ⊆ V \ {v}, σI(S ∪ {v}) − σI(S) = 0 if S ⊇ R, and σI(S ∪ {v}) − σI(S) = 1 if S (cid:54)⊇ R. For a ran- dom permutation π, the event R ⊆ Sπ,v is the event that all nodes in R are ordered before v in π, which has probability 1/(R + 1). Then we have that for v (cid:54)∈ R, ψShapley = Pr(R ⊆ Sπ,v)Eπ[σI(Sπ,v ∪ {v}) − σI(Sπ,v) R ⊆ Sπ,v]+ (I) = Eπ[σI(Sπ,v ∪ {v}) − σI(Sπ,v)] Pr(R (cid:54)⊆ Sπ,v)Eπ[σI(Sπ,v ∪ {v}) − σI(Sπ,v) R (cid:54)⊆ Sπ,v] v = Pr(R (cid:54)⊆ Sπ,v) = R R + 1 . Therefore, Shapley centrality ψShapley is a solution consistent with Axioms 1-5. Completeness (or Uniqueness) We now prove the uniqueness of axiom set A. Fix a set V . For any R, U ⊆ V with R (cid:54)= ∅ and R ⊆ U , we define the critical set instance IR,U , an extension to the critical set instance IR,v defined for Axiom 5. Definition 2 (General Critical Set Instances). For any R, U ⊆ V with R (cid:54)= ∅ and R ⊆ U , the critical set instance IR,U = (V, E, PIR,U ) is the following influence instance: (1) The network G = (V, E) contains a complete directed bipartite sub-graph from R to U \ R, together with isolated nodes V \U . (2) For all S ⊇ R, PIR,U (S, U∪S) = 1, and (3) For all S (cid:54)⊇ R, PIR,U (S, S) = 1. For this instance, R is called the critical set, and U is called the target set. Intuitively, in the critical set instance IR,U , once the seed set contains the critical set R, it guarantees to activate target set U together with other nodes in S; but as long as some nodes in R is not included in the seed set S, only nodes in S can be activated. These critical set instances play an important role in the uniqueness proof. Thus, we first study their properties. To study the properties of the critical set instances, it is helpful for us to introduce a special type of sink nodes called isolated nodes. We say v ∈ V is an isolated node in I = (V, E, PI), if ∀S, T ⊆ V \ {v} with S ⊆ T , PI(S ∪ {v}, T ∪ {v}) = PI(S, T ). In the extreme case, PI({v},{v}) = PI(∅,∅) = 1, meaning that v only acti- vates itself, No seed set can influence v unless it contains v: For any S, T ⊆ V \ {v} with S ⊆ T , PI(S, T ∪ {v}) ≤ T (cid:48)⊇S,T (cid:48)⊆V \{v} PI(S∪ {v}, T (cid:48) ∪ {v}) = 0. The role of v in any seed set is just to activate itself: The probability of activating other nodes is unchanged if v is removed from the seed set. It is easy to see that by definition an isolated node is a sink node. T (cid:48)⊇S,T (cid:48)⊆V \{v} PI(S, T (cid:48)) = 1−(cid:80) 1−(cid:80) Lemma 7 (Sinks and Isolated Nodes). In the criti- cal set instance IR,U , every node in V \ U is an isolated node, and every node in V \ R is a sink node. Proof. We first prove that every node v ∈ V \ U is an isolated node. Consider any two subsets S, T ⊆ V \ {v} with S ⊆ T . We first analyze the case when S ⊇ R. By Definition 2, PI(S∪{v}, T∪{v}) = 1 iff T∪{v} = U∪S∪{v}, which is equivalent to T = U ∪ S since v (cid:54)∈ U . This implies that PI(S ∪ {v}, T ∪ {v}) = PI(S, T ). We now analyze the case when S (cid:54)⊇ R. By Definition 2, PI(S ∪ {v}, T ∪ {v}) = 1 iff T ∪ {v} = S ∪ {v}, which is equivalent to T = S. This again implies that PI(S ∪ {v}, T ∪ {v}) = PI(S, T ). Therefore, v is an isolated node. Next we show that every node v (cid:54)∈ R is a sink node. Con- sider any two subsets S, T ⊆ T \{v} with S ⊆ T . In the case when S ⊇ R, PI(S∪{v}, T∪{v}) = 1 iff T∪{v} = U∪S∪{v}, which is equivalent to T = U ∪ S \ {v}. Depending on whether v ∈ U , T = U ∪ S \ {v} is equivalent to exactly one of T = U ∪ S or T ∪ {v} = U ∪ S being true. This im- plies that PI(S ∪{v}, T ∪{v}) = PI(S, T ) + PI(S, T ∪{v}). In the case when S (cid:54)⊇ R, PI(S ∪ {v}, T ∪ {v}) = 1 iff T ∪ {v} = S ∪ {v}, which is equivalent to T = S. This also implies that PI(S∪{v}, T ∪{v} = PI(S, T )+PI(S, T ∪{v}). Therefore, v is a sink node by definition. Lemma 8 (Projection). In the critical set instance IR,U , for any node v ∈ V \ U , the projected influence in- stance of IR,U on V \ {v}, IR,U \ {v}, is a critical set in- stance with critical set R and target U , in the projected graph G \ {v} = (V \ {v}, E \ {v}). For any node v ∈ U \ R, the projected influence instance of IR,U on V \{v}, IR,U \{v}, is a critical set instance with critical set R and target U \ {v}, in the projected graph G \ {v} = (V \ {v}, E \ {v}). Proof. First let v ∈ V \ U and consider the projected instance IR,U \ {v}. If S ⊆ V \ {v} is a subset with S ⊇ R, then by the definition of projection and critical sets: PIR,U \{v}(S, S ∪ U ) = PIR,U (S, S ∪ U ) + PIR,U (S, S ∪ U ∪ {v}) = 1 + 0 = 1. If S ⊆ V \ {v} is a subset with S (cid:54)⊇ R, similarly, we have: PIR,U \{v}(S, S) = PIR,U (S, S) + PIR,U (S, S ∪ {v}) = 1 + 0 = 1. Thus by Definition 2, IR,U \{v} is still a critical set instance with R as the critical set and U as the target set. Next let v ∈ U \ R and consider the projected instance IR,U \ {v}. If S ⊆ V \ {v} is a subset with S ⊇ R, then by the definition of projection and critical sets: PIR,U \{v}(S, S ∪ (U \ {v})) = PIR,U (S, S ∪ (U \ {v})) + PIR,U (S, S ∪ (U \ {v}) ∪ {v}) = 0 + 1 = 1. If S ⊆ V \ {v} is a subset with S (cid:54)⊇ R, similarly, we have: PIR,U \{v}(S, S) = PIR,U (S, S) + PIR,U (S, S ∪ {v}) = 1 + 0 = 1. Thus by Definition 2, IR,U \{v} is still a critical set instance with R as the critical set and U \ {v} as the target set. Lemma 9 (Uniqueness in Critical Set Instances). Fix a set V . Let ψ be a centrality measure that satisfies axiom set A. For any R, U ⊆ V with R (cid:54)= ∅ and R ⊆ U , the centrality ψ(IR,U ) of the critical set instance IR,U must be unique. Proof. Consider the critical set instance IR,U . First, it is easy to check that all nodes in R are symmetric to one another, all nodes in U \ R are symmetric to one another, and all nodes in V \ U are symmetric to one another. Thus, by the Anonymity Axiom (Axiom 1), all nodes in R have the same centrality measure, say aR,U , all nodes in U \ R have the same centrality measure, say bR,U , and all nodes in V \ U have the same centrality measure, say cR,U . By the Normalization Axiom (Axiom 2), we have aR,U · R + bR,U · (U − R) + cR,U · (V − U) = V . (8) Second, we consider any node v ∈ V \U . By Lemma 7, v is an isolated node, which is also a sink node. By Lemma 8, we can iteratively remove all sink nodes in U \ R, which would not change the centrality measure of v by the Independence of Sink Nodes Axiom (Axiom 3). Moreover, after removing all nodes in U \ R, the projected instance I(cid:48) is on set R ∪ (V \U ), with R as both the critical set and the target set. In this projected instance I(cid:48), it is straightforward to check that for every S ⊆ R∪ (V \ U ), PI(cid:48) (S, S) = 1, which implies that every node in R ∪ (V \ U ) is an isolated node. Then we can apply the Anonymity Axiom to know that every node in I(cid:48) has the same centrality, and together with the Normalization Axiom, we know that every node in I(cid:48) has centrality 1. Since by the Independence of Sink Nodes Axiom removing nodes in U \ R does not change the centrality of nodes in V \ U , we know that cR,U = 1. Third, if U = R, then we do not have parameter bR,U and aR,U is determined by Eq. (8). If U (cid:54)= R, then by Lemma 7, any node v ∈ V \ R is a sink node. Then we can apply the Sink Node Axiom (Axiom 3) to iteratively remove all nodes in V \ (R ∪ {v}) (which are all sink nodes), such that the centrality measure of v does not change after the removal. By Lemma 8, the remaining instance with node set R ∪ {v} is still a critical set instance with critical set R and target set R ∪ {v}. Thus we can apply the Critical Set Axiom (Axiom 5) to this remaining influence instance, and know that the centrality measure of v is R/(R + 1), that is, bR,U = R/(R + 1). Therefore, aR,U is also uniquely de- termined, which means that the centrality measure ψ(IR,U ) for instance IR,U is unique, for every nonempty subset R and its superset U . for any S,(cid:80) The influence probability profile, (PI(S, T ))S⊆T⊆V , of each social-influence instance I can be viewed as a high- dimensional vector. Note that in the boundary cases: (1) when S = ∅, we have PI(S, T ) = 1 iff T = ∅; and (2) when S = V , PI(S, T ) = 1 iff T = V . Thus, the influence-profile vector does not need to include S = ∅ and S = V . Moreover, T⊇S PI(S, T ) = 1. Thus, we can omit the entry associated with one T ⊇ S from influence-profile vector. In our proof, we canonically remove the entry associated with T = S from the vector. With a bit of overloading on the no- tation, we also use PI to denote this influence-profile vector for I, and thus PI(S, T ) is the value of the specific dimension of the vector corresponding to S, T . We let M denote the di- mension of space of the influence-profile vectors. M is equal to the number of pairs (S, T ) satisfying (1) S ⊂ T ⊆ V , and (2) S (cid:54)∈ {∅, V }. S ⊂ T means S ⊆ T but S (cid:54)= T . We stress that when we use PI as a vector and use linear combinations of such vectors, the vectors have no dimension corresponding to (S, T ) with S ∈ {∅, V } or S = T . For each R and U with R ⊂ U and R (cid:54)∈ {∅, V }, we consider the critical set instance IR,U and its corresponding vector PIR,U . Let V be the set of these vectors. Lemma 10 (Linear Independence). Vectors in V are linearly independent in the space RM . we have a number αR,U ∈ R, such that(cid:80) Proof. Suppose, for a contradiction, that vectors in V are not linearly independent. Then for each such R and U , R(cid:54)∈{∅,V },R⊂U αR,U · PIR,U = (cid:126)0, and at least some αR,U (cid:54)= 0. Let S be the smallest set with αS,U (cid:54)= 0 for some U ⊃ S, and let T be any superset of S with αS,T (cid:54)= 0. By the critical set instance definition, we have PIS,T (S, T ) = 1. Also since the vector does not contain any dimension corresponding to PI(S, S), we know that T ⊃ S. Then by the minimality of S, we have αS,U · PIS,U (S, T )+ αR,U · PIR,U (S, T ) (cid:88) U :U⊃S,U(cid:54)=T αR,U · PIR,U (S, T ) αS,U · PIS,U (S, T )+ αR,U · PIR,U (S, T ). (9) (cid:88) (cid:88) (cid:88) (cid:88) 0 = R,U :R(cid:54)∈{∅,V },R⊂U = αS,T · PIS,T (S, T ) + R,U :R≥S,R(cid:54)=S,U⊃R = αS,T + U :U⊃S,U(cid:54)=T R,U :R≥S,R(cid:54)=S,U⊃R For the third term in Eq.(9), consider any set R with R ≥ S and R (cid:54)= S. We have that S (cid:54)⊇ R, and thus by the critical set instance definition, for any U ⊃ R, PIR,U (S, S) = 1. Since T ⊃ S, we have T (cid:54)= S, and thus PIR,U (S, T ) = 0. This means that the third term in Eq.(9) is 0. For the second term in Eq.(9), consider any U ⊃ S with U (cid:54)= T . By the critical set instance definition, we have PIS,U (S, U ) = 1 (since S is the critical set and U is the target set). Then PIS,U (S, T ) = 0 since T (cid:54)= U . This means that the second term in Eq.(9) is also 0. Therefore, vectors in V are linearly independent. Then we conclude that αS,T = 0, which is a contradiction. The following basic lemma is useful for our uniqueness proof. (cid:33) ψ((cid:126)v) = ψ i=1 αi = 1, ψ((cid:80)s Lemma 11. Let ψ be a mapping from a convex set D ⊆ RM to Rn satisfying that for any vectors (cid:126)v1, (cid:126)v2, . . . , (cid:126)vs ∈ D, i=1 αi · i=1 αi · ψ((cid:126)vi). Suppose that D contains a set of lin- early independent basis vectors of RM , {(cid:126)b1,(cid:126)b2, . . . ,(cid:126)bM} and also vector (cid:126)0. Then for any (cid:126)v ∈ D, which can be represented for any α1, α2, . . . , αs ≥ 0 and (cid:80)s (cid:126)vi) =(cid:80)s as (cid:126)v =(cid:80)M (cid:32) M(cid:88) (cid:32) (cid:33) i=1 λi · (cid:126)bi for some λ1, λ2, . . . , λM ∈ R, we have 1 − M(cid:88) M +1 ((cid:80)M i=1 Proof. We consider (cid:126)v = (cid:80)M formed by convex hull {(cid:126)b1,(cid:126)b2, . . . ,(cid:126)bM} together with (cid:126)0. Let (cid:126)v(0) = 1 (cid:126)bi + (cid:126)0), which is an interior point in the convex hull. For any (cid:126)v ∈ D, since {(cid:126)b1,(cid:126)b2, . . . ,(cid:126)bM} is a set of basis, we have i=1 λi · (cid:126)bi for some λ1, λ2, . . . , λM ∈ R. Let (cid:126)v(1) = ρ(cid:126)v(0) + (1 − ρ)(cid:126)v with ρ ∈ (0, 1) be a convex combination of (cid:126)v(0) and (cid:126)v. Then we have ψ((cid:126)v(1)) = ρψ((cid:126)v(0)) + (1 − ρ)ψ((cid:126)v), or equivalently λi· ψ((cid:126)bi) + M(cid:88) i=1 · ψ((cid:126)0). λi λi · (cid:126)bi = the i=1 i=1 ψ((cid:126)v) = 1 1 − ρ ψ((cid:126)v(1)) − ρ 1 − ρ ψ((cid:126)v(0)). (10) ρ ρ ρ i=1( M +1 + (1 − ρ)(1 −(cid:80)M We select a ρ close enough to 1 such that for all i ∈ [M ], Then (cid:126)v(1) =(cid:80)M i=1 λi) ≥ 0. M +1 + (1 − ρ)λi ≥ 0, and (cid:80)M M +1 + (1 − ρ)(1− i=1 λi))(cid:126)0 is in the convex hull of {(cid:126)b1,(cid:126)b2, . . . ,(cid:126)bM ,(cid:126)0}. Then (cid:32) M(cid:88) from Eq.(10), we have M +1 + (1 − ρ)λi)(cid:126)bi + ( (cid:33) (cid:18) ρ (cid:32) M(cid:88) λi · (cid:126)bi ψ((cid:126)v) =ψ i=1 1 ρ (cid:33) − (cid:126)bi+ (cid:126)0 i=1 λi + (1 − ρ)λi (cid:19) (cid:32) (cid:33)(cid:33) 1 − M(cid:88) (cid:32) M(cid:88) (cid:33)(cid:33) (cid:19) (cid:33)(cid:33) (cid:33)(cid:33) + (1 − ρ)λi (cid:126)bi + (cid:126)0 1 − M(cid:88) (cid:33) ψ((cid:126)bi) + ψ((cid:126)0) λi i=1 i=1 · ψ((cid:126)0). λi 1 − ρ ψ (cid:32) M + 1 i=1 + (1 − ρ) M + 1 ρ (cid:32) (cid:32) M(cid:88) (cid:18) ρ ψ 1 M + 1 ρ 1 − ρ 1 − ρ 1 (cid:32) i=1 ρ (cid:32) M + 1 ρ 1 − ρ M(cid:88) λiψ((cid:126)bi) + + (1 − ρ) M + 1 (cid:32) (cid:32) M(cid:88) (cid:32) 1 − M(cid:88) i=1 1 M + 1 = = = ψ((cid:126)bi)+ (cid:33) − ψ((cid:126)0) i=1 i=1 Lemma 12 (Completeness). The centrality measure satisfying axiom set A is unique. set A. Proof. Let ψ be a centrality measure that satisfies axiom Fix a set V . Let the null influence instance I N to be the instance in which no seed set has any influence except to it- self, that is, For any S ⊆ V , PIN (S, S) = 1. It is straightfor- ward to check that every node is an isolated node in the null instance, and thus by the Anonymity Axiom (Axiom 1) and the Normalization Axiom (Axiom 2), we have ψv(I N ) = 1 for all v ∈ V . That is, ψv(I N ) is uniquely determined. Note that, by our canonical convention of influence-profile vector space, PIN (S, S) is not in the vector representation of PIN . Thus vector PIN is the all-0 vector in RM . By Lemma 10, we know that V is a set of basis for RM . Then for any influence instance I, (cid:88) PI = R(cid:54)∈{∅,V },R⊂U λR,U · PIR,U , where parameters λR,U ∈ R. Because of the Bayesian Influ- ence Axiom (Axiom 4), and the fact that the all-0 vector in RM is the influence instance I N , we can apply Lemma 11 and obtain: ψ(PI) = (cid:88) 1 − (cid:88) R(cid:54)∈{∅,V },R⊂U + λR,U · ψ(PIR,U )  ψ(PIN ), λR,U R(cid:54)∈{∅,V },R⊂U where the notation ψ(PI) is the same as ψ(I). By Lemma 9 we know that all ψ(PIR,U )'s are uniquely determined. By the argument above, we also know that ψ(PIN ) is uniquely determined. Therefore, ψ(PI) must be unique. Independence An axiom is independent if it cannot be implied by other axioms in the axiom set. Thus, if an axiom is not inde- pendent, the centrality measure satisfying the rest axioms should still be unique by Lemma 12. Therefore, to show the independence of an axiom, it is sufficient to show that there is a centrality measure different from the Shapley centrality that satisfies the rest axioms. We will show the indepen- dence of each axiom in A in the next series of lemmas. Lemma 13. The Anonymity Axiom (Axiom 1) is inde- pendent. Proof. We consider a centrality measure ψ(1) defined as follows. Let Π(cid:48) be a nonuniform distribution on all permu- tations over set V , such that for any node v ∈ V , the prob- ability that v is ordered at the last position in a random permutation π drawn from Π(cid:48) is 1/V , but the probabilities of v in other positions may not be uniform. Such a nonuni- form distribution can be achieved by uniformly pick v ∈ V and put v in the last position, and then apply an arbitrary nonuniform distribution for the rest V − 1 positions. We then define ψ(1) as: v (I) = Eπ∼Π(cid:48) [σI(Sπ,v ∪ {v}) − σI(Sπ,v)]. ψ(1) Since Π(cid:48) is nonuniform, the above defined ψ(1) is not Shapley centrality, although it has the same form. We now verify that ψ(1) satisfies Axioms 2–5. Actually, since ψ(1) follows the same form as ψShapley, one can easily check that it also satisfies Axioms 2, 3 and 4. In particular, for Axiom 3, one can check the proof of Lemma 6 and see that the proof for Shapley centrality satisfying Axiom 3 does not rely on whether random permutation π is drawn from a uniform or nonuniform distribution of permutations. Thus the same proof works for the current ψ(1). For Axiom 5, following the same proof as in the proof of Lemma 6, we v (IR,v) = Prπ∼Π(cid:48) (R (cid:54)⊆ Sπ,v) = 1 − Prπ∼Π(cid:48) (R = have ψ(1) Sπ,v). As we know, for distribution Π(cid:48), node v appearing as the last node in a random permutation π drawn from Π(cid:48) is 1/V = 1/(R + 1), which is exactly Prπ∼Π(cid:48) (R = Sπ,v). v (IR,v) = R/(R + 1). Axiom 5 also Therefore, we have ψ(1) holds. As a remark, the Anonymity Axiom 1 does not hold for ψ(1): Consider the influence instance I where every subset deterministically influences all nodes. In this case, for any permutation π, π(I) is the same as I, because every node is symmetric. Axiom 1 says in this case all nodes should have the same centrality. Notice that by our definition of is exactly V times the probability of v being ψ(1), ψ(1) ranked first in a random permutation π drawn from Π(cid:48). But since Π(cid:48) is nonuniform, some node u would have higher probability to be ranked first than some other node v, and thus ψ(1) v (I), and Axiom 1 does not hold. u (I) > ψ(1) v Lemma 14. The Normalization Axiom (Axiom 2) is in- dependent. (11) Proof. For this axiom, we define ψ(2) first on the critical set instances in V, and then use their linear independence to define ψ(2) on all instances. For every instance IR,U ∈ V, we define Then by Eq. (14),  aR,U,V R R+1 c v ∈ R, v ∈ U \ R, v ∈ V \ U. (12) v (IR,U ) = ψ(2) ψ(2)(IB({Iη},λ)) = R,U · ψ(2)(IR,U ) Iη R(cid:54)∈{∅,V },R⊂U r(cid:88) (cid:88) r(cid:88) ληλ η=1 ληλ Iη R,U  ψ(2)(I N )   ψ(2)(I N ) R,U · ψ(2)(IR,U ) Iη λ Iη R,U λ R(cid:54)∈{∅,V },R⊂U η=1 1 − (cid:88)  (cid:88) 1 − (cid:88) λη · R(cid:54)∈{∅,V },R⊂U R(cid:54)∈{∅,V },R⊂U = + r(cid:88) η=1 + r(cid:88) We can show that by Axioms 1, 3 and 5, the above centrality assignments are the only possible assignments. In fact, for every v ∈ V \U , we can repeatedly apply Axiom 3 to remove nodes in U \R first, and then remove all but v to get a single node instance, which must have one centrality value, and we denote it c. For every v ∈ U \ R, we apply Axiom 3 again to remove all nodes in V \ (R ∪ {v}). and then apply Axiom 5 to show that v must have centrality R/(R + 1). For every node v ∈ R, by Anonymity Axiom, they must have the same centrality within the same instance IR,U , and then further apply Anonymity Axiom between two instances with the same size of V , R and U, we know that they all have the same value, and thus we can use aR,U,V to denote it. Thus, for a fixed V , totally the degree of freedom for ψ(2) is (V − 2)(V − 1)/2 + 1. For the null instance I N defined in the proof of Lemma 12 (where every node is an isolated node), applying Axiom 3 repeatedly we know that the centrality of every node in I N must be c. For an arbitrary instance I, by Lemma 10 we have (cid:88) PI = R(cid:54)∈{∅,V },R⊂U R,U · PIR,U , I λ (13) where λI R,U ∈ R. Now we define ψ(2)(I) as ψ(2)(I) = (cid:88) 1 − (cid:88) R(cid:54)∈{∅,V },R⊂U + R(cid:54)∈{∅,V },R⊂U R,U · ψ(2)(IR,U ) I λ  ψ(2)(I N ). I R,U λ (14) It is obvious that the definition of ψ(2) does not depend on node labeling, and thus it satisfies Axiom 1. Since v (IR,v) = R/(R + 1), it satisfies Axiom 5. Since all ψ(2) construction is linear, it is not hard to see that it satisfies Axiom 4, and we provide the complete derivation below. For any Bayesian instance IB({Iη},λ), by definition we have η=1 ληPIη (S, T ). Note that here λη's and λI PIB({Iη},λ) (S, T ) =(cid:80)r R,U 's are different sets of parameters. Thus we have PIB({Iη},λ) = ληPIη η=1 r(cid:88) (cid:88) (cid:32) r(cid:88) R(cid:54)∈{∅,V },R⊂U r(cid:88) η=1 = = λη (cid:88) R(cid:54)∈{∅,V },R⊂U η=1 λ R,U · PIR,U Iη (cid:33) Iη R,U ληλ · PIR,U . = ληψ(2)(I η), η=1 where the second equality uses the fact that (cid:80)r η=1 λη = 1. Therefore, the Bayesian Axiom (Axiom 4) holds. Finally, we verify that the Independence of Sink Node Ax- iom (Axiom 3) holds. Suppose u and v are two sink nodes of an instance I = (V, E, PI). Since we need to work on pro- R,U and I V,N to jection, we clarify the notation and use I V represent the critical instance and the null instance, respec- tively, in set V . By the definition of projection and Eq. (13), we have for any ∅ ⊂ S ⊂ T ⊆ V \ {v}, PI\{v}(S, T ) = PI(S, T ) + PI(S, T ∪ {v}) = R,U ·(cid:16) (S, T ∪ {v}) (cid:88) (S, T ) + PIV (cid:17) PIV λ I . R,U R,U ∅⊂R⊂U⊆V (15) R,U R,U R,U R,U For v (cid:54)∈ R, v is a sink node in I V (S, T ) + PIV For v ∈ R, by the definition of I V R,U , we know that for (S, T ∪ {v}) = S ⊂ T ⊆ V \ {v}, PIV (S, T ) = PIV R,U , and thus 0. (S, T ∪ {v}) = PIV R,U \{v}(S, T ). From PIV R,U \ {v} = I V \{v} Lemma 8, we know that the projection I V R,U\{v} when if v ∈ U\R. when v ∈ V \U , and I V In particular, if U = R ∪ {v}, then I V \{v} R,U\{v} is the null in- stance where every node is an isolated node, in which case (S, T ) = 0 for any S ⊂ T . Combining the above, PIV \{v} we continue Eq. (15) to have R,U \{v} = I V \{v} R,U\{v} R,U PI\{v}(S, T ) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) λ R,U R,U\{v} ∅⊂R⊂U⊆V,v(cid:54)∈U R,U · PIV \{v} I R,U · PIV \{v} I λ ∅⊂R⊂U⊆V,v∈U\R R,U · PIV \{v} I R,U∪{v} · PIV \{v} I (cid:16) λ I R,U + λ ∅⊂R⊂U⊆V \{v} ∅⊂R⊂U⊆V \{v} I R,U∪{v} R,U R,U λ λ = = = ∅⊂R⊂U⊆V \{v} (S, T )+ (S, T ) (S, T )+ (S, T ) (cid:17) · PIV \{v} R,U (S, T ). Since PI\{v} has unique PIV \{v} 's, for all ∅ ⊂ R ⊂ U ⊆ V \ {v} we have linear representation from R,U ∅⊂R⊂U⊆V \{v},u∈R I\{v} R,U = λ λ I R,U + λ I R,U∪{v}. (16) = ∅⊂R⊂U⊆V,u∈R,v(cid:54)∈U We now derive the relation between ψ(2) u (I). By Eqs. (12), (14) and (16), ψ(2) u (I \ {v}) and  ψ(2) u (I \ {v}) = ψ(2) I\{v} R,U ψ(2) λ u (I V \{v} R,U ) ∅⊂R⊂U⊆V \{v} I\{v} R,U λ u (I V \{v},N ) ∅⊂R⊂U⊆V \{v} I R,U + λ λ I R,U∪{v} (cid:88) (cid:16) ∅⊂R⊂U⊆V \{v},u(cid:54)∈U 1 − (cid:88) (cid:88) (cid:88) (cid:88) 1 − (cid:88) (cid:88) (cid:88) + + + + = = ∅⊂R⊂U⊆V \{v},u∈U\R ∅⊂R⊂U⊆V \{v},u∈R (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) λ ∅⊂R⊂U⊆V \{v} ∅⊂R⊂U⊆V \{v},u∈U\R + + 1 − ∅⊂R⊂U⊆V \{v},u∈R (cid:88) (cid:16) ∅⊂R⊂U⊆V \{v},u∈U I I R,U∪{v} R,U + λ λ I I R,U∪{v} R,U + λ λ I R,U + λ λ I R,U∪{v} I R,U + λ I R,U∪{v} I I R,U∪{v} R,U + λ λ R R + 1 (cid:17) · c (cid:17) · (cid:17) · aR,U,V −1 (cid:17) · c (cid:17) · (cid:17) · aR,U,V −1 (cid:17) · c (17) R + 1 R I I R,U∪{v} R,U + λ λ ∅⊂R⊂U⊆V \{v},u∈U Note that(cid:88) (cid:88) (cid:88) = = ∅⊂R⊂U⊆V,u∈U,v(cid:54)∈R (cid:16) λ I R,U + λ I R,U∪{v} (cid:17) (cid:88) I λ R,U + I R,U . λ ∅⊂R⊂U⊆V,u∈U,v(cid:54)∈U ∅⊂R⊂U⊆V,u∈U,v∈U\R I λ R,U (cid:17) · R R + 1 (cid:16) λ I R,U + λ R,U · I λ I R,U∪{v} R R + 1 (cid:88) (cid:88) (cid:88) (cid:88) ∅⊂R⊂U⊆V \{v},u∈U\R ∅⊂R⊂U⊆V,u∈U\R,v(cid:54)∈U = = + ∅⊂R⊂U⊆V,u∈U\R,v∈U\R R,U · I λ ∅⊂R⊂U⊆V,u∈U\R,v(cid:54)∈R R R + 1 λ R,U · I R R + 1 . (cid:88) (cid:88) (cid:88) (cid:16) λ I R,U + λ I R,U∪{v} (cid:17) · aR,U,V −1 R,U · aR,U,V −1 I λ R,U · aR,U−1,V −1 I λ Plugging the above three results into Eq. (17), we have + + ∅⊂R⊂U⊆V,u∈R,v∈U\R u (I \ {v}) = ψ(2) (cid:88) (cid:88) ∅⊂R⊂U⊆V,u∈R,v(cid:54)∈U ∅⊂R⊂U⊆V,u∈R,v∈U\R (cid:88) R,U · I λ R R + 1 ∅⊂R⊂U⊆V,u∈U\R,v(cid:54)∈R R,U · aR,U,V −1 I λ (cid:88) R,U · aR,U−1,V −1 I λ  · c I λ R,U (18) Similarly, we expand ψ(2) u (I) = ψ(2) ∅⊂R⊂U⊆V,u∈U,v(cid:54)∈R u (I): u (I V I R,U ψ(2) λ R,U )  ψ(2) I R,U λ u (I V,N ) (cid:88) ∅⊂R⊂U⊆V R,U · c + I λ ∅⊂R⊂U⊆V,u(cid:54)∈U ∅⊂R⊂U⊆V,u∈U\R λ R,U · I 1 − (cid:88) ∅⊂R⊂U⊆V R R + 1  · c I R,U λ + + ∅⊂R⊂U⊆V 1 − (cid:88) 1 − (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) 1 − (cid:88) ∅⊂R⊂U⊆V,u∈R ∅⊂R⊂U⊆V,u∈R ∅⊂R⊂U⊆V,u∈U\R = = + + + + + − − − R,U · aR,U,V + I λ R,U · I R R + 1 λ R,U · aR,U,V I λ  · c (cid:88) I R,U λ ∅⊂R⊂U⊆V,u∈U Subtracing Eq. (18) from Eq. (19), we have u (I) − ψ(2) ψ(2) u (I \ {v}) = ∅⊂R⊂U⊆V,u∈R ∅⊂R⊂U⊆V,u∈R,v(cid:54)∈U ∅⊂R⊂U⊆V,u∈U\R,v∈R R,U · aR,U,V I λ R,U · aR,U,V −1 I λ R,U · aR,U−1,V −1 I λ (cid:88) (cid:88) (cid:88) (cid:88) ∅⊂R⊂U⊆V,u∈R,v∈U\R R,U · c. I λ ∅⊂R⊂U⊆V,u∈U,v∈R (19) R,U · I λ R R + 1 (20) We want the above difference to be zero, but so far we have not used the property that both u and v are sink nodes in I yet, except that the project I \ {v} is defined when v is a sink node. Next, suppose that v is a sink node in I and we would derive some properties on λI R,U based on this fact. By the definition of sink nodes, we have for all ∅ ⊆ S ⊂ T ⊆ V \ {v}, (cid:88) I R,U PIV λ (S ∪ {v}, T ∪ {v}) R,U ∅⊂R⊂U⊆V = PI(S ∪ {v}, T ∪ {v}) = PI(S, T ) + PI(S, T ∪ {v}) (cid:88) = ∅⊂R⊂U⊆V (cid:16) I R,U λ PIV R,U (S, T ) + PIV R,U (S, T ∪ {v}) (cid:17) . (S, T ) + PIV Note that when v (cid:54)∈ R, v is a sink node in I V R,U , and so we (S, T ∪ {v}) = PIV (S ∪ {v}, T ∪ have PIV {v}). When v ∈ R, since v (cid:54)∈ S and S ⊂ T , we have (S, T ∪{v}) = 0. Thus the above implies PIV that (S, T )+PIV R,U R,U R,U R,U R,U I R,U PIV R,U λ (S ∪ {v}, T ∪ {v}) = 0. (cid:88) ∅⊂R⊂U⊆V,v∈R Note that if R (cid:54)⊆ S ∪ {v}, then PIV (S ∪ {v}, T ∪ {v}) = 0. (S ∪ {v}, T ∪ {v}) = 1 if and only When R ⊆ S ∪ {v}, PIV if T ∪ {v} = S ∪ {v} ∪ U ; otherwise it is 0. Thus the above is equivalent to (cid:88) R,U R,U I λ R,U = 0. (21) R,U :∅⊂R⊂U,v∈R,R⊆S∪{v},T∪{v}=S∪{v}∪U Set S = ∅ first. Then R must be {v}. For any T such that S ⊂ T ⊆ V \ {v}, we see that U must be T ∪ {v} to satisfy the constraint in the above summation. Thus we have {v},U = 0 for any {v} ⊂ U ⊆ V . Now, let S = 1. In this λI case R = {v} or R = S ∪ {v}. We already know from above that if R = {v}, λI {v},U = 0. Thus in Eq. (21), what are left are the terms with R = S ∪{v}. When R = S ∪{v}, we can see that U must be T ∪ {v}. Then we obtain that λI R,U = 0 for every R = 2, v ∈ R, and R ⊂ U ⊆ V . Repeating the above argument for R = 3, 4, . . . (or S = 2, 3, . . . ,), we eventually conclude that for every R and U such that v ∈ R and R ⊂ U ⊆ V , λI R,U = 0. u (I) = ψ(2) With the above important property, we look back at Eq. (20). Notice that in all the five summation terms, we have either v ∈ R or u ∈ R, and both u and v are sink nodes. Therefore, all these five summation terms are 0, and u (I \ {v}) for an ar- finally we conclude that ψ(2) bitrary instance I. This means that the Independence of Sink Nodes Axiom (Axiom 3) always holds for the definition of ψ(2) (Eq. (12)) with any possible parameters aR,U,V 's and c. Hence, we have many degree of freedom to chose these parameters other than the ones determined by the Shapley centrality, and thus the Normalization Axiom (Axiom 2) is independent. Lemma 15. The Independence of Sink Nodes Axiom (Ax- iom 3) is independent. Proof. Similar to the proof of Lemma 14, we first define ψ(3) on critical set instances in V, and then use their linear independence to extend the definition to all instances. For every instance IR,U ∈ V, we know that R ⊂ U and v (IR,U ) = R (cid:54)∈ {∅, V }. When R = V − 1, we define ψ(3) R/(R + 1) for the unique v ∈ V \ R, and ψ(3) u (IR,U ) = (V −R/(R+1))/R for every u ∈ R. When R (cid:54)= V −1, v (IR,U ) = 1 for all v ∈ V . It is straight- we simply define ψ(3) forward to see that for every IR,U ∈ V, ψ(3) is anonymous (not depend on the label of a node), and normalized (cen- trality summed up to V ), and for the critical set instance IR,v, it satisfies the requirement of Axiom 5. For the null instance I N defined in the proof of Lemma 12, we define v (I N ) = 1 for every v ∈ V . Thus for I N ψ(3) is also ψ(3) anonymous and normalized. Now for an arbitrary influence instance I, by Lemma 10 we have PI = (cid:88) R,U · PIR,U , I λ R(cid:54)∈{∅,V },R⊂U where λI by Eq. (11): R,U ∈ R. Then we define ψ(3)(I) below patterned ψ(3)(I) = (cid:88) 1 − (cid:88) R(cid:54)∈{∅,V },R⊂U + R(cid:54)∈{∅,V },R⊂U R,U · ψ(3)(IR,U ) I λ  ψ(3)(I N ). I R,U λ (22) It is straightforward to verify that when all IR,U ∈ V and I N are anonymous and normalized, I is also anonymous and normalized, and thus ψ(3) satisfies Axioms 1 and 2. More- over, ψ(3) also satisfies the Bayesian Axiom (Axiom 4), with the same proof as the one in the proof of Lemma 14. By its definition, we already know that ψ(3) satisfies Axiom 5. Obviously, ψ(3) is different from ψShapley and it does not satisfies the Independence of Sink Nodes Axiom (Axiom 3). Hence, Axiom 3 is independent. Lemma 16. The Bayesian Axiom (Axiom 4) is indepen- dent. (cid:80) Proof. We define a centrality measure ψ(4) as follows. Given an influence instance I = (V, E, PI), for every sink node v in I, if there is a set R ⊆ V \ {v}, such that T :R∪{v}⊆T PI(R, T ) = 1, we let Rv be the smallest such set (tie is broken with some arbitrary deterministic rule); if such R does not exists, then we let Rv = ∅. Intu- itively, Rv is the smallest set that can influence v with probability 1. Then, for every sink node v, we define v (I) = Rv/(Rv + 1); for non-sink nodes, we let them ψ(4) equally devide the rest share so that the total centrality is V . The definition of ψ(4) does not depend on node labeling, so Axiom 1 is clearly satisfied. The definition enforces that the sum of all centralities is V , so Axiom 2 is satisfied. For the critical set infance IR,v, v is a sink node, and by definition R is the smallest one such that PIR,v (R, R∪{v}) = 1, so by v (IR,v) = R/(R + 1), the definition of ψ(4), we have ψ(4) thus satisfying Axiom 5. For the Independence of Sink Nodes Axiom (Axiom 3), consider an influence instance I and its two sink nodes u and (cid:80) v. We claim that the projection I \ {v} does not change set Ru. In fact, suppose first that in I there is a set R ⊆ V \{u} T :R∪{u}⊆T PI(R, T ) = 1. If v (cid:54)∈ R, then by the definition of projection,(cid:88) (cid:88) (cid:88) = = T :R∪{u}⊆T⊆V \{v} T :R∪{u}⊆T⊆V \{v} T :R∪{u}⊆T⊆V PI\{v}(R, T ) PI(R, T ) + PI(R, T ∪ {v}) PI(R, T ) = 1. PI(R \ {v}, T \ {v}) + PI(R \ {v}, T ) PI(R, T ) = 1, where the second to last equality is by the definition of sink node. Thus, the above equation implies that R \ {v} is a smaller set that influences u with probability 1. The cases of v ∈ R and v (cid:54)∈ R together imply that Ru for instance I still works for instance I \ {v}. bility 1 in I \ {v}. Then we have Conversely, suppose R is a set influencing u with proba- (cid:88) (cid:88) (cid:88) PI(R, T ) T :R∪{u}⊆T⊆V = = T :R∪{u}⊆T⊆V \{v} T :R∪{u}⊆T⊆V \{v} PI(R, T ) + PI(R, T ∪ {v}) PI\{v}(R, T ) = 1. Therefore, R is still a set influencing u with probability 1 in I. Hence, from the above argument from both sides, we know that, either there does not exists a set R that influences u with probability 1 in I or I \ {v}, or the smallest such sets in I and I \{v} are the same. Therefore, we have ψ(4) u (I) = u (I \ {v}) = Rv/(Rv + 1), where Rv is the smallest ψ(4) such set or ∅. This means, ψ(4) satisfies the Independence of Sink Nodes Axiom (Axiom 3). Since ψ(4) is clearly different from ψShapley, we know that Axiom 3 is independent. Lemma 17. The Bargaining with Critical Sets Axiom (Axiom 5) is independent. Proof. We construct ψ(5) by trivially assigning every node with centrality 1. It is obvious that this constant ψ(5) satisfies Axioms 1–4, and it is different from ψShapley. Thus Axiom 5 is independent. Lemmas 13–17 together implies the following: Lemma 18 set A are independent. (Independence). All axioms in the axiom Proof of Theorem 1. The theorem is proved by com- bining Lemmas 6, 12 and 18. Thus R is still a set in I \ {v} that influence v with proba- bility 1. If v ∈ R, since v is a sink node, we have PI(R \ {v}, T ) = η=1 λησIη ({v}), (cid:88) (cid:88) (cid:88) T :R\{v}∪{u}⊆T⊆V = = T :R∪{u}⊆T⊆V T :R∪{u}⊆T⊆V A.2 On Theorem 2 The proof of Theorem 2, the axiomatic characterization of SNI centrality, follows the same structure as the proof of Theorem 1. For soundness, it is easy to verify that SNI centrality satisfies Axioms 4, 6, and 7. In particular, for the Bayesian Influence Axiom (Axiom 4), we can verify that σIB({Iη},λ) ({v}) = r(cid:88) PIB({Iη},λ) ({v}, T ) · T (cid:88) r(cid:88) (cid:88) T⊆V,v∈T ληPIη ({v}, T ) = PIη ({v}, T ) η=1 η=1 λη T⊆V,v∈T = (cid:88) r(cid:88) T⊆V,v∈T and thus Bayesian Influence Axiom also holds for SNI cen- trality. For completeness, since SNI centrality also satisfies the Bayesian Influence Axiom, we following the same proof structure as Lemma 12, which utilizes the linear mapping lemma 11. All we need to show is that for all critical set in- stances IR,U as well as the null instance I N , Axioms 6, and 7 dictate that their centrality measure is unique. For the null instance, by the Uniform Sink Node Axiom (Axiom 6), we know that all nodes have centrality value of 1. For the critical set instance IR,U , again by the Uniform Sink Node Axiom, we know that all nodes in V \ R are sink nodes and thus have centrality value of 1. Finally, Axiom 7 uniquely determines the centrality value of all critical nodes in R. Therefore, Axioms 4, 6, and 7 uniquely determines the cen- trality measure, which is SNI centrality. This proves Theo- rem 2. For independence, the independence of Axiom 7 can be shown by considering the uniform centrality measure where every node is assigned centrality of 1. We can see that the uniform centrality satisfies Axiom 4 and 6 but not 7. For the independence of Axiom 6, we can see that Axiom 7 only restricts the nodes in R in the critical set instances IR,U . Then we can assign arbitrary values, for example 0, to nodes not in R in IR,U , and thus obtaining a centrality measure defined on the critical set instances that is consistent with Axiom 7 but different from ψSNI. Next, we use the linearity (Eq. (11)) to extend the centrality measure to arbitrary in- stances. Bayesian Axiom (Axiom 4) holds because our way of linear extension. Finally, for the independence of Ax- iom 4, we notice that with Axioms 6 and 7, the centrality for all critical set instances are uniquely determined, but we have the freedom to define other instances, as long as the sink nodes always have centrality of 1. This means we can easily find a centrality measure that is different from ψSNI but satisfies Axioms 6 and 7. Therefore, Axioms 4, 6 and 7 are all independent of one another. B. SHAPLEY SYMMETRY OF SYMMET- RIC IC MODELS In this appendix section, we formally prove the Shapley symmetry of the symmetric IC model stated in Section 3. We restate it in the following theorem. Theorem 5 (Shapley Symmetry of Symmetric IC). In any symmetric IC model, the Shapley centrality of every node is the same. We first prove the following basic lemma. Lemma 19 (Deterministic Undirected Influence). Consider an undirected graph G = (V, E), and the IC in- stance I on G in which for every undirected edge (u, v) ∈ E, pu,v = pv,u = 1. Then, ψShapley (I) = 1, ∀v ∈ V , v Proof. Let C be the connected component containing node v. For any fixed permutation π of V , if some other node u ∈ C appears before v in π - i.e. u ∈ Sπ,v - then because all edges have influence probability 1 in both directions, u influences every node in C. For this permutation, v has no marginal influence: σI(Sπ,v∪{v})−σI(Sπ,v) = 0. If v is the first node in C that appears in π, then v activates every node in C, and its marginal spread is C. The probability that v appears first among all nodes in C in a random permutation π is exactly 1/C. Therefore: ψShapley (I) = Eπ[σI(Sπ,v∪{v})−σI(Sπ,v)] = 1/C·C = 1. v Proof of Theorem 5. We will use the following well- known but important observation about symmetric IC mod- els: We can use the following undirected live-edge graph model to represent its influence spread. For every edge (u, v) ∈ E, since we have pu,v = pv,u, we sample an undi- rected edge (u, v) with success probability pu,v. The result- ing undirected random live-edge graph is denoted as ¯L. For any seed set S, the propagation from the seed set can only pass through each edge (u, v) at most once, either from u to v or from v to u, but never in both directions. Therefore, we can apply the Principle of Deferred Decision and only de- cide the direction of the live edge (u, v) when the influence process does need to pass the edge. Hence, the set of nodes reachable from S in the undirected graph ¯L, namely Γ( ¯L, S), is the set of activated nodes. Thus, σI(S) = E ¯L[Γ( ¯L, S)]. For each "deferred" realization ¯L of ¯L, the propagation on ¯L is the same as treating every edge in ¯L having influence probability 1 in both directions. Then, by Lemma 19, the Shapley centrality of every node on the fixed ¯L is the same. Finally, by taking expectation over the distribution of ¯L, we have: ψShapley v (I) =Eπ[σI(Sπ,v ∪ {v}) − σI(Sπ,v)] =Eπ[E ¯L[Γ( ¯L, Sπ,v ∪ {v}) − Γ( ¯L, Sπ,v)]] =E ¯L[Eπ[Γ( ¯L, Sπ,v ∪ {v}) − Γ( ¯L, Sπ,v)]] =E ¯L[1] = 1. C. ANALYSIS OF ASV-RR In this appendix, we provide a complete proof of The- orem 3, and briefly extend the discussion to the proof of Theorem 6. In the discussion below, we will use v ∼ V to denote that v is drawn uniformly at random from V . We will use π ∼ Π(V ) to denote that π is a uniform random permutation of V . Let I{E} be the indicator function for event E. Let m = E and n = V . C.1 Unbiasedness and Absolute Normaliza- tion of the Shapley Estimator of ASV-RR We first build connections between random RR sets and the Shapley value computation. The following is a straight- forward proposition to verify: Proposition 20. Fix a subset R ⊆ V . For any v ∈ R, Pr(R ∩ Sπ,v = ∅) = 1/R, where π ∼ Π(V ) and Sπ,v is the subset of nodes preceding v in π. Proof. The event R∩Sπ,v = ∅ is equivalent to π placing v ahead of other nodes in R. Because π ∼ Π(V ), this event happens with probability exactly 1/R. Proposition 21. A random RR set R is equivalently generated by first (a) generating a random live-edge graph L, and (b) selecting v ∼ V . Then, R is the set of nodes that can reach v in L. Lemma 22 (Marginal Contribution). Let R be a random RR set. For any S ⊆ V and v ∈ V \ S: σ(S) = n · Pr(S ∩ R (cid:54)= ∅), (23) σ(S ∪ {v}) − σ(S) = n · Pr(v ∈ R ∧ S ∩ R = ∅).(24) Proof. Let L be a random live-edge graph generated by the triggering model (see Section 2.1). Recall that Γ(L, S) denote the set of nodes in graph L reachable from set S. Then: σ(S) =EL[Γ(L, S)] =EL I{u ∈ Γ(L, S)} (cid:35) (cid:34)(cid:88) (cid:34)(cid:88) u∈V 1 n u∈V · I{u ∈ Γ(L, S)} =n · EL =n · EL [Eu∼V [I{u ∈ Γ(L, S)}]] =n · Pr L,u∼V {u ∈ Γ(L, S)}, (cid:35) (cid:35) (cid:35) Note that for any function f , and random variables x, y: Ex [Ey[f (x, y)]] = E [E[f (x, y) x = x]] . In other words, we can evaluate the expectation as the fol- lowing: (1) fix the value of random variable x to x first, then (2) take the conditional expectation of f (x, y) conditioned upon x = x, and finally (3) take the expectation according to x's distribution. By Proposition 21, event u ∈ Γ(L, S) is the same as the event S ∩ R (cid:54)= ∅. Hence we have σ(S) = n · Pr(S ∩ R (cid:54)= ∅). Similarly, σ(S ∪ {v}) − σ(S) = EL[Γ(L, S ∪ {v}) \ Γ(L, S)] = EL I{u ∈ Γ(L, S ∪ {v}) \ Γ(L, S)} (cid:34)(cid:88) (cid:34)(cid:88) u∈V 1 n u∈V · I{u ∈ Γ(L, S ∪ {v}) \ Γ(L, S)} = n · EL = n · EL [Eu∼V [I{u ∈ Γ(L, S ∪ {v}) \ Γ(L, S)}]] = n · Pr L,u∼V {u ∈ Γ(L, S ∪ {v}) \ Γ(L, S)}. By a similar argument, event u ∈ Γ(L, S ∪ {v}) \ Γ(L, S) is the same as the event v ∈ R ∧ S ∩ R = ∅. Hence we have σ(S ∪ {v}) − σ(S) = n · Pr(v ∈ R ∧ S ∩ R = ∅). For a fixed subset R ⊆ V and a node v ∈ V , define: (cid:26) 0 1R if v (cid:54)∈ R; if v ∈ R. XR(v) = If R is a random RR set, then XR(v) is a random variable. The following is a restatement of Lemma 1 using the XR(v) random variable. Lemma 23 (Shapley Value Identity). Let R be a random RR set. Then, for all v ∈ V , the Shapley centrality of v is ψv = n · ER[XR(v)]. Proof. Let R be a random RR set. We have {by Eq. (2)} {by Lemma 22} ψv =Eπ[σ(Sπ,v ∪ {v}) − σ(Sπ,v)] =Eπ[n · Pr(v ∈ R ∧ Sπ,v ∩ R = ∅)] =n · Eπ[ER[I{v ∈ R ∧ Sπ,v ∩ R = ∅}]] =n · ER[Eπ[I{v ∈ R ∧ Sπ,v ∩ R = ∅}]]. By Proposition 20, for any realization of R: Eπ∼Π(V )[I{v ∈ R ∧ Sπ,v ∩ R = ∅}] = This means that Eπ∼Π(V )[I{v ∈ R∧Sπ,v∩R = ∅}] = XR(v). Therefore, ψv = n · ER[XR(v)]. if v (cid:54)∈ R, if v ∈ R. (cid:26) 0 1R After the above preparation, we are ready to show the unbiasedness of our Shapley estimator. Lemma 24 (Unbiased Estimator). For any v ∈ V , the estimated value ψv returned by Algorithm 1 satisfies E[ ψv] = ψv, where the expectation is taken over all ran- domness used in Algorithm ASV-RR. Proof. In Phase 2 of Algorithm ASV-RR, when θ is fixed to θ, the algorithm generates θ independent random RR sets R1, . . . , Rθ. Let est θ v be the value of est v at the end of the for-loop in Phase 2, when θ = θ. It is straightforward to see that est θ i=1 XRi (v). Therefore, by Lemma 23: v =(cid:80)θ E[ ψv θ = θ] = E[n · est θ XRi (v)/θ] = ψv. v/θ] = E[n · θ(cid:88) i=1 Since this is true for every fixed θ, we have E[ ψv] = ψv. run of ASV-RR, we have(cid:80) Lemma 25 v∈V ψv = n. (Absolute Normalization). In every Proof. According to line 22 of the algorithm, for every RR set R generated in Phase 2, each node u ∈ R increases its estimate est u by 1/R and no other nodes increase their estimates. Thus the total increase in the estimates of all nodes for each R is exactly 1. Then after generating θ RR sets, the sum of estimates is θ. According to line 24, we conclude that(cid:80) ψv = n. v∈V C.2 Robustness of the Shapley Estimator of ASV-RR The analysis on the robustness and time complexity is similar to that of IMM in [33], but since we are working on Shapley values while IMM is for influence maximization, there are also a number of differences. In what follows, we provide an indepdent and complete proof for our algorithm, borrowing some ideas from [33]. We will use the following basic Chernoff bounds [20, 13] in our analysis. Fact 26 (Chernoff Bounds). Let Y be the sum of t i.i.d. random variables with mean µ and value range [0, 1]. For any δ > 0, we have: Pr{Y − tµ ≥ δ · tµ} ≤ exp − δ2 2 + 2 3 δ tµ . For any 0 < δ < 1, we have Pr{Y − tµ ≤ −δ · tµ} ≤ exp − δ2 2 tµ . (cid:18) (cid:18) (cid:19) (cid:19) Let ψ(k) be the k-th largest value among all shapley values in {ψv}v∈V , as defined in Theorem 3. The following lemma provides a condition for robust Shapley value estimation. Lemma 27. At the end of Phase 2 of Algorithm ASV-RR, (cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). holds with probability at least 1 − 1 alization θ of θ satisfies: 2n(cid:96) , provided that the re- θ ≥ n(((cid:96) + 1) ln n + ln 4)(2 + 2 3 ε) ε2ψ(k) . (25) Proof. Let R1, R2, . . . , Rθ be the θ independent and random RR sets generated in Phase 2. Let est θ v be the value of est v at the end of the for-loop in Phase 2, when i=1 XRi (v), ∀v ∈ V . By Lemma 23, θ = θ. Then, est θ E[XRi (v)] = ψv/n. For every v ∈ V with ψv > ψ(k), we apply the Chernoff v =(cid:80)θ bounds (Fact 26) and have: Pr{ ψv − ψv ≥ εψv} = Pr{n · est θ = Pr{est θ ≤ 2 exp v/θ − ψv ≥ εψv} (cid:19) (cid:18) v − θ · ψv/n ≥ ε · (θ · ψv/n)} − ε2 (cid:18) 2 + 2 3 ε − ε2 · ψv · θ · ψv/n (2 + 2 3 ε) · n ≤ 2 exp ≤ 2 exp (−((cid:96) + 1) ln n + ln 4) ≤ 1 2n(cid:96)+1 . (cid:19) · n(((cid:96) + 1) ln n + ln 4)(2 + 2 3 ε) ε2ψ(k) {since ψv > ψ(k)} For every v ∈ V with ψv ≤ ψ(k), we also apply the Chernoff bound and have: Pr{ ψv − ψv ≥ εψ(k)} = Pr{n · est θ = Pr{est θ ≤ 2 exp v/θ − ψv ≥ εψ(k)} v − θ · ψv/n ≥ (εψ(k)/ψv) · (θ · ψv/n)} − (εψ(k)/ψv)2 · θ · ψv/n (cid:19) (cid:18) (cid:18) (cid:18) = 2 exp ≤ 2 exp ≤ 1 2n(cid:96)+1 . (cid:19) · θ 2 + 2 3 (εψ(k)/ψv) ε2(ψ(k))2 − n(2ψv + 2 − ε2ψ(k) n(2 + 2 3 ε) (cid:19) 3 εψ(k)) · θ {since ψv ≤ ψ(k)} {use Eq. (25)} Finally, we take the union bound among all n nodes in V to obtain the result. every i ∈ [t]. Let Y =(cid:80)t and for some µ ∈ [0, 1], E[X i X 1, X 2, . . . , X i−1] = µ for i=1 X i. For any δ > 0, we have: Pr{Y − tµ ≥ δ · tµ} ≤ exp − δ2 2 + 2 3 δ tµ . For any 0 < δ < 1, we have Pr{Y − tµ ≤ −δ · tµ} ≤ exp − δ2 2 tµ . (cid:18) (cid:18) (cid:19) (cid:19) Since there are numerous variants of Chernoff and martin- gale tail bounds in the literature, and the ones we found in [20, 13, 33] are all slightly different from the above, in Appendix E we provide a pair of general martingale tail bounds that cover Facts 26 and 29 we need in this paper, with a complete proof. For each i = 1, 2, . . . ,(cid:98)log2 n(cid:99) − 1, let xi = n/2i, and let i be the value of est (k) set in line 11 in the i-th iteration Lemma 30. For each i = 1, 2, . . . ,(cid:98)log2 n(cid:99) − 1, (1) If xi = n/2i > ψ(k), then with probability at least 1 − est (k) of the for-loop (lines 4–16) of Phase 1. (2) If xi = n/2i ≤ ψ(k), then with probability at least 1 − 1 2n(cid:96) log2 n , n · est (k) 2n(cid:96) log2 n , n · est (k) 1 , R(1) 1 i /θi < (1 + ε(cid:48)) · xi. i /θi < (1 + ε(cid:48)) · ψ(k). 2 , . . . , R(1) θi Proof. Let R(1) be the θi generated RR sets by the end of the i-th iteration of the for-loop (lines 4– 16) of Phase 1. For every v ∈ V , let est v,i be the value of est v at line 11 in the i-th iteration of the same for-loop. Then we have est v,i =(cid:80)θi (v). By Lemma 28, we have for every 1 ≤ j ≤ θi, E[X (v) (v)] = ψv/n. Then we can apply the (v), . . . , X (1) R i j=1 X R (1) j X R (1) 1 R (1) i−1 martingale tail bound of Fact 29 on the sequence. For the Statement (1) of the lemma, we consider xi = n/2i > ψ(k), and for every v ∈ V such that ψv ≤ ψ(k), we obtain (cid:19) (cid:48) · xi/ψv) · θi · ψv/n} · θi · ψv/n ) · xi} ) · θi · xi/n} (cid:48) (cid:48) Pr{n · est v,i/θi ≥ (1 + ε = Pr{est v,i ≥ (1 + ε ≤ Pr{est v,i − θi · ψv/n ≥ (ε ≤ exp (cid:18) − (ε(cid:48) · xi/ψv)2 (cid:18) 3 (ε(cid:48) · xi/ψv) ε(cid:48)2 · x2 − (cid:19) (cid:18) 3 · ε(cid:48) · xi − ε(cid:48)2 · xi 3 ε(cid:48) · θi/n = exp ≤ exp 2ψv + 2 2 + 2 2 + 2 i (cid:19) · θi/n ≤ 1 2n(cid:96)+1 log2 n . For Phase 1, we need to show that with high probability LB ≤ ψ(k), and thus Eq.(25) hold for the random θ set in line 17 of the algorithm. The structure of the Phase 1 of ASV-RR follows the Sampling() algorithm in [33] (Algo- rithm 2, lines 1-13), with the difference that our Phase 1 is to estimate a lower bound for ψ(k), while their purpose is to estimate a lower bound for OP Tk, the maximum influence spread of any k seed nodes. The probabilistic analysis fol- lows the same approach, and for completeness, we provide an independent analysis for our algorithm. Let θ(cid:48) be the number of RR sets generated in Phase 1, and R(1) θ(cid:48) be these RR sets. Note that these random RR sets are not mutually independent, because ear- lier generated RR sets are used to determine if more RR sets need to be generated (condition in line 12). However, once RR sets R(1) 1 , . . . , R(1) i−1 are generated, the generation of RR set R(1) follows the same random behavior for each i, which means we could use martingale approach [20] to analyze these RR sets and Phase 1 of Algorithm ASV-RR. 2 , . . . , R(1) 1 , R(1) i Definition 3 (Martingale). A sequence of random variables Y 1, Y 2, Y 3, . . . is a martingale, if and only if E[Y i] < +∞ and E[Y i+1 Y 1, Y 2, . . . , Y i] = Y i for any i ≥ 1. In our case, let Y i(v) =(cid:80)i j=1(X R (v) − ψv/n), for any (1) j v ∈ V and any i. (v) X Lemma 28. For every v ∈ V and every i ≥ 1, E[X (v)] = ψv/n. As a con- sequence, for every v ∈ V , the sequence of random variables {Y i(v), i ≥ 1} is a martingale. (v), . . . , X (1) R i (1) i−1 (1) 1 R R 1 , . . . , R(1) Proof. Consider a node v and an index i ≥ 1. Note that RR sets R(1) i+1 should be generated, but the actual random generation process of R(1) i+1, i.e. selecting the random root and the random live edge graph, is independent of R(1) . Therefore, by Lemma 23 we have i determines whether R(1) 1 , . . . , R(1) i (cid:21) X R (1) i (v) R (1) 1 (v), . . . , X R (1) i−1 (v) = ψv/n. (26) From the definition of Y i(v), it is straightforward to see that the value range of Y i(v) is [−i, i], and thus E(Y i(v)] < +∞. Second, by definition Y i+1(v) = X (v) − ψv/n + Y i(v). With the similar argument as for Eq. (26), we have (1) i+1 R E[Y i+1(v) Y 1(v), . . . , Y i(v)] = E[X (v) − ψv/n Y 1(v), . . . , Y i(v)] R (1) i+1 + E[Y i(v) Y 1(v), . . . , Y i(v)] (cid:20) E (cid:12)(cid:12)(cid:12)(cid:12)X {use ψv ≤ ψ(k) < xi} (27) = 0 + Y i(v) = Y i(v). Therefore, {Y i(v), i ≥ 1} is a martingale. tingale, but notice that if we define Y i = (cid:80)i Martingales have similar tail bounds as the Chernoff bound given in Fact 26, as we give below. For convenience, we did not explicitly refer to the sequence below as a mar- j=1(X i − µ), then {Y 1, Y 2, . . . , Y t} is indeed a martingale. Fact 29 (Martingale Tail Bounds). Let X 1, X 2, . . . , X t be random variables with range [0, 1], i Note that est (k) is the k-th largest among est v,i's, while there are at most k − 1 nodes v with ψv > ψ(k). This means that there is at least one node v with ψv ≤ ψ(k) and est v,i ≥ est (k) . Thus, by taking union bound on Eq. (27), we have i (cid:48) ) · xi} i /θi ≥ (1 + ε Pr{n · est (k) ≤ Pr{∃v ∈ V, ψv ≤ ψ(k), n · est v,i/θi ≥ (1 + ε ≤ 1 . 2n(cid:96) log2 n (cid:48) ) · xi} For the Statement (2) of the lemma, we consider xi = n/2i ≤ ψ(k), and for every v ∈ V with ψv ≤ ψ(k), we have ) · ψ(k)} ) · θi · ψ(k)/n} (cid:48) (cid:48) 2 + 2 Pr{n · est v,i/θi ≥ (1 + ε = Pr{est v,i ≥ (1 + ε (cid:18) ≤ Pr{est v,i − θi · ψv/n ≥ (ε − (ε(cid:48) · ψ(k)/ψv)2 (cid:32) ≤ exp 3 (ε(cid:48) · ψ(k)/ψv) ε(cid:48)2 · ψ(k)2 − (cid:19) (cid:18) 3 · ε(cid:48) · ψ(k) − ε(cid:48)2 · ψ(k) · θi/n (cid:19) (cid:18) 3 ε(cid:48) 2 + 2 − ε(cid:48)2 · xi 3 ε(cid:48) · θi/n ≤ exp ≤ exp 2ψv + 2 = exp 2 + 2 (cid:19) (cid:48) · ψ(k)/ψv) · θi · ψv/n} · θi · ψv/n (cid:33) · θi/n {use ψv ≤ ψ(k)} {use xi ≤ ψ(k)} ≤ 1 2n(cid:96)+1 log2 n . Similarly, by taking the union bound, we have (cid:48) ) · ψ(k)} i /θi ≥ (1 + ε Pr{n · est (k) ≤ Pr{∃v ∈ V, ψv ≤ ψ(k), n · est v,i/θi ≥ (1 + ε ≤ 1 . 2n(cid:96) log2 n (cid:48) ) · ψ(k)} Thus the lemma holds. 1 Lemma 31. Suppose that ψ(k) ≥ 1. In the end of Phase 1, with probability at least 1 − 1 Proof. Let LB i = n · est (k) 2n(cid:96) , LB ≤ ψ(k). i /(θi · (1 + ε(cid:48))). Suppose first that ψ(k) ≥ x(cid:98)log2 n(cid:99)−1, and let i be the smallest index such that ψ(k) ≥ xi. Thus, for each i(cid:48) ≤ i − 1, ψ(k) < xi(cid:48) . By Lemma 30 (1), for each i(cid:48) ≤ i − 1, with probability at most i(cid:48) /θi(cid:48) ≥ (1 + ε(cid:48)) · xi(cid:48) . Taking union bound, 2n(cid:96) log2 n , n · est (k) we know that with probability at least 1 − i−1 2n(cid:96) log2 n , for all i(cid:48) /θi(cid:48) < (1 + ε(cid:48)) · xi(cid:48) . This means, with i(cid:48) ≤ i − 1, n · est (k) probability at least 1− i−1 2n(cid:96) log2 n , that the for-loop in Phase 1 would not break at the i(cid:48)-th iteration for i(cid:48) ≤ i− 1, and thus LB = LB i(cid:48)(cid:48) for some i(cid:48)(cid:48) ≥ i, or LB = 1. Since for every i(cid:48)(cid:48) ≥ i, we have xi(cid:48)(cid:48) ≤ ψ(k), by Lemma 30 (2), for every such i(cid:48)(cid:48), with probability at most 2n(cid:96) log2 n , LB i(cid:48)(cid:48) > ψ(k). Taking union bound again, we know that with probability at most 2n(cid:96) , LB > ψ(k). Finally, if ψ(k) < x(cid:98)log2 n(cid:99)−1, use the similar argument as the above, we can show that, with probability at least 1− 1 2n(cid:96) , the for-loop would not break at any iteration, which means LB = 1, which still implies that LB ≤ ψ(k) since ψ(k) ≥ 1. 1 1 Lemma 32 (Robust Estimator). Suppose ψ(k) ≥ 1. With probability at least 1 − 1 ASV-RR returns { ψv}v∈V that satisfy: that n(cid:96) , Algorithm (cid:26) ψv − ψv ≤ εψv ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). Proof. By Lemma 31, we know that at the end of Phase 1, with probability at least 1 − 1 2n(cid:96) , LB ≤ ψ(k). Then by Lemma 27, we know that when we fix LB to any fixed value LB with LB ≤ ψ(k), with probability at least 1 − 1 ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). 2n(cid:96) , we have (cid:26) ψv − ψv ≤ εψv (cid:26) ψv − ψv ≤ εψv n(cid:96) , we have Taking the union bound, we know that with probability at least 1 − 1 ∀v ∈ V with ψv > ψ(k), ψv − ψv ≤ εψ(k) ∀v ∈ V with ψv ≤ ψ(k). C.3 Time Complexity of ASV-RR Finally, we argue about the time complexity of the algo- rithm. For this purpose, we need to refer to the martingale stopping theorem, explained below. A random variable τ is a stopping time for martingale {Y i, i ≥ 1} if τ takes positive integer values, and the event τ = i depends only on the values of Y 1, Y 2, . . . , Y i. The following martingale stopping theorem is an important fact for our analysis. Fact 33 (Martingale Stopping Theorem [20]). Suppose that {Y i, i ≥ 1} is a martingale and τ is a stopping time for {Y i, i ≥ 1}. If τ ≤ c for some constant c independent of {Y i, i ≥ 1}, then E[Y τ ] = E[Y 1].2 Given a fixed set R ⊆ V , let the width of R, denoted ω(R), be the total in-degrees of nodes in R. By Assumption 1, the time complexity to generate the random RR set R is Θ(ω(R) + 1). We leave the constant 1 in the above formula because ω(R) could be less than 1 or even o(1) when m < n, while Θ(1) time is needed just to select a random root. The expected time complexity to generate a random RR set is Θ(E[ω(R)] + 1). Let EP T = E[ω(R)] be the expected width of a random RR set. Let θ(cid:48) be the random variable denoting the number of RR sets generated in Phase 1. Lemma 34. Under Assumption 1, the expected running time of ASV-RR is Θ((E[θ(cid:48)] + E[θ]) · (EP T + 1)). j j 1 , R(1) 2 , . . . , R(1) Proof. Let R(1) is Θ(ω(R(1) θ(cid:48) be the RR sets gener- ated in Phase 1. Under Assumption 1, for each RR set R(1) , the time to generate R(1) j ) + 1), where the con- stant Θ(1) term is to accommodate the time just to select a random root node for the RR set, and it is not absorbed by Θ(ω(R(1) j ) because the width of an RR set could be less than 1. After generating R(1) , ASV-RR also needs to go through j all entries u ∈ R(1) to update est u (line 9), which takes Θ(R(1) ) time. Note that for every random RR set R, we have R ≤ ω(R) + 1, because the RR set generation process guarantees that the induced sub-graph of any RR set must be weakly connected. Thus, for each RR set R(1) , ASV-RR takes Θ(ω(R(1) j ) + 1) time, and summing up for all θ(cid:48) RR sets, the total running time of j ) + 1 + R(1) ) = Θ(ω(R(1) j j j j Phase 1 is Θ((cid:80)θ(cid:48) j=1(ω(R(1) j ) + 1)). 2There are two other alternative conditions besides that τ is bounded by a constant, but they are not needed in our analysis and thus are omitted. We define W i =(cid:80)i j=1(ω(R(1) j )− EP T ), for i ≥ 1. By an argument similar to Lemma 28, we know that {W i, i ≥ 1} is a martingale. Moreover, θ(cid:48) is a stopping time of {W i, i ≥ 1} because its value is only determined by the RR sets already generated. The value of θ(cid:48) is upper bounded by θ(cid:98)log2 n(cid:99)−1, which is a constant set in line 6. Therefore, we can apply the martingale stopping theorem (Fact 33) and obtain 0 = E[W 1] = E[W θ(cid:48) ] = E ω(R(1) j ) − θ (cid:48) · EP T   θ(cid:48)(cid:88)  − E[θ j=1 (cid:48)  θ(cid:48)(cid:88) = E ω(R(1) j ) ] · EP T. j=1 This implies that the expected running time of Phase 1 is Θ(E[θ(cid:48)] · (EP T + 1)). For Phase 2, all θ RR sets are independently generated, and thus the expected running time of Phase 2 is Θ(E[θ] · (EP T + 1)). Together, we know that the expected running time of ASV-RR is Θ((E[θ(cid:48)] + E[θ]) · (EP T + 1)). We now connecting EP T with the influence spread of a single node, first established in [34] (Lemma 7). For com- pleteness, we include a proof here. Lemma 35 (Expected Width of Random RR Sets). Let v be a random node drawn from V with probability proportional to the in-degree of v. Let R be a random RR set. Then: EP T = ER[ω(R)] = Ev[σ({v})]. m n Proof. For a fixed set R ⊆ V , let p(R) be the probability that a randomly selected edge (from E) points to a node in R. Since R has ω(R) edges pointing to nodes in R, we have p(R) = ω(R)/m. Let dv denotes the in-degree of node v. Let v be a random node drawn from v with probability proportional to the in- degree of v. We have: (cid:88) (cid:88) (u,v)∈E dv m v∈V p(R) = = · I{v ∈ R} 1 m · I{v ∈ R} = Ev[I{v ∈ R}]. Let R be a random RR set. Then, we have: ER[ω(R)] =m · ER[p(R)] =m · ER[Ev[I{v ∈ R}]] =m · Ev[ER[I{v ∈ R}]] =m · Ev[Pr (v ∈ R)] =m · Ev[σ({v})/n], R where the last equality is by Lemma 22. Next we need to bound E[θ(cid:48)] and E[θ]. Lemma 36. For each i = 1, 2, . . . ,(cid:98)log2 n(cid:99) − 1, if ψ(k) ≥ 2n(cid:96)+1 log2 n , i /θi > (1 + ε(cid:48)) · xi. be the θi generated RR sets by the end of the i-th iteration of the for-loop (lines 4– 16) of Phase 1. For every v ∈ V , let est v,i be the value (1 + ε(cid:48))2 · xi, then with probability at least 1 − n · est (k) i /θi > ψ(k)/(1 + ε(cid:48)), and n · est (k) Proof. Let R(1) 2 , . . . , R(1) θi 1 , R(1) k Then we have est v,i =(cid:80)θi of est v at line 11 in the i-th iteration of the same for-loop. j=1 X R (1) j (v). (cid:48) Suppose that ψ(k) ≥ (1 + ε(cid:48))2 · xi. For every node v ∈ V with ψv ≥ ψ(k), we apply the lower tail of the martingale tail bound (Fact 29) and obtain Pr{n · est v,i/θi ≤ ψv/(1 + ε )} ≤ Pr{est v,i ≤ 1 1 + ε(cid:48) · θi · ψv/n} = Pr{est v,i − θi · ψv/n ≤ − ε(cid:48) (cid:18) (cid:19) 1 + ε(cid:48) · θi · ψv/n} ε(cid:48)2 (cid:18) − 2(1 + ε(cid:48))2 · θi · ψv/n ≤ exp − ε(cid:48)2 · xi {use xi ≤ ψ(k) (1+ε(cid:48))2 ≤ ψv (1+ε(cid:48))2 } ≤ exp (cid:19) · θi 2n ≤ 1 2n(cid:96)+1 log2 n . i is Note that est (k) the k-th largest value among {est v,i}v∈V , or equivalently (n − k + 1)-th smallest among {est v,i}v∈V . But there are at most n − k nodes v with ψv < ψ(k), which means that there is at least one node v with ψv ≥ ψ(k) and est v,i ≤ est (k) . To be precise, such a v has ψv ranked before or the same as ψ(k), and thus there are at most k such nodes. Then we have i (cid:48) )} i /θi ≤ ψ(k)/(1 + ε Pr{n · est (k) ≤ Pr{∃v ∈ V, ψv ≥ ψ(k), n · est v,i/θi ≤ ψv/(1 + ε ≤ k Pr{n · est v,i/θi ≤ ψv/(1 + ε ≤ )} k . (cid:48) 2n(cid:96)+1 log2 n (cid:48) )} Since ψ(k) ≥ (1 + ε(cid:48))2 · xi, we have that n · est (k) i /θi > (1 + ε(cid:48)) · xi. ψ(k)/(1 + ε(cid:48)) implies that n · est (k) i /θi > Lemma 37. For both θ(cid:48) and θ, we have E[θ(cid:48)] = O((cid:96)n log n/(ψ(k)ε2)) and E[θ(cid:48)] = O((cid:96)n log n/(ψ(k)ε2)), when (cid:96) ≥ (log2 k − log2 log2 n)/ log2 n. Proof. If ψ(k) < (1 + ε(cid:48))2 · x(cid:98)log2 n(cid:99)−1, then ψ(k) < 4(1 + θ ≤ = θ(cid:98)log2 n(cid:99)−1 ε(cid:48))2. In this case, in the worst case, (cid:48) (cid:24) n · (((cid:96) + 1) ln n + ln log2 n + ln 2) · (2 + 2 (cid:38) n · (((cid:96) + 1) ln n + ln log2 n + ln 2) · (2 + 2 ε(cid:48)2 3 ε(cid:48)) (cid:25) 3 ε(cid:48)) · 4(1 + ε(cid:48))2 (cid:39) ≤ ε(cid:48)2 · ψ(k) = O((cid:96)n log n/(ψ(k)ε2)), 2 · ε, and where the last equality uses the fact that ε(cid:48) = the big O notation is for sufficiently small ε. Similarly, for θ, since LB ≥ 1, we have √ (cid:24) n(((cid:96) + 1) ln n + ln 4)(2 + 2 (cid:38) n(((cid:96) + 1) ln n + ln 4)(2 + 2 ε2 (cid:25) 3 ε) θ ≤ ≤ 3 ε) · 4(1 + ε(cid:48))2 ε2 · ψ(k) (cid:39) = O((cid:96)n log n/(ψ(k)ε2)). Therefore, the lemma holds when ψ(k) < (1+ε(cid:48))2·x(cid:98)log2 n(cid:99)−1. Now suppose that ψ(k) ≥ (1 + ε(cid:48))2 · x(cid:98)log2 n(cid:99)−1. Let i be the smallest index such that ψ(k) ≥ (1 + ε(cid:48))2 · xi. Thus ψ(k) < (1 + ε(cid:48))2 · xi−1 = 2(1 + ε(cid:48))2xi (denote x0 = n). For E[θ(cid:48)], by Lemma 36, with probability at least 1 − 2n(cid:96)+1 log2 n , n · est (k) i /θi > (1 + ε(cid:48)) · xi, which means that (cid:25) (cid:38) 3 ε(cid:48)) · 2(1 + ε(cid:48))2 (cid:24) n · (((cid:96) + 1) ln n + ln log2 n + ln 2) · (2 + 2 n · (((cid:96) + 1) ln n + ln log2 n + ln 2) · (2 + 2 Phase 1 would stop in the i-th iteration, and thus ε(cid:48)2 · xi 3 ε(cid:48)) = θi = θ k (cid:48) (cid:39) ≤ ε(cid:48)2 · ψ(k) = O((cid:96)n log n/(ψ(k)ε2)), where the last equality uses the fact that ε(cid:48) = LB = n · est (k) Lemma 36. Then, for Phase 2 we have When Phase 1 stops at the i-th iteration, we know that i /(θi · (1 + ε(cid:48))) ≥ ψ(k)/(1 + ε(cid:48))2, again by (28) √ 2 · ε. (cid:38) θ ≤ n(((cid:96) + 1) ln n + ln 4)(2 + 2 3 ε) · (1 + ε(cid:48))2 ε2 · ψ(k) (cid:39) = O((cid:96)n log n/(ψ(k)ε2)). (29) With probability at most 2n(cid:96)+1 log2 n , Phase 1 does not stop at the i-th iteration and continues to iterations i(cid:48) > i. In the worst case, it continues to iteration (cid:98)log2 n(cid:99) − 1, and θ(cid:48) = O((cid:96)n log n/ε2). Combining with Eq. (28), we have k (cid:48) E[θ ] = O((cid:96)n log n/(ψ(k)ε2)) + k 2n(cid:96)+1 log2 n · O((cid:96)n log n/ε2) = O((cid:96)n log n/(ψ(k)ε2)), where the last equality uses the fact that ψ(k) ≤ n, and the condition that (cid:96) ≥ (log2 k − log2 log2 n)/ log2 n. Similarly, for Phase 2, in the worst case LB = 1 and we have θ = O((cid:96)n log n/ε2). Combining with Eq. (29), we have E[θ] = O((cid:96)n log n/(ψ(k)ε2)) + 1 2n(cid:96)+1 log2 n · O((cid:96)n log n/ε2) = O((cid:96)n log n/(ψ(k)ε2)). This concludes the lemma. We remark that, the setting of ε(cid:48) = 2·ε is to balance the complicated terms appearing in the upper bound of E[θ(cid:48)] + E[θ], as suggested in [33]. √ Lemma 38 (Shapley Value Estimators: Scalability). 1 the condition ≥ Under Assumption and (log2 k − log2 log2 n)/ log2 n, the expected running time of ASV-RR is O((cid:96)n log n · (EP T + 1)/(ψ(k)ε2)) = O((cid:96)(m + n) log n · E[σ(v)]/(ψ(k)ε2)), where EP T = E[ω(R)] is the expected width of a random RR set, and v is a random node drawn from V with probability proportional to the in-degree of v. (cid:96) Proof. The result is immediate from Lemmas 34, 35, and 37. Together, Lemmas 24, 25, 32 and 38 establish Theorem 3. C.4 On Theorem 4 The proof of Theorem 4 follows the same proof structure as the proof of Theorem 3. The only difference is to replace the definition XR(v) with the following X(cid:48) R(v): (cid:26) 0 if v (cid:54)∈ R; if v ∈ R. 1 (cid:48) R(v) = X v Then, by Eq. (23) in Lemma 22, we know that the SNI cen- (I) = σI({v}) = n · ER[XR(v)]. trality of node v is: ψSNI This replaces the corresponding Lemma 23 for Shapley cen- trality. In the rest of the proof of Theorem 3, we replace every occurrence of XR(v) with X(cid:48) R(v) and keep in mind that now ψ refers to the SNI centrality, then all the analy- sis went through without change for SNI centrality, and thus Theorem 4 holds (Lemma 25 for absolute normalization does not apply to SNI centrality and no other result depends on this lemma). C.5 Adaptation to Near-Linear-Time Algo- rithm We remark that, if we want a near-linear time algorithm for Shapley or SNI centrality with some relaxation in robust- ness, we can make an easy change to Algorithm ASV-RR or ASNI-RR as follows. In line 9, replace est u = est u + 1/R with est u = est u + 1, and use parameter k = 1. What this does is to estimate a lower bound LB of the largest single 1 = maxv∈V σ({v}). Note that for line 22, it node influence σ∗ is still the case that, for ASV-RR we use est u = est u +1/R while for ASNI-RR we use est u = est u + 1. Then we have this alternative result: Theorem 6. If we use k = 1 and replace est u = est u + 1/R in line 9 of Algorithm 1 with est u = est u + 1, while keeping the rest the same for ASV-RR and ASNI-RR respec- tively, then the revised algorithm guarantees that with prob- ability at least 1 − 1 n(cid:96) ; ψv − ψv ≤ εσ 1 ,∀v ∈ V, ∗ with expected running time O((cid:96)(m + n) log n/ε2). Note that ψv above represents Shapley centrality of v for ASV-RR and SNI centrality of v for ASNI-RR, and ψv is the algorithm output for the corresponding estimated centrality of v. The proof of Theorem 6 would follow exactly the same structure as the proof of Theorem 3. To complete the proof, one only needs to observe that by Lemma 22, σ({u}) = n · E[I{u ∈ R}] with a random RR set R, and thus after changing to est u = est u + 1 in line 9, n · est u/θi provides an estimate of σ({u}) at the end of the i-th iteration of Phase 1. This means that LB obtained in Phase 1 is an estimate of the lower bound of the largest single node influ- ence σ∗ 1 . Thus, essentially we only need to replace ψ(k) with σ∗ 1 everywhere in the proof and the theorem statement. Fi- nally, because E[σ(v)] ≤ σ∗ 1 , the time complexity no longer has the extra ratio term E[σ(v)]/σ∗ 1 . The detailed proof is thus omitted. D. EXTENSION TO WEIGHTED INFLU- ENCE MODELS In this section, we extend our results to models with weighted influence-spread functions. We focus on the ex- tension of Shapley centrality, and results on SNI centrality For any subset S ⊆ V , let w(S) =(cid:80) can be similarly derived. The extended model uses node weights to capture the practical "nodes are not equal when activated" in network influence. Let w : V → R be a non- negative weight function over V , i.e., w(v) ≥ 0, ∀v ∈ V . v∈S w(v). We can ex- tend the cardinality-based influence spread σ(S) to weighted influence spread: σw(S) = E[w(I(S))]. Here, the influence spread is weighted based on the value of activated nodes in I(S). Note that, in the equivalent live-edge graph model for the triggering model, we have: σw(S) = EL[w(Γ(L, S))]. Note also that set function σw(S) is still monotone and sub- modular. The influence instance I is extended to include weight w. D.1 Algorithm ASV-RR-W Our Algorithm ASV-RR can be extended to the trigger- ing model with weighted influence spreads. Algorithm ASV- RR-W follows essentially the same steps of ASV-RR. The only exception is that, when generating a random RR set R rooted at a random node v (either in Phase 1 or Phase 2), we select the root v with probability proportional to the weights of nodes. To differentiate from random v ∼ V , we use vw ∼w V to denote a random node vw selected from V according to node weights. The random RR set generated from root vw is denoted as R(vw). All the other aspects of the algorithm remains exactly the same. In particular, the statement of Theorem 3 remains essentially the same, except that ψ is now the Shapley centrality of the weighted influence instance I = (G, E, PI, w). The proof of Lemma 22 is changed accordingly to: σw(S) =n · EL [Euw [I{uw ∈ Γ(L, S)}]] =n · EL,uw [I{Γ − (L, uw) ∩ S (cid:54)= ∅}], where Γ−(L, u) is the set of nodes in graph L that can reach u, and uw is a random node drawn proportionally accord- ing to weight function w. With random live-edge graph L, Γ−(L, u) is the same as the RR set generated from root u, which is denoted as R(u). Thus, we have: σw(S) =n · ER(),uw [I{R(uw) ∩ S (cid:54)= ∅}] =n · Pr (R(uw) ∩ S (cid:54)= ∅), R(),uw where the notation R() means the randomness is only on the random generation of reversed reachable set, but not on the random choice of the root node. We use R() to distinguish it from R, which include the randomness of selecting the root node. Weighted marginal spread σw(S ∪ {v}) − σw(S) can be similarly argued. The rest of the proof, including the proof on robustness and time complexity, essentially remains the same as given in Appendix C. D.2 Centrality Axioms for Weighted Influ- ence Models In this section, we presented our axiomatic analysis for weighted influence models. Weighted Social-Influence Instances Mathematically, a weighted social-influence instance is a 4- tuple IW = (V, E, PI, W ), where (1) the influence instance I = (V, E, PI) characterizes the probabilistic profile of the influence model. (2) W is a non-negative weight function over V , i.e., W (v) ≥ 0,∀v ∈ V . Although W does not impact the influence process, it defines the value of the acti- vated set, and hence impacts the influence-spread profile of the model: The weighted influence spread σIW is then given by: (cid:88) T⊇S σIW (S) = E[W (II(S))] = PI(S, T )W (T ). Note that here we use the capital letter W as the weight function that is integrated into the weighted influence in- stance IW = (V, E, PI, W ). The capital letter W is used to differentiate from the small letter w used later as the parametrized weight function outside the influence instance. Because I and W address different aspects of the weighted influence model, IW = (V, E, PI, W ), we assume they are independent of each other. We also extend the definition of centrality measure (Definition 1) to weighted centrality mea- sure, which is a mapping from a weighted influence instance IW = (V, E, PI, W ) to a real vector in RV . We use ψW to denote such a mapping. Extension of Axioms 1-5 • Axiom 1 (Anonymity) has a natural extension, if when we permute the influence-distribution-profile I with a π, we also permute weight function W by π. We will come back to this if-condition shortly. • Axiom 2 (Normalization) is slightly changed such that the sum of the centrality measures is the total weights of all nodes: (Weighted Normalization). For ev- ery weighted influence instance IW = (V, E, PI, W ), (cid:80) Axiom 8 v∈V ψv(I) = W (V ). • Axiom 3 (Independence of Sink Nodes) remains the same. • Axiom 4 (Bayesian Influence) remains the same. • Axiom 5 (Bargaining with Critical Sets) is replaced by the following natural weighted version: Axiom 9. (Weighted Bargaining with Criti- cal Sets) For the weighted critical set instance IW R,v = (R ∪{v}, E, PIR,v , W ), the weighted centrality measure of v is RW (v) R+1 , i.e. ψW RW (v) R+1 . v (IW R,v) = The justification of the above axiom follows the same Nash bargaining argument for the non-weighted case. Now the threat point is (W (R), 0) and the slack is W (v). The solution of (x1, x2) ∈ argmax x1≥r,x2≥0,x1+x2=r+1 (x1 − W (R))1/r · x2 gives the fair share of v as RW (v) R+1 . Characterization of Weighted Social Influence Model Let AW denote the set of Axioms 1, 3, 4, , 8 and 9. Let weighted Shapley centrality, denoted as ψW,Shapley, be the Shapley value of the weighted influence spread σIW , i.e., ψW,Shapley(IW ) = φShapley(σIW ). We now prove the follow- ing characterization theorem for weighted influence models: Theorem 7. (Shapley Centrality of Weighted So- cial Influence) Among all weighted centrality measures, the weighted Shapley centrality ψW,Shapley is the unique weighted centrality measure that satisfies axiom set AW (Ax- ioms 1, 3, 4, 8, and 9). The proof of Theorem 7 follows the same proof structure of Theorem 1, and the main extension is on building a new full-rank basis for the space of weighted influence instances {IW}, since this space has higher dimension than the un- weighted influence instances {I}. Lemma 39 Shapley centrality ψW,Shapley satisfies all Axioms in AW . (Weighted Soundness). The weighted Proof Sketch. The proof essentially follows the same proof of Lemma 6, after replacing unweighted influence spread σI with weighted influence spread σIW . Note that the proof of Lemma 6 relies on earlier lemmas on the proper- ties of sink nodes, which would be extended to the weighted version. In particular, the result of Lemma 2 (a) is extended to: σIW (S ∪ {v}) − σIW (S) = Pr(v (cid:54)∈ II(S)) · W (v). Lemma 4 is extended to: σI\{v}W (S) = σIW (S) − Pr(v ∈ II(S)) · W (v). All other results in Lemmas 2–5 are either the same, or ex- tended by replacing σI and σI\{v} to σIW and σI\{v}W , re- spectively. With the above extension, the proof of Lemma 39 follows in the same way as the proof of Lemma 6. To prove the uniqueness, consider the profile of a weighted influence instance IW = (V, E, PI, W ). Comparing to the corresponding unweighted influence instance I = (V, E, PI), IW has n = V additional dimensions for the weights of the nodes, and thus we need n additional parameters to specify node weights. Recall that in the proof of Theorem 1, we overload the notation PI as a vector of M dimensions to represent the influence probability profile of unweighted influence instance I = (V, E, PI). Similarly, we overload W to represent a vector of n dimensions for the weights of n nodes. Together, we use vector (PI, W ) to represent a vector of M(cid:48) = M + n dimensions that fully determines a weighted influence instance I. We now need to construct a set of basis vectors in RM(cid:48) , each of which corresponds to a weighted influence instance. The construction is still based on the critical set instance defined in Definition 2. For every R ⊆ V with R (cid:54)∈ {∅, V } and every U ⊃ R, we consider the critical set instance IR,U with uniform weights (i.e. all nodes have weight 1). We use (cid:126)1 to denote the uniform weight vector. Then vector (PIR,U ,(cid:126)1) ∈ RM(cid:48) is the vector specifying the corresponding weighted critical set influence instance, denoted as I(cid:126)1 R,U . Let (cid:126)ei ∈ Rn be the unit vector with i-th entry being 1 and all other entries being 0, for i ∈ [n]. Then (cid:126)ei corresponds to a weight assignment where the i-th node has weight 1, and all other nodes have weight 0. Consider the null influence instance I N , in which every node is an isolated node, same as defined in Lemma 12. We add weight vector (cid:126)ei to the null instance I N , to construct a unit-weight null instance I N,(cid:126)ei , where every node is an isolated node, the i-th node has weight 1, and the rest have weight 0, for every i ∈ [n]. The vector representation of I N,(cid:126)ei is (PIN , (cid:126)ei). Note that, as already argued in the proof of Lemma 12, vector PIN is the all-0 vector in RM . Given the above preparation, we now define V(cid:48) as the set containing all the above vectors, that is: V(cid:48) ={(PIR,U ,(cid:126)1) R, U ⊆ V, R /∈ {∅, V }, R ⊂ U} ∪ {(PIN , (cid:126)ei) i ∈ [n]}. We prove the following lemma: Lemma 40 (Independence of Weighted Influence). Vectors in V(cid:48) are linearly independent in the space RM(cid:48) . Proof. Our proof extends the proof of Lemma 10. Sup- pose, for a contradiction, that vectors in V(cid:48) are not linearly independent. Then for each R and U with R, U ⊆ V, R /∈ {∅, V }, R ⊂ U , we have a number αR,U ∈ R, and for each i we have a number αi ∈ R, such that: αR,U · (PIR,U ,(cid:126)1) + αi · (PIN , (cid:126)ei) = (cid:126)0, (cid:88) R(cid:54)∈{∅,V },R⊂U (cid:88) i∈[n] (30) and at least some αR,U (cid:54)= 0 or some αi (cid:54)= 0. Suppose first that some αR,U (cid:54)= 0. Let S be the smallest set with αS,U (cid:54)= 0 for some U ⊃ S, and let T be any superset of S with αS,T (cid:54)= 0. By the critical set instance definition, we have PIS,T (S, T ) = 1. Since the vector does not contain any dimension corresponding to PI(S, S), we know that T ⊃ S. Moreover, since PIN is an all-0 vector, we know that PIN (S, T ) = 0. (cid:88) Then by the minimality of S, we have: αR,U · PIR,U (S, T ) 0 = (cid:88) (cid:88) (cid:88) R,U :R(cid:54)∈{∅,V },R⊂U = αS,T · PIS,T (S, T ) + (cid:88) U :U⊃S,U(cid:54)=T αR,U · PIR,U (S, T ) αS,U · PIS,U (S, T )+ R,U :R≥S,R(cid:54)=S,U⊃R = αS,T + U :U⊃S,U(cid:54)=T αS,U · PIS,U (S, T )+ R,U :R≥S,R(cid:54)=S,U⊃R αR,U · PIR,U (S, T ). are 0, what left in Eq. (30) is (cid:80) Following the same argument as in the proof of Lemma 10, we have αS,T = 0, which is a contradiction. Therefore, we know that αR,U = 0 for all R, U pairs, and there must be some i with αi (cid:54)= 0. However, when all αR,U 's i∈[n] αi · (cid:126)ei = (cid:126)0. But since vectors (cid:126)ei's are obviously linearly independent, the above cannot be true unless all αi's are 0, another contradiction. Therefore, vectors in V(cid:48) are linearly independent. Lemma 41 (Centrality Uniqueness of the Basis). Fix a set V . Let ψW be a weighted centrality measure that satisfies axiom set AW . For any instance IW that corresponds to a vector in V(cid:48), the centrality ψ(IW ) is unique. Proof. Suppose first that IW is a weighted critical set instance I(cid:126)1 R,U has the same weight for all nodes, its weighted centrality uniqueness can be argued in the exact R,U . Since I(cid:126)1 same way as in the proof of Lemma 9 (except that the un- weighted Axioms 2 and 5 are replaced by the corresponding weighted Axioms 8 and 9). Now suppose that IW is one of the instances I N,(cid:126)ei , for some i ∈ [n]. Since in instance I N,(cid:126)ei all nodes are isolated nodes, and thus sink nodes, for each node v, we can repeat- edly apply the Sink Node Axiom (Axiom 3) to remove all other nodes until v is the only node in the graph, and this re- peated projection will not change the centrality of v. When v is the only node in the graph, by the Weighted Normaliza- tion Axiom (Axiom 8), we know that v's weighted centrality measure is W (v). Since the weights of all nodes are deter- mined by the vector (cid:126)ei, the weighted centrality of I N,(cid:126)ei is fully determined and is unique. Lemma 42 (Weighted Completeness). The weighted centrality measure satisfying axiom set AW is unique. Proof Sketch. The proof follows the proof structure of Lemma 12. Lemma 40 already show that V(cid:48) is a set of basis vectors in the space RM(cid:48) , and Lemma 41 further shows that instances corresponding to these basis vectors have unique weighted centrality measures. In addition, we define the 0-weight null instance I N,(cid:126)0 to be an instance in which all nodes are isolated nodes, and all nodes have weight 0. Then the vector corresponding to I N,(cid:126)0 is the all-0 vector in RM(cid:48) . Moreover, similar to I N,(cid:126)ei , the weighted centrality of I N,(cid:126)0 satisfying axiom set AW is also uniquely determined. With the above preparation, the rest of the proof fol- lows exactly the same logic as the one in the proof of Lemma 12. Proof of Theorem 7. Theorem 7 follows from Lem- mas 39 and 42. Axiom Set Parametrized by Node Weights The above axiomatic characterization is based on the di- rect axiomatic extension from unweighted influence models to the weighted influence models, where node weight func- tion W is directly added as part of the influence instance. One may further ask the question: "What if we treat node weights as parameters outside the influence instance? Is it possible to have an axiomatic characterization on such parametrized influence models, for every weight function?" The answer to the above question would further highlight the impact of the weight function to the influence model. Since our goal is to achieve axiomatization that works for every weight function, we may need to seek for stronger ax- ioms. To achieve the above goal, for a given set V , we as- sume that the node weight function cannot be permuted. To differentiate parametrized weight function from the inte- grated weight function W discussed before, we use small letter w to represent the parametrized weight function: w : V → R+ ∪ {0}. The weight parameter w appearing on the superscripts of notations such as influence instance I and influence spread σ denotes that these quantities are parametrized by weight function w. The influence spread σwI in influence instance I = (V, E, PI) parametrized by weight w is defined as: (cid:88) T⊇S σwI (S) = E[w(II(S))] = PI(S, T )w(T ). We would like to provide a natural axiom set Aw parametrized by w : V → R+ ∪ {0}, such that the Shap- ley value for the weighted influence spread σw, denoted as ψw,Shapley(I) = φShapley(σwI ), is the unique weighted cen- trality measure satisfying the axiom set Aw, for every such weight function w. Recall that the weight function w satis- fies that w(v) ≥ 0 for all v ∈ V . Let ψw denote a centrality measure satisfying the axiom set Aw. Our Axiom set Aw contains the weighted version of Ax- ioms 2–5, namely Axioms 3, 4, 8, and 9 (of course, notation W (v) is replaced by w(v)). But it also needs an replacement of the Anonymity Axiom (Axiom 1). By making w "independent" of the distribution profile of the influence model I = (V, E, PI), the extension of Axiom Anonymity does not seem to have a direct weighted version. Conceptually, Axiom Anonymity is about node symmetry in the influence model. However, when influence instance is parametrized by node weights, which cannot be permuted and may not be uniform, even if the influence instance I has node symmetry, it does not imply that their weighted centrality is still the same. This is precisely the reason we assume w can not be permuted. Therefore, we are seeking a new property about node sym- metry in the influence model parametrized by node weights to replace Axiom Anonymity. We first define node pair sym- metry as follows. We denote πuv as the permutation in which u and v are mapped to each other while other nodes are mapped to themselves. Definition 4. A node pair u, v ∈ V is symmetric in the influence instance I if for every S, T ⊆ V , PI(S, T ) = PI(πuv(S), πuv(T )), where πuv(S) = {πuv(v(cid:48)) v(cid:48) ∈ S}. We now give the axiom about node symmetry in the weighted case, related to sink nodes and social influence pro- jections. Axiom 10 (Weighted Node Symmetry). In an in- fluence instance I = (V, E, PI), let S be the set of sink nodes. If every pair of none-sink nodes are symmetric, then for any v ∈ S and any u (cid:54)∈ S, ψw u (I \ {v}) + 1V \S (w(v) − ψw u (I) = ψw v (I)). We justify the above axiom as follows. Consider a sink node v ∈ S. ψw v (I) is its fair share to the influence game. Since v cannot influence other nodes but may be influenced by others, its fair share is at most its weight w(v) (can be for- v (I), mally proved). Thus the leftover share of v, w(v) − ψw is divided among the rest nodes. Since sink nodes do not influence others, they should have no contribution for the above leftover share from v. Thus, the leftover share should be divided only among the rest non-sink nodes. By the as- sumption of the axiom, all non-sink nodes are symmetric to one another, therefore they equally divide w(v) − ψw v (I), v (I)) contribution from each non- leading to sink node. Here an important remark is that, the weights of the non-sink nodes do not play a role in dividing the leftover share form v. This is because, the weight of a node is an indication of the node's importance when it is influenced, but not its power in influencing others. In other words, the influence power is determined by the influence instance I, in particular PI, and it is unrelated to node weights. Therefore, the above equal division of the leftover share is reasonable. After this division, we can apply the influence projection to remove sink node v, and the remaining share 1V \S (w(v)− ψw of a non-sink node u is simply the share of u in the projected instance. The parametrized weighted axiom set Aw is formed by Axioms 2, 3, 4, 8, 9, and 10 (after replacing the weight notation W () with w() in the corresponding axioms). We define the weighted Shapley centrality ψw,Shapley(I) as the Shapley value of the weighted influence spread φShapley(σw). Note that this definition coincides with the definition of ψW,Shapley(IW ), that is, whether or not we treat the weight function as an outside parameter or integrated into the influ- ence instance, the weighted version of Shapley centrality is the same. The following theorem summarizes the axiomatic characterization for the case of parametrized weighted influ- ence model. Theorem 8. (Parametrized Weighted Shapley Centrality of Social Influence) Fix a node set V . For any normalized and non-negative node weight func- tion w : V → R+ ∪ {0}, the weighted Shapley centrality ψw,Shapley is the unique weighted centrality measure that satisfies axiom set Aw (Axioms 2, 3, 4, 8, 9, and 10). Lemma 43. If v is a sink node in I, then for any S ⊆ V \ {v}, (a) σwI (S ∪ {v}) − σwI (S) = w(v) Pr(v (cid:54)∈ II(S)); and (b) σwI\{v}(S) = σwI (S) − w(v) Pr(v ∈ II(S)). Proof. The proof follows the proofs of Lemma 2 (a) and Lemma 4, except replacing 1 with weight w(v). Lemma 44. If node pair u, u(cid:48) are symmetric in I, then for any v ∈ V \{u, u(cid:48)}, (a) for any S ⊆ V , Pr(v ∈ II(S)) = Pr(v ∈ II(πuu(cid:48) (S)))' (b) for any random permutation π(cid:48) on V \ {v}, Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ))], and Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u}))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ∪ {u(cid:48)}))]. Proof. For (a), by the definition of symmetric node pair (Definition 4), we have Pr(v ∈ II(S)) = (cid:88) PI(S, T ) T⊇S∪{v} PI(πuu(cid:48) (S), πuu(cid:48) (T )) T⊇S∪{v} (cid:88) (cid:88) (cid:88) = = = −1 uu(cid:48) (T )⊇S∪{v} π T⊇πuu(cid:48) (S)∪{v} PI(πuu(cid:48) (S), T ) PI(πuu(cid:48) (S), T ) = Pr(v ∈ II(πuu(cid:48) (S))). The equality Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u}))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ∪ {u(cid:48)}))] can be argued in the same way. Lemma 45 ley centrality ψw,Shapley(I) satisfies all axioms in Aw. (Weighted Soundness). Weighted Shap- Proof. Satisfaction of Axioms 2, 3, 4, 8 and 9 can be similarly verified as in the proof of Lemma 39. We now verify Axiom 10. Let v be a sink node and u be a non-sink node. Let π(cid:48) be a random permutation on node set V \ {v}. We have u (I) = Eπ[σwI (Sπ,u ∪ {u}) − σwI (Sπ,u)] ψw,Shapley = Pr(u ≺π v)Eπ[σwI (Sπ,u ∪ {u}) − σwI (Sπ,u) u ≺π v]+ Pr(v ≺π u)Eπ[σwI (Sπ,u ∪ {u}) − σwI (Sπ,u) v ≺π u] = Eπ(cid:48) [σwI (Sπ(cid:48),u ∪ {u}) − σwI (Sπ(cid:48),u)]/2+ Eπ[σwI (Sπ,u ∪ {u}) − σwI (Sπ,u) v ≺π u]/2 = Eπ(cid:48) [σwI (Sπ(cid:48),u ∪ {u}) − σwI (Sπ(cid:48),u)]/2+ Eπ[σwI (Sπ,u \ {v} ∪ {u}) − σwI (Sπ,u \ {v}) v ≺π u]/2 + w(v)Eπ[Pr(v (cid:54)∈ II(Sπ,u \ {v} ∪ {u}))− Pr(v (cid:54)∈ II(Sπ,u \ {v})) v ≺π u]/2 = Eπ(cid:48) [σwI (Sπ(cid:48),u ∪ {u}) − σwI (Sπ(cid:48),u)] + w(v)Eπ(cid:48) [Pr(v (cid:54)∈ II(Sπ(cid:48),u ∪ {u}))− Pr(v (cid:54)∈ II(Sπ(cid:48),u))]/2 = Eπ(cid:48) [σwI\{v}(Sπ(cid:48),u ∪ {u}) − σwI\{v}(Sπ(cid:48),u)] + w(v)Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u}))− Pr(v ∈ II(Sπ(cid:48),u))] + w(v)Eπ(cid:48) [Pr(v (cid:54)∈ II(Sπ(cid:48),u ∪ {u}))− Pr(v (cid:54)∈ II(Sπ(cid:48),u))]/2 = ψw,Shapley u (I \ {v}) + w(v)Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u}))− Pr(v ∈ II(Sπ(cid:48),u))]/2. (31) (32) (33) For (b), we use (a) and obtain Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u))] = Eπ(cid:48) [Pr(v ∈ II(πuu(cid:48) (Sπ(cid:48),u)))]. Note that πuu(cid:48) (Sπ(cid:48),u) is a random set obtained by first gen- erating a random permutation π(cid:48), then selecting the prefix node set Sπ(cid:48),u before node u in π(cid:48), and finally replacing the possible occurrence of u(cid:48) in Sπ(cid:48),u with u (u cannot oc- cur in Sπ(cid:48),u so there is no replacement of u with u(cid:48)). This random set can be equivalently obtained by first generating the random permutation π(cid:48), then switching the position of u and u(cid:48) (denote the new random permutation πuu(cid:48) (π(cid:48))), and finally selecting the prefix node set Sπuu(cid:48) (π(cid:48)),u(cid:48) before u(cid:48) in πuu(cid:48) (π(cid:48)). We note that random permutations π(cid:48) and πuu(cid:48) (π(cid:48)) follow the same distribution, and thus we have Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ))]. Eq.(31) is by Lemma 43 (a), and Eq.(32) is by Lemma 43 (b). For v's weighted Shapley centrality, we have ψw,Shapley v (I) = Eπ[σwI (Sπ,v ∪ {v}) − σwI (Sπ,v)] = w(v)Eπ[Pr(v (cid:54)∈ II(Sπ,v))], (34) where the last equality above is also by Lemma 43 (a). Recall that in Axiom 10 S is the set of sink nodes and u ∈ V \ S is a non-sink node. Then we have 1 = Eπ[Pr(v ∈ II(V ))] = Eπ[ (Pr(v ∈ II(Sπ,u(cid:48) ∪ {u (cid:48)})) − Pr(v ∈ II(Sπ,u(cid:48) )))] (cid:88) (cid:88) u(cid:48)∈V (cid:88) (cid:88) = Pr(u (cid:48) ≺π v)Eπ[Pr(v ∈ II(Sπ,u(cid:48) ∪ {u (cid:48)}))− u(cid:48)∈V \{v} Pr(v ∈ II(Sπ,u(cid:48) )) u Eπ[Pr(v ∈ II(Sπ,v ∪ {v})) − Pr(v ∈ II(Sπ,v))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ∪ {u (cid:48) ≺π v]+ (cid:48)}))− u(cid:48)∈V \{v} Pr(v ∈ II(Sπ(cid:48),u(cid:48) ))]/2 + Eπ[Pr(v (cid:54)∈ II(Sπ,v))] = Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u(cid:48) ∪ {u (cid:48)}))− u(cid:48)∈V \S Pr(v ∈ II(Sπ(cid:48),u(cid:48) ))]/2 + Eπ[Pr(v (cid:54)∈ II(Sπ,v))] = V \ S · Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u}))− Pr(v ∈ II(Sπ(cid:48),u))]/2 + Eπ[Pr(v (cid:54)∈ II(Sπ,v))] (35) (36) (37) (38) Eq. (35) is a telescoping series where all middle terms are canceled out. Eq. (36) is because when v ≺π u(cid:48), v ∈ Sπ,u(cid:48) and thus Pr(v ∈ II(Sπ,u(cid:48)∪{u(cid:48)})) = Pr(v ∈ II(Sπ,u(cid:48) ))) = 1. Eq. (37) is by Lemma 2 (b), and Eq. (38) is by Lemma 44 (b). Therefore, from Eq. (38), we have Eπ(cid:48) [Pr(v ∈ II(Sπ(cid:48),u ∪ {u})) − Pr(v ∈ II(Sπ(cid:48),u))]/2 = 1 V \ S (1 − Eπ[Pr(v (cid:54)∈ II(Sπ,v))]). Plugging the above equality into Eq. (33), we obtain ψw,Shapley u (I) = ψw,Shapley u (I \ {v}) + = ψw,Shapley u (I \ {v}) + w(v)(1 − Eπ[Pr(v (cid:54)∈ II(Sπ,v))]) V \ S 1 V \ S (w(v) − ψw,Shapley v (I)), where the last equality above uses Eq. (34). The above equality is exactly the one in Axiom 10. For the uniqueness of the parametrized axiom set Aw, the proof follows the same structure as the proof for A. The only change is in the proof of Lemma 9, which we provide the new version for the weighted case below. Lemma 46 (Weighted Critical Set Instances). Fix a V . For any normalized and non-negative node weight function w : V → R+ ∪ {0}, let ψw be a weighted centrality measure that satisfies axiom set Aw. For any R, U ⊆ V with R (cid:54)= ∅ and R ⊆ U , and the critical set instance IR,U as defined in Definition 2, its weighted centrality ψw(IR,U ) must be unique. Proof. Consider the critical set instance IR,U . We first consider a node v ∈ V \ U . By Lemma 7, every node u ∈ V \ R is a sink node. Then we can apply the Sink Node Axiom (Axiom 3) to iteratively remove all nodes in U \ R without changing v's centrality measure. After removing nodes in U \ R, we know that in the remaining projected instance, every node becomes an isolated node. Then we can further remove all other nodes and only leave v in the graph, still not changing v's centrality measure. When v is the only node left in the graph, we then apply the Weighted Normalization v (IR,U ) = w(v). Thus, Axiom (Axiom 8) and obtain that ψw every node v ∈ V \ U has uniquely determined centrality measure w(v). Next, we consider a node v ∈ U \ R. By Lemma 7, every node v ∈ V \ R is a sink node. Then we can apply the Sink Node Axiom (Axiom 3) to iteratively remove all these sink nodes except v, such that the centrality measure of v does not change after the removal. By Lemma 8, the remaining instance with node set R∪{v} is still a critical set instance with critical set R and target set R∪{v}. Thus we can apply the Weighted Bargaining with Critical Set Axiom (Axiom 9) to this remaining influence instance, and know v (IR,U ) = Rw(v)/(R + 1), for every node v ∈ U \ R. ψw Finally, we consider a node v ∈ R. Again we can remove all sink nodes in V \R iteratively by influence projection until we only have nodes in R left, which is the instance IR,R in the graph with node set R. It is straightforward to verify that every pair of nodes in R are symmetric. Therefore, we can apply the Weighted Node Symmetry Axiom (Axiom 10) to node v ∈ R. In particular, for every isolated node u ∈ V \ U , since we have ψw u (IR,U ) = w(u), there is no leftover share from u that v could claim. For every node u(cid:48) ∈ U \ u(cid:48) (IR,U ) = Rw(u(cid:48))/(R + 1), and thus the R, we have ψw leftover share from u(cid:48) is w(u(cid:48))/(R + 1). By Axiom 10, node v ∈ R would obtain w(u(cid:48))/(R(R + 1)) from u(cid:48). In the final projected instance IR,R with node set R, it is easy to check that every node is an isolated node. Thus by a similar argument of removing all other nodes and applying Weighted Normalization Axiom, we know that in this final projected instance v's weighted centrality is w(v). Summing v (IR,U ) = w(v)+ w(U\R) them up by Axiom 10, we have ψw R(R+1) . Therefore, the weighted centrality measure for instance ψw(IR,U ) is uniquely determined. Once we set up the uniqueness for the critical set instances in the above lemma, the rest proof follows the proof for the unweighted axiom set A. In particular, the linear indepen- dence lemma (Lemma 10) remains the same, since it only concerns about influence instances and is not related to node weights. Lemma 12 also follows, excepted that when arguing the centrality uniqueness for the null influence instance I N , we again use repeated projection and apply the Weighted Normalization Axiom (Axiom 8) to show that each node v in the null instance has the unique centrality measure of w(v). Therefore, Theorem 8 holds. E. MARTINGALE TAIL BOUNDS There are numerous variants of Chernoff bounds and the more general martingale tail bounds in the literature (e.g. [20, 13, 33]). However, they either cover the case of independent variables, or Bernoulli variables, or a bit looser bounds, or some general cases with different conditions. In this section, for completeness, we state the general martin- gale tail bounds that we need for this paper, and provide a complete proof. The proof structure follows that of [20] for Chernoff bounds, but the result is more general. Theorem 9 (Martingale Tail Bounds). Let X 1, X 2, . . . , X t be random variables with range [0, 1]. (1) Suppose that E[X i X 1, X 2, . . . , X i−1] ≤ µi for every i=1 µi. For any i=1 X i, and µ =(cid:80)t i ∈ [t]. Let Y =(cid:80)t Finally, by setting α = ln(δ + 1), we obtain Inequality (39). for any For (2), the analysis follows the same strategy: α > 0, we have −αY ≥ e −α(1−δ)µ} Pr{Y − µ ≤ −δ · µ} = Pr{e ≤ E[e−αY ] e−α(1−δ)µ E[e−αY t−1E[e−αXt X 1, . . . , X t−1]] = e−α(1−δ)µ ≤ e(e−α−1)µtE[eαY t−1 ] ≤ e(e−α−1)µ e−α(1−δ)µ . e−α(1−δ)µ (39) Finally, by setting α = − ln(1 − δ), we obtain Inequal- ity (40). δ > 0, we have: Pr{Y − µ ≥ δ · µ} ≤ exp for every i ∈ [t]. Let Y =(cid:80)t (2) Suppose that E[X i X 1, X 2, . . . , X i−1] ≥ µi, µi ≥ 0, i=1 µi. For any 0 < δ < 1, we have: . µ − δ2 2 + 2 3 δ (cid:18) (cid:19) i=1 X i, and µ =(cid:80)t (cid:19) (cid:18) − δ2 2 µ . Pr{Y − µ ≤ −δ · µ} ≤ exp δ > 0, we have: Lemma 47. Let X 1, X 2, . . . , X t be Bernoulli random variables with range {0, 1}. (1) Suppose that E[X i X 1, X 2, . . . , X i−1] ≤ µi for every i=1 µi. For any i=1 X i, and µ =(cid:80)t i ∈ [t]. Let Y =(cid:80)t (cid:18) i=1 X i, and µ =(cid:80)t for every i ∈ [t]. Let Y =(cid:80)t (cid:19)µ (cid:18) (2) Suppose that E[X i X 1, X 2, . . . , X i−1] ≥ µi, µi ≥ 0, i=1 µi. For any 0 < δ < 1, we have: Pr{Y − µ ≤ −δ · µ} ≤ Pr{Y − µ ≥ δ · µ} ≤ (1 + δ)(1+δ) (cid:19)µ e−δ eδ . . Proof. Let Y i =(cid:80)i (1 − δ)(1−δ) (40) j=1 X j, for i ∈ [t]. For (1), applying Markov's inequality, for any α > 0, we have Pr{Y − µ ≥ δ · µ} = Pr{eαY ≥ eα(1+δ)µ} ≤ E[eαY ] eα(1+δ)µ E[eα(Xt+Y t−1)] = = = eα(1+δ)µ E[E[eα(Xt+Y t−1) X 1, . . . , X t−1]] E[eαY t−1E[eαXt X 1, . . . , X t−1]] eα(1+δ)µ , eα(1+δ)µ (41) (42) where Inequality (41) is by the Markov's inequality. Next, for the term E[eαXt X 1, . . . , X t−1], we utilize the fact that Xt is a Bernoulli random variable and have E[eαXt X 1, . . . , X t−1] = Pr{X t = 0 X 1, . . . , X t−1}E[1 X 1, . . . , X t−1] + Pr{X t = 1 X 1, . . . , X t−1}E[eα X 1, . . . , X t−1] = (eα − 1)E[X t X 1, . . . , X t−1] + 1 ≤ (eα − 1)µt + 1 ≤ e(eα−1)µt , where the last inequality uses the fact that 1 + z ≤ ez for any z. Plugging the above inequality into Eq.(42), we have eα(1+δ)µ Pr{Y − µ ≥ δ · µ} ≤ e(eα−1)µtE[eαY t−1 ] ≤ e(eα−1)(µt−1+µt)E[eαY t−2 ] ≤ ··· ≤ e(eα−1)µ eα(1+δ)µ . eα(1+δ)µ Recall that a function f is convex if for any x1 and x2 and for any 0 ≤ λ ≤ 1, f (λx1 + (1 − λ)x2) ≤ λf (x1) + (1 − λ)f (x2). Lemma 48. Let X be a random variable with range [0, 1], and let p = E[X]. Let Z be the Bernoulli random variable with Pr(Z = 1) = p. For any convex function f , we have E[f (X)] ≤ E[f (Z)]. Proof. Let X be a random variable defined on the prob- ability space (Ω, Σ, P ), then we have (cid:90) p = E[X] = X(ω)P (dω), Ω and (cid:90) Ω E[f (X)] = f (X(ω))P (dω). Applying the convexity of f , together with the assumption that the range of X is [0, 1], we have (cid:90) (cid:90) (cid:18) Ω Ω (cid:90) E[f (X)] = ≤ = f (X(ω))P (dω) ((1 − X(ω))f (0) + X(ω)f (1))P (dω) (cid:19) 1 − (cid:18)(cid:90) + X(ω)P (dω) Ω X(ω)P (dω) (cid:19) f (0) · f (1) Ω = (1 − p)f (0) + pf (1) = E[f (Z)]. Lemma 49. Let X 1, X 2, . . . , X t be random variables with range [0, 1]. Items (1) and (2) in Lemma 47 still hold. Proof. For item (1), following the proof of Lemma 47, we can still obtain Inequality (42). For the term E[eαXt X 1, . . . , X t−1], notice that function f (x) = eαx is con- vex for any α > 0. Therefore, we can apply Lemma 48. In particular, let Z t be a Bernoulli random variable with For Inequality (44), we take logarithm of both sides, and obtain the following equivalent inequality: g(δ) = −δ − (1 − δ) ln(1 − δ) + δ2 2 ≤ 0. Taking the derivatives of g(δ), we have: (cid:48) (δ) = −1 + ln(1 − δ) + g = ln(1 − δ) + δ; 1 − δ 1 − δ + δ (cid:48)(cid:48) g (δ) = − 1 1 − δ + 1. Looking at g(cid:48)(cid:48)(δ), it is clear that g(cid:48)(cid:48)(δ) < 0 for δ ∈ (0, 1), and g(cid:48)(cid:48)(0) = 0. This implies that g(cid:48)(δ) is monotonically decreasing in (0, 1). Since g(cid:48)(0) = 0, we have g(cid:48)(δ) ≤ 0 for δ ∈ (0, 1). This implies that g(δ) is monotonically increasing in (0, 1). Finally, since g(0) = 0, we have g(δ) ≤ 0 for all δ ∈ (0, 1). E[Z t] = E[X t X 1, . . . , X t−1]. By Lemma 48, we have E[eαXt X 1, . . . , X t−1] ≤ E[eαZt ] = Pr{Z t = 0} · 1 + Pr{Z t = 1} · eα = (eα − 1)E[Z t] + 1 ≤ (eα − 1)µt + 1 ≤ e(eα−1)µt . The rest of the proof of item (1) is the same. For item (2), the treatment is the same, as long as we notice that function g(x) = e−αx is also convex for any α > 0. Lemma 50. For δ > 0, we have (cid:18) eδ (1 + δ)(1+δ) ≤ exp − δ2 2 + 2 3 δ For 0 < δ < 1, we have e−δ (1 − δ)(1−δ) ≤ exp (cid:18) − δ2 2 (cid:19) . (43) . (44) (cid:19) Proof. For Inequality (43), we take logarithm of both sides, and obtain the following equivalent inequality: f (δ) = δ − (1 + δ) ln(1 + δ) + δ2 2 + 2 3 δ ≤ 0. Taking the derivatives of f (δ), we have: f (cid:48) (δ) = 1 − 1 + δ 1 + δ − ln(1 + δ) + − 2δ 2 + 2 3 δ 2 3 δ2 (2 + 2 3 δ)2 = − ln(1 + δ) + = − ln(1 + δ) + = − ln(1 + δ) + 18δ + 3δ2 2 · (3 + δ)2 3 · ((3 + δ)2 − 9) 2 · (3 + δ)2 − 27 2 · (3 + δ)2 ; 3 2 27 f (cid:48)(cid:48) (δ) = − 1 1 + δ + (3 + δ)3 ; (cid:48)(cid:48)(cid:48) f (δ) = 1 (1 + δ)2 − 81 (3 + δ)4 . at f(cid:48)(cid:48)(0) = 0 f(cid:48)(cid:48)(δ), we have When δ ≥ 0, f(cid:48)(cid:48)(cid:48)(δ) = 0 has exactly two solutions at δ1 = 0 and δ2 = 3. When δ ∈ (0, 3), f(cid:48)(cid:48)(cid:48)(δ) < 0, and when δ > 3, f(cid:48)(cid:48)(cid:48)(δ) > 0. Looking and limδ→+∞ f(cid:48)(cid:48)(δ) = 0. When δ increases from 0 to 3, since f(cid:48)(cid:48)(cid:48)(δ) < 0, f(cid:48)(cid:48)(δ) decreases, which means f(cid:48)(cid:48)(δ) < 0; when δ increases from 3, since f(cid:48)(cid:48)(cid:48)(δ) > 0, f(cid:48)(cid:48)(δ) keeps increasing, but never increases above 0. Thus, for all δ ≥ 0, f(cid:48)(cid:48)(δ) ≤ 0. Looking at f(cid:48)(δ), we have f(cid:48)(0) = 0. Since f(cid:48)(cid:48)(δ) ≤ 0 for all δ ≥ 0, f(cid:48)(δ) is monotonically non-increasing, and thus f(cid:48)(δ) ≤ 0 for all δ ≥ 0. Finally, looking at f (δ), we have f (0) = 0. Since f(cid:48)(δ) ≤ 0 for all δ ≥ 0, f (δ) is monotonically non-increasing, and thus f (δ) ≤ 0 for all δ ≥ 0.
1511.00906	1	1511	2015-11-03T13:36:52	Reinventing the Triangles: Rule of Thumb for Assessing Detectability	[ "cs.SI", "cs.IR", "physics.soc-ph" ]	Statistical significance of network clustering has been an unresolved problem since it was observed that community detection algorithms produce false positives even in random graphs. After a phase transition between undetectable and detectable cluster structures was discovered, the connection between spectra of adjacency matrices and detectability limits were shown, and both were calculated for a wide range of networks with arbitrary degree distributions and community structure. In practice the full eigenspectrum is not known, and whether a given network has any communities within detectability regime cannot be easily established. Based on the global clustering coefficient we construct a criterion telling whether in an undirected, unweighted network there is some/no detectable community structure, or if the network is in a transient regime. The method is simple and faster than methods involving bootstrapping.	cs.SI	cs	Reinventing the Triangles: Rule of Thumb for Assessing Detectability Jeremi K. Ochab M. Smoluchowski Institute of Physics Jagiellonian University Cracow, Poland Email:[email protected] 5 1 0 2 v o N 3 ] I S . s c [ 1 v 6 0 9 0 0 . 1 1 5 1 : v i X r a Abstract -- Based on recent advances regarding eigenspectra of adjacency matrices, derived from the random matrix theory, the paper reintroduces the global clustering coefficient as a simple means of assessing if nonrandom modules can be detected in undirected, unweighted networks. I. INTRODUCTION Across disciplines one of the most relevant features of com- plex networks is their modular structure. It is crucial in social analysis, it defines or represents the function of biological networks, it affects the processes taking place on networks. The modules, also called communities, bring information: they allow to visualize and understand data, and to predict behavior. Consequently, community detection is at the forefront of data mining techniques. Even though numerous -- and at times surprisingly effective -- algorithms of community detection have been appearing now for a decade, there remains a long- standing question of statistical significance of the modules discovered in networks. Several approaches were already de- scribed in [1] (see Sec. 14, Significance of clustering therein), but new techniques and theoretical advances, notably those stemming from the random matrix theory, have since been brought into play. As described in the following section, the key problem with assessing significance is caused by the detectability limit. Our aim is to propose a fast and efficient rule-of-thumb to discriminate between networks in which there can or cannot hide significant communities. II. RATIONALE -- LIMITS OF DETECTABILITY It has been known that community detection algorithms tend to find some clusters even in random graphs (RG), that they fail to detect clusters in graphs which are not fully random, and that in RG ensembles themselves there are instances of graphs in which communities appear solely by chance. In order to alleviate these problems one may apply bootstrap procedures [2], which significantly enhance the partitioning. Alternatively, the problem has been tackled with the aid of extreme value statistics [3], so that graph partitions can be assigned a statistical significance score. However, Reichardt and Leone [4] have shown that a sharp phase transition exists between undetectable and detectable cluster structures. Such a divide means that below the transi- tion community structure found by an algorithm is unreliable. Further studies by Decelle et al. [5] confirm that the transi- tion exists, in some cases however it is possible -- although exponentially hard -- to find true clusters below this limit. In a series of papers that followed [6] -- [8] Nadakuditi, et al., have demonstrated a connection between spectra of adjacency matrices of graphs and detectability. Namely, the eigenvalues of the adjacency matrix either belong to a bulk (a continuous spectrum) or are isolated; in graphs without a community structure the only isolated eigenvalue is the largest one; any new community introduces a new isolated eigenvalue. Conversely, whenever an eigenvalue associated with a given community hides into the bulk, it becomes undetectable. From this perspective, the knowledge of a graph's eigenspectrum should be sufficient to answer the question of significance of clusters, and it seems that the spectral methods of community detection ought to be favored. That being said, one issue remains: calculating the spectra is computationally expensive. Thus, the aim is to quantify detectability by means of a single parameter which can be computed fast and easily for any given network. We argue that global clustering coefficient can be applied for that purpose. III. REMINDER -- CLUSTERING COEFFICIENTS It has been long years since the clustering coefficient has become widely used to benchmark empirical networks and new -- small-world [9] and scale-free networks -- against random Erdos-R´enyi (ER) graphs. Unsurprisingly, one such number cannot describe a real complex network; later studies focused on degree distributions, path lengths, node correlations or quantities describing percolation, epidemic spread, and many others. Although the clustering coefficient(s) gradually ceased to be interesting to measure, databases such as Stanford Large Network Dataset Collection [10] still provide it, along the number of triangles, as a fundamental graph statistic. Indeed, our paper advocates this measure as a simple and fast (especially that often it is precalculated) benchmark for assessing the significance of community structure. Since there has been some confusion about what precisely is the clustering coefficient, let us define it: by local clustering coefficient Cl of node l we mean 2(cid:19), Cl = Nl,△/(cid:18)kl (1) CER = hki N − 1 , (4) where Nl,△ is the number of triangles formed by the edges (l, m), (l, n) and (m, n), where m and n are nodes adjacent to l; kl is the degree of l (i.e., the number of edges incident to l), and the denominator is the number of all pairs of edges incident to node l. The average or mean local clustering coefficient often found in the literature is simply C = 1 N N Xl=1 Cl, (2) where N is the size of the network (i.e., the number of nodes). As argued by Bollob´as [11], it "is often not very informative", therefore we do not use it further. Rather, we choose to work with what we will refer to as global clustering coefficient (GCC), customarily defined as C = 3 × number of triangles number of pairs of adjacent edges . (3) These definitions have been further extended to the cases of directed [12] and weighted [13], [14] networks. As mentioned above, these coefficients have been used as a way of comparison with Erdos-R´enyi graphs, for which it is known to be where h·i denotes averaging, and hki is simply the average de- gree. For uncorrelated graphs with a given degree distribution the formula [15] -- [18] Cuc ≈ 1 N (hk2i − hki)2 hki3 (5) is widely used. As an example of a scale-free network, Bollob´as [11] derives and proves a corrected (after [19]) estimate on the clustering coefficient for Barab´asi-Albert (BA) networks CBA = m − 1 8 (log N )2 N , (6) where m is the number of edges that a new node appears with in the process of network growth.We use BA graphs in Fig. 3 as an example of unclustered scale-free networks to show behavior of various estimates of C. IV. OBSERVATIONS A. GCC vs Mixing Parameter The critical observation is that for ensembles of benchmark graphs with tunable strength of community structure the theoretically predicted limit of detectability coincides with the region where GCC saturates, as shown in Fig. 1. Each of the points in the plot represents a single graph. There are two benchmark ensembles used: Newman-Girvan (NG) [20], whose graphs consist of 4 equally-sized communi- ties (ER type with the probability of intercommunity linking a pair of nodes pout and intracommunity linking probability pin), Fig. 1. Log plot of GCC versus the mixing parameter (µ for LFR, and kOU T /hki for NG networks; the ensembles are described in text). Dashed horizontal lines indicate mean ± standard deviation of GCC for respective samples of random graphs. The dotted vertical line shows the theoretical detectability limit for NG graphs as given by [6]. The undetectability regime overlaps with the region where GCC curves saturate. mean degree hki = 16, and N = 256 nodes; Lancichinetti- Fortunato-Radicchi (LFR) [21], with hki = 20, maximal de- gree kMAX = 50, power-law degree distribution with exponent γ = 2, power-law community size distribution with exponent γc = 1, allowed range of community sizes 20 − 100 nodes, and the network size N = 1000 (we retain the parameters used originally in [22] for the sake of comparison). In the case of LFR graphs, the parameter quantifying community strength is the mixing parameter µ, which is a fraction of links of a node pointing outside of its community. In the case of NG graphs, this role is taken by kOU T /hki, where kOU T is the number of links pointing outside of a node's community; it is simpler than µ, since the degree distribution is well peaked around its mean value. For each mixing parameter value, 100 networks were sampled (clouds of points visible in the plot). For these ensembles, the fully random connections form for kOU T /hki ≈ 12/16 for NG and µ = 1 for LFR. In both cases, community detection algorithms fail much earlier [22]. The theoretical position of detectability limit as calculated in [6] is in this case kOU T /hki ≈ 9/16; unfortunately, the calculations for LFR benchmark are much harder to obtain. The reason why it was possible to find the detectability limit theoretically, is the fact that we already know the structure of NG networks. In the limit of large network size and, what is more important, large average degree, Nadakuditi et al. have shown how to calculate the eigenspectra of adjacency matrices of stochastic block model [6], configuration model [7], and a generalization of the two [8], which allows to find the detectability limits for networks with arbitrary de- gree distributions and community structure. Nevertheless, in practice we do not know the structure in the first place, hence we do not know the full eigenspectrum, and so we cannot decide whether a given network has any communities within detectability regime or not. Now, let us try to deduce that from the value of GCC. B. GCC vs Eigenspectrum The reason that the global clustering coefficient can be an index of detectability is fairly straightforward. Let us rewrite (3) as follows , (7) = PN C = PN l=1 Nl,△/N (hk2i − hki)/2 l=1 A3 ll/N hk2i − hki where Aij is an element of the adjacency matrix A, λi is its l=1 λ3 l /N hk2i − hki eigenvalue, and where we have used PN 2(cid:1) = N (hk2i − ll/2. The second equality in (7)comes hki)/2 and Nl,△ = A3 from the fact that (An)ij encodes the number of paths of length n between nodes i and j; the third equality comes from the fact that thanks to the cyclic property of the trace of a matrix, it is invariant to orthogonal transformations. = PN l=1(cid:0)kl Now, the simple fact that clustering coefficients are con- nected to the third moment of the eigenspectrum was known [23]. Yet, as explained in Sec. II, in undirected and unweighted networks communities appear as isolated eigenvalues to the right of the bulk of the spectrum. This means that the spectrum becomes more skewed to the right; in other words, the third moment of the spectrum -- and consequently GCC -- increases. Such a behavior is consistent with the observation in Sec. IV-A. C. Estimating Baseline GCC The forthcoming issue is how to calculate the reference value of GCC, i.e., the value for a network without communi- ties. One straightforward approach is to randomize the network under study, while keeping constant other of its characteristics that can affect GCC, mainly the degree distribution (or even degree sequence). Notwithstanding, this method is computa- tionally costly, since a whole ensemble of randomized graphs is needed to make the comparison with the original empirical network meaningful. Another approach is to use a theory to reproduce the eigenspectrum of graphs with a given degree distribution, calculate its third moment, and obtain GCC. Indeed, it can be accomplished, e.g., following [8]. Under assumptions on the community structure, e.g., as obtained from a community detection algorithm, the isolated eigenvalues can also be sepa- rately calculated. For large scale-free graphs, however, solving numerically the resulting equation for a specified degree sequence can become increasingly infeasible, as the number of terms (and solutions) also increases with the number of distinct degrees in the network. More importantly, and to our demise, the bulk of the spectrum thus obtained is strictly symmetric. Consequently, the whole third spectral moment of a random graph will come from the largest eigenvalue. Unfortunately, we observed that for ER, BA networks, and even Watts-Strogatz [9] networks with sufficiently high rewiring probability the third moment of the bulk is negative1, see Fig. 2. Consequently, for a wide range of real-world networks we can expect C < 1 N λ3 1 hk2i − hki 1It should be noted that in graphs which are bipartite, or even partly so -- e.g., WS graphs with small p -- negative isolated eigenvalues can appear and thwart our efforts. Hence, we assume only positive isolated eigenvalues. Fig. 2. Log-log plot of minus third moment of the bulk of eigenspectrum for Barab´asi-Albert, Erdos-R´enyi, and Watts-Strogatz networks with the given parameters. Error bars are standard deviations for samples of 100 graphs each. Fig. 3. Comparison of global clustering coefficients C (average over a sample of 100 BA graphs; standard deviations are of the size of the points), theoretical predictions obtained for uncorrelated networks with a given degree distribution Cuc (5) (shaded orange region), for BA (6) (red dashed line), and the contribution of the largest eigenvalue to GCC (red squares). When the network structure is unknown, the predictions are overestimated by more than the sample error of GCC. to hold. (This inequality can also be seen as a lower bound for λ1 > 3√number of triangles which is usually stronger than the known bound λ1 > 3p2 maxl Nl,△.) It follows that the theoretical prediction on GCC as cal- culated using [8] is largely overestimated (i.e., an order of magnitude larger than the standard deviation of an ensemble with a given degree distribution). An example of such a be- havior is shown for BA networks in Fig.3. The only theoretical prediction that could be useful for our purposes is (6); it is, however, a prediction specifically derived for this particular type of network. When analyzing real-world graphs, we lack such conveniences. V. RESULTS At this point we ought to recapitulate our aims, collect the pieces of information we have, and formulate the assumptions under which we will be able to proceed. The aim is: based on the global clustering coefficient (3) to construct a criterion that decides whether (8) 1) a given network has a significant community structure, 2) a given partition of the network is statistically signifi- cant, 3) a given community of the partition is statistically signif- icant. In this paper we will propose an answer to 1) and validate it, as well as suggest a way to develop a criterion for 2). Let us list the (empirical) facts we have gathered up to this point: 1) GCC is proportional to the third moment of the eigen- spectrum, 2) (empirical) for graphs with a community structure GCC is greater than for random graphs, but it reaches the RG value before its connectivity is fully random, see Fig. 1, 3) (empirical) GCC is strictly lower than what can be predicted by the largest eigenvalue alone, see (8), 4) GCC for the uncorrelated graphs can be approximated by (5). After GCC, the degree distribution and its two moments, hki and hk2i, have been calculated, which is the only compu- tational burden, we may proceed further. One has to keep in mind, nevertheless, following assumptions: 1) the network is not bipartite, see Footnote 1., 2) the network is uncorrelated, so that (5) is legitimate, 3) λ1 ≥ hk2i/hki − 1, 4) N is large, as always. Let us note that for the Poisson degree distribution the Assumption 3) is always true, since λ1 ≥ phk2i = phki2 + hki > hki = hk2i/hki − 1, which comes from Rayleigh's inequalities (see e.g. [24]). For scale-free networks, however, the second inequality is inverted and there may appear graphs breaking Assum. 3). (We did find such instances for BA networks but, surprisingly, not for LFR which are scale-free as well.) A. Procedure for an unpartitioned graph In the case of Aim 1), the way to proceed is embarrassingly simple: if C ≤ Cuc (9) is true, then the network is outside the detectability regime, i.e., there are no significant communities. A remark is necessary: let us notice that the above equation is a conservative criterion. The condition that we should have checked is in fact (8). But it can be easily checked that the approximation Cuc corresponds to λ1 ≈ hk2i/hki − 1. However, since the true largest eigenvalue is unknown, we can utilize the well-known approximation by Chung et al. [25], λ1 = (1 + o(1)) max(cid:26)hk2i hki ,pkMAX(cid:27) , (10) which in most cases we are interested in yields hk2i/hki. Consequently, the criterion (8) should be given by C < 1 N . (11) hk2i3 hki3(hk2i − hki) By now, it should be clear why the Assumption 3) was listed before: as long as it is fulfilled the criterion given in (9) is stronger than (8). Though (10) is proved to be the upper bound of λ1, in Fig. 3 the approximation would be even higher than the one Fig. 4. Log plot of GCC for LFR N = 1000, B benchmark, as in Fig. 1. The red shaded strip shows the contribution of true λ1 to C; the orange shaded strip shows the region between conditions (9) and (11). The red dots represent graphs fulfilling (9); the black dots fulfill (11), but not (9); the green dots fulfill neither. The grey region indicates where all known community detection methods gradually begin to fail. The inset magnifies the region of (un)detectability transition. resulting from true value of λ1 (red squares). This means that in some cases the factor o(1) is of a magnitude which can affect our inference. Finally, the ordering of the theoretical predictions, as depicted in Fig. 3, will be of use to us. Even though, condition (9) does not provide us with any error estimates, we still believe that the approximate position of λ1 is given by (10). As a result we can use (11) as an upper bound for an error estimate. Thus, the rule-of-thumb is as follows: • if condition (9) is true, there is no detectable community structure, • if condition (9) is false, but (11) is true, either there is no detectable community structure or there are some detectable and some undetectable communities, • if both are false, there is some detectable community structure. To validate above criteria we again resort to LFR benchmark graphs, since all of the state of the art community detection methods have been tested on them, including Infomap [26], the Louvain method [27] or Oslom [28]. Tests on these graphs, with exactly the same parameters, have shown [2], [22] that all known methods begin to decrease their accuracy at µ ≈ 0.6, and fail after crossing µ ≈ 0.7 − 0.8. As illustrated in Fig. 4, this is in accord with the criteria given above. To address the Aim 2) above, let us assume that we already have obtained a partition of the graph into q clusters. For instance, it might be a result of any of the algorithms we have mentioned, or it might be the planted partition, e.g., in LFR benchmarks (note that by construction these graphs have a partition planted even for µ = 1). Following [8], it is possible to calculate the q discrete eigenvalues that are theoretically predicted to appear in such a partitioned graph. We are then able to compute their contribution to C in (3), just as we did for λ1, and accordingly revise the criteria (9)-(11). Note, that instead of using approximation (10) for λ1 in this Section, we may use [8] for the largest eigenvalue only. Depending on the computational resources one has, it might be feasible to simply calculate λ1 numerically -- although the theoretical approximations seem sufficient, an improvement on the error estimate of roughly 20% can be expected (at least on graphs akin to LFR), as can be seen in the inset of Fig. 4. VI. CONCLUSION We have proposed a simple method to deduce whether it is possible to detect communities in a given network. Since the only computations needed are: calculating triangles and the degree distribution, we claim this method is fast and computationally inexpensive. The state of the art algorithm for an exact triangle enumeration runs in O(E 1.41), where E is the number of edges in the graph. Faster approximate algorithms do exist [29], whose accuracy of around 95% are sufficient, although introduce an additional error. Although there exist algorithms of community detection linear in time, assessing statistical significance of the partitions obtained usually involves bootstrapping which is not favorable. The error bound in procedure we suggested is rather heuris- tic; however, the approximations of λ1 (10) and GCC (5) can be improved, e.g., if we have ansatz on degree-degree correlations, which in some cases are known [30]. Still, if the assumption 3) holds, the criteria are conservative and on the safe side. The natural extension of the scheme presented here is: directed and weighted networks, for which generalized ver- sions of clustering coefficients have been proposed [12] -- [14]. Initial computations on directed LFR benchmarks show that analogous criteria are plausible. For weighted networks, however, the clustering coefficients behave nonmonotonously as a function of the mixing parameter, which introduces additional complications. Further systematic comparison with other methods is needed and, more importantly, attacking real-world networks. ACKNOWLEDGMENT My thanks go to Santo Fortunato for the discussions at the early stage of the study and a week's stay at BECS, Aaalto University and to Zdzisław Burda for consultations throughout. I also thank an Anonymous Reviewer for useful insights about bounds for the principal eigenvalue. The work has been funded by Grant No. DEC-2013/09/N/ST6/01419 of the National Science Centre of Poland. REFERENCES [1] S. Fortunato, "Community detection in graphs," Physics Reports, [Online]. Available: vol. 486, no. 35, pp. 75 -- 174, 2010. http://www.sciencedirect.com/science/article/pii/S0370157309002841 [2] A. Lancichinetti and S. Fortunato, "Consensus clustering in complex networks," Scientific reports, vol. 2, 2012. [3] A. Lancichinetti, significance F. Radicchi, of tical E, http://link.aps.org/doi/10.1103/PhysRevE.81.046110 communities 046110, Apr in 2010. vol. 81, p. and J. J. Ramasco, "Statis- networks," Phys. Rev. [Online]. Available: [4] J. Reichardt and M. Leone, "(Un)detectable cluster structure in sparse networks," Phys. Rev. Lett., vol. 101, p. 078701, Aug 2008. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevLett.101.078701 [5] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, "Inference and phase transitions in the detection of modules in sparse networks," Phys. Rev. Lett., vol. 107, p. 065701, Aug 2011. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevLett.107.065701 [6] R. R. Nadakuditi and M. E. J. Newman, "Graph spectra and in networks," Phys. [Online]. Available: the Rev. Lett., vol. 108, p. 188701, May 2012. http://link.aps.org/doi/10.1103/PhysRevLett.108.188701 detectability community structure of [7] -- -- , "Spectra of random graphs with arbitrary expected degrees," Phys. Rev. E, vol. 87, p. 012803, Jan 2013. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.87.012803 [8] X. Zhang, R. R. Nadakuditi, and M. E. J. Newman, "Spectra of random graphs with community structure and arbitrary degrees," Phys. Rev. E, vol. 89, p. 042816, Apr 2014. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.89.042816 [9] D. J. Watts and S. H. Strogatz, "Collective dynamics of small- worldnetworks," Nature, vol. 393, no. 6684, pp. 440 -- 442, 1998. [10] J. Leskovec and A. Krevl, "SNAP Datasets: Stanford large network dataset collection," http://snap.stanford.edu/data, Jun. 2014. [11] B. Bollob´as and O. M. Riordan, "Mathematical results on scale-free random graphs," in Handbook of graphs and networks: from the genome to the internet, S. Bornholdt and H. G. Schuster, Eds. Weinheim: Wiley- VCH, 2003, pp. 1 -- 34. [12] G. Fagiolo, "Clustering in complex directed networks," Phys. [Online]. Available: Rev. E, vol. 76, p. 026107, Aug 2007. http://link.aps.org/doi/10.1103/PhysRevE.76.026107 [13] A. Barrat, M. Barthelemy, R. Pastor-Satorras, and A. Vespignani, "The architecture of complex weighted networks," Proceedings of the Na- tional Academy of Sciences of the United States of America, vol. 101, no. 11, pp. 3747 -- 3752, 2004. [14] J. Saramaki, M. Kivela, J.-P. Onnela, K. Kaski, and J. Kert´esz, "Generalizations of the clustering coefficient to weighted complex networks," Phys. Rev. E, vol. 75, p. 027105, Feb 2007. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.75.027105 [15] M. E. J. Newman, "Random graphs as models of networks," in Handbook of graphs and networks: from the genome to the internet, S. Bornholdt and H. G. Schuster, Eds. Weinheim: Wiley-VCH, 2003, pp. 35 -- 68. [16] M. Bogun´a and R. Pastor-Satorras, "Class of correlated random networks with hidden variables," Phys. Rev. E, vol. 68, p. 036112, Sep 2003. [On- line]. Available: http://link.aps.org/doi/10.1103/PhysRevE.68.036112 [17] Z. Burda, J. Jurkiewicz, and A. Krzywicki, "Network transitivity and matrix models," Phys. Rev. E, vol. 69, p. 026106, Feb 2004. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.69.026106 [18] S. N. Dorogovtsev, "Clustering of Rev. E, vol. 69, p. 027104, Feb 2004. http://link.aps.org/doi/10.1103/PhysRevE.69.027104 correlated networks," Phys. [Online]. Available: [19] K. Klemm and V. M. Egu´ıluz, "Growing scale-free networks with small- world behavior," Phys. Rev. E, vol. 65, p. 057102, May 2002. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.65.057102 [20] M. Girvan and M. E. Newman, "Community structure in social and biological networks," Proceedings of the national academy of sciences, vol. 99, no. 12, pp. 7821 -- 7826, 2002. [21] A. Lancichinetti, S. Fortunato, community graphs detection Rev. E, vol. 78, p. 046110, Oct 2008. http://link.aps.org/doi/10.1103/PhysRevE.78.046110 testing for and F. Radicchi, "Benchmark algorithms," Phys. [Online]. Available: [22] A. Lancichinetti and S. Fortunato, "Community detection algorithms: A comparative analysis," Phys. Rev. E, vol. 80, p. 056117, Nov 2009. [On- line]. Available: http://link.aps.org/doi/10.1103/PhysRevE.80.056117 [23] I. J. Farkas, I. Der´enyi, A.-L. Barab´asi, graphs: Beyond the "Spectra Phys. Rev. E, vol. 64, p. 026704, Jul 2001. http://link.aps.org/doi/10.1103/PhysRevE.64.026704 'real-world' of and T. Vicsek, semicircle law," [Online]. Available: [24] P. Van Mieghem, Graph spectra for complex networks. Cambridge: Cambridge University Press, 2010. [25] F. Chung, L. Lu, and V. Vu, "Spectra of random graphs with given expected degrees," Proceedings of the National Academy of Sciences, vol. 100, no. 11, pp. 6313 -- 6318, 2003. [26] M. Rosvall and C. T. Bergstrom, "Maps of random walks on complex the National networks reveal community structure," Proceedings of Academy of Sciences, vol. 105, no. 4, pp. 1118 -- 1123, 2008. [27] V. D. Blondel, vre, "Fast J.-L. Guillaume, R. Lambiotte, and E. Lefeb- networks," communities large of in unfolding of 2008, Journal vol. http://stacks.iop.org/1742-5468/2008/i=10/a=P10008 Statistical Mechanics: no. P10008, Theory 2008. 10, p. and Experiment, [Online]. Available: [28] A. Lancichinetti, F. Radicchi, "Finding statistically significant communities ONE, vol. 6, no. 4, p. e18961, 04 2011. http://dx.doi.org/10.1371%2Fjournal.pone.0018961 J. J. Ramasco, and S. Fortunato, in networks," PLoS [Online]. Available: [29] C. E. Tsourakakis, M. N. Kolountzakis, and G. L. Miller, "Approximate triangle counting," CoRR, vol. abs/0904.3761, 2009. [Online]. Available: http://arxiv.org/abs/0904.3761 [30] P. L. Krapivsky and S. Redner, "Organization of growing random networks," Phys. Rev. E, vol. 63, p. 066123, May 2001. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.63.066123
1912.00463	1	1912	2019-12-01T18:16:18	Generalizable prediction of academic performance from short texts on social media	[ "cs.SI", "cs.CL" ]	It has already been established that digital traces can be used to predict various human attributes. In most cases, however, predictive models rely on features that are specific to a particular source of digital trace data. In contrast, short texts written by users $-$ tweets, posts, or comments $-$ are ubiquitous across multiple platforms. In this paper, we explore the predictive power of short texts with respect to the academic performance of their authors. We use data from a representative panel of Russian students that includes information about their educational outcomes and activity on a popular networking site, VK. We build a model to predict academic performance from users' posts on VK and then apply it to a different context. In particular, we show that the model could reproduce rankings of schools and universities from the posts of their students on social media. We also find that the same model could predict academic performance from tweets as well as from VK posts. The generalizability of a model trained on a relatively small data set could be explained by the use of continuous word representations trained on a much larger corpus of social media posts. This also allows for greater interpretability of model predictions.	cs.SI	cs	GENERALIZABLE PREDICTION OF ACADEMIC PERFORMANCE FROM SHORT TEXTS ON SOCIAL MEDIA A PREPRINT Ivan Smirnov Institute of Education National Research University Higher School of Economics Moscow 101000, Russia [email protected] December 3, 2019 ABSTRACT It has already been established that digital traces can be used to predict various human attributes. In most cases, however, predictive models rely on features that are specific to a particular source of digital trace data. In contrast, short texts written by users -- tweets, posts, or comments -- are ubiquitous across multiple platforms. In this paper, we explore the predictive power of short texts with respect to the academic performance of their authors. We use data from a representative panel of Russian students that includes information about their educational outcomes and activity on a popular networking site, VK. We build a model to predict academic performance from users' posts on VK and then apply it to a different context. In particular, we show that the model could reproduce rankings of schools and universities from the posts of their students on social media. We also find that the same model could predict academic performance from tweets as well as from VK posts. The generalizability of a model trained on a relatively small data set could be explained by the use of continuous word representations trained on a much larger corpus of social media posts. This also allows for greater interpretability of model predictions. Keywords academic performance · prediction · social media · text · transferability Introduction It is already well-known that various human attributes, from socio-economic status to sexual preferences, can be predicted from digital traces on social media platforms. However, in most cases, predictive models rely on specific features of a particular source of digital trace data. For instance, it was shown in Ref. [1] that a wide range of human traits could be effectively predicted from likes on Facebook, while in Ref. [2] the number of Twitter followers was used (along with other features) to predict users' income. The use of features such as likes or number of followers means that the application of the resulting model is limited to a particular platform, e.g. Facebook or Twitter. Some forms of digital traces are, in contrast, ubiquitous across multiple platforms. For instance, short texts written by users can be found in the form of tweets, posts on Facebook, comments on message boards, reviews on e-commerce sites, etc. In this paper, we explore the predictive power of short texts with respect to the academic performance of users. First, we build a model that predicts the educational outcomes of users from their posts on a popular Russian social networking site, VK, and then test its predictive performance in different contexts. In particular, we use it to predict the rankings of schools and universities based on the public posts of their students on VK. We also test the generalizability of the model by applying it to users' tweets rather then VK posts. We use data from a representative Russian panel study titled "Trajectories in Education and Careers" (TrEC) [3] that tracks 4,400 students who participated in the Programme for International Student Assessment (PISA) [4] in 2012. In addition to survey data, this data set contains information about public posts on VK for those participants who agreed to share their VK data (N = 3,483). We use this data to build a model that predicts the PISA scores of users based on the A PREPRINT - DECEMBER 3, 2019 content of their posts. We combine unsupervised learning of word embeddings on a large corpus of VK posts (1.9B tokens) with a simple supervised model trained on individual posts (see Methods for details). In the next step, we collect data about the academic performance of students from the 100 largest Russian universities and from 914 schools from 3 cities (see Methods). The performance is measured as results of the Unified State Examination (USE), a mandatory examination for all high school graduates in Russia. This performance is available in aggregated form, i.e. for each school the average USE score of its graduates is known and for each university the average USE score of its enrollees is known. We then find VK users who indicated in their profiles that they are from these schools and universities (N = 154, 637) and apply our model to their public posts. This allows us to rank schools and universities according to their estimated scores from social media posts and then compare the result with an actual ranking based on USE scores. To further test the generalizability of the model, we select users who indicate a link to their Twitter account in their VK profiles. We download the tweets of these users and check if their academic performance can still be predicted by using the same model applied to tweets instead of VK posts. Finally, we use our model to explore the variation in the language of users with different educational outcomes. To this end, we compute predicted scores for individual words. This approach is similar to the open-vocabulary method [5], however, the use of continuous word representations makes the results less sensitive to the frequency at which a word appears in the training data set. In fact, it even allows computing of reliable scores for words that are absent in the training data set as long as they are frequent enough in the corpus on which word embeddings are trained. Methods TrEC data We used data from the Russian Longitudinal Panel Study of Educational and Occupational Trajectories (TrEC) [3]. The study tracks 4,400 students from 42 Russian regions who took the PISA test in 2012 [4]. We used PISA reading scores as a measure of students' academic performance. PISA defines reading literacy as "understanding, using, reflecting on and engaging with written texts in order to achieve one's goals, to develop one's knowledge and potential, and to participate in society" and considers it a foundation for achievement in other subject areas within the educational system and also a prerequisite for successful participation in most areas of adult life [6]. PISA scores are scaled so that the OECD average is 500 and the standard deviation is 100, while every 40 score points roughly correspond to the equivalent of one year of formal schooling [4]. In 2018, publicly available information from the social networking site VK was collected for 3,483 TrEC participants who provided informed consent for the use of this data for research purposes. Note that while the initial sample was representative of the 9th-grade high school Russian students in 2012, the social network data is not necessarily representative. There were no publicly available posts for 498 users. The median number of public posts for remaining users was 35. We removed posts that contain URLs from our data set to account for potentially automated postings, and we also excluded re-posts and posts with no text. This resulted in the final data set of 130,575 posts from 2,468 users. Model We trained the fastText model [7] on the VK corpus (1.9B tokens, vocabulary size is 2.5M) to obtain vector representa- tions of Russian words (the model is available at [8]). We then represented each post as a 300-dimensional vector by averaging over all its constituent words. We used this representation to train a linear regression model to predict the PISA scores of the posts' authors. The model was trained on individual posts and the predicted value for users was computed as an average predicted value of all their posts. Schools and university data VK provides an application programming interface (API) that enables the downloading of information systematically from the site. In particular, downloading user profiles from particular educational institutions and within selected age ranges is possible. For each user, obtaining a list of their public posts is also possible. According to the VK Terms of Service: "Publishing any content on his/her own personal page, including personal information, the User understands and accepts that this information may be available to other Internet users, taking into account the architecture and functionality of the Site." The VK team confirmed that their public API could be used for research purposes. We created a list of high schools in Saint Petersburg (N = 601), Samara (N = 214), and Tomsk (N = 99) and then accessed the IDs of users who indicated on VK that they graduated from one of these schools. We removed profiles 2 A PREPRINT - DECEMBER 3, 2019 with no friends from the same school, profiles that already belong to the TrEC data set, and users who indicated several schools in their profiles. The public posts of the remaining users were downloaded and our model was applied to them to get a prediction of the users' academic performance. We then estimated the educational outcomes of a school by averaging the predictions of its students' performance. Overall, 1,064,371 posts from 38,833 users were used at this stage of analysis. The same procedure was performed to obtain a prediction of academic performance for students from the 100 largest universities in Russia (Nusers = 115, 804; Nposts = 6, 508, 770). Moscow State University was excluded from the analysis as it is known to be a default choice for bots and fake profiles. Even the application of the aforementioned filtering method does not allow reliable data to be obtained for this university, i.e. there is still an order of magnitude more user profiles than the real number of students for a given cohort. We used data from the web portal "Schools of Saint Petersburg" [9] to obtain the average performance of schools' graduates in the Unified State Examination (USE). This is a mandatory state examination that all school graduates should pass in Russia. The USE scores for Samara were provided by the web portal "Zeus" [10]. The USE scores of the Tomsk schools along with the data on university enrollees [11] were collected by the Higher School of Economics. This information was used to check if the scores predicted from social media data correspond to the ranking of schools and universities based on their USE results. Note that this means that the model was tested not only on a different set of users but that the measure of academic performance was also different from the training settings. Twitter data VK allows users to indicate links to other social media accounts, including Twitter, in their profiles. Only a small proportion of users provide links to their Twitter accounts. In our sample information, about 665 Twitter accounts were available for the Saint-Petersburg data set (less for other cities) and 2,836 Twitter accounts were available for the university data set. This allowed the analysis to be performed only for the university data set. The latest tweets of these 2,836 users were downloaded via Twitter's API. Note that unlike tweets, VK posts are not limited to 140 or 280 characters. However, most VK posts are short texts (85.2% of the posts are less than 140 characters and 92.6% of the posts are less than 280 characters in our sample). Exploring differential language use Thanks to the simplicity of our model, it is possible to obtain meaningful scores for individual words. By construction, the post score is equal to the average scores of its constituent words. This means that it is enough to explore the individual word scores to interpret the model. We ranked all 2.5M words from our dictionary based on the predicted scores and used this ranking to compare the language of users with different educational outcomes. This approach is similar to the open-vocabulary approach [5, 12] because it operates on a level of individual words and allows one to find relations that are not captured with traditional closed dictionaries. The main difference is that our approach uses continuous, rather than discrete, word representations and incorporates rich knowledge about the language structure learned from the training of unsupervised word embeddings. Results Prediction We first explored the predictive power of common text features with respect to academic performance. We found a small negative effect for the use of capitalized words (P = 2 × 10−3), emojis (P = 7 × 10−3), and exclamations (P = 0.05), as seen in Fig. 1. The use of Latin characters (P = 5 × 10−3), average post length (P = 2 × 10−4), word length (P = 4 × 10−4), and vocabulary size (P < 10−10) are positively correlated with academic performance. The strongest correlation was found for the information entropy of users' texts (Pearson's r = 0.20, P < 10−15). We used a TF-IDF model to obtain a base-line prediction of academic performance from the users' posts. We selected the 1000 most common unigrams and bigrams from our corpus, excluding stop words, for the Russian language. We then applied a TF-IDF transformation to represent posts as 1000-dimensional vectors and then trained a linear regression model on individual posts to predict the academic performance of their authors. The correlation between predicted and real scores is r = 0.285. Here, and for the following models, we report results on the user level obtained using leave-one-out cross-validation, i.e. scores for posts of a certain user were obtained from the model trained on posts of all other users. We obtained significantly better results with a model that used word-embeddings (see Methods). We also find that embeddings trained on the VK corpus outperform models trained on the Wikipedia and Common Crawl corpora (Table 1). 3 A PREPRINT - DECEMBER 3, 2019 Figure 1: Pearson correlation between common text features and academic performance. The use of capitalized words, emojis and exclamations (average number per post normalized by the post length in tokens) is negatively correlated with performance. The use of Latin characters, average post and word length, vocabulary size and entropy of users' texts are positively correlated with academic performance. Table 1: Predictive power of the models measured as Pearson correlation between real and predicted outcomes. The results were computed using leave-one-out cross-validation. Correlation coefficient (LOOCV) TF-IDF 0.284 fastText (Wiki) 0.335 fastText (CC) 0.359 fastText (VK) 0.420 The predictive power of a model depends on the number of posts available for each user (see Fig. 2). If only one post is available per user, the predictive power is rather low (r = 0.237). However, it increases with the number of posts available, reaching r = 0.541 for 20 posts per user. Transfer Figure 3 shows the correlation between the predicted performance of schools (a-c) and universities (d) and the USE scores of their graduates or enrollees. In all four cases, we find a relatively strong signal despite the fact that the VK sample might not be representative and that academic performance was measured differently than in training settings, being available only in aggregated form and from secondary sources. Intriguingly, we find that the substitution of VK posts by tweets doesn't substantially alter the resulting performance (see Fig. 4)). For fair comparison, we use VK data only for those users for whom Twitter data was also available. This is why the performance of the model is substantially lower than in Fig. 3d, where all available VK data was used. Differential language use We explored the resulting model by selecting 400 words with the highest and lowest scores that appear at least 5 times in the training corpus. A t-SNE representation of the embeddings produced by our model helped [13] to identify several thematic clusters (Fig. 5). High performing clusters include English words (above, saying, yours, must), words related to literature (Dandelion, Bradbury, Fahrenheit, Orwell, Huxley, Faulkner, Nabokov, Brodsky, Camus, Mann, Shelley, Shakespeare), words related to reading (read, reread, published, book, volume), words related to physics (universe, hole, string, theory, quantum, Einstein, Newton, Hawking), and words related to thinking processes including various synonyms of "thinking" and "remembering." Low performing clusters include common spelling errors and typos, names of popular computer games, words related to military service (army, to serve, military oath), horoscopes (Aries, Sagittarius), and cars and road accidents (traffic collision, General Administration for Traffic Safety, wheels, tuning). 4 entropyvocabulary sizeword lengthpost lengthlatinexclamationsemojiscapitalized0.200.160.090.090.070.040.060.08correlation with performanceA PREPRINT - DECEMBER 3, 2019 Figure 2: The predictive power of the model depending on the number of posts used per user. First, users with at least 20 posts were selected. Then, for each user, N of their posts were selected to predict their academic performance (N = 1, ..., 20). The shaded region corresponds to the bootstrapped 90% confidence interval. Figure 3: Correlation between the predicted and real performance of schools and universities. Pearson's cor- relation coefficients between predicted school scores and the USE scores of their graduates were computed for Saint-Petersburg (a), Samara (b) and Tomsk (c). The correlation between predicted university scores and the USE scores of their enrollees was also computed for the 100 largest Russian universities (d). The use of continuous word representations allows one to compute scores even for words that were not present in the training data set. We computed the scores for all 2.5M words in our vector model and made it available for further exploration [8]. This could be used for exploratory analysis to get insights into differential use of language with respect to academic performance and could be applied to various domains, from literature to politics or food (Fig. 6) 5 15101520number of posts0.00.20.40.6correlation495505515525406080r = 0.54a495505515525507090r = 0.49b495505515507090r = 0.60c505510515520507090r = 0.83dpredicted PISA scoreUSE scoreA PREPRINT - DECEMBER 3, 2019 Figure 4: Comparison of predictions based on VK and Twitter data. While estimates from Twitter and VK vary for individual universities the overall performance of the model is similar for both cases. Note that the performance of the model is rather low due to the limited number of users per university for whom both VK and Twitter data is available. Figure 5: t-SNE representation of the words with the highest and lowest scores from the training data set. High performing clusters (orange) include English words and words related to literature, physics, or thinking processes. Low performing clusters (green) include spelling errors and words related to horoscopes, military service, or cars and road accidents. Discussion In traditional cross-validation settings, one data set is randomly split into two parts: the model is trained on one part, and then its performance is assessed using the other part. In our case, we use two different data sources (VK and Twitter) and different measures of academic performance (PISA scores and USE scores). Some additional limitations were that 6 500510520predicted PISA score507090USE scorevk, r = 0.49twitter, r = 0.51military servicecars and road accidentshoroscopesenglish wordsspelling errorsliteraturephysicscomputer gamesreadingrememberingA PREPRINT - DECEMBER 3, 2019 Figure 6: Ranking of selected words by their predicted score (translated from Russian). The application of the model to words from different domains confirms its face validity. the PISA scores were collected in 2012, while most of the posts were written much later, and that the USE scores were available only at the aggregated level. Despite these limitations, we found a relatively strong signal in the data. For instance, we were able to explain 69% of the variation in universities' scores using information about VK posts of users from these universities. While the result for tweets was significantly lower (R2 = 0.26), this was probably at least partly due to the smaller sample size. Also, note that the prediction of individual scores substantially depends on the number of available posts. If only one post per user is used for prediction, then 6% of the variation in their academic performance could be explained by our model. This number rises to 29% if 20 posts are used. The ability to predict the ranking of educational organizations might seem trivial given that direct ranking information is readily available. However, USE scores are measured only once per cohort, making it extremely hard to estimate any added value provided by an educational organization. A comparison of predicted scores for the same users at different points in time might shed light on factors contributing to the students' progress. We demonstrate how domain-specific unsupervised learning of word embeddings allows predictive models to be trained using relatively small labeled data sets. One reason is that even words that are rare in the training or testing data sets could be valuable for prediction. For instance, even if the word "Newt" never occurs in the training data set, the model could assign a higher score to posts containing it. This would happen if the model learns from the training data set that words from the Harry Potter universe are indicative of high performing students and learns from the unsupervised training that "Newt" belongs to this category, i.e. this word is close to other Harry Potter related words in the vector space. This might make the use of continuous word representation preferable to common approaches relying on counting word frequencies. As our approach does not depend on a particular language, source of texts, or target variable (i.e. academic performance could be substituted by income or depression), it could be applied to a wide variety of settings. Our results also suggest that models trained on text data could be successfully transferred from one data source to another. While this certainly might be useful in some applications, it also means there is a greater risk to users' privacy. If users of platform A do not disclose an attribute X on it, then there is no data to train a model to predict X from digital traces on platform A. However, if X is disclosed on platform B, and both platforms collect short texts from users, then it becomes possible to predict X from digital traces on A given access to data from B. In recent years, face recognition technology has raised particular privacy concerns because of its potential omnipresence and the inability of people to hide from it. In a similar way, digital traces in the form of short texts are ubiquitous, and our results suggest that they allow, if not to identify a person, then at least to predict potentially sensitive private attributes. 7 020406080100percentileBourdieuAriesNabokovhoroscopeLibraHarryNewtDumbledoreMcGonagallvodkabeerwineNavalnyPutinAvailability of data and material A PREPRINT - DECEMBER 3, 2019 and the word rankings are The models http://doi.org/10.17605/OSF.IO/9PBKR TrEC data cannot be publicly shared but is available to interested researchers upon request https://trec.hse.ru/en/data in the Open Science Framework repository available Competing interests The authors declare that they have no competing interests. Funding This work was supported by a grant from the Russian Science Foundation (project №19-18-00271). Acknowledgements Author thanks the Open Data University Research Consortium [14] for providing data on public posts of VK users. Author thanks Ilyuhin B.V., vice rector for informatization and education quality assessment, TOIPKRO, and the Centre of General and Extracurricular Education, HSE University, for providing data on USE scores of Tomsk schools. Author thanks Yulia Torgasheva, head of Zeus web-portal for providing data on USE scores of Samara schools. The TrEC project is supported by the Basic Research Programme of the National Research University Higher School of Economics. References [1] Michal Kosinski, David Stillwell, and Thore Graepel. Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 110(15):5802 -- 5805, 2013. [2] Daniel Preo¸tiuc-Pietro, Svitlana Volkova, Vasileios Lampos, Yoram Bachrach, and Nikolaos Aletras. Studying user income through language, behaviour and affect in social media. PloS one, 10(9):e0138717, 2015. [3] Valeriya Malik. The russian panel study'trajectories in education and careers'. Longitudinal and Life Course Studies, 10(1):125 -- 144, 2019. [4] OECD. PISA 2012 Results: What Students Know and Can Do -- Student Performance in Mathematics, Reading and Science. OECD Publishing, 2014. [5] H Andrew Schwartz, Johannes C Eichstaedt, Margaret L Kern, Lukasz Dziurzynski, Stephanie M Ramones, Megha Agrawal, Achal Shah, Michal Kosinski, David Stillwell, Martin EP Seligman, et al. Personality, gender, and age in the language of social media: The open-vocabulary approach. PloS one, 8(9):e73791, 2013. [6] Andreas Schleicher, Karin Zimmer, Juliet Evans, and Niccolina Clements. PISA 2009 assessment framework: Key competencies in reading, mathematics and science. OECD Publishing (NJ1), 2009. [7] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135 -- 146, 2017. [8] Predicting academic performance from short texts on social media. https://doi.org/10.17605/OSF.IO/ 9PBKR. [9] Schools of Saint Petersburg. https://shkola-spb.ru/. [10] Zeus. http://zeus.volgamonitor.com/. [11] Quality of university admission. https://ege.hse.ru/. [12] Vivek Kulkarni, Margaret L Kern, David Stillwell, Michal Kosinski, Sandra Matz, Lyle Ungar, Steven Skiena, and H Andrew Schwartz. Latent human traits in the language of social media: An open-vocabulary approach. PloS one, 13(11):e0201703, 2018. [13] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9:2579 -- 2605, 2008. [14] Open data university research consortium. https://opendata.university/en/. 8
1712.08712	2	1712	2018-10-21T15:13:17	Persistence of the Jordan center in Random Growing Trees	[ "cs.SI", "cs.DS" ]	The Jordan center of a graph is defined as a vertex whose maximum distance to other nodes in the graph is minimal, and it finds applications in facility location and source detection problems. We study properties of the Jordan Center in the case of random growing trees. In particular, we consider a regular tree graph on which an infection starts from a root node and then spreads along the edges of the graph according to various random spread models. For the Independent Cascade (IC) model and the discrete Susceptible Infected (SI) model, both of which are discrete time models, we show that as the infected subgraph grows with time, the Jordan center persists on a single vertex after a finite number of timesteps. Finally, we also study the continuous time version of the SI model and bound the maximum distance between the Jordan center and the root node at any time.	cs.SI	cs	Persistence of the Jordan center in Random Growing Trees Sarath Pattathil, Nikhil Karamchandani and Dhruti Shah ∗ Abstract The Jordan center of a graph is defined as a vertex whose max- imum distance to other nodes in the graph is minimal, and it finds applications in facility location and source detection problems. We study properties of the Jordan center in the case of random growing trees. In particular, we consider a regular tree graph on which an in- fection starts from a root node and then spreads along the edges of the graph according to various random spread models. For the In- dependent Cascade (IC) model and the discrete Susceptible Infected (SI) model, both of which are discrete time models, we show that as the infected subgraph grows with time, the Jordan center persists on a single vertex after a finite number of timesteps. Finally, we also study the continuous time version of the SI model and bound the maximum distance between the Jordan center and the root node at any time. 1 Introduction There are several notions of node centrality in graphs that have been pro- posed in the literature, such as distance centrality, betweenness centrality, degree centrality, and eigenvalue centrality (see for example [1, 2]). These centrality measures find application in a wide variety of contexts, such as identifying influential/critical entities in social/communication networks [15, 16], source/root detection in diffusion/growing networks [4, 10, 12, 13], and facility location problems [14, 22, 23]. ∗S. Pattathil is with the department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology. E-mail: [email protected] N. Karamchandani and D. Shah are with the department of Electrical Engineer- ing, [email protected], [email protected] Indian Institute of Technology, Bombay. E-mail: 1 Recently, there has been some work on analyzing the movement of graph centers in randomly growing graphs. In particular, the question of interest is whether a given notion of a graph center persists in a random growth model, i.e., does a particular vertex emerge and remain as the center after a finite number of steps, even as the graph continues to grow or does the center move around indefinitely. [17] considered the preferential attachment model of growth [11], where at each step a new node connects to existing nodes with probability proportional to the degrees of the existing nodes in the graph, and showed that with probability 1 the degree center (node with the maximum degree) of the graph persists in this random growth model. In other words, after a finite number of timesteps, there is a single node which remains as the most-connected node in the network. In more recent work, [7] considered the centroid or 'balancedness center' [18] of a tree graph where the score of a given vertex is the maximum size of the subtrees rooted at its neighbors, and the center corresponds to the vertex with the minimum score. [7] proved the persistence of the balancedness center in both the uniform as well as preferential attachment tree growth models, as well as the random growing tree arising from an infection spread according to the popular Susceptible-Infected (SI) model on an underlying regular tree. Some follow-up work from the authors [19] studied the same question in sublinear preferential attachment trees. In this paper, we focus on the Jordan center of random growing trees. The Jordan center of a graph is defined as a vertex whose maximum distance to other nodes in the graph is minimal, and it finds applications in facility location [14, 22, 23] and source detection problems [10, 21]. We consider an underlying regular tree on which an infection starts from a root node and spreads along the edges of the graph according to various random spreading models. For the Independent Cascade (IC) model and the discrete Suscep- tible Infected (SI) model, both of which are discrete time models, we show that as the infected subgraph grows with time, the Jordan center persists on a single vertex after a finite number of timesteps. Finally, we also study the continuous time version of the SI model and while we are unable to prove persistence in this case, we bound the maximum distance between the Jordan center and the root node at any time. To the best of our knowledge, there has been no prior study on the persistence properties of the Jordan center. In terms of previous results which are relevant to the contents of this paper, [10] showed that when the infection spreads according to the discrete SI model on an underlying regular tree, the distance between the Jordan center of the infected subtree and the root node is bounded by a fi- nite constant. We extend this result to demonstrate that under the discrete 2 SI model, the Jordan center in fact reaches persistence in a finite number of timesteps, and thereafter does not move. Apart from being a very natural property to examine, persistence of cen- trality measures also has important implications for root-finding algorithms in growing graphs, which have applications such as identifying the center of an epidemic or the source of a rumor. It has been shown [13, 20] that selecting the top nodes according to appropriate centrality measures can yield a confidence set for the root node such that it belongs to this set with high probability. Persistence of centrality measures in this context has been recently used [7] to show that the confidence set thus generated stabilizes af- ter some finite time, implying that the construction is in some sense robust, which is a desirable property. This also suggests the possibility of savings in terms of computational costs, since while the size of the underlying graph and the corresponding complexity of running the root-finding algorithm in- crease with time, the output of the algorithm does not change beyond a certain threshold. We realize that while our results hold for trees, real- world networks are not tree-like and thus any application of these results to such settings would need an extension to more general network topologies. However, this is technically a really challenging problem and we consider the study of regular trees to be an important first step in this direction. Finally, while this paper mainly focuses on regular trees, some of our results do generalize to irregular trees and we discuss these briefly in Section 8. The rest of the paper is organized as follows. We describe the graph and infection models we work with in Section 2. Section 3 describes some preliminary results for the Jordan center under the infection models we study. In Sections 4 and 5, we show that the Jordan center is persistent in the IC model and the discrete SI models respectively. Section 6 bounds the maximum distance between the Jordan center and the root node in the continuous time SI model. Section 7 provides some simulation results and Section 8 discusses generalizations of the results that we have proved and also some open problems. 2 Problem Setup Let G = (V (G), E(G)) denote a tree graph. Define the function ψG : V (G) → N as ψG(u) = depth(G, u) (1) 3 where depth(·,·) is a function which takes as input a tree graph G and a vertex u, and returns the depth of G when rooted at u i.e. the maximum distance of any vertex from the root u. Definition 2.1. A node v∗ (2) G is the Jordan center of a tree G if: G ∈ argmin v∗ v∈V (G) ψG(v) and let ψG = ψG(v∗ G) (3) ψG is called the centrality of the Jordan center of tree G. For any two nodes u and v, if ψG(u) ≤ ψG(v), we say that u is at least as central as v. It can be easily verified that in a tree, there can be at most two Jordan centers and they have to be neighbors. Figure 1: 3 regular tree. v∗ G3. 3 denotes the Jordan center of the infected graph The basic underlying tree model that we consider for our results is a d + 1-regular tree for some d ≥ 2, where the root u(cid:63) has d + 1 children and every other vertex has exactly d children. See Figure 1 for an illustration. Starting from the root node u(cid:63), an infection spreads according to a random growth model (these will be discussed later in the section). At any time t, let the tree subgraph formed by infected nodes be denoted by Gt = (Vt, Et). We study two properties of the Jordan center in random growing trees: 1. Distance from the root: We study how the distance of the root u(cid:63) to of the tree Gt changes with time t. In particular, the Jordan center v∗ Gt can the distance become very large or does the Jordan center remain close to the root at all times? 4 2. Persistence of the Jordan center: Here we study if the Jordan center is persistent. A Jordan center is said to be persistent if ∃ t0 < ∞ such that ∀ t ≥ t0, v∗ Gt i.e. the Jordan center does not change after time t0. = v∗ Gt0 We study three different random growth models in this paper: Definition 2.2. (Independent Cascade (IC) Model): This is a discrete time model. The root node u(cid:63) is infected at time 0. At each time step, an infected node infects each of its as yet uninfected neighbors independently with probability p. Each node can infect its neighbors for exactly one time step, after which it cannot infect and becomes sterile thereafter. We assume1 pd > 1. Definition 2.3. (Discrete Susceptible Infected (SI) Model): This is a discrete time model. The root node u(cid:63) is infected at time 0. At each time step, an infected node infects each of its as yet uninfected neighbors independently with probability p. An infected node can infect its neighbors at every following time step, unlike the case of the IC model. We assume pd > 1. Definition 2.4. (Susceptible Infected (SI) Model): This is a continu- ous time model. The root node u(cid:63) is infected at time 0. Each infected node can infect its as yet uninfected neighbors independently, and the time taken for the infection to spread along each edge is an independent and identically distributed random variable which is exponentially distributed with parameter λ (which we take as 1, without loss of generality). Main results: In a nutshell, the main results of this paper demonstrate the persistence of the Jordan center of an infected subtree growing according to the IC and the discrete SI models on an underlying d + 1-regular tree. As corollaries, we show that for the IC and the discrete SI models, the distance between the root of the underlying tree and the Jordan center of the infected subtree is finite. Finally, we also consider the the continuous time SI model. While we are unable to prove persistence of the Jordan center for this case, we show that if the depth of the infected subtree is n, then the distance between the root of the underlying tree and the Jordan center of the infected subtree is at most O(log n). While this paper mainly focuses on regular trees, some of our results do generalize to irregular trees and we discuss these briefly in Section 7. 1If pd ≤ 1, the infected subtree will be finite under the IC model with probability 1 and the problem is uninteresting. 5 We define some more notation which will be used throughout the paper. A rooted tree is denoted by (G, u), where u ∈ V (G) is the root node. For a rooted tree (G, u), let (G, u)v denote the subtree rooted at v which consists of all vertices w ∈ V (G) such that the path from u to w passes through v. Alternatively, if we interpret the rooted tree (G, u) as a tree of descendants of the ancestor u, then (G, u)v denotes the subtree rooted at v and consisting of v and the descendants of v. For any positive integer k, let [1 : k] denote the set {1, 2, . . . , k}. 3 Preliminaries v v v is the neighbor which has the deepest subtree, (G, v)n1 In this section, we will present a few general properties regarding the move- ment of the Jordan center in growing trees, which hold true for all the spread models we consider in this paper. For a node v , define ni v to be the ith neigh- bor of v in the rooted subtree (G, v), for some ordering of the neighboring nodes. Without loss of generality, assume that the ordering of the neighbors is such that n1 and v is the neighbor which has the second deepest subtree, (G, v)n2 n2 . In case both the depths are the same, the numbering can be arbitrary. See Figure 2 for an illustration. Lemma 3.1. Let the Jordan center of the tree G be the node v∗ depth of the deepest subtree (G, v∗ deepest subtree (G, v∗ is ψG − 1 follows from the Proof. The fact that the depth of (G, v∗ definition of the Jordan center in Definition 2.1. The second part can be proved by contradiction. Suppose the second deepest subtree (G, v∗ has depth ψG − k for some k > 2. Then, on moving the Jordan center by (cid:98)k/2(cid:99) in the direction of the deepest subtree, we reach a node whose centrality is ψG − (cid:98)k/2(cid:99) < ψG, which contradicts the fact that v∗ G. Then the is ψG − 1 and the depth of the second is either ψG − 1 or ψG − 2. G is the Jordan center. G)n2 v∗ G G)n1 v∗ G G)n1 v∗ G G)n2 v∗ G Consider a sequence of growing random trees {Gt, t ≥ 0}. For a discrete time model, we say that the Jordan center changes from time t0 to time t0+1, if ∃ v∗ ) i.e. we assume that the center changes only when the centrality of the new center is strictly lesser than the centrality of the current center. For ease of notation, define t as the Jordan center of the infection graph at time t, Gt. v∗ ∈ V (Gt0+1) s.t. ψGt0+1(v∗ Gt0 ) > ψGt0+1(v∗ Gt0+1 Gt0+1 6 Figure 2: Subtree rooted at node v. The neighbors and the subtrees corre- sponding to this root are shown in this figure. Lemma 3.2. The Jordan center moves from time t to time t + 1 if and only if the second deepest subtree has a depth ψGt − 2 at time t, and at time step t + 1, the depth of the second deepest subtree does not grow, while the deepest subtree grows by 1. In other words, the Jordan center changes from time t to t + 1 if and only if the following equations are satisfied: depth((Gt, v∗ t )n2 v∗ t ) = depth((Gt+1, v∗ ) = ψGt − 2, t t )n2 v∗ ) = ψGt − 1, ) = ψGt. depth((Gt, v∗ depth((Gt+1, v∗ t t )n1 v∗ t )n1 v∗ t (4) (5) (6) Furthermore, in case the center moves, the new Jordan center at time t + 1 will be the neighbor n1 v∗ . t Proof. For sake of simplicity, we assume here that the deepest and second deepest subtrees at time t are unique. By this we mean that ∀ i ≥ 3 depth((Gt, v∗ t )ni v∗ t ) < depth((Gt, v∗ t )n2 v∗ t ). The proof can be easily adapted to the more general scenario. From Lemma 3.1, we know that the second deepest subtree (Gt, v∗ can only have two possible depths ψGt − 1 or ψGt − 2. Now, consider all scenarios other than the one mentioned in the lemma: t )n2 v∗ t 7 (cid:19) (cid:18) (cid:19) (cid:19) (Gt, v∗ t )n1 v∗ t 1. depth 2. depth t (Gt, v∗ (cid:19) t )n2 v∗ (cid:19) ψGt − 1 at time t, (cid:19) (cid:18) (cid:18) (cid:18) (Gt, v∗ t )n2 v∗ t = ψGt − 1: the second deepest subtree has depth t t )n2 v∗ (Gt+1, v∗ = ψGt − 2 and depth = ψGt − 1: the second deepest subtree has depth ψGt − 2 at time t, but the depth (cid:18) (cid:18) grows at time t + 1, and : the second deepest subtree has depth ψGt − 2 at time t, but neither the deepest nor the second deepest trees grow in depth at time t + 1. = ψGt−2, and depth t )n2 v∗ (Gt+1, v∗ (Gt+1, v∗ (Gt, v∗ = depth = depth t )n2 v∗ t )n1 v∗ (cid:18) (cid:19) t t t 3. depth It can be verified that in all the above cases, the Jordan center will not move from time t to t + 1. For example, in case (1) the deepest and second deepest subtrees have the same depth at time t. Note that under all the growth models we consider in this paper, namely the IC, and the discrete SI models, the depth of either subtree can increase by at most 1 at time t + 1. Thus, no node can have a strictly lower centrality than the Jordan center at time t and therefore, the center will not move at time t + 1. Next, we will argue that if the center moves at time t+1, it can only move in the direction of the deepest subtree to the neighbor n1 . See Figure 2. v∗ Note that from time t to t + 1, the centrality of any node increases by at most 1. Thus, it is evident that the Jordan center, if it shifts at time t + 1, will move to one of the neighboring vertices (cid:110) (cid:111) . t ni v∗ t i∈[1:d+1] Now say the Jordan center moves at time t + 1, see Figure 2. From the above argument, this implies that the second deepest subtree has a depth ψGt − 2 at time t, and at time step t + 1, the depth of the second deepest subtree does not grow, while the deepest subtree grows by 1. For any i (cid:54)= 1, the centrality of node ni at time t+1 will be more than the centrality of the v∗ original Jordan center v∗ t and therefore it cannot be the new Jordan center. It can be seen that if the Jordan center moves to node n1 , the centrality v∗ will reduce by 1 as compared to the centrality of the original Jordan center v∗ t . This shows that the Jordan center will move here at time t + 1. t t The above lemma gives rise to the following corollaries. 8 Corollary 3.1. If at least two of the deepest subtrees rooted at the neighbors of the Jordan center grow in depth from time t to t + 1, the center will not change in this time step. Corollary 3.2. If the Jordan center shifts from time t to t + 1, we have ψGt = ψGt+1. 4 Jordan center in the IC model Recall the IC model from Definition 2.2. In this section, we will consider the scenario where the infection starts from the root node u(cid:63) of a d + 1-regular tree G and then spreads along the edges according to the IC model. For the sequence of infected trees {Gt, t ≥ 0} growing according to the IC model, it is easy to see that all new vertices added at timestep t will be at distance t from the root node u(cid:63). This implies that as long as the infected tree is growing, the centrality of u(cid:63) at any time t is ψGt(u(cid:63)) = t. Also, we will say that a tree (or subtree) T is 'dead' at time t0 if no new vertices are added to it at t0, which implies the same for all t > t0 as well because of the nature of the IC model. We begin with a few preliminary lemmas for the IC model, before presenting the main theorem of this section. Lemma 4.1. The Jordan center in the IC model changes from time t to time t + 1 if all except one of the subtrees rooted at the neighbors of the Jordan center v∗ t at time t die, i.e., (Gt, v∗ dies for all i ≥ 2. t )ni v∗ t Proof. Consider the infected tree at time 0, when only the root node u(cid:63) is infected. Trivially, u(cid:63) is the Jordan center at time 0 denoted by v∗ 0. Say the Jordan center shifts from the root at time t0 + 1, for some t0 ≥ 0. From − 2) = t0 − 2, Lemma 3.2, the subtrees (Gt0, v∗ 0)ni v∗ ∀i ≥ 2 . From the definition of the IC model in Definition 2.2, this means that all the subtrees (Gt0, v∗ for i ≥ 2 are dead by time t0 − 1. This proves the lemma for the first Jordan center i.e. the root node u(cid:63). must have depth ≤ (ψGt0 0)ni v∗ 0 0 Assume the statement of the lemma is true for the first n Jordan centers. Consider the (n + 1)th Jordan center and say this center moves from time tn to tn + 1. At time tn, the subtrees for this Jordan center v∗ tn are given by (Gtn, v∗ . Again, using Lemma 3.2, we have that the center will change at time tn + 1 if the deepest subtree grows by 1, and the second deepest tree(s) does not grow, which for the IC model means that it is dead. tn)ni v∗ tn 9 2) to v∗ 1 and v∗ Figure 3: The Jordan center changes from u∗ (which is also v∗ 3 3) to v∗ from time 2 to 3 and from v∗ 5 from timestep 4 to 5. Note that the center changes from v∗ 5 since all except one 4 got infected at timestep 1 and v∗ subtree of v∗ 5 got infected at timestep 2. 4 (which is the same as v∗ 4 to v∗ 4 are dead. Also note that v∗ tn)nc v∗ tn be the parent of v∗ For some c ∈ [1 : d + 1], let nc tn in the original tree v∗ (G, u(cid:63)) and thus, also the previous Jordan center since the center moves at tn most one hop at a time. Then from the induction hypothesis, we have that (Gtn, v∗ is dead, see Figure 3 for an illustration. Since the deepest subtree must have grown from time tn to tn + 1, this implies that c (cid:54)= 1 and (Gtn, v∗ is not the deepest subtree. Next, from the properties of the IC , i (cid:54)= c have the model, all the other growing subtrees amongst (Gtn, v∗ same depth. Since we know that the second deepest subtree did not grow from time tn to tn + 1, this implies that for the center to move, all subtrees (Gt, v∗ , i ≥ 2 must be dead. This completes the proof. tn)ni v∗ tn tn)nc v∗ tn t )ni v∗ t t precedes that of v∗ Lemma 4.2. If the Jordan center changes from time t to time t + 1, then the time of infection of v∗ Proof. From Lemma 4.1, we know that the Jordan center will move at time t only if the deepest subtree grows, and all other subtrees are dead. From Lemma 3.2, in this case the new center will move in the direction of the be the parent of v∗ deepest subtree to n1 v∗ in the original tree (G, u(cid:63)). As argued in the proof of Lemma 4.1 above, the . For some c ∈ [1 : d + 1], let nc v∗ t+1. t t t 10 is dead at time t. Since the deepest subtree must have subtree (Gt, v∗ grown from time t to t + 1, this implies that (Gt, v∗ t )nc v∗ t subtree and thus the next Jordan center cannot be nc . Finally, it is easy to v∗ see from the properties of the IC model that the time of infection of all the other neighbors of v∗ t is one more than v∗ t . This completes the proof. is not the deepest t )nc v∗ t t The above lemma gives rise to the following corollary. Corollary 4.1. A node v which was the Jordan center at time t0, but shifted at time t0 + 1, can never be the Jordan center for t ≥ t0 + 1. We now present our main result for the Jordan center under the IC model. Theorem 4.1. The Jordan center of an infected subtree growing accord- ing to the IC model on an underlying d + 1-regular tree is persistent with probability 1. Proof. Let the set A be defined as follows: A = {v : v = v∗ t for some time t} i.e. A consists of all nodes that have been the Jordan center of the graph at some point of time. We will show that the size of the set A is finite with probability 1. This, along with Corollary 4.1 will complete the proof. Consider any node v which gets infected, say at time t0 ≥ 1. Recall that Gt denotes the infected subgraph at time t and for any t ≥ t0, (Gt, u(cid:63))v denotes the subtree rooted at v and consisting of v and the infected de- scendants of v. It is easy to see that {(Gt, u(cid:63))v}t≥t0 forms a Galton-Watson (GW) branching process [3], which grows independent of the rest of the in- fection tree. A GW branching process is defined as follows. Let Z0 denote the number of nodes at time 0. Then, the number of nodes at time n is given by: Zn−1(cid:88) ξi Zn = i=1 is are i.i.d. where ξ(cid:48) random variables denoting the number of children spawned by each node in the previous generation. For our setup, Z0 = 1 and ξi ∼ bin(p, d). We say that a branching process is 'dead' if ∃ n0 such that Zn = 0 ∀n ≥ n0. For pd > 1, there is a positive probability, say ¯p, that the GW branching process will not die [3]. We assume pd > 1 which implies 11 (cid:26) will not die and grow forever. that for any infected node v, there is a positive probability ¯p that the rooted growing subtree {(Gt, u(cid:63))v}t≥t0 For any i ≥ 2, let Ei be the event Ei := {A ≥ i}. Lemma 4.1 shows that the Jordan center v∗ t moves at the next step only when all except one of the subtrees t )ni rooted at its neighbors are dead and the other v∗ subtree survives. Instead, consider the event F that for the Jordan center v∗ rooted at its neighbors all except one of the subtrees (Gt, v∗ i∈[1:d+1] (cid:27) (cid:27) (cid:26) t t (Gt, v∗ t )ni v∗ t i∈[1:d+1] are dead and the remaining neighbor gets infected. For example, in Figure 3 all except one rooted subtrees of v∗ 4 die and the remaining neighbor gets infected, which in fact becomes the Jordan center v∗ 5. Note that occurrence of the event F is a necessary condition for the Jordan center v∗ t to move. Let the probability of this event be q0 when the root node is the Jordan center, and q when the center resides elsewhere. Since ¯p > 0 and 0 < p < 1, we can see that 0 < q0, q < 1. Let Qi := {event F happens at least i − 1 times}. As can be seen from Figure 3 and Definition 2.2, the events F are indepen- dent across different Jordan centers, and thus we have P(Qi) = q0 × qi−2. P(Ei) ≤(cid:80)∞ that P(Ei) ≤ P(Qi). Now,(cid:80)∞ Also, we see that the event Ei happens only if Qi happens, thereby showing P(Qi) < ∞ and thus from the Borel-Cantelli lemma, we have that with probability 1, only finitely many of these events can happen. This shows that the size of the set A is finite with probability 1 and together with Corollary 4.1, completes the proof of the theorem. i=1 i=1 Corollary 4.2. The distance between the root of a d + 1-regular tree and the Jordan center of an infected subtree growing according to the IC model on the underlying regular tree is finite. Proof. This follows immediately from Theorem 4.1 which shows that the number of distinct centers is finite, and the fact that the Jordan center in the IC model can move at most one hop at a time, from Lemma 3.2. 5 Jordan Center in the discrete SI Model Recall the discrete SI model from Definition 2.3. In this section, we will consider the scenario where the infection starts from the root node u(cid:63) of 12 a d + 1-regular tree G and then spreads along the edges according to the discrete SI model. We borrow some notation and proof ideas from [10], which considered the maximum distance between the Jordan center and the root node. For the infected subtree Gt at any time t, we say that a node is at level l if its distance from the root is l. Let Zl denote the set of infected nodes2 at level l. Z τ is the set of infected nodes at level l, whose parents are in Z τ l−1 and who were infected within τ time slots after their parents were infected. This implies that all nodes in Z τ l are infected by time t ≤ lτ . The 0 for any τ ≥ 0 are singletons and consist of the root node sets Z0 and Z τ l . It is not difficult to see that the evolution of Z τ l = Z τ u(cid:63). Also, let Zτ l with level l forms a Galton-Watson (GW) branching process [6], which we denote by Bτ i.e. Bτ (l) = Zτ l . l Lemma 5.1. Given any > 0, we can find sufficiently large τ and l, inde- pendent of time and the number of infected nodes, such that the probability that at least two B1 branching processes starting from Z τ l survive is at least (1 − ). Proof. See proof of Theorem 5 in [10]. The above lemma shows that starting from nodes at some finite distance l from the root node u(cid:63), there are at least two subtrees in the infected subgraph, whose depths increase by 1 at every time step. Next, we present the main result of this section. Theorem 5.1. Given any > 0, the Jordan center of an infected subtree growing according to the discrete SI model on an underlying d + 1-regular tree is persistent with probability at least (1 − ). Proof. We prove this result by showing that ∃ t0 < ∞ such that the Jordan center of the infection subtree at time t0, v∗ t0, satisfies the following property: the two deepest subtrees (Gt0, v∗ rooted at neighbors t0 increase in depth by one ∀t ≥ t0. From Corollary 3.1, this implies that of v∗ the Jordan center will not move from v∗ t0 and thus establishes its persistence. Lemma 5.1 tells us that with probability at least (1 − ), there are at least two surviving B1 processes starting from Z τ l , i.e., their depths increase by 1 at each timestep. Let dt denote the distance from the Jordan center at time t, v∗ t , to the farthest node in one of the B1 processes. As defined 2Note that the set Zl grows over time, we have suppressed the time dependence in the and (Gt0, v∗ t0)n1 v∗ t0 t0)n2 v∗ t0 notation for convenience. 13 earlier, the centrality of the tree Gt is given by ψGt. Define the function f (t), for t ≥ lτ , as follows: f (t) = ψGt − dt. (7) Some of the properties of f are: 1. f (lτ ) ≤ lτ : Consider time lτ . Starting from the root node u(cid:63), the depth of the infected subtree can increase by at most one in each time unit, and thus the centrality of the root node ψG(u(cid:63)) is at most lτ . From the definition of the Jordan center, this implies that the centrality of the Jordan center ψGt ≤ lτ . We also have dlτ ≥ 0 which gives the claimed property. 2. f is non-increasing: From time t to t+1, if the Jordan center does not change, dt+1 = dt + 1 since the B1 process increases in depth by one at each timestep. Also, ψGt+1 ≤ ψGt + 1 and thus f is non-increasing in this case. On the other hand, if the center changes from time t to t+1, we have ψGt+1 = ψGt from Corollary 3.2. Either the center moves in the direction of the B1 process or away from it, which leads to dt+1 = dt and dt+1 = dt + 2 respectively. In both cases, f is non-increasing. 3. f (·) ≥ 0: Follows from the definition of the Jordan center, see Defini- tion 2.1. T )n1 v∗ T Consider the Jordan center v∗ Recall that our goal is to show that after some finite time, the Jordan center will be such that the two deepest subtrees rooted at its neighbors increase in depth at each timestep. We begin by proving that ∃ t < ∞ such T , is such that for any t ≥ T , the deepest that ∀ T ≥ t, the Jordan center, v∗ rooted subtree (Gt, v∗ T at some time T ≥ lτ . If the height of the increases by 1 at every time step after T , we deepest subtree, (Gt, v∗ stop. If not, then ∃ t1 < ∞ such that the deepest subtree (Gt1, v∗ not increase in height at time t1 + 1. Then from (7), we have + dt1 − dt1+1 increases in depth by 1 from t to t + 1. T )n1 v∗ T T )n1 v∗ T does (8) f (t1 + 1) − f (t1) = ψGt1+1 − ψGt1 = dt1 − dt1+1 = −1, (a) (b) 14 where (a) follows from Lemma 3.2, which dictates that since the deepest subtree3 does not grow, the Jordan center will not move from time t1 to t1 + 1, and furthermore, the centrality will remain the same; and (b) follows from the fact that the Jordan center does not move from time t1 to t1 + 1, while the depth of the B1 process increases by 1. Following t1 + 1, we continue the process and again wait for the next timestep when the deepest subtree does not grow in depth. Let {t1, t2, . . .} denote the timesteps when the function f (·) decreases by 1. Thus, we have f (ti) = lτ − i. t )n1 v∗ t Since lτ < ∞, we have that the process will stop at some t < ∞ when either the height of the deepest subtree (Gt, v∗ increases by 1 at every time step following t or f (t) = 0. If it is the former, our claim is proved. Say it is the latter, then from the properties of the function f detailed above, we have that f (t) = 0 ∀t ≥ t. From (7), this implies that ψGt = dt, which shows that the farthest node in one of the B1 processes, is also the farthest node t at any time t ≥ t. Since (or at least one of them) for the Jordan center v∗ the B1 process grows in depth by 1 at each timestep, we have that ∀ T ≥ t, T , is such that for any t ≥ T , the deepest rooted subtree the Jordan center, v∗ (Gt, v∗ T )n1 v∗ Now, what remains is to show that ∃ t such that t ≤ t < ∞ and ∀ T (cid:48) ≥ t, T (cid:48), is such that for any t ≥ T (cid:48) ≥ t, the second deepest the Jordan center, v∗ rooted subtree (Gt, v∗ also increases in depth by 1 at every time step. T (cid:48))n2 v∗ T(cid:48) Recall that Lemma 5.1 shows that with probability at least (1 − ), there are at least 2 surviving B1 processes starting from Z τ l . Thus, we know that there is at least 1 B1 process which is not the deepest B1 process considered in the first part of the proof above. Let this process be called B1 2. increases in depth by 1 from t to t + 1. T We split the proof into 2 cases: Case 1: When B1 2 is a branch of the deepest subtree, see Figure 4. Consider time t as derived in the proof of the first claim above. Consider the Jordan center v∗ 2 be vx and the farthest t node in the deepest subtree be vy, see Figure 4. Consider a node ¯v in the deepest subtree (Gt, v∗ and let the farthest node in B1 such that: t )n1 v∗ t dist(¯v, vx) = dist(¯v, vy). 3The deepest rooted subtree for the Jordan center v∗ (9) t1 also does not grow since v∗ t1 belongs to (Gt, v∗ T )n1 v∗ T , from Lemma 3.2 15 Figure 4: Case 1 From the definition of a Jordan center in Definition 2.1, we have dist(v∗ t , ¯v) ≤ ψGt < ∞. Let g(·) be defined as follows: g(t) = dist(v∗ t , ¯v), ∀t ≥ t. (10) (11) Now, consider time t and the second deepest subtree (Gt, v∗ neighbor of the Jordan center v∗ . If this subtree continues to grow in depth by 1 at every following time step, then we are done. If not, ∃ ¯t < ∞ such t that it does not grow at time ¯t+1. If this happens, then from Lemma 3.2 we have that the Jordan center will change4 from time ¯t to ¯t + 1. Furthermore, since the Jordan center moves in the direction of the deepest subtree, we have: rooted at a )n2 v∗ t t g(¯t + 1) − g(¯t) = −1. (12) t )n2 v∗ Now, continue this process and stop when either the height of the subtree (Gt, v∗ grows by 1 at every following time step, or when g(·) = 0. Let If g(·) = 0, the Jordan center has reached ¯v, for which this time be t(cid:48) 0. the two deepest subtrees are (a subtree of) the deepest subtree at time t and B1 2, both of which grow by one at every following time step. Thus we t 4Note that it is possible that the depth of the second deepest is the same as the deepest and in this case from lemma 3.2, the center will change only when the second deepest does not grow for two distinct time steps. This technicality does not alter the flow of the proof. 16 have shown that after some finite time, the Jordan center will be such that the two deepest subtrees rooted at its neighbors increase in depth at each timestep. This completes the proof of the claim in Case 1. Figure 5: Case 2 Case 2: When B1 2 is not a branch of the deepest subtree, see Figure 5 for an illustration. The proof for this case is similar to the proof for the deepest subtree considered earlier, and so we skip it here. Finally, tying together all the pieces described above, we have that there exists some time t0 < ∞ such that ∀t > t0, the two deepest subtrees (Gt, v∗ rooted at the neighbors of the Jordan center v∗ t increase in depth from t to t + 1. From Corollary 3.1, this implies that the Jordan center will not move and thus establishes its persistence. and (Gt, v∗ t )n1 v∗ t t )n2 v∗ t Corollary 5.1. Given any > 0, the distance between the root of a d + 1- regular tree and the Jordan center of an infected subtree growing according to the discrete SI model on the underlying regular tree is finite with probability at least (1 − ). Proof. This follows immediately from Theorem 5.1 which shows that the number of distinct centers is finite, and the fact that the Jordan center in the SI model can move at most one hop at a time, from Lemma 3.2. This result has also been proved in [10, Theorem 5]. 6 Jordan center in the SI model Recall the continuous time SI model from Definition 2.4. In this section, we will consider the scenario where the infection starts from the root node 17 u(cid:63) of a d + 1-regular tree G and then spreads along the edges according to the SI model. In this case, we show that when the height of the tree is n, the Jordan center is within a distance of O(log n) from the root u(cid:63) of the tree. u(cid:63) u(cid:63) Lemma 6.1. For any > 0 and n large enough, when the depth of a deepest subtree (Gt, u(cid:63))n1 rooted at a neighbor of u(cid:63) is n, the depth of every other subtree (Gt, u(cid:63))ni , i > 1, is at least n − c3 log n for some constant c3 > 0 with probability at least (1 − ). Proof. Consider the d+1 subtrees of the root u(cid:63). Each of them gets infected according to an independent SI model. For some i ∈ [1 : d + 1], consider any neighbor of the root ni n denote the first time a node at distance n from the root node ni = n. Now, u(cid:63), k ∈ [1 : d + 1] be denoted by the random let the time of infection of nk variable Tk. Since this is the SI model, we have Tk ∼ exp(1). Therefore, the total time for the infection to reach level (n + 1) through ni n. We use results on branching random walks from [9] to show that u(cid:63) and let Bi u(cid:63) gets infected, i.e., depth(GBi u(cid:63) is Ti + Bi , u(cid:63))ni u(cid:63) n P(Bi n ≤ γn + c1 log n − x) ≤ e−δx (13) for some constants γ, c1, δ > 0, details are provided below in Section 6.1. For any n ≥ 1, let E i δ ) log n ∆= t1. Taking x = 2 n denote the event that Bi n ≤ γn + (c1 − 2 δ log n in (13), we get (cid:1) ≤ 1 P(cid:0)E i (cid:1) < ∞ and thus, applying the Borel-Cantelli n2 . (14) n P(cid:0)E i n It is easy to see that(cid:80)∞ n=1 lemma, we get that with probability 1, only finitely many of these events can occur. Therefore, for n large enough n > γn + c1 log n ∆= t1 Bi w.p. 1 (15) where c1 = c1 − 2 δ . Next, consider some j (cid:54)= i and the subtree rooted at the correspond- n−c3 log n denote the first time a node at distance i.e., u(cid:63), gets infected, = n − c3 log n. Again, using [9] we also have ing neighbor nj n − c3 log n from the root node of the subtree, nj depth(Gj u(cid:63). Let Bj , u(cid:63))nj u(cid:63) Bn−c3 log n P(Bj n−c3 log n ≥ γ(n − c3 log n)+c2 log(n − c3 log n) + x) ≤ e−δx (16) 18 for some γ, c2, δ > 0 and any c3 ≥ 0. Once again, applying the Borel Cantelli lemma, we know that for n large enough n−c3 log n ≤γ(n − c3 log n) Bj + c2 log(n − c3 log n) ∆= t2 w.p. 1 (17) δ . The above arguments hold true for any neighbor nj where c2 = c2 + 2 u(cid:63), j (cid:54)= i. Next, note that by choosing c3 large enough, one can make t2 smaller than t1. Now using (15) and (17), we can choose c3 large enough such that for n large enough and any j (cid:54)= i, with probability at least (1 − ) we have: Tj − Ti ≤ t1 − t2 ⇒ Bj n−c3 log n + Tj ≤ Bi n + Ti. (18) This says that by the time any subtree (Gt, u(cid:63))ni reaches a depth n, all other subtrees have depths greater than n− c3 log n with probability (1− ) and this proves the lemma. See Figure 6 for an illustration of the result. u(cid:63) Using this lemma, we have the following theorem: Theorem 6.1. Given any > 0, when the depth of an SI infected tree is n, the Jordan center of the tree is within a distance of C log n from the root with probability at least (1 − ), for some constant C. Proof. This follows from Lemma 6.1. When the depth of the tree is n, the original root, u(cid:63), has a centrality of n. Consider any node v which is at a distance of more than c3 log n from the root, where the constant c3 has been defined in Lemma 6.1. From Lemma 6.1, for n large enough the depth of all the subtrees rooted at the neighbors of the root node is at least c3 log n. Thus, the centrality of node v is greater than n, which is the centrality of the root. This shows that the node with minimum centrality i.e. the Jordan center, has to lie within a distance of c3 log n from the root. 6.1 Proofs of (13) and (16) In the proof of Lemma 6.1, what remains is to show how we obtain (13) and (16). We use Theorems 1 and 2 from [9] in order to establish these in- equalities. Here, we show that our problem setting satisfies all the necessary conditions for these theorems to be applied. Let Zn(t) denotes the number of nodes v in the nth generation (nodes which are at a distance of n from the root node), which are born before time t, and let Zn = supt→∞ Zn(t). Since our underlying tree model is a d + 1-regular 19 Figure 6: Tree in the SI model tree, we have Zn = dn. Bn is defined as the first time of a birth in the nth generation. The time of infection of the nodes in the nth generation are denoted by zn1, zn2,··· . F (t) = E[Z1(t)] and α = inf{t : F (t) > 0}. Since we are working with the SI model, Definition 2.4, where the infection spread from one node to its neighbour is exponentially distributed with mean 1, we can see that α = 0. Now, define: (cid:88) φ(θ) = E[ e−θz1r ]. Since z1r ∼ exp(1) we have: r φ(θ) = d 1 + θ . Again, define: µ(a) = inf{eθaφ(θ) : θ ≥ 0}. On substituting φ(·) we get: µ(a) = ade1−a. 20 (19) (20) (21) (22) (a) Distance from the root (b) Number of center changes Figure 7: Jordan center in the IC model (p = 0.4, d = 4) Time constant γ is defined as: γ = inf{a : µ(a) ≥ 1}. Since µ(0) = 0 and µ(1) = d, we have by the continuity of µ(·) that 0 < γ < 1. (23) (24) We have that the conditions of Theorem 2 [9] are satisfied which gives for some constants γ > 0, c1, δ > 0 P(Bn ≤ γn + c1 log n − x) ≤ e−δx. (25) Since Z1 = d < ∞, we apply part (b) of the theorem on Bn−c3 log n to get: P(Bn−c3 log n ≥ γ(n − c3 log n) + c2 log(n − c3 log n) + x) ≤ e−δx. (26) 7 Simulations In this section, we simulate infection spread on a regular tree using the IC model and the SI model and track the movement of the Jordan center of the infected subtree. In particular, we consider the maximum distance between the Jordan center and the root node and the number of times the center changes over the course of the simulation. For the IC model, we created an underlying 4-regular tree and used p = 0.4 for the probability of an infected node spreading the infection to its neighbor. Starting with the root node, we spread the infection for 40 21 (a) Distance from the root (b) Number of center changes Figure 8: Jordan center in the IC model on an irregular tree (p = 0.4, d ∈ {3, 4}) (a) Distance from the root (b) Number of distinct centers Figure 9: Jordan center in the SI model (d = 4) 22 timesteps5 and evaluate the Jordan center of the infected subtree at each timestep. We repeated this experiment 100 times and present our results in Figures 7a and 7b. We see that for a majority of the simulations, the Jordan center remains close to the root, as illustrated in Figure 7a. Also, from Figure 7b we see that in almost all the experiments, the Jordan center did not change in the last 10 out of the 40 timesteps. This suggests that the Jordan center is close to attaining persistence, as suggested by our theoretical results in Section 4. For the continuous time SI model, we again considered a 4-regular tree and in each run, ran the simulation till the infection spread to 100 vertices. We tracked the movement of the Jordan center each time a new node is added. We repeated the experiment 100 times and present our results in Figures 9a and 9b. As for the IC model, the Jordan center remains quite close to the root over the course of the simulation, and the center only visits a few nodes during the last 30 out of the 100 timesteps. While our theoretical results for the continuous time SI model in Section 6 did not prove persistence of the Jordan center or indeed even bounded distance from the root, empirical results strongly suggest these hold true. 8 Discussion The theme of this paper broadly is how centers of an evolving graph change over time. We have looked at the specific case where an infection spreads on a regular tree following the IC and the discrete SI models and we track the Jordan center of the infected subtree. There are several possible directions we can consider here: 1) Other graph topologies: In this paper, we have considered in- fection spreads on regular trees. Our results for the IC and the discrete SI models directly extend to irregular trees, with minimum degree of each node dmin > 3, and pd > 1. This follows from the following observation: for any such irregular tree, we can consider the embedded dmin-regular tree. Our proofs for the case of an underlying regular tree are based largely on the observation made in Corollary 3.1, which states that the Jordan center does not move as long as the two deepest subtrees continue to grow in depth. If this property holds true for the embedded dmin-regular tree, then it is 5We were able to run the IC model simulation for only 40 timesteps of the simulation because it became computationally infeasible beyond that. This is also why we were not able to run the simulations for the discrete SI model long enough and hence do not present the results here. 23 (a) Distance from the root (b) Number of center changes Figure 10: Jordan center in the Preferential Attachment model (a) Distance from the root (b) Number of center changes Figure 11: Balancedness Center in the IC Model also true for the original irregular tree graph and thus, we have that the Jordan center will not move here as well. Similar to the previous section, we simulate the IC model with p = 0.4 on an irregular tree where each node has a degree uniformly chosen between 3 and 4. These results are shown in Figures 8a and 8b. Also, while we have considered infection spreading models for the growth of the random tree, one can consider other evolution models as well. For example, we can consider a tree graph growing according to the Preferential Attachment model [11] in which each new node connects to existing nodes with probability proportional to the degrees of the nodes in the previous time step. We track the Jordan center of the graph at each timestep as it grows. Some preliminary simulation results reported in Figures 10a and 10b do seem to indicate that the Jordan center is persistent in this model as well and indeed remains very close to the origin node. 24 2) Other graph centers: While we have only considered the Jordan center in this paper, there are various other popular notions of a graph cen- ter in the literature, such as the distance center, betweenness center etc. An obvious research direction is to explore the question of persistence for dif- ferent centers under various growth models. Recently, [7] considered scoring vertices in a tree graph based on the maximum size of the subtrees rooted at the neighbors of the considered vertex, and then choosing the vertex with the minimum score as the tree center. We will refer to this center as the "balancedness center". They proved the persistence of this center in both the uniform as well as preferential attachment tree growth models, as well as the (continuous-time) SI infection spread model on a regular tree. In each of these models, one new node is added to the tree at any time. On the other hand, for the IC and discrete SI models studied in this paper, several nodes can simultaneously be added to the tree and this can intro- duce much more variation in the movement of the balancedness center. For example, Figures 11a and 11b indicate that the balancedness center in this case demonstrates much more movement and it will be interesting to check if the balancedness center is indeed persistent in this case. References [1] Borgatti, Stephen P. "Centrality and network flow." Social networks 27.1: 55-71, 2005. [2] Jackson, Matthew O. Social and Economic Networks. Princeton Uni- versity Press, 2010. [3] Dawson, Donald A. "Introductory lectures on stochastic population systems." arXiv preprint arXiv:1705.03781 (2017). [4] Fanti, Giulia, Peter Kairouz, Sewoong Oh, Kannan Ramchandran, and Pramod Viswanath. "Rumor source obfuscation on irregular trees". In ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Science, pp. 153-164, 2016. [5] Fanti, Giulia, Peter Kairouz, Sewoong Oh, Kannan Ramchandran, and Pramod Viswanath. "Metadata-conscious anonymous messaging." In International Conference on Machine Learning, pp. 108-116, 2016. [6] Haccou, Patsy, Peter Jagers, and Vladimir A. Vatutin. Branching pro- cesses: variation, growth, and extinction of populations. Cambridge uni- versity press, 2005. 25 [7] Jog, Varun, and Po-Ling Loh. "Persistence of centrality in random growing trees." Random Structures & Algorithms, 2018. [8] Kingman, John F. C. "The first birth problem for an age-dependent branching process." The Annals of Probability 790-801, 1975. [9] McDiarmid, Colin. "Minimal positions in a branching random walk." The Annals of Applied Probability 5.1: 128-139, 1995. [10] Zhu, Kai, and Lei Ying. "Information source detection in the SIR model: A sample-path-based approach." IEEE/ACM Transactions on Networking 24.1: 408-421, 2016. [11] Barabasi, Albert-Laszlo, and Reka Albert. "Emergence of scaling in random networks." Science 286.5439 : 509-512, 1999. [12] Shah, Devavrat and Tauhid Zaman. "Rumors in a network: Who's the culprit?" IEEE Transactions on Information Theory, 57(8):5163-5181, 2011. [13] Bubeck, Sebastien, Luc Devroye, and Gabor Lugosi. "Finding Adam in random growing trees. Random Structures and Algorithms". Random Structures & Algorithms, 50.2: 158-172, 2017. [14] Slater, Peter J. "Maximin facility location". Journal of National Bureau of Standards B, 79:107-115, 1975. [15] Tan, Chee Wei, Pei-Duo Yu, Chun-Kiu Lai, Wenyi Zhang, and Hung- Lin Fu. "Optimal detection of influential spreaders in online social networks." Annual Conference on Information Science and Systems (CISS), pp. 145-150, 2016. [16] Hwang, Woochang, Young-rae Cho, Aidong Zhang, and Murali Ra- manathan. "Bridging centrality: identifying bridging nodes in scale-free networks." In ACM SIGKDD International conference on Knowledge discovery and data mining, pp. 20-23, 2006. [17] Galashin, Pavel. "Existence of a persistent hub in the convex prefer- ential attachment model". Probability and Mathematical Statistics, pp. 59-74, 2016. [18] Mitchell, Sandra L. "Another characterization of the centroid of a tree." Discrete Mathematics, 24(3):277-280, 1978. 26 [19] Jog, Varun, and Po-Ling Loh. "Analysis of centrality in sublinear pref- erential attachment trees via the Crump-Mode-Jagers branching pro- cess." IEEE Transactions on Network Science and Engineering 4, no. 1: 1-12, 2017. [20] Khim, Justin, and Po-Ling Loh. "Confidence sets for the source of a diffusion in regular trees." IEEE Transactions on Network Science and Engineering 4, no. 1: 27-40, 2017. [21] Zhu, Kai, and Lei Ying. "Information source detection in networks: Possibility and impossibility results." In Annual IEEE International Conference on Computer Communications (INFOCOM) pp. 1-9. IEEE, 2016. [22] Hedetniemi, S. Mitchell, E. J. Cockayne, and S. T. Hedetniemi. "Linear algorithms for finding the Jordan center and path center of a tree." Transportation Science 15, no. 2: 98-114, 1981. [23] Handler, Gabriel Y. "Minimax location of a facility in an undirected tree graph." Transportation Science 7, no. 3: 287-293, 1973. 9 Appendix We use Theorems 1 and 2 from [9] in order to establish the concentration inequalities used in Section 6. Here, we show that our problem setting satisfy all the necessary conditions for these theorems to be applied. Let Zn(t) denotes the number of nodes v in the nth generation (nodes which are at a distance of n from the root node), which are born before time t. Zn = supt→∞ Zn(t). Since our underlying tree model is a d + 1-regular tree, we have Zn = dn. Bn is defined as the first time of a birth in the nth gener- ation. The time of infection of the nodes in the nth generation are denoted by zn1, zn2,··· . F (t) = E[Z1(t)] and α = inf{t : F (t) > 0}. Since we are working with the SI model, Definition 2.4, where the infection spread from one node to its neighbour is exponentially distributed with mean 1, we can see that α = 0. Now, define: φ(θ) = E[ e−θz1r ] (27) (cid:88) r 27 Since z1r ∼ exp(1) we have: Again, define: φ(θ) = d 1 + θ µ(a) = inf{eθaφ(θ) : θ ≥ 0} On substituting φ(·) we get: µ(a) = ade1−a Time constant γ is defined as: γ = inf{a : µ(a) ≥ 1} Since µ(0) = 0 and µ(1) = d, we have by the continuity of µ(·) that 0 < γ < 1 (28) (29) (30) (31) (32) We have that the conditions of Theorem 2 [9] are satisfied which gives for some constants γ > 0, c1, δ > 0 P(Bn ≤ γn + c1 log n − x) ≤ e−δx (33) Since Z1 = d < ∞, we apply part (b) of the theorem on Bn−c3 log n to get: P(Bn−c3 log n ≥ γ(n − c3 log n) + c2 log(n − c3 log n) + x) ≤ e−δx (34) 28
1512.06880	1	1512	2015-12-21T21:31:19	Where Chicagoans tweet the most: Semantic analysis of preferential return locations of Twitter users	[ "cs.SI" ]	Recent studies on human mobility show that human movements are not random and tend to be clustered. In this connection, the movements of Twitter users captured by geo-located tweets were found to follow similar patterns, where a few geographic locations dominate the tweeting activity of individual users. However, little is known about the semantics (landuse types) and temporal tweeting behavior at those frequently-visited locations. Furthermore, it is generally assumed that the top two visited locations for most of the users are home and work locales (Hypothesis A) and people tend to tweet at their top locations during a particular time of the day (Hypothesis B). In this paper, we tested these two frequently cited hypotheses by examining the tweeting patterns of more than 164,000 unique Twitter users whom were residents of the city of Chicago during 2014. We extracted landuse attributes for each geo-located tweet from the detailed inventory of the Chicago Metropolitan Agency for Planning. Top-visited locations were identified by clustering semantic enriched tweets using a DBSCAN algorithm. Our results showed that although the top two locations are likely to be residential and occupational/educational, a portion of the users deviated from this case, suggesting that the first hypothesis oversimplify real-world situations. However, our observations indicated that people tweet at specific times and these temporal signatures are dependent on landuse types. We further discuss the implication of confounding variables, such as clustering algorithm parameters and relative accuracy of tweet coordinates, which are critical factors in any experimental design involving Twitter data.	cs.SI	cs	Where Chicagoans tweet the most: Semantic analysis of preferential return locations of Twitter users Aiman Soliman, Junjun Yin, Kiumars Soltani, Anand Padmanabhan, Shaowen Wang CyberGIS Center for Advanced Digital and Spatial Studies, University of Illinois at UrbanaChampaign, Urbana, IL 61801 {asoliman, jyn, soltani2, apadmana, shaowen}@illinois.edu ABSTRACT Recent studies on human mobility show that human movements are not random and tend to be clustered. In this connection, the movements of Twitter users captured by geo-located tweets were found to follow similar patterns, where a few geographic locations dominate the tweeting activity of individual users. However, little is known about the semantics (landuse types) and temporal tweeting behavior at those frequently-visited locations. Furthermore, it is generally assumed that the top two visited locations for most of the users are home and work locales (Hypothesis A) and people tend to tweet at their top locations during a particular time of the day (Hypothesis B). In this paper, we tested these two frequently cited hypotheses by examining the tweeting patterns of more than 164,000 unique Twitter users whom were residents of the city of Chicago during 2014. We extracted landuse attributes for each geo-located tweet from the detailed inventory of the Chicago Metropolitan Agency for Planning. Top-visited locations were identified by clustering semantic enriched tweets using a DBSCAN algorithm. Our results showed that although the top two locations are likely to be residential and occupational/educational, a portion of the users deviated from this case, suggesting that the first hypothesis oversimplify real-world situations. However, our observations indicated that people tweet at specific times and these temporal signatures are dependent on landuse types. We further discuss the implication of confounding variables, such as clustering algorithm parameters and relative accuracy of tweet coordinates, which are critical factors in any experimental design involving Twitter data. Categories and Subject Descriptors J.4 [Social and Behavioral Sciences]: Sociology General Terms Measurement, Experimentation, Human Factors Keywords Twitter, Big Data, social media, human mobility, semantic trajectories, urban activity INTRODUCTION Related Work 1. 1.1 Understanding human mobility patterns in urban environments is critical for organizing transportation, urban planning and studying the spread of infectious diseases [1-3]. Recently, the availability of new sources of geo-located Big Data has shown promise in studying the complex human mobility patterns [4]. For example, users’ trajectories, which were extracted from phone call history [1] revealed that human movements are not random and characterized by a preferential return to few frequently-visited locations [2, 5]. Identifying these frequently- visited locations and their semantics (e.g., home, work) is essential for understanding the origin and destination of each trip and the purpose of human mobility at both the individual and collective levels. Although human movements could be extracted from different data sources (e.g. phone call data, GPS trajectories, survey data), there are advantages of using Twitter data over other Big data sources in studying human mobility patterns. For example, geo-located Twitter records provide a consistent data at regional scales, which is available to public without infringement on individual privacy in case of using phone call records, and the positional accuracy of each geo-locate tweets, approximately 10 meters, provides adequate resolution to resolve the associated landuse type particular in mixed landuse neighbors [5-7]. Analysis of geo-located Twitter data showed the presence of preferential return locations, in which most of the uses’ broadcasting activities take place near few geographic locations [6]. Different methods were used to identify preferential return locations for Twitter users or spatial clustering of big data. The simplest method is to impose a fixed grid over the study area and identify the cells with largest number of geo-located tweets [8]. More advanced techniques include bivariate normal density kernel [9], modified mean shift algorithm [10] or applying DBSCAN algorithm to geo-located tweets collected during night (7:00 P.M. to 4:00 A.M.) [11]. Both mean shift and DBSCAN algorithms do not require the number of clusters in advance, which is advantageous because the number of frequently-visited locations could vary from one user to another. 1.2 Although the literatures provide various examples of detecting preferential return locations of Twitter users, the semantic composition (i.e. landuse type) of these top-visited locations was not examined in detail. For example, the top two visited locations, where users spent most of their time, are usually assumed to be home and work locales without verifications [5, 6, 11]. These assumptions could result in a higher uncertainty in identifying the purpose of individual and collective mobility patterns and understanding intercity dynamics. In this context, we present a better method for extracting and labeling frequently-visited locations of Twitter users. We examined the semantic composition of top-visited locations of residents of Contribution Chicago by matching them to a detailed landuse map, which is to the best of our knowledge, the first attempt to incorporate and analyze the semantics of top tweeted locations. In addition, we introduce a semantic uncertainty index as an objective method for assessing the quality of identified preferential return locations and to account for issues such as GPS accuracy and the geographic spread of geo-located tweets as factors that could impact the quality of spatial clustering. Finally, we demonstrate that the distribution of hourly volume of tweets at these preferential return locations have distinct time signatures that are dependent on the dominant landuse type and discuss its potential application for labeling frequently visited locations. DATA AND METHODS 2. Geo-located Twitter data was collected over North America using Twitter streaming API from January 1st through December 31st, 2014 within a bounding box with lower left and upper right corners' coordinates 41.201577N, -88.707599W, 42.495775N, -87.524535W respectively. Each tweet in the final data set contained a geo-tag and a timestamp indicating local time in CDT, which was corrected for daylight saving calendar. The trajectory of each unique Twitter user (x, y, time) was constructed by arranging a user’s locations recorded during 2014 in chronological order. We selected the data above the median value, which yielded 164,409 (48.8%) unique users with an average of 64 tweets per user, a maximum of 2683 tweets and minimum of 4 tweets (median value). The landuse inventory for Northeastern Illinois is one of the most detailed and updated landuse maps of Chicago [12]. We retained 23 classes from the original landuse classes that were found to be informative, while merging the remaining excessively detailed classes into seven meaningful classes. Following, we constructed the semantic trajectories by associating each geo-located tweet with one of the thirty landuse types (x, y, time, landuse). We decided to reassign geo-located tweets that fall on the road networks to the nearest landuse polygon since it is unlikely that streets could count as top-visited locations. We applied the DBSCAN clustering algorithm [13] to each unique user trajectory Ts to identify clusters of semantic-enriched tweets, which should corresponded to the top-visited locations for that user. We defined a search window (ε) of approximately 250 meters (0.0025 degrees) [5] to account for variability of GPS accuracy between devices and the fact that top-visited locations are not infinitesimal points on a map. The minimum number of points to form a cluster was selected to be four to ensure that it is a true location and not merely a coincidence. After applying the DBSCAN clustering algorithm we were able to associate most of the tweets to one of the top-visited clusters, which were ordered in a descending fashion based on the number of associated tweets (x, y, time, landusj, cluster rank). The accuracy of any identified top visited location (named here semantic uncertainty) was assessed using the ratio of tweets that belongs to the dominant landuse type to the total number of tweets at each frequently-visited location. We pooled the dominant landuse types for top-visited locations for all unique users residing of Chicago to test the hypothesis that the top two visited locations are home and work locales (Hypothesis A). Furthermore, we tested the hypothesis that users are likely to tweet from their top-visited location at a specific time of the day (Hypothesis B). We tested Hypothesis B by combining timestamps from all tweets and unique users that belong to a certain rank of top-visited locations into a single pool, which was then examined for the existence of significant features. RESULTS AND DISCUSSIONS Semantics of Frequently-Visited 3. 3.1 Locations The absolute count of frequently-visited locations decreased exponentially by increasing the rank of top-visited locations (Figure 1). The power distribution signifies that the majority of Chicago residents tweet near one or the top two locations at the most and it is less likely to find users who frequently tweet from a large number of locations. These results are in agreement with the findings of Song et al. [2] which suggested that the top two locations explained more than 60% of the total frequent visitations’ pattern. The most significant result is that, although the most likely location for Chicagoans to tweet from is their home (i.e. residential land use type), there are significant cases in which home did not count as the most frequently-visited location. The results summarized in Figure 1 indicate that Hypothesis A is oversimplifying the real world. We also speculate that the percentage of residential landuse types identified at higher ranks exists to complete the cases in which home were not at the top tweeting location. However, a complete understanding of this effect needs a full analysis of the transition matrix between different landuse types. Figure.1 Semantic composition of top-visited location for all residents of Chicago Nevertheless, the likelihood of finding a frequently-visited location with a particular landuse type is still dependent on the rank of frequency of visitation. For example, it is most likely to find dominant landuse type such as schools, post secondary education and office at the second top location and similarly hotels are likely to be found on the top-visited local (Figure 2). We plotted the distribution of uncertainty ratio for all identified top-visited locations at the top five ranks for all landuse types (Figure 3). The distribution of semantic uncertainty shows that the majority of clusters identified using DBSCAN have a purity ratio that is no less than 65%, as indicated by the lower 25 percentile edge for all five ranks. However, the semantic purity decreased with increasing the rank, which could be attributed to the spread of tweets as the frequency of visitation decrease. Table 1. The impact of DBSCAN parameters on the percentage of dominant landuse type at Top Location 1; Exp.1 (ε=250 m, minimum of 4 points) and Exp.2 (ε=500 m, minimum of 10% of total user tweets) Top Location 1 Residential Hotel Urban mix Com. Office Urban mix Resi. Cultural Post Sec. Educ. Exp. 1 61.1 4.5 4.3 3.8 3.3 3.3 2.9 Exp. 2 76.1 1 2.9 2.5 2.8 0.7 2.9 Temporal Tweeting Patterns 3.3 We examined the hourly patterns of tweeting in order to test if people tend to tweet at their top-visited locations at particular time of the day (Hypothesis B). Figure 4 depicts the hourly distribution of tweets at the most visited location (Top Location 1) classified by different land use types. The observations clearly demonstrate that tweeting intensity at top-visited Location 1 varies considerably between different landuse types. We also noticed that the observed variations of tweeting intensity for different landuse types are tightly coupled with their function. For example, K-12 educational landuse is associated with a peak in the morning and a significant drop after 3:00 P.M. local time, which is consistent with students’ daily activity. Similarly, the temporal signature associated with Office landuse type showed a similar diurnal pattern like schools, but with less decline in the afternoon hours as people continue to work after 3 P.M. In contrast, residential, hotel/motel and urban mix with residential units showed all a significant increase of nocturnal activity. Remarkably, the temporal patterns bear clear resemblance to spectral signatures, which are common in remote sensing analysis. Figure 4. Hourly tweeting intensity for most frequently- visited location (Top Location 1) normalized by the total clusters count for each landuse category Our results depicting the changes of hourly tweeting intensity for different landuse types (Figure 4) compared well with the travel diary survey collected in the city of Chicago, which was measured using survey data collected from 30,000 individuals, refer to Figure 4-a from Jiang et al.[14]. In addition, the comparison between Twitter and survey-based data indicated that tweeting activity might be slightly decoupled from daily activity Figure 2. Dependency of landuse type on the rank of frequency of visitation, excluding residential landuse for visualization; clusters count was normalized to the total number of clusters in the rank Sensitivity of DBSCAN Parameters Figure 3. Boxplot of semantic uncertainty range for all top- visited locations arranged in a descending order, where rank 1 represents the most visited location; the semantic uncertainty is scaled between 0 and 1. For example 0.7 means that 70 % of the tweets were identified with the dominant landuse type 3.2 DBSCAN algorithm requires selecting two parameters, the size of searching window (ε) and the minimum number of points in any cluster. In order to investigate the impact of parameters’ sensitivity on our results, we compared our initial selection (a 250 meter search window and a minimum of 4 points) with a more restrictive set of parameters (500 meters and a minimum of 10% of the users’ tweets), hereafter named Experiment 2. The results (Table 1) indicated that restricting clustering parameters resulted in a smaller number of clusters for most landuse types, with exception to residential and K12 educational. Remarkably, the number of hotels and cultural/entertainment clusters dropped significantly for Experiment 2 (Table 1). We speculate that relaxing the clustering parameters in Experiment 1 allowed detection of the behavior of temporary residents and tourists, which is characterized by a single top location (hotel) with fewer tweets and scattered visitation near touristic attraction sites. By restricting the clustering parameters in Experiment 2, the Post-Secondary Educational landuse type became the second most common landuse type found at top-visited location 1 after residential, which is likely because university students spent most of their time on campus. In summary, the results in Table 1 suggest that tuning DBSCAN parameters could allow filtering some generic human movement’s patterns. [5] (Nov. 2014), e112608- e112608. DOI= http://dx.doi.org/10.1371/journal.pone.0112608 Jurdak, R., Zhao, K., Liu, J., AbouJaoude, M., Cameron, M. and Newth, D. 2014. Understanding Human Mobility from Twitter. arXiv:1412.2154 [physics]. (Dec. 2014). [6] Frank, M.R., Mitchell, L., Dodds, P.S. and Danforth, C.M. 2013. Happiness and the Patterns of Life: A Study of Geo- located Tweets. Scientific Reports. 3, (Sep. 2013). DOI = http://dx.doi.org/10.1038/srep02625 [7] Liu, J., Zhao, K., Khan, S., Cameron, M. and Jurdak, R. 2014. Multi-scale population and mobility estimation with geo-tagged Tweets. arXiv preprint arXiv:1412.0327. [8] Cao, G., Wang, S., Hwang, M., Padmanabhan, A., Zhang, Z. and Soltani, K. 2015. A scalable framework for spatiotemporal analysis of location-based social media data. Comput. Environ. Urban Syst.51, (2015), 70–82. DOI= http://dx.doi.org/10.1016/j.compenvurbsys.2015.01.002 [9] Li, Z., Ding, B., Han, J., Kays, R. and Nye, P. 2010. Mining periodic behaviors for moving objects. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining (Washington, DC, USA, on July 24-28, 2010), 1099–1108. DOI= http://dx.doi.org/10.1007/s10618-011-0227-9 [10] Cao, L., Luo, J., Gallagher, A., Jin, X., Han, J. and Huang, T.S. 2010. A worldwide tourism recommendation system based on geotagged web photos. In Proceedings of 2010 IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP), (Dallas, Texas, USA, March 14-19, 2010), 2274–2277. DOI= http://dx.doi.org/10.1109/ICASSP.2010.5495905 [11] Bora, N., Chang, Y.-H. and Maheswaran, R. 2014. Mobility patterns and user dynamics in racially segregated geographies of US cities. In Social Computing, Behavioral- Cultural Modeling and Prediction, W. G. Kennedy, N. Agarwal and S.J. Yang, Eds. Springer. 11–18. DOI= http://dx.doi.org/10.1007/978-3-319-05579-4_2 [12] CMAP 2014. Chicago Metropolitan Agency for Planning’s 2010 Land Use Inventory for Northeastern Illinois. [13] Ester, M., Kriegel, H.P., Sander, J. and Xu, X. 1996. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD-96 (Portland, Oregon, USA, August 2– 4, 1996), 226-231. DOI = http://dx.doi.org/10.1023/A:1009745219419 [14] Jiang, S., Ferreira, J. and González, M.C. 2012. Clustering daily patterns of human activities in the city. Data Mining and Knowledge Discovery. 25, 3 (Nov. 2012), 478–510. DOI = http://dx.doi.org/10.1007/s10618-012-0264-z CONCLUSION AND FUTURE WORK during certain periods of the day. For example, most of the surveyed individuals indicated the high likelihood to find them at their Home location during early hours of the day (i.e. 1 A.M.to 5 A.M.). However, this period showed a significant drop of tweeting intensity because people are likely to be found at home sleeping. Another example is the apparent peak of tweeting associated with Office landuse type around 12:00 P.M., which could be linked to the possible increase of tweeting activity during lunch breaks. In general, the distinct diurnal variability of tweeting is a very promising approach for distinguishing the semantics of frequently-visited locations of individual users. However, the ultimate success of classifying a cluster to one of the typical landuse types is dependent on the semantic purity of the cluster, as well as the number of tweets involved. 4. We demonstrated the potential of using Twitter data in analyzing urban dynamics. In this context, we examined the semantic composition of preferential return locations for Twitter users. We found that home (residential landuse) explained 60-75% of the semantics of all top-visited locations, while offices, schools and universities explained approximately 7, 10 and 6% respectively of the semantics of the second most visited location for all users. Our results suggested that generic assumptions, such as top two visited locations are home and work locales, are oversimplifications of the real world and therefore they are not reliable. However, the hourly volumes of tweets provide a more generic method to label these frequently-visited locations because they are tightly coupled with the landuse types of preferential return locations. The presented case study is a work in progress to build a scalable workflow for the synthesis of geo- located Twitter data with more authoritative data sources. Pinpointing the frequently-visited locations is the first step to analyze the origin-destination of users’ trips and to understand the purpose of these trips as well as the function of different neighbor within metropolitan cities based on the collective mobility patterns of its residents. ACKNOWLEDGMENTS 5. This research is based in part upon work supported by the U.S. National Science Foundation under grant numbers: 1047916, 1443080, and 1354329. Insightful comments were received from members of the CyberGIS Center for Advanced Digital and Spatial Studies. 6. [1] González, M.C., Hidalgo, C.A. and Barabási, A.-L. 2008. REFERENCES Understanding individual human mobility patterns. Nature. 453, 7196 (Jun. 2008), 779–782. DOI= http://dx.doi.org/10.1038/nature06958 [2] Song, C., Qu, Z., Blumm, N. and Barabasi, A.-L. 2010. Limits of Predictability in Human Mobility. Science. 327, 5968 (Feb. 2010), 1018–1021. DOI= http://dx.doi.org/10.1126/science.1177170 [3] Padmanabhan, A., Wang, S., Cao, G., Hwang, M., Zhang, Z., Gao, Y., Soltani, K. and Liu, Y. 2014. FluMapper: A cyberGIS application for interactive analysis of massive location-based social media. Concurr. Comp- Pract. 26, 13 (Sep. 2014), 2253–2265. DOI= http://dx.doi.org/10.1002/cpe.3287 [4] Wang, Q. and Taylor, J.E. 2014. Quantifying human mobility perturbation and resilience in Hurricane Sandy. PLoS ONE
1501.07203	1	1501	2015-01-28T17:23:59	Structural Patterns of the Occupy Movement on Facebook	[ "cs.SI", "cs.HC", "physics.soc-ph" ]	In this work we study a peculiar example of social organization on Facebook: the Occupy Movement -- i.e., an international protest movement against social and economic inequality organized online at a city level. We consider 179 US Facebook public pages during the time period between September 2011 and February 2013. The dataset includes 618K active users and 753K posts that received about 5.2M likes and 1.1M comments. By labeling user according to their interaction patterns on pages -- e.g., a user is considered to be polarized if she has at least the 95% of her likes on a specific page -- we find that activities are not locally coordinated by geographically close pages, but are driven by pages linked to major US cities that act as hubs within the various groups. Such a pattern is verified even by extracting the backbone structure -- i.e., filtering statistically relevant weight heterogeneities -- for both the pages-reshares and the pages-common users networks.	cs.SI	cs	Structural Patterns of the Occupy Movement on Facebook Michela Del Vicario IMT Lucca, Italy Qian Zhang Northeastern University, Boston, USA Alessandro Bessi IUSS Pavia, Italy Fabiana Zollo IMT Lucca, Italy Antonio Scala ISC CNR,Rome, Italy Guido Caldarelli IMT Lucca, Italy Walter Quattrociocchi IMT Lucca, Italy [email protected] 5 1 0 2 n a J 8 2 ] I S . s c [ 1 v 3 0 2 7 0 . 1 0 5 1 : v i X r a Abstract In this work we study a peculiar example of social organi- zation on Facebook: the Occupy Movement -- i.e., an interna- tional protest movement against social and economic inequal- ity organized online at a city level. We consider 179 US Face- book public pages during the time period between September 2011 and February 2013. The dataset includes 618K active users and 753K posts that received about 5.2M likes and 1.1M comments. By labeling user according to their interaction pat- terns on pages -- e.g., a user is considered to be polarized if she has at least the 95% of her likes on a specific page -- we find that activities are not locally coordinated by geograph- ically close pages, but are driven by pages linked to major US cities that act as hubs within the various groups. Such a pattern is verified even by extracting the backbone structure -- i.e., filtering statistically relevant weight heterogeneities -- for both the pages-reshares and the pages-common users net- works. Introduction Social media have proved to foster aggregation of people around shared interests such as political ideas, narratives and worldviews [1, 2, 3, 4]. An interesting scenario to be ex- plored is the case of online political movements that coor- dinate and interact through social media [5]. To which ex- tent microblogging platforms such as Twitter or Facebook played a crucial role in protest dynamics has been a matter of debate in [6, 7, 8]. Occupy Wall Street, an international movement against social and economic inequality, is a pecu- liar example of the use of Information and Communication Technologies (ICTs). There, the diffusion of ideas, the re- cruitment of people, as well as the promotion of the protest, are ascribed to online social media [9, 10]. Particular em- phasis is given to the simplification gained in the communi- cation and coordination paradigm of protest's activities. The reason why Twitter and Facebook appeared to be particu- larly suited in supporting the diffusion of political ideas and socio-economic objectives of the movement resides in their structure. Many works analyzed the topology and dynam- ics on the Occupy Movement's network [11, 12, 13, 14] on Twitter. Copyright c(cid:13) 2015, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. However, users tend to be more active on Facebook. In- deed, the most of online activity during the Arab Spring has been done on Facebook (the total accesses to Twitter were just the 1% of the entire population) [15]. The collective framing -- the processes that, out of the essential features of the movement's purpose and struggle, establish its narra- tives, language, and imagery [6, 16] -- better deal with the Facebook interaction paradigm. On Facebook users can post information (videos, pictures, etc.) without particular limits and information can be liked (positive feedback), shared, or commented by other users. In [17, 18] the diffusion of the Occupy movement on Facebook has been explored. Find- ings indicate that major uses for Facebook within the move- ment include the recruitment of people and resources to lo- cal occupations, information sharing and story telling, and across-group exchanges. Taking into account all the posts from 179 public pages of the Occupy Movement inside the US (US Occupy pages for short) during the period September 2011-February 2013, we analyze users' activity on pages and posts. We find high lev- els of correlation for the number of users, posts, likes, com- ments and reshares1 (which are all power law distributed) on each page, and different activity patterns for posts with different post types or shared by users from different cat- egories. Then, we divide users in two categories, habitual and occasional, depending on their total number of likes. Furthermore, habitual users having at least 95% of their lik- ing activity on one particular page are defined as polarized in that page. Finally, we extract the multi-scale backbone structures for two networks, the pages-reshares network and the pages-common users network. We want to analyze ge- ographical patterns in the information diffusion and polar- ized users' external activity -- i.e., the activity on all pages but the one where the user is polarized. Our findings show that the most of the activity, in terms of likes on each page, is made by occasional and not polarized users, and more- over only 21% of polarized users commented at least once. However, polarized users tend to increase their probability of commenting multiple times, with respect to general users. Our analysis points out that, although the majority of posts is shared by common users, they attract little attention i.e., 1The two terms shares and reshares are interchangeable throughout the manuscript. about 6% of total likes. On the contrary, posts shared by page administrators, which represent only 27% of the total, get all the remaining likes. These evidences emphasize the cen- tral role of the pages in promoting the debate and driving its direction, and at the same time reshape the idea of on- line debate participation. We find that online activities are not locally coordinated by geographically close pages, in- deed pages linked to major US cities -- e.g., New York, Los Angeles, Chicago, Boston, San Francisco -- drive the diffu- sion of contents online and serve as coordination points for all other pages, that perform a minor activity in the system. Related Work Sociologists and political scientists explored the environ- ment of Internet based social movements focusing on a set of common themes [6, 7, 8, 19, 20, 21, 22, 23, 16, 24]. In the general frame of communicational and organizational issues, several relevant topics emerged, such as the afore- mentioned collective framing [6, 16] and the resources mo- bilization, which refers to all those processes exploited by social movements in order to arrange financial, material, and human resources required to sustain their activities in an efficient way [23]. The Internet is commonly perceived as a flexible and relatively fluid media, in terms of reor- ganization. However, at the same time its flexibility may reduce the coherence of the movement's ideological defi- nition. In [24] three communication eras have been iden- tified, each one dominated by a different communication medium -- i.e., print media and word of mouth, television and the Internet. It is argued that each of those communi- cation technologies implies important differences in the net- work structures under which potential protestors may be in- fluenced to act and, hence, in the resulting protest waves. In the particular case of Occupy Wall Street, many works investigated the role of Twitter in organizing and promot- ing the protest [11, 12, 13, 14]. The relationship between the geospatial dimensions of social movements communica- tion networks and the organizational pressures facing such movements, was extensively investigated in [11]. High lev- els of geospatial concentration in the attention allocation is- sue and a differentiation of contents from the national to the international level were found. The evolution of the com- munication activity and of the topics under discussion was analyzed in [12, 13]. A study of the evolution of the Occupy Movement on Facebook was presented in [17]. Twitter, in addition to Occupy, has also played a prominent role in sev- eral other social movements such as the Egyptian revolution- ary protests of 2011 [25, 26, 27, 28] and the Arab Spring in general [29, 30]. The most investigated aspects concern in- formation flows and relationships between news media and information sources [25, 29], analysis of tweets' contents [26], impact of media disruption on the dispersion of the protest [28], presence of social influence and complex con- tagion in the recruitment patterns [27]. The role of social ties in movement recruitment were extensively analysed also of- fline [31]. In [15] evidences regarding the prominent role of Facebook with respect to Twitter are reported. A set of 1K interviews of people who took part in the protest was ana- lyzed, with the interesting finding that, for about 50% of the interviews, the news of the first demonstrations came from Facebook, face-to-face interaction and telephone. Moreover about 50% of the interviewed declared to have access to Facebook, while it is known that less than 1% had access to Twitter. Data Description We collected all posts from 179 Facebook public US pages about the Occupy Movement during the time span Septem- ber 2011-February 2013. Data have been collected using the Facebook Graph API [32] on public accessible pages, for each of which we have a geographical reference located in the US. A total of 618K users is active, in terms of liking activity, on a set of 753K posts2. The total number of likes and comments on the downloaded posts are about 5.2M and 1.1M, respectively. The dataset represents a complete screenshot of the Occupy Movement in the period imme- diately following the outbreak of the protest on September 17th, 2011 in the Zuccotti Park of New York. The dataset covers all the posts until the end of February 2013, at the time when all the major protests were no more active. Af- ter the Zuccotti occupation, in fact, an October full of simi- lar occupational events followed, leading to an international protest movement that extended itself until the end of 2012, when the movement was principally an online collective protest. Preliminaries and Definitions In this section we introduce the basic notions and definitions that will be used throughout the paper. We refer to the set of US Occupy pages and their respective active users as the Occupy System or Network, where a user is said to be active if she liked at least one post belonging to the Occupy pages. Statistical Tools. We use several figures showing PDFs, CDFs, and CCDFs of some metrics related to activity and consumption patterns. We remind that the probability den- sity function (PDF) of a real -- valued random variable is a function fX that describes the probability of the random variable falling within a given range of values, so that Pr[a ≤ X ≤ b] = fX (x)dx. (cid:90) b a The cumulative distribution function (CDF) of a real -- valued random variable X is defined as FX (x) = Pr(X ≤ x) = fX (u)du. (cid:90) x −∞ Similarly, the complementary cumulative distribution func- tion (CCDF) is defined as one minus the CDF, so that CX (x) = 1 − FX (x) = Pr(X > x) = (cid:90) ∞ fX (u)du. x 2 For each post we have the following data: post ID, sharing user ID, sharing time, post type -- e.g., photo, status, video or link -- number of received likes and comments, sharing page ID and object ID, that is a unique reference to the post in case of resharing. Classification of users activity. We assume the following definitions for the users categories in the Occupy System. A first distinction is between habitual and occasional users: habitual users are those who liked at least 5 posts within the Occupy System, while occasional users are those who liked less than 5 posts. Being commonly perceived as posi- tive feedbacks, likes are used as a determinant to identify the membership in one page. Polarized users have been identi- fied by a thresholding technique, using again the like as a discriminant. Habitual users having 95% or more of their liking activity on a particular page are said to be polarized on that page. Notice that a polarized user in the occupy sce- nario is a user having the most of her liking activity on a specific geolocated community. Bipartite networks and backbone filter. A bipartite graph is a triple G = (A, B, E) where A = {ai i = 1 . . . nA} and B = {bj j = 1 . . . nB} are two dis- joint sets of vertices, and E ⊆ A×B is the set of edges -- i.e. edges exist only between vertices of the two different sets A and B. The bipartite graph G is described by the matrix M defined as Mij = if an edge exists between ai and bj 0 otherwise (cid:26) 1 The co-occurrence matrices C A = M M T and C B = M T M count, respectively, the number of common neigh- bors between two vertices of A or B. C A is the weighted adjacency matrix of the co-occurrence graph CA with ver- tices on A. Each non-zero element of C A corresponds to an edge among vertices ai and aj with weight P A ij . The co- occurrence graph CB is built in the same way from the co- occurrence matrix C B. Let A be the set of the 179 Occupy US pages, B1 the set of all posts equivalence classes' representatives (rep- resentatives posts for short)3, and B2 the set of all polar- ized users active on A; the pages-posts bipartite network is then defined as the triple G1 = (A, B1, E1), where an edge ij ∈ E1 exists if representative post b1 j is shared on page a1 i , e1 while the pages-polarized users bipartite network is defined ij ∈ E2 ex- as the triple G2 = (A, B2, E2), where an edge e2 ists if polarized user b2 i . For our analysis we used two networks derived as co-occurrence networks of the pages-posts and the pages-polarized users bipartite net- 1 and C A works. Considering the co-occurrence matrices C A 2 we get two co-occurrence networks on the vertex set A: the pages-reshares (CA 2 ) net- works. In CA 1 an edge between two pages exists if at least one representative post is shared on both pages, while in CA 1 there is a link between two pages if they share at least one user who is polarized in either page. 1 ) and the pages-common users (CA j is active on page a2 In the last section we show results about the Backbone Extraction presented in [33]. Such a method apply a thresh- olding filter based on the local identification of the statisti- cally relevant weight heterogeneities. This kind of approach 3We can consider the equivalence relation of having the same object ID; two posts are equivalent, and hence belong to the same equivalence class, if they have the same object ID. Figure 1: Fractions of number of posts and activity on posts (likes and comments) divided by post types. The lev- els of commenting activities are proportional to those of the number of posts for each different post type. Photos get most part of the likes, although they represent less than a fifth of the total number of posts. For other post types the fraction of likes is one half of the number of posts. is able to filter out the backbone of dominant connections in weighted networks with strong disorder, preserving the structural properties and hierarchies at all scales. The Back- bone structure has been extracted for the two aforemen- tioned co-occurrence real networks. The discrimination of the right trade-off between the level of network reduction and the amount of relevant information preserved in the new representation involve additional issues. In many cases, the probability distribution P (x) that any given link is carrying a weight x is broadly distributed, spanning several orders of magnitude. Such a problem is addressed by using the afore- mentioned method presented in [33]. Results and Discussion Activity on Pages Our analysis targets information consumption patterns of the US Occupy Movement. The focus is on the relationship be- tween geographic distance and the spreading of information across pages. In addition we characterize the role of polar- ized users promoting and bridging inter-page connections. As a first step we look at the users' activity on the different city related pages. Figure 1 reports the fractions of number of posts, likes, and comments for the four kind of posts: sta- tus, photo, link, and video. The commenting activity is cor- related to the number of posts, while more than a half of the likes is concentrated on photos. Statuses, that account for al- most two fifths of the total number of posts, get less than a fifth of the total number of likes. Such a disproportion in the liking activity might be due to the users limited attention [34, 35, 36] that preferentially tend to select easy to han- dle post types. Figure 2 shows the empirical complementary cumulative distribution function (CCDF) of the number of users, posts, likes, comments, and reshares for all pages. All the five distributions are heavy-tailed. To better visualize the geographical distribution of the Occupy Movement, activity statusphotolinkvideoposts0.00.10.20.30.40.50.6statusphotolinkvideolikes0.00.10.20.30.40.50.6statusphotolinkvideocomments0.00.10.20.30.40.50.6Figure 2: Empirical CCDF for number of users, posts, likes, comments, and shares on pages. Taking all the users (purple), posts (light blue), likes (red), comments (sky blue), and shares (pink) on each page we plotted the CCDF of the five quantities and we got a power law distribution for all the measures, with similar shapes and reasonable scale dif- ferences. patterns are shown on the US map (Figure 3). Notice that there exists a high correlation between all pairs of measures -- i.e., users, posts, likes, comments, and shares. In Figure 4 similar results to those depicted in Figure 3 hold for resharing patterns. 4 Pearson correlation coefficient shows high level of correlation between all considered mea- sures -- i.e., number of users, posts, likes, comments and shares -- on the different pages, as shown in Table 1. Sum- marizing, the activity of the users on the Occupy pages is comparable across pages; number of users, posts, likes, com- ments and shares for each page are highly correlated. Users Posts Likes Comments Shares Users 1 0.974 0.997 0.991 0.739 Posts 0.974 1 0.963 0.987 0.840 Likes Comments 0.997 0.963 0.991 0.987 0.987 1 0.987 0.714 1 0.769 Shares 0.739 0.840 0.714 0.769 1 Table 1: Correlations between the number of users, posts, likes, comments, and shares. Pearson correlation for num- ber of users, posts, likes, comments, and shares is high for all different combinations. There are high correlation values among the five considered measures. 4In order to identify a reshare we need a reference to the previ- ous posts through an Object ID. This is a fundamental information for the reconstruction of resharing patterns, although we had no in- formation about the original posts and no references to the directly reshared posts. Figure 5: Empirical CCDF for likes and comments on posts. The CCDF of the number of likes (red) and comments (sky blue) on the totality of posts, shows power law distribu- tions for the two measures, with a significant difference in the scale. Activity on Posts The CCDF of the number of likes and comments for each post shows power law distributions (Figure 5). Regardless of the page, information can be shared by its administrators, and hence appears as shared by the page it- self, or by common users. Posts shared by page administra- tors -- i.e. admin posts -- are 27% of the total (202K) against the 73% (551K) from common users -- i.e., not admin posts -- and, on a page scale, there is a weak correlation (0.37) between the number of posts made by admin and not admin users (Figure 6). The number of likes to a post from the two categories is power law distributed with a quantitative difference: while the majority of posts is in the not admin category, they get less than 6% of total likes, as shown in Figure 7. Number of posts by each category and number of likes reached are not proportional. This disproportion is due to the fact that, on Facebook public pages, posts by common users become soon hidden, while those from pages admin get higher vis- ibility. Activity on posts is power law distributed. An inter- esting quantitative difference between posts shared by page administrators and common users emerges. Users Activity Our dataset includes 618K active users. A user is considered to be active if she liked at least one post. We identified two main users categories: occasional -- i.e., users that made less than 5 total likes -- and habitual -- i.e., users that made at least 5 total likes. Such a preliminary distinction, coherently with the heavy- tailed distributions of the users activities, counts 183K ha- bitual users (about 30% of the total). As a further step we llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# users, posts, likes and commentsCCDF10010110210310410510610-210-1100 ****************************************************************************************************************************************************************** luserspostslikescommentsresharesllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# likes and commentsCCDF10010110210310410510-610-510-410-310-210-1100 llikescommentsFigure 3: Geospatial distributions of activities. The number of users (purple), posts (light blue), likes (red), and comments (sky blue) is plotted in the form of dots on the US map. The bigger the dot the higher the number of elements in the corresponding category (on the page corresponding to the geographical spot on the map). We can see how the size of the dots, across the different measures, is comparable for all the geographical points. This is a first empirical evidence of the high correlation among the four considered measures. Figure 4: Resharing patterns. Results similar to the ones in Figure 3 hold for the resharing activity. We computed the number of reshares on a subset representing about the 14% of all posts. llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# of Reshareslll5003,0006,0008,500Figure 6: Posts by admin and not admin. On the x-axis there are the US Occupy pages and on the y-axis the corre- sponding number of posts shared by the page administrators (blue) and by common users (green). On the total of posts, 27% is made by page administrators and 73% by common users. However there is low correlation (0.37) between the two sharing patterns. Figure 8: CCDF of different users' categories on all pages. The CCDF of the number of active users for each categories, computed for every single page, shows how the majority of the liking activity is made by occasional (red) and not polarized (purple) users. Remember that a user is active on a page if she liked at least one post from that page. However the activity of polarized users (blue) is still signifi- cant, while users polarized on other pages (violet) represent a marginal portion of the total likes. define polarized users -- i.e., a habitual user is polarized on a given page if she has at least 95% of her likes on that page. Such a thresholding classification detects about 128K polar- ized users (21% of the total), that produced 3.2M likes and 198K comments -- i.e., respectively the 62% and 15% of the overall likes and comments. Figure 8 shows the CCDF of the number of users from all different categories on a given page -- i.e., occasional, polarized, not polarized, and polarized on a different page -- , they all show a power law decay. Moreover, the size of polarized communities, for each page, is shown on the US map in Figure 9. Regarding the activity of polarized users we observe a power law distribution for both number of likes and com- ments (Figure 10). Notice that only about 21% of polar- ized users commented at least once. The Pearson correlation coefficient computed on the number of polarized users, to- tal users, posts, likes, comments, and shares, provides high correlations levels for all the combinations of measures, as shown in Table 2. Our findings, coherently with the heavy-tailed distribu- tions of users activities, show a dominant role of occasional and not polarized users in the information consumption and a differentiation of the probability of liking and commenting for polarized users depending on their level of activity. Figure 7: Likes activity on posts by admin, not admin, and all users. The CCDF for the number of likes reached by all users posts (red), admin posts (blue), and not admin posts (green) shows a significantly bigger probability of high number of likes for the admin posts rather than not admin posts. Considering that just about one third of all the posts is shared by page administrators, we got an evident dispropor- tion, that may be explained by the higher visibility reserved to admin posts. lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllpages# posts100101102103104105106llPosts by AdminPosts by Not Admin Correlation = 0.37llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# likesCCDF10010110210310410-510-410-310-210-1100lalladminnot adminlllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# of Users for each CategoryCCDF10010110210310410510-210-1100lOccasionalPolarizedNot−PolarizedOtherFigure 9: Geospatial distributions of the polarized users on pages. Representation of polarized users' communities for all pages: the size of the dots is proportional to the size of the communities and emphasizes the correlation between the number of polarized users and the number of users, posts, likes, comments, and reshares for each page (see also Figure 3 and 4). Polarized Users Posts Likes Comments Shares Polarized 1 0.993 0.948 0.998 0.978 0.675 Users 0.993 1 0.974 0.997 0.991 0.739 Posts 0.948 0.974 1 0.963 0.987 0.840 Likes 0.998 0.997 0.963 1 0.987 0.714 Comments 0.978 0.991 0.987 0.987 1 0.769 Shares 0.675 0.739 0.840 0.714 0.769 1 Table 2: Correlations among number of polarized users, total users, posts, likes, comments, and reshares. Pearson correlation coefficient measures high levels of correlation within the considered elements -- i.e. number of total users, polarized users, posts, likes, comments, and reshares. Backbone of Interaction Patterns We want to understand if the geographical affiliation of pages or polarized users in the Occupy movement influ- ence the diffusion of information. Hence, we have applied a Backbone Extraction Algorithm [33] to two different net- works, as mentioned before. Figure 11 shows on the US map the results for the backbone extraction applied to the pages-reshares network, for two different levels of signifi- cance α = {0.01, 0.05}. These results provide a clear im- age of the absence of geographical correlation in the reshar- ing patterns. Moreover, pages corresponding to the major US cities5 emerge as leaders in the information spreading. Figure 12 illustrates the multi-scale backbone structure for the pages-common users network, for the two levels of sig- nificance α = {0.01, 0.05}. The links correspond to the ac- tivity of users polarized on one page inside another page. Again, no geographic correlation is present, and five main cities showing an exchange of polarized users' activity may be identified: New York, Los Angeles, Boston, Portland and San Francisco. Hubs corresponding the US major cities drive the overall 5For both levels of significance, the following cities emerge as information spreading leaders: New York, Los Angeles, Chicago, Boston, Portland, Phoenix and Denver. Figure 10: Activity patterns for polarized users. The CCDF of liking (red) and commenting (sky blue) activity for polarized users shows again power law distributions. Com- paring it with Figure 5, where the number of likes and com- ments are computed on all users, we can notice that the quantitative difference between red and sky blue curves is smaller and that, when a level of about 1000 likes (com- ments) is reached, the trend starts to revert leading to a higher probability of commenting rather than liking (for po- larized users). llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# of Polarizedll10,00035,00070,000llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll# likes and commentsCCDF10010110210310410-510-410-310-210-1100 llikescomments(a) α = 0.01 (a) α = 0.01 (b) α = 0.05 (b) α = 0.05 Figure 11: Backbone structure for the pages-reshares net- work, for α = {0.01, 0.05}. The Backbone Extraction Al- gorithm, applied to the pages-reshares network, preserves 26.58% of the total edge weight for α = 0.01 and 42.7% for α = 0.05. We can notice the absence of geographical cor- relation in the information spreading and the emergence of US major cities -- New York, Los Angeles, Chicago, Boston, Portland, Phoenix and Denver -- as information diffusion leaders. Figure 12: Backbone structure for the pages-common users network, for α = {0.01, 0.05}. The Backbone Ex- traction Algorithm, applied to the pages-common users net- work, preserves 64.95% of the total edge weight for α = 0.01 and 70.65% for α = 0.05. We can notice the absence of geographical correlation in the activity of polarized users across pages and the emergence of US major cities -- New York, Los Angeles, Boston, Portland and San Francisco -- as polarized users' activity exchanger. activity of the movement, the diffusion of contents online and serve as coordination points for all other pages. Conclusion In this paper we explore the case of online political move- ments, that coordinate and interact through social media even more often with the advent of the World Wide Web [5]. Our focus is to characterize information consumption patterns by identifying different actors according to their interaction patterns with pages. As the Occupy movement online presents a geographical diversification of groups we address geographical patterns behind information diffusion. Taking into account all the posts from 179 US public Face- book pages about the Occupy Movement during the time span September 2011-February 2013, we analyze users ac- tivity on pages and posts. We find high levels of correlation for the number of users, posts, likes, comments, and reshares on each page (that are all power law distributed) and differ- ent activity patterns for posts with different post types or shared by users from different categories. Then we divide the users in two categories, habitual and occasional, accord- ing to the total number of likes they made, and we further label habitual users with at least 95% of their liking activity on one particular page as polarized in that page. We find that the number of polarized users for each page is positively cor- related with all the measures mentioned above and that the most of the liking activity is performed by occasional or not polarized users. Moreover, we extract the multi-scale back- bone for two networks, the pages-reshares network and the pages-common users network, in order to analyze geograph- ical patterns in the information diffusion and polarized users activity. Our analysis reveals that activities online are not locally coordinated by geographically close pages. Indeed, pages linked to major US cities -- e.g., New York, Los Ange- les, Chicago, Boston, San Francisco -- drive the diffusion of contents online and serve as coordination points for all other pages, which perform a minor activity in the system. Acknowledgments Funding for this work was provided by EU FET project MULTIPLEX nr. 317532 and SIMPOL nr. 610704. The fun- ders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. References [1] Delia Mocanu, Luca Rossi, Qian Zhang, M`arton Kar- sai, and Walter Quattrociocchi. Collective attention in the age of (mis) information. Computers in Human Behavior, Elsevier, 2014. [2] Alessandro Bessi, Mauro Coletto, George Alexandru Davidescu, Antonio Scala, Guido Caldarelli, and Wal- ter Quattrociocchi. Science vs conspiracy: collective narratives in the age of (mis) information. Plos One, 2014. [3] Aris Anagnostopoulos, Alessandro Bessi, Guido Cal- darelli, Michela Del Vicario, Fabio Petroni, Antonio Scala, Fabiana Zollo, and Walter Quattrociocchi. Viral misinformation: the role of homophily and polariza- tion. arXiv preprint arXiv:1411.2893, 2014. [4] Adrien Friggeri, Lada A Adamic, Dean Eckles, and Justin Cheng. Rumor cascades. ICWSM, 2014. [5] Martha McCaughey and Michael D Ayers. Cyberac- tivism: Online activism in theory and practice. Psy- chology Press, 2003. [6] R. Kelly Garrett. Protest in an information society: a review of literature on social movements and new icts. Information, Communication and Society, 9(2):202 -- 224, 2006. [7] Winston Bennett. Communicating global activism. In- formation, Communication & Society, 6(2):143 -- 168, 2003. [8] Daniel J Myers. Communication technology and social movements: Contributions of computer networks to ac- tivism. Social Science Computer Review, 12(2):250 -- 260, 1994. [9] Noam Chomsky. Occupy. Occupied Media Pamphlet Series, 1, 2012. [10] J Byrne. Occupy the media: Journalism for (and by) the 99 percent. The Occupy Handbook. Little, Brown, pages 256 -- 264, 2012. [11] Michael D. Conover, Clayton Davis, Emilio Ferrara, Karissa McKelvey, Filippo Menczer, and Alessandro Flammini. The geospatial characteristics of a so- cial movement communication network. PLoS ONE, 8(3):e55957, 03 2013. [12] Michael D Conover, Emilio Ferrara, Filippo Menczer, and Alessandro Flammini. The digital evolution of oc- cupy wall street. PloS one, 8(5):e64679, 2013. [13] Floriana Gargiulo, Jacopo Bindi, and Andrea Apolloni. The topology of a discussion: the# occupy case. arXiv preprint arXiv:1412.4639, 2014. [14] Kjerstin Thorson, Kevin Driscoll, Brian Ekdale, Stephanie Edgerly, Liana Gamber Thompson, Andrew Schrock, Lana Swartz, Emily K Vraga, and Chris Wells. Youtube, twitter and the occupy movement: Connecting content and circulation practices. Informa- tion, Communication & Society, 16(3):421 -- 451, 2013. [15] Zeynep Tufekci and Christopher Wilson. Social me- dia and the decision to participate in political protest: Observations from tahrir square. Journal of Communi- cation, 62(2):363 -- 379, 2012. [16] Robert D Benford and David A Snow. Framing pro- cesses and social movements: An overview and as- sessment. Annual review of sociology, pages 611 -- 639, 2000. [17] Neal Caren and Sarah Gaby. Occupy online: Face- book and the spread of occupy wall street. Social Sci- ence Research Network Working Paper Series, October 2011. [18] Sarah Gaby and Neal Caren. Occupy online: How cute old men and malcolm x recruited 400,000 us users [34] Reuven Dukas and Alan C Kamil. Limited attention: the constraint underlying search image. Behavioral Ecology, 12(2):192 -- 199, 2001. [35] Lilian Weng, Alessandro Flammini, Alessandro Vespignani, and Filippo Menczer. Competition among memes in a world with limited attention. Scientific Reports, 2, 2012. [36] Alessandro Bessi, Antonio Scala, Luca Rossi, Qian Zhang, and Walter Quattrociocchi. The economy of attention in the age of (mis) information. Journal of Trust Management, 2014. to ows on facebook. Social Movement Studies, 11(3- 4):367 -- 374, 2012. [19] Jeroen Van Laer and Peter Van Aelst. Cyber-protest and civil society: the internet and action repertoires in social movements. Handbook on internet crime, pages 230 -- 254, 2009. [20] Stefan Wray. On electronic civil disobedience. Peace Review, 11(1):107 -- 111, 1999. [21] W Lance Bennett. Changing citizenship in the digital age. Civic life online: Learning how digital media can engage youth, 1:1 -- 24, 2008. [22] Devin Gaffney. #iranelection: quantifying online ac- tivism. Web Science Conf., 2010. [23] John D McCarthy and Mayer N Zald. Resource mobi- lization and social movements: A partial theory. Amer- ican journal of sociology, pages 1212 -- 1241, 1977. [24] Daniel J Myers. Media, communication technology, and protest waves. In Trabalho apresentado na con- ferencia Social Movement Analysis: The Network Per- spective, Ross Priory, Scotland, pages 22 -- 25, 2000. [25] Gilad Lotan, Erhardt Graeff, Mike Ananny, Devin Gaffney, Ian Pearce, and Danah Boyd. The revolu- tions were tweeted: Information flows during the 2011 tunisian and egyptian revolutions. International Jour- nal of Communications, 5:1375 -- 1405, 2011. [26] Alok Choudhary, William Hendrix, Kathy Lee, Diana Palsetia, and Wei-Keng Liao. Social media evolution of the egyptian revolution. Communications of the ACM, 55(5):74 -- 80, 2012. [27] Sandra Gonz´alez-Bail´on, Javier Borge-Holthoefer, Alejandro Rivero, and Yamir Moreno. The dynamics of protest recruitment through an online network. Sci- entific reports, 1, 2011. [28] Navid Hassanpour. Media disruption exacerbates rev- olutionary unrest: Evidence from mubarak's natural experiment. American Political Science Association (APSA), Annual Meeting Paper, 2011. [29] Philip N Howard, Aiden Duffy, Deen Freelon, Muzam- mil Hussain, Will Mari, and Marwa Mazaid. Opening closed regimes: what was the role of social media dur- ing the Arab Spring? Working Paper, PIPTI, 2011. [30] Habibul Haque Khondker. Role of the new media in the arab spring. Globalizations, 8(5):675 -- 679, 2011. [31] Doug McAdam and Ronnelle Paulsen. Specifying the relationship between social ties and activism. Ameri- can journal of sociology, pages 640 -- 667, 1993. [32] Facebook. Using the graph api. Website, 8 2013. last checked: 19.01.2014. [33] M ´Angeles Serrano, Mari´an Bogun´a, and Alessandro Vespignani. Extracting the multiscale backbone of complex weighted networks. Proceedings of the na- tional academy of sciences, 106(16):6483 -- 6488, 2009.
1202.0617	2	1202	2012-02-17T05:14:17	Classification of Flames in Computer Mediated Communications	[ "cs.SI", "cs.CL" ]	Computer Mediated Communication (CMC) has brought about a revolution in the way the world communicates with each other. With the increasing number of people, interacting through the internet and the rise of new platforms and technologies has brought together the people from different social, cultural and geographical backgrounds to present their thoughts, ideas and opinions on topics of their interest. CMC has, in some cases, gave users more freedom to express themselves as compared to Face-to-face communication. This has also led to rise in the use of hostile and aggressive language and terminologies uninhibitedly. Since such use of language is detrimental to the discussion process and affects the audience and individuals negatively, efforts are being taken to control them. The research sees the need to understand the concept of flaming and hence attempts to classify them in order to give a better understanding of it. The classification is done on the basis of type of flame content being presented and the Style in which they are presented.	cs.SI	cs	Classification of Flames in Computer Mediated Communications Nitin Department of Computer Science, The Peter Kiewit Institute, College of Information Science and Technology, University of Nebraska at Omaha, Omaha-68182-0116, Nebraska, USA [email protected] ABSTRACT Computer Mediated Communication (CMC) has brought about a revolution in the way the world communicates with each other. With the increasing number of people, interacting through the internet and the rise of new platforms and technologies has brought together the people from different social, cultural and geographical backgrounds to present their thoughts, ideas and opinions on topics of their interest. CMC has, in some cases, gave users more freedom to express themselves as compared to Face-to-face communication. This has also led to rise in the use of hostile and aggressive language and terminologies uninhibitedly. Since such use of language is detrimental to the discussion process and affects the audience and individuals negatively, efforts are being taken to control them. The research sees the need to understand the concept of flaming and hence attempts to classify them in order to give a better understanding of it. The classification is done on the basis of type of flame content being presented and the Style in which they are presented. General Terms Classification, Computer Mediated Communication, Flames. Keywords indirect flames, straightforward Direct flames, satirical flames, flaming. flames, 1. INTRODUCTION With the advent of computer mediated communications (CMC), human beings have started to communicate with people who are not only their friends, relatives and acquaintances but also with strangers from different social and cultural backgrounds. Information, ideas, thoughts and opinions are exchanged through this medium. The growth in the use of computers to form global networks, which is commonly labeled as the internet, has led to the research being conducted to analyze the benefits and effects of using it as an emerging communication tool. Both modes of internet communication, asynchronous and synchronous, have been able to influence the lives of people around the world. This information is characteristically transmitted to audiences around the world, which shows traits of both written and oral language. The users identified by their ‘nickname’ adopted by them in a CMC and all messages that transmit appear along with this ‘nickname’. At times, the email address of the user is used to identify the user if they have not provided with a desired nickname. Ankush Bansal, Siddhartha Mahadev Sharma, Kapil Kumar, Anuj Aggarwal, Sheenu Goyal, Kanika Choudhary, Kunal Chawla, Kunal Jain and Manav Bhasin Jaypee University of Information Technology, Waknaghat-173234, via Kandaghat, Dumehar Bani, Himachal Pradesh, India [email protected] The most common and widely used asynchronous CMC is the e-mail. It is an electronic adaptation of the real life mailing system where a sender leaves a message or ‘mail’ in a receiver’s electronic mailbox (commonly called ‘Inbox’) and the receiver must open this mailbox and the letter before he can read the messages received [1]. Other more complex and sophisticated example of an asynchronous CMC is ‘Usenet Newsgroup’ which is an electronic pin-up notice board on which users can post messages regarding certain topics or areas of interest. Other users can read and reply to these messages by opening the notice board and post their own messages in turn. Similar to the e-mail system, there is as such, no real timeline between the users interacting. On the other hand, synchronous CMC provides a real timeline of interactions between users’ computers. This is most important feature of synchronous CMC in comparison to asynchronous CMC. The most widespread system, which is an example of synchronous CMC, is Internet Relay chat, or IRC. IRC is a form of CMC, which allows users to ‘Chat’ by exchanging written messages. They interact in two different ways: By sending their intended messages to a specified user or to all members of the group who are online. CMC is being largely used by small and large organizations for varied activities such as "group problem solving and forecasting, consensus development, coordination and operation of group projects, sharing ideas and information, and mobilizing organizational action within special forums, interest groups or workforces" [2]. It has been observed that these discussions and conversations sometimes lead to altercations, which are not conducive for the well being of users or organizations per se. Such altercations are also reported in CMC, which has distinct characteristics like that of real life altercations. Such venues of CMC interactions can include examples like forums, blogs, UseNet, IRC chat, micro blogging sites and social networking sites which allow people to come together and discuss on the topics of their interest. As people started to get more acquainted with CMC, they started to use these heated conversation for their own benefit. Whenever people find that the discussion (mainly debates and arguments) is not going in their favor or they are about to be proved wrong , in that case people use flaming to make the conversation go their way [3]. The use of hostile language in CMC has been reported in most chat forms of synchronous and asynchronous forms of communications. Earlier studies and meta-analysis showed that instances of offensive conduct are highly exaggerated [4] but present scenario of CMC to quite an extent is characterized by intense language, swearing, and hostile communication [3]. 2. FLAMING The term “flaming” originates from the early computing community, and The Hacker’s Dictionary [5] defines it as “to speak rabidly or incessantly on an uninteresting topic or with a patently ridiculous attitude”. Flaming is described as anti-normative hostile and insulting interaction between users [6]. O'Sullivan and Flanagin (2003) define flaming as "a concept emerged from popular discourse surrounding the online community to describe aggressive, hostile, profanity-laced interactions"[7]. Landry (2000) refers to the phenomenon as "uninhibited and aggressive communication" [8]. In general, flaming means to attack with an intention to offend someone through e-mail, posting, commenting or any statement using insults, swearing and hostile, intense language, trolling, etc.[9] . However, this definition is not considered to be completely appropriate as there are many incidences found where people do not use insulting words or aggressive language directly. Users make use of tools of language such sarcasm, references and figures of speech to flame other user(s) [10, 11]. According to Aiken and Waller, flaming constitutes of "comments intended to offend others. While it is subjective in certain instances, extreme cases of flaming include obscenities and other inappropriate comments." [12] The term ‘Flaming’ has also been equated with ‘disinhibited’ behavior, which is in fact a theorized cause of the action the behavior (flaming), rather itself [13]. Some than researchers have specifically included words such as “electronically” in its definition and used “Computer Mediated” to explain it [14]. Though the term has been adopted from CMC, it has been debatable that the defining flaming only as an online phenomenon is assuming it as technological determinism from a few points of view, confusing the behavior with its theorized causes [7, 13 and 15]. In that respect, several studies have compared flaming on CMC to Face-to-Face communication and have tried to identify the similarities. While some research studies argue that flaming is more apparent in CMC [2], others claimed flaming to be less frequent or rare in both situations [16]. These studies are only considered relevant if flaming by the understood definition is not an online phenomenon [17].The term “flaming” is mainly used in electronic contexts and rarely in non-electronic ones, example the classroom [18]. Flaming has different implication in different scenarios as it has been seen that sometimes the user who resorts to flaming has some advantages whereas in many other case studies the user being flamed has the distinct advantage [19]. Many a times it has been observed that, a user for their redirecting the argument or for forcing one’s opinion uses the flaming intentionally. Flaming is used deliberately by the flame sender as a means of diverting the other factions from the original discussion, by sending flames so that they can use it for their own benefit in an attempt to agitate and make the other factions change from present topic of discussion or to remain on a certain topic or point which is preferred by the flame sender. The essence of this topic is that flaming is a very real phenomenon and to some people, it is even an actual problem. There are reported cases where several distinguished individuals have terminated or abandoned maintaining their weblogs (Online diaries that is open for net users to read and comment on), due to excessive negative or hateful feedback they received on their weblogs [20]. Comments and reader messages to online newspaper and magazine articles have also been strongly criticized for been unnecessarily rude and unparliamentary [21]. Some research suggests the law to provide for the protection of net users against flaming and other misuses of the anonymity inherent on the internet [22, 23]. Some network groups have taken a step in this direction by establishing netiquette – rules of network usage – which specially addresses users on how they can write and post their messages. These rules emphasize on obligations for group and self-monitoring to ensure decorum to be maintained throughout discussions by use of correct language, respect for other users and communicative relevance. These netiquette rules differ on different sites. Some offenses are likely to be considered more hostile than others and there is a possibility that what certain groups condemns, others are indifferent to or condones. For example, while posting a certain hostile message directed to another user, it is acceptable on some newsgroups (Examples: alt.fan.warlords and alt.flame), whereas this is usually not allowed and censured by social oriented newsgroups. 3. COMPUTER MEDIATED COMMUNICATION VS. FACE-TO- FACE COMMUNICATION Although, According to researchers, flaming is generally more in computer mediated communication as compare to the face-to-face communication, but such Incidences of flaming are quite low in both [13]. In the same line, initial research comparing CMC and face-to-face communication showed that ‘‘flaming’’ and other antisocial interactions were more common in electronic interactions [14]. The reason behind such hostile behavior lies within the fact that in computer mediated communication the identity of the users in most cases remains hidden and salient. This anonymity associated with computer mediation allows the users to be more expressive and less concerned about the results of their actions [24]. Another reason for increased cases of flaming in CMC is the factions that are involved in sending and receiving flames do not share eye contact or physical presence during communication [1]. Hence, much of the expressions, gestures and body language that might normally be present in face-to- face conversations is lost in CMC interactions. It has been documented face-to-face though that communication and CMC can both be used to discuss any topic; there are marked variations due to the medium of communication. Physical presence plays as an important aspect as an inhibitor to volatile behaviour and as an incentive to be acceptably assertive in the very least. Such “conversations occur in a cooperative environment constantly regulated by mutual adjustment and correction.” [25, 26] 4. MOTIVATION The research saw the need to classify flames as it has not been attempted before on the basis of type of content presented and style of language it is presented in. Upon classification, the research aims to help provide with a better understanding of the flaming process and hence would give a better idea for developers to design efficient solutions to the subject matter. 5. CLASSIFICATION OF FLAMES Up until now, much research has been conducted in the field of flaming for the last 25 years but, there have been little attempts to classify flames. The study categorizes on the basis
1706.04110	1	1706	2017-06-13T14:59:53	Compressing networks with super nodes	[ "cs.SI", "physics.soc-ph" ]	Community detection is a commonly used technique for identifying groups in a network based on similarities in connectivity patterns. To facilitate community detection in large networks, we recast the network to be partitioned into a smaller network of 'super nodes', each super node comprising one or more nodes in the original network. To define the seeds of our super nodes, we apply the 'CoreHD' ranking from dismantling and decycling. We test our approach through the analysis of two common methods for community detection: modularity maximization with the Louvain algorithm and maximum likelihood optimization for fitting a stochastic block model. Our results highlight that applying community detection to the compressed network of super nodes is significantly faster while successfully producing partitions that are more aligned with the local network connectivity, more stable across multiple (stochastic) runs within and between community detection algorithms, and overlap well with the results obtained using the full network.	cs.SI	cs	Compressing Networks with Super Nodes Natalie Stanley1 Roland Kwitt2 Marc Niethammer1 Peter J. Mucha1 1University of North Carolina, Chapel Hill 2University of Salzburg June 14, 2017 Abstract Community detection is a commonly used technique for identifying groups in a network based on similarities in connectivity patterns. To facilitate community detection in large networks, we recast the network to be partitioned into a smaller network of 'super nodes', each super node comprising one or more nodes in the original network. To define the seeds of our super nodes, we apply the 'CoreHD' ranking from dismantling and decycling. We test our approach through the analysis of two common methods for community detection: modularity maximization with the Louvain algorithm and maximum likelihood optimization for fitting a stochastic block model. Our results highlight that applying community detection to the compressed network of super nodes is significantly faster while successfully producing partitions that are more aligned with the local network connectivity, more stable across multiple (stochastic) runs within and between community detection algorithms, and overlap well with the results obtained using the full network. Introduction 1 Networks appear across disciplines as natural data structures for modeling relations between entities, such as regulatory interactions between genes and proteins, and social connections between people. In practice, most networks are large and have intricate substructures that can be interrogated for further insights about the underlying data. A common practice for extracting subgraphs of interest is community detection [1, 2, 3], which aims to partition network nodes into groups based on the group-level connectivity patterns. For example, assortative communities are typically defined through some measure of a greater weight of within-group edges compared to those between groups. Communities can be identified through various optimization problems, including likelihood maximization [4, 5, 6] and quality function optimization [7, 8, 9], or from spectral properties of associated matrices [10]. In this paper, we specifically examine how the compressed version of a network can be handled by modularity maximization and likelihood maximization. To maximize modularity in a network with N nodes, one seeks to find the node-to-community assignment vector, z = [z1, z2, . . . zN ] that maximizes modularity, Q, defined by [aij − γ kikj 2M i,j Q = ]δ(zi, zj), where aij encodes presence and weight of a possible edge between i and j, ki =(cid:80) j aij gives the strength of node i, δ(zi, zj) = 1 if nodes i and j have the same community assignment and 0 otherwise, M denotes the number of edges and γ is the resolution parameter controlling community sizes. Alternatively, to find community structure maximizing likelihood, one can fit a stochastic block model (SBM), which assumes that connectivity patterns of N nodes in a network with K communities can be modeled according to their community memberships through a probability matrix π, where πkl gives the edge connection probability between communities k and l. Assuming no corrections due to node degrees (degree-corrected versions also exist), one seeks to (z) maximize, (1) (cid:88) 1 2M p(a z, π) = kl (1 − π1−aij (πaij kl ))δ(zi=k,zj =l), (2) N(cid:89) K(cid:89) i(cid:54)=j k,l 1 where δ(zi = k, zj = l) is 1 if zi = k and zj = l, and 0 otherwise. One can maximize this objective with the expectation maximization (EM) algorithm, belief propagation, and efficiently with an iterative approach which, similar to Louvain, agglomerates blocks of nodes at each iteration [6]. Specifically, at each iteration a Metropolis-Hastings accept-reject sampling approach is used to compute a probability of merging blocks based on how the merge affects the likelihood. These agglomerative heuristics simplify a computationally challenging task but can still be time consuming for large networks and often give rise to large variability in returned partitions across multiple runs of the algorithms. We seek to explore how a compressed network representation can improve some of these issues. The identification of communities in networks is in some ways similar to multi-label image segmentation, which aims to partition a grid of pixels into contiguous regions corresponding to objects in the image. In this sense, each segmented region can be viewed as a community [11]. To speed up segmentation for large images, a popular approach is to avoid computing segmentations at the pixel level and instead reformulate the segmentation problem based on larger-scale image primitives that are likely part of the same partition. Specifically, this can be accomplished by super pixels that aggregate pixels together in a way that faithfully adheres to image boundaries, maintaining or improving segmentation accuracy [12]. The SLIC super pixel method [12] chooses seed pixels across the image's pixel grid to serve as the super pixel centers and then iteratively grows out and recomputes based on aggregation with neighboring pixels with similar visual features. Motivated by super pixels, we wish to define seed nodes in networks that can be used as a starting point to grow out 'super nodes' to define a new, smaller network upon which we apply standard community detection algorithms. Creating a direct analog of super pixels in networks is challenging because the inherent geometry of a network can be quite different from the grid layout of an image (where simple neighborhood structures such as 4- or 8-neighborhoods are typically used), and we need ensure seeds are well distributed across the network. Further, while super pixels are largely constrained by the structure of the pixel grid (i.e. proximity between pixel pairs matter), their definition also incorporates extra image features to refine members of a super pixel set, whereas in network community detection we typically only have the edges of the network to work with. Finally, the performance of a super pixel representation of an image can be objectively validated from the quality of the corresponding segmentation result, with reference to human-specified objects in images; in contrast, community detection is typically an unsupervised, exploratory data analysis technique with limited available notions of 'ground truth' [13]. As such, we must develop measures that can be used to validate the quality of the super node representation. Our approach in defining a smaller network of super nodes is a form of network compression. A useful approach developed previously is the 'KeepAll' method [14], which seeks to prioritize a set of nodes and retain only the smallest set of additional nodes required for the induced subgraph of prioritized nodes to be connected. Two other network compression methods related to our objective in defining super nodes are explored in Refs. [15, 16]; but these approaches differ from our proposal in that they seek to define super nodes along with additional side information about relationships between node pairs. Ref. [15] describes 'super genomic network compression' to reduce the number of edges in a large protein interaction network. To do this, the authors identify 'clusters of orthologous groups' of proteins, or proteins that give rise to similar functions in different species and originated from a common ancestor. Members of an orthologous group are connected as a star network, with the center node as one member of the orthologous group. Furthermore, edges between orthologous groups are replaced by a single weighted link reflecting the pairwise group evolutionary similarity. Ref. [16] defines super nodes by defining 'must link' and 'cannot link' constraints between pairs of nodes, agglomerating as many nodes as possible sharing must link constraints while being cautious about agglomerating nodes that cannot link. To define the super node representation of an N-node network, we first select S (cid:28) N seed nodes, agglomerate the remaining N − S nodes with the seeds to create super nodes, and specify the network between these super nodes. Community detection can then be applied to the S-node network representation. Figure 1 visualizes this approach, with details provided in Sec. 2. The aims of this work are twofold: First, we seek an effective way to define a super node representation of a network that can then be used in standard community detection algorithms. Second, we wish to establish a principled approach to validate the quality of a super node representation. Our approach to defining super nodes aims to accomplish the following. 1. Decreased runtime for community detection: Even though recently developed heuristics for maximizing modularity [7] and fitting SBMs [6] are highly efficient relative to previous approaches for 2 Figure 1: Defining super nodes. To define the super node representation of a network, we select S seeds and agglomerate local regions around them to create super nodes. This then leads to a new network with weighted edges between the S super nodes upon which community detection can be more efficiently applied. performing the same computational optimizations, these methods can still be time consuming for large networks. We aim to reduce runtime for large networks, moving most of the computational cost in practice from tasks scaling with the size of the network to alternatives scaling with the (much smaller) size of the super node representation. 2. Decreased stochastic variability of community detection algorithm output: In large net- works, there is often significant variability across multiple runs of the same algorithm (employing computational heuristics to solve NP-Complete optimizations), as well as differences between various community detection algorithms. We expect applying community detection to a well-chosen super node representation to decrease the observed variability. 3. High local agreement: In defining super nodes, we inherently assume that the identified communities should agree with the local network connectivity in that members of a neighborhood should be more likely to have the same community assignment, provided that the super nodes were constructed to minimize aggregation across community boundaries. 4. Consistent with communities found using the full network: The identified community structure should be relatively similar to the distributions of results that would have been obtained through applying community detection to the full network. Organization. In section 2, we describe the specific algorithms we employ for choosing seeds, growing these out to super nodes, and specifying the network of super nodes. Our method of selecting seed nodes is a novel application of the CoreHD algorithm [17], previously used for network decycling and dismantling. Section 3 presents experiments used to understand and quantify the quality and usefulness of the super node representation, addressing the four objectives listed above. Finally, in section 4, we discuss other possible applications of super nodes and directions for future work. 2 Method for Generating a Super Node Representation of a Net- work Identify Seeds in the Network We create a super node representation through three steps, outlined in Figure 1. First, we define seeds. Next, we 'grow' super nodes by assigning the remaining nodes to seeds. Finally, the network of super nodes is defined by agglomerating edges, and can then be used in a community detection algorithm. 2.1 To define S seeds, we aim to identify a set of nodes S∗ that are individually central to the network and to their communities, and that are well separated from one another. Such problems are related to influence maximization, where one identifies a small number of nodes in the network from which to effectively spread influence or diffuse messages across the network [18, 19]. The most naive approach is to select nodes with highest degree, and this might be perfectly reasonable under various circumstances. Importantly, the selection of nodes with highest degree is computationally fast, requiring O(M ) operations, summing over the M edges to calculate the degrees of the N nodes. At slightly higher computational cost, we employ the CoreHD algorithm, which nearly optimally identifies nodes in network decycling and dismantling [17]. CoreHD recursively identifies the highest degree node in 3 Input: NetworkGrow out super nodesDe(cid:1)ne SeedsOutput: Super Node NetworkFigure 2: Choosing seeds in a synthetic network. The identification of 20 seeds with the CoreHD algorithm in a network generated from a stochastic block model with 8 communities. Seeds (black nodes) are well distributed across communities. the 2-core (the maximal connected subgraph in which all nodes have degree at least 2 within the subgraph, obtained efficiently by recursively pruning away all nodes of degree less than 2). At each iteration, the removed node is added to our seed set and the now smaller 2-core and their degrees are recomputed. The difference between selecting highest degree nodes and CoreHD for our present task may be small, both in terms of result and computational cost. In particular, because we will only select S (cid:28) N seeds, there is reduced opportunity for the removals to lead to subgraphs with substantial differences between degree order in the graph and its 2-core [cf. selection of O(N ) removals in network dismantling]. Indeed, in our experience, simply selecting the highest degree nodes as the seeds often works well in practice. Because of the minimal extra computational cost for computing the 2-core, we use CoreHD for all of our results shown here. In Figure 2, we plot a sample network generated from the stochastic block model, with nodes colored according to community assignments and black nodes indicating seeds, showing the seeds are distributed across all 8 communities. In supplementary figures 2 and 3, we compare the communities obtained selecting seeds by different methods with the full network communities. 2.2 Grow Super Nodes Around Seeds Once the set of seeds, S∗, is defined, we 'grow' them out agglomerating nearby nodes to build the super nodes. We formally define a super node as a subset of one or more nodes from the original network, X . To do this, seeds are grown out to engulf nodes in increasing neighborhood orders until either all nodes are assigned to a super node center or until a user-defined number of neighborhood orders has been considered. The maximum order, omax, can be specified to control the maximum order neighborhood to consider in building the super nodes. If after omax, there are still unassigned nodes, the unassigned nodes are not used to build the new super node network and are all ultimately assigned to the same periphery community as they are not considered relevant to the network core. Depending on the number of chosen super nodes, S, the degree distribution of the original network and the quality of the chosen seeds at collapsing the network, different networks will require repeating the agglomeration process for different neighborhood orders if one wants to ensure every node is assigned to a super node. The output is a vector, s of length N, which gives the node-to-supernode assignments for the nodes in the original network. 2.3 Create Network of Super Nodes Finally, after growing the super nodes, we must create a new network representation of the super nodes. To do this, we create a weighted network, W, where each super node is a node and the weight of the edge between a pair of distinct super nodes is the total weight of edges in the original network X between pairs of nodes assigned to the respective super nodes. For pairs of super nodes whose members have no edges between them in X , there is no corresponding edge in W. By definition, we construct W with no self loops. After applying community detection to the super nodes, their community assignments are mapped back to their constituent N nodes of the original network, X . In section 3, we use this final N-length matched community assignment, which we denote as z. We consider the node-to-community assignment zF ull obtained by applying community detection to the full network and the mapped result zSN obtained by applying community detection to the super node representation. 4 Project Dblp, Email, (Enron, Amazon, database1 3 Results To demonstrate the effectiveness of super nodes, we performed experiments to analyze the runtime, partition variability, alignment of communities with local network connectivity, and agreement with the communities ob- tained from the full network. We considered 9 network data sets (see Table 1) from the Stanford Network Analy- sis BrightKite, Stanford, Notre Dame) and Newman's collection2 (As22, CMatter). We treat all networks as undi- rected. To explore a range of network sizes, in some of these cases we used large subgraphs defined by the union of all nodes of degree ≥ 2, their neighbors, and next nearest neighbors. We use the Louvain algorithm3 for modularity maximization [7] and the stochastic block model (SBM) inference4 described in [6]. In all experiments, we created the super node representation with S = 500 super nodes. 3.1 Objectively Comparing Partitions on Possibly Different Scales A challenge in directly comparing the community partitions on the full and super node network representations is the difference in scales between the partitions. For example, using the full network typically produces significantly more communities than under the super node representation. In an attempt to compare community partitions with similar size distributions in the subsequent experiments, we can adapt the scales obtained from the Louvain algorithm and SBM fitting. # Nodes # Edges 22,801 48,270 178,195 32,374 17,816 83,337 639,330 150,801 209,887 77,463 420,045 265,214 2,312,497 281,903 325,729 1,497,134 214,078 58,228 Table 1: Network data characteristics. CMatter* (Condensed matter 2003 collab.) Notre Dame (web-Notre Dame) BrightKite (loc-BrightKite) Dataset (* indicates subgraphs) Dblp* (com-DBLP) Amazon* (com-Amazon) Email (email-EuAll) Stanford (web-Stanford) As22* (Internet) Enron* In community detection with the Louvain algorithm, we identified comparable resolution parameters (controlling community size) to apply to the full network that would produce a size distribution agreeing as much as possible with the community partition in the super node network. Further, we compute all following experiments using both the default resolution parameter and the 'matched' parameter. While the default resolution parameter is γ = 1, in our analyses we computed partitions of the full network using several different γ ∈ [0.05, 2.5]. To choose the matched resolution parameter on the full network, we first find the community partition using the super node representation. For each partition, we then rank nodes based on the sizes of the communities to which they belong. For each partition of the full network (at different resolution parameters), we then consider the similarity of this ranking with that from the super node communities, measuring this similarity by Kendall's tau correlation. We identify the resolution parameter producing the highest Kendall's tau correlation, referring to this resolution parameter as the 'matched parameter' in the remainder of the text, while we refer to the standard γ = 1 as the 'default' resolution parameter. In fitting SBMs, we chose to fit a model with the same number of blocks that was found in the super node representation using the standard optimization and model selection strategies discussed in [6]. We refer to the 'matched' version as that using the number of blocks identified by the model selection on the super node representation, while the 'default' result is obtained using the model selection strategy on the full network. In subsequent experiments, we compare both the 'matched' and 'default' versions to ensure our results are not artificially influenced by the scale of the community sizes. 3.2 Experiments For each of the 9 networks, we recorded the runtime required to identify communities with the Louvain algorithm and stochastic block model inference procedure under the full and super node representations (Figure 3). The Louvain algorithm is fast and scales well, at O(M ) per iteration for M edges, with its relative speed and high modularity values contributing to its popularity. While the reported runtimes may seem quite modest, in practice it is common to run many realizations of the algorithm (hundreds, thousands, or even more for large networks) to explore resolution parameters and stochastic variation due to node order in 1https://snap.stanford.edu/data 2http://www-personal.umich.edu/~mejn/netdata/ 3as implemented in https://github.com/vtraag/louvain-igraph 4as implemented in https://graph-tool.skewed.de/ 5 Figure 3: Runtimes. We compare community detection runtimes (in seconds) with the Louvain algorithm and by fitting an SBM on the full networks and super node representations for the 9 data sets. (A) Louvain on the full network. (B) Louvain on the super nodes. (C) SBM on the full network. (D) SBM on the super nodes. the heuristic. We note a large increase in runtime for the full Stanford network, with over 2 million edges. As also observed in the figure, fitting a stochastic block model, at O(N ln2N ) for sparse networks in this implementation [6], becomes significantly slower on the full networks with more than 200,000 edges. While we see a significant improvement in community detection runtime from using super nodes for both methods, the benefit in the SBM fitting is particularly large, especially for the bigger networks (DBLP, Stanford, Email). In moving to the super node representation, we traded out large-coefficient scaling-with-N community detection computations for those scaling with S (cid:28) N (with possible increases due to the increased density of the super node representation), at the cost of constructing the super node representation. In particular, we observe the SBM runtimes on super nodes appear to be relatively independent of M. We note that each of the three steps building our super-node representation is O(M ), so in the large graph limit the expected gain of our approach may be only a constant factor over Louvain iterations and the SBM fitting (up to logarithmic factors). In the present calculations, we have not endeavored to optimize the runtime to build our super node representations; even so, the three steps building the S = 500 super node representation of the Stanford network in our current implementation together take ∼ 350 sec with CoreHD and ∼ 200 sec using highest degree nodes. While this alone might not seem like a large improvement compared to a single realization of running Louvain or fitting an SBM, the computational gains compared to generating multiple community partitions can be quite significant. Next, we used normalized mutual information (NMI) [20] to study similarities in partitions within and between community detection algorithms and between the full and super node representations. With the Louvain algorithm and with fitting SBMs according to the inference described in Ref. [6], sampling and non-convexity in the optimization landscape lead to large variability in partitions. In Figure 4 we computed normalized mutual information (NMI) between all pairs of partitions within and between multiple runs of Louvain and SBM fitting. A high normalized mutual information between a pair of partitions indicates that the algorithms identified similar community structures. The distribution of these pairwise NMI values for the full network and super node representation are visualized with box plots in Figure 4A and B, respectively. For each network, the purple, orange and maroon boxes show the distribution of pairwise NMIs between 10 runs of the Louvain, SBM fitting, and all individual runs of Louvain against all runs of SBM fitting, respectively. The most significant improvement we observe under the super node representation is between runs of the Louvain algorithm and SBM fitting, suggesting that the new compressed representation of the network has prominent structural features that are robustly identified with both approaches. The Louvain algorithm is generally less variable than fitting an SBM, but we also observed decreased variability in the fitting of stochastic block models on the super node representation. While we see a reduction in variability in partitions on the super node representation, we also quantified the similarity in partitions on the full network under default parameters to those on the super node network by computing the NMI between their respective partitions, NMI(zF ull, zSN ). In Figure 4, we indicated the mean pairwise NMI between a full network and 5 super node partitions (again with S = 500 super nodes) with a star. That is, a star in the Louvain (SBM) column indicates the mean pairwise NMI computed from NMI(zF ull, zSN ) using Louvain (SBM). As previously mentioned, a challenge in objective evaluation of partition similarity between representations is the difference in community size distributions. As a result, it is quite difficult to achieve a high NMI using default parameters. However, even these NMIs indicate structural similarity between partitions. Given the high variability within and between community detection algorithms indicated by moderate-to-low NMI values, one does not lose significantly more information by simply using 6 A.B.C.D.0100200300400As22CMatterEnronAmazonDblpBrightKiteEmailNotreDameStanfordFull NetworkRuntime (secs)Louvain0510As22CMatterEnronAmazonDblpBrightKiteEmailNotreDameStanfordSN NetworkRuntime (secs)02505007501000As22CMatterEnronAmazonDblpBrightKiteEmailNotreDameStanfordFull NetworkRuntime (secs)SBM0204060As22CMatterEnronAmazonDblpBrightKiteEmailNotreDameStanfordSN NetworkRuntime (secs)Figure 4: Quantifying partition variability. For each of the 9 networks, we obtained 10 different partitions using Louvain and fitting SBMs under default settings. We computed pairwise normalized mutual information (NMI) between partitions within and between algorithms. For each network, the purple, orange and maroon boxplots (and jittered points) show the pairwise NMIs between the 10 runs of Louvain, the 10 SBM fits, and each individual run of Louvain against the SBM fitting (Louvain + SBM). (A) Within- and between-algorithm variability on the full networks. Stars in the Louvain and SBM column indicate the mean pairwise NMI computed between 5 super node partitions and a full network representation under Louvain and SBM, respectively [i.e., NMI(zF ull, zSN ) for multiple super node representations. (B) Within- and between-algorithm variability on the super node representations. Overall, we observe decreased variability within and between runs of community detection with the super node representations compared to the full networks, while in the context of this general variability the super node results are comparable to the 'default' communities from the full networks. the compressed super node network representation, even without matching scales. While these experiments are performed with S = 500, we study the effects from using other S values in supplementary tables 1 and 2. While we have emphasized benefits in the mechanics and usability of running standard community detection algorithms, we now seek to address whether the communities that we find using the super node representation align with local network connectivity so that neighbors are more likely to have similar community assignments and how this alignment compares with what we would have found by community detection on the full network. While we visualize this qualitatively for the As22 network in Figure 5E-H, we also designed a prediction task to quantify this alignment. In this prediction task, we seek to take a node-to-community partition (from either the full or super node network representations zF ull or zSN, respectively) and the full network X to see how accurately we can predict members of a community for different neighborhood sizes. For network X with node-to-community assignments z, we assign a probability distribution to each node over all of the communities under z. For a neighborhood order o (x-axis in Figure 5), we say that node i has probability of being in community k, based on what fraction of its neighbors belong to that community under z. Then for each community in z, we perform a binary prediction task for whether each node of X should be assigned to that community, according to the computed probability distributions for all nodes with respect to that community. We sweep the probability parameter, p, representing the required threshold probability for a node to achieve in order to be assigned to a community in this binary classification task. By sweeping p for each of the communities, we compute an ROC curve for each community and the corresponding areas under the curve (AUC). Finally, we use the minimum AUC value as our summary statistic of this task, with a high AUC value indicating that the neighboring regions of a node were strong predictors of community assignments, as shown in Figure 5A-D. All experiments are performed on 5 networks (As22, Enron, CMatter, Dblp, Amazon) and for both the matched and default parameters (indicated by line type) in the full network. (Recall from section 3.1 that the matched parameters for the full network were chosen based on the super node partition results under default settings; hence, there is no corresponding 'matched' set for the super nodes in these plots.) We observe in most cases using the super node representation improves the minimum AUC value, indicating that communities obtained from this representation have higher agreement with local connectivity by this measure. To qualitatively evaluate how the super node representation is able to ultimately partition 7 ●●●●**●●●●●●●●●●●****StanfordNotreDameBrightKiteDblpAmazonEmailAs22EnronCMatterLouvainLouvain+SBMSBMLouvainLouvain+SBMSBMLouvainLouvain+SBMSBM0.000.250.500.751.000.000.250.500.751.000.000.250.500.751.00NMIFull Network●●●●●●●●●●StanfordNotreDameBrightKiteDblpAmazonEmailAs22EnronCMatterLouvainLouvain+SBMSBMLouvainLouvain+SBMSBMLouvainLouvain+SBMSBM0.000.250.500.751.000.000.250.500.751.000.000.250.500.751.00MethodNMISuper Node NetworkA.B.SBMLouvainLouvain+SBMSBMMethodLouvainLouvain+SBMSBMFigure 5: Agreement of community assignments with local connectivity. We study how consistent partitions are within local neighborhood regions of the network by examining how well a node's neighbors (for various order neighborhoods) can be used to predict its community assignment, under some community partition z. For each community in a partition, we give a binary prediction of whether a node is assigned to that community, based on probabilities we compute for a node from its neighbors. Sweeping the parameter p that sets the probability required for a node to be assigned to a community, we compute ROC curves for each community and report the minimum AUC value observed. Panels A-D show minimum AUC values observed as a function of neighborhood order for communities obtained from the full networks and super node representations by Louvain and by SBM. Line color indicates network and line type indicates communities obtained from the matched and default parameters used by the algorithms on the full networks. Panels E-H visualize the communities obtained in the As22 data on the full network (default parameters) and super node representation (SN) under Louvain and SBM, with node colors indicating community memberships. the network into communities that are locally relevant, we visualize the As22 network, with nodes colored by communities identified under the full network (E,G) and super node representation (F,H) under Louvain (E,F) and SBM (G,H). Consistent with the quantitative results in Figure 5A-D, we observe that the super node representation leads to an effective coarse-graining, with the community labels appearing to be qualitatively consistent across large regions of the network. 4 Conclusion and Future Work We developed an approach for compressing a network into a super node representation that can be used in standard community detection algorithms. Using the smaller super node network reduces runtime and the variability between multiple runs of the same community detection algorithm. Our results also demonstrate that the communities in the super node network are better aligned with local network neighborhoods in a predictive sense, while still being in relatively good alignment with the partitions obtained using the full network. Super nodes may be useful in a variety of contexts where large datasets are otherwise difficult to mine and interpret. For example, one might visualize the super node version of the network rather than the entire network, or use the members of a super node to identify redundant information in the network. Future work on super node representations could include the extension of this method to directed, signed, attributed, or bipartite networks. Additionally, one might consider a probabilistic model framework, attempting to infer latent super node assignments. 8 ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●0.000.250.500.751.00123456Neighborhood OrderMin AUCLouvain,Full●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●0.000.250.500.751.00123456Neighborhood OrderMin AUCLouvain,SN●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●0.000.250.500.751.00123456Neighborhood OrderMin AUCSBM,Full●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●0.000.250.500.751.00123456Neighborhood OrderMin AUCSBM,SNA.B.C.D.0.000.250.500.75123Neighborhood OMin AUCNetworkAs22EnronCMatterDblpAmazonParameterMatchedDefaultLouvain, FullLouvain, SNSBM, FullSBM, SNF.G.H.E.References [1] Santo Fortunato. Community detection in graphs. Physics Reports, 486(3):75–174, 2010. [2] Santo Fortunato and Darko Hric. Community detection in networks: A user guide. Physics Reports, 659:1–44, 2016. [3] Mason A Porter, Jukka-Pekka Onnela, and Peter J Mucha. Communities in networks. Notices of the AMS, 56(9):1082–1097, 2009. [4] Tom AB Snijders and Krzysztof Nowicki. Estimation and prediction for stochastic blockmodels for graphs with latent block structure. Journal of classification, 14(1):75–100, 1997. [5] Jaewon Yang and Jure Leskovec. Overlapping community detection at scale: a nonnegative matrix factorization approach. In Proceedings of the sixth ACM international conference on Web search and data mining, pages 587–596. ACM, 2013. [6] Tiago P Peixoto. Efficient Monte Carlo and greedy heuristic for the inference of stochastic block models. Physical Review E, 89(1):012804, 2014. [7] Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. Fast unfolding of communities in large networks. Journal of statistical mechanics: theory and experiment, 2008(10):P10008, 2008. [8] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103(23):8577–8582, 2006. [9] David F Gleich and C Seshadhri. Vertex neighborhoods, low conductance cuts, and good seeds for local community methods. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 597–605. ACM, 2012. [10] Mark EJ Newman. Spectral methods for community detection and graph partitioning. Physical Review E, 88(4):042822, 2013. [11] Arnaud Browet, P-A Absil, and Paul Van Dooren. Community detection for hierarchical image segmentation. In International Workshop on Combinatorial Image Analysis, pages 358–371. Springer, 2011. [12] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence, 34(11):2274–2282, 2012. [13] Jaewon Yang and Jure Leskovec. Defining and evaluating network communities based on ground-truth. Knowledge and Information Systems, 42(1):181–213, 2015. [14] Anna C Gilbert and Kirill Levchenko. Compressing network graphs. In Proceedings of the LinkKDD workshop at the 10th ACM Conference on KDD, volume 124, 2004. [15] Andreas Martin Lisewski, Joel P Quiros, Caroline L Ng, Anbu Karani Adikesavan, Kazutoyo Miura, Nagireddy Putluri, Richard T Eastman, Daniel Scanfeld, Sam J Regenbogen, Lindsey Altenhofen, et al. Supergenomic network compression and the discovery of exp1 as a glutathione transferase inhibited by artesunate. Cell, 158(4):916–928, 2014. [16] Liang Yang, Di Jin, Dongxiao He, Huazhu Fu, Xiaochun Cao, and Francoise Fogelman-Soulie. Improving the efficiency and effectiveness of community detection via prior-induced equivalent super-network. Scientific Reports, 7(1):634, 2017. [17] Lenka Zdeborová, Pan Zhang, and Hai-Jun Zhou. Fast and simple decycling and dismantling of networks. Scientific Reports, 6, 2016. 9 [18] Kyomin Jung, Wooram Heo, and Wei Chen. Irie: Scalable and robust influence maximization in social networks. In Data Mining (ICDM), 2012 IEEE 12th International Conference on, pages 918–923. IEEE, 2012. [19] David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 137–146. ACM, 2003. [20] Leon Danon, Albert Diaz-Guilera, Jordi Duch, and Alex Arenas. Comparing community structure identification. Journal of Statistical Mechanics: Theory and Experiment, 2005(09):P09008, 2005. 10
1903.03036	2	1903	2019-05-02T11:17:22	HEAT: Hyperbolic Embedding of Attributed Networks	[ "cs.SI" ]	Finding a low dimensional representation of hierarchical, structured data described by a network remains a challenging problem in the machine learning community. An emerging approach is embedding these networks into hyperbolic space because it can naturally represent a network's hierarchical structure. However, existing hyperbolic embedding approaches cannot deal with attributed networks, in which nodes are annotated with additional attributes. These attributes might provide additional proximity information to constrain the representations of the nodes, which is important to learn high quality hyperbolic embeddings. To fill this gap, we introduce HEAT (Hyperbolic Embedding of ATributed networks), the first method for embedding attributed networks to a hyperbolic space. HEAT consists of 1) a modified random walk algorithm to obtain training samples that capture both topological and attribute similarity; and 2) a learning algorithm for learning hyperboloid embeddings from the obtained training samples. We show that by leveraging node attributes, HEAT can outperform a state-of-the-art Hyperbolic embedding algorithm on several downstream tasks. As a general embedding method, HEAT opens the door to hyperbolic manifold learning on a wide range of attributed and unattributed networks.	cs.SI	cs	HEAT: Hyperbolic Embedding of Attributed Networks David McDonald1[0000−0002−0540−8254] and Shan He1 The University of Birmingham, Edgebaston, Birmingham, B15 2TT, UK {dxm237,s.he}@cs.bham.ac.uk Abstract. Finding a low dimensional representation of hierarchical, struc- tured data described by a network remains a challenging problem in the machine learning community. An emerging approach is embedding these networks into hyperbolic space because it can naturally represent a network's hierarchical structure. However, existing hyperbolic embed- ding approaches cannot deal with attributed networks, in which nodes are annotated with additional attributes. These attributes might pro- vide additional proximity information to constrain the representations of the nodes, which is important to learn high quality hyperbolic em- beddings. To fill this gap, we introduce HEAT (Hyperbolic Embedding of ATributed networks), the first method for embedding attributed net- works to a hyperbolic space. HEAT consists of 1) a modified random walk algorithm to obtain training samples that capture both topological and attribute similarity; and 2) a learning algorithm for learning hyper- boloid embeddings from the obtained training samples. We show that by leveraging node attributes, HEAT can outperform a state-of-the-art Hyperbolic embedding algorithm on several downstream tasks. As a gen- eral embedding method, HEAT opens the door to hyperbolic manifold learning on a wide range of attributed and unattributed networks. Keywords: Complex networks · Hyperbolic embedding 1 Introduction The success of machine learning algorithms often depends upon data representa- tion [3]. Representation learning -- where we learn alternative (low dimensional) representations of data -- has become common for processing information on non-Euclidean domains, such as the domain of nodes and edges that comprise complex networks. Prediction over nodes and edges, for example, requires care- ful feature engineering [10] and representation learning leads to the extraction of features from a graph that are most useful for downstream tasks, without careful design or a-priori knowledge. An emerging representation learning approach for complex networks is hy- perbolic embedding. This approach is based on compelling evidence that the underlying metric space of many complex networks is hyperbolic [12,1,2] . A hyperbolic space can be interpreted as a continuous representation of a discrete 2 David McDonald and Shan He tree structure that captures the hierarchical organisation of elements within a complex system [12]. Since real world networks are often hierarchical -- where nodes are grouped hierarchically into groups in an approximate tree structure [16]. Hyperbolic spaces therefore offer a natural continuous representations of hierarchical complex networks [13,12]. However, existing hyperbolic embedding approaches cannot deal with at- tributed networks, of which nodes (entities) are richly annotated with attributes [9,14]. For example, a paper within a citation network may be annotated with the presence of keywords, and the people in a social network might have additional information such as interests, hobbies, and place of work. These attributes might provide additional proximity information to constrain the representations of the nodes. Therefore, incorporating this node attributes can improve the quality of the final embedding, with respect to many different downstream tasks [14]. This paper proposes the first hyperbolic embedding method for attributed networks called HEAT. The method uses a modified random walk algorithm to obtain training samples that capture both topological and attribute similarities. Inspired by the concept of homophily -- that similar nodes are more likely to interact [14] -- the random walk algorithm uses a teleport procedure to capture "phantom links" between nodes that do not share a topological link, but have highly similar attributes. We then further formulate a learning algorithm to learn the embeddings from the training samples based on stochastic gradient descent in hyperboliod space. Our results show that, by leveraging node attributes, HEAT can achieve bet- ter performance on several downstream tasks compared with a state-of-the-art hyperbolic embedding algorithm. As a general framework, HEAT can embed both unattributed and attributed networks with continuous and discrete at- tributes, which opens the door to hyperbolic manifold learning for a wide range of complex networks. 1.1 Related Work Many hyperbolic embedding algorithms are based on two-dimensional hyper- bolic disk, which was popularised by the Popularity-Similarity (or PS) model [16]. In [16], Maximum Likelihood (ML) was used in to search the space of all PS models with similar structural properties as the observed network, to find the one that fit it best. This was extended by the authors in [17]. Due to the com- putationally demanding task of maximum likelihood estimation, often heuristic methods are used. For example, [1] used Laplacian Eigenmaps to estimate the angular coordinates of nodes in the PS model. The authors then combined both approaches to leverage the performance of ML estimation against the efficiency of heuristic search with a user controlled parameter in [2]. Additionally, [21] pro- pose the use of classical manifold learning techniques in the PS model setting with a framework that they call coalescent embedding. Beyond the two-dimensional hyperbolic disk, an n-dimensional Poincar´e ball can give more degrees of freedom to the embedding and capture further dimen- sions of attractiveness than just "popularity" and "similarity" [15,6]. By embed- HEAT: Hyperbolic Embedding of Attributed Networks 3 ding general graphs to trees, [8] were able to achieve state-of-the-art results by extending the work of [20]. 2 Background: Hyperbolic Geometry In this section, we give some background knowledge about Hyperbolic Geometry to make our paper self-contained. 2.1 Models of Hyperbolic Geometry Hyperbolic geometry is what one obtains by negating Euclid's fifth postulate, the parallel postulate. In hyperbolic geometry there exists a line l and a point P not on l such that at least two distinct lines parallel to l pass through P [12]. It has a negative Gaussian curvature K for the entire space and that implies that every single point in the Hyperbolic space is a saddle point. Furthermore, hyperbolic spaces cannot be embedded into a Euclidean space without distortion. [12] informally explain hyperbolic spaces are "larger" and have more "space" than Euclidean spaces. This is mathematically reflected by the fact that the area of a circle of radius r, does not grow quadratically with r, as we are used to in Euclidean space1, but grows exponentially with A = 2π cosh(ζr−1).2 Because of this property, we observe an exponential expansion of space [4]. Because of the fundamental difficulties in representing spaces of constant negative curvature as subsets of Euclidean spaces, there are not one but many equivalent models of hyperbolic spaces. We say the models are equivalent be- cause all models of hyperbolic geometry can be freely mapped to each other by an isometry. Each model (the most popular of which is the Poincar´e ball [12,15] ) emphasizes different aspects of hyperbolic geometry, but no model simultane- ously represents all of its properties. We are free to chose the model that best fits our need. The most popular model in the network embedding literature is the Poincar´e disk (or Poincar´e ball for n > 2 dimensions.) Here, we have the entire hyperbolic plane represented as the interior of a Euclidean unit ball [12,15]. , sitting in a Euclidean ambient space, where the boundary of the ball represents infinity [12,15]. The boundary of the ball represents infinity in the hyperbolic system that it is modelling. Euclidean and hyperbolic distances, re and rh from the disk centre, or the origin of the hyperbolic plane, are related by re = tanh(rh/2) and the shortest path between points (geodesics) are given by the diameters of the circle or Euclidean circle arcs that intersect the boundary of the ball perpendicularly. This model has the main advantage that it is conformal : the Euclidean angles in the model are equal to the hyperbolic angles in hyperbolic geometry that the model is representing. This makes the model a popular choice for embedding methods that abstract node similarity as the angular distance between them, for 1 Recall that the area of a circle is given by A = πr2 in Euclidean geometry. 2 ζ = √−K and cosh is Hyperbolic cosine function: cosh(x) = (ex + e−x)/2 4 David McDonald and Shan He example: [1,2]. We have also the much less popular Klein model of hyperbolic geometry. This model also represents hyperbolic geometry as a unit disk (or ball) in an ambient Euclidean space. This model preserves straight lines: straight Euclidean lines map to straight hyperbolic lines. However, unlike the Poincar´e ball, the Klein model is not conformal, and the Euclidean angles in the model are not equal to hyperbolic angles. 2.2 The Hyperboloid Model of Hyperbolic Space We consider the Hyperboloid model primarily in this work. Unlike disk models, that sit in an ambient Euclidean space of dimension n, the Hyperboloid model of n-dimensional hyperbolic geometry sits in n + 1-dimensional Minkowski space- time. Furthermore, the set of hyperboloid points do not form a disk in this ambient space, but in n-dimensional hyperboloid. We view both the Poincar´e and Klein models as (stereographic and orthographic) projections of the points from the hyperboloid to disks orthogonal to the main axis of the hyperboloid [12]. Informally, we can see this relationships as analogous to the relationship between a projected map and a globe [19]. Specifically, n + 1-dimensional Minkowski spacetime is defined as the com- bination of n-dimensional Euclidean space with an additional time co-ordinate t. Commonly, we denote this set Rn:1. We say that point x ∈ Rn:1 has time co-ordinate x0 i and spacial coordinates xk i for k = 1, 2, ..., n. The Minkowski bilinear form is defined as (cid:104)xi, xj(cid:105)Rn:1 = −ψ2x0 i x0 j + n(cid:88) xk i xk j k=1 where ψ is the speed of information flow in our system (normally set to 1 for simplified calculations). Further details have been omitted for brevity, however the reader is directed to [7] for more details. This bilinear form functions as an inner product (like the Euclidean dot product) and allows use to compute (cid:104)x, x(cid:105)Rn:1 < 0 and so norms may be imaginary. norms in a familiar way: xRn:1 := (cid:112)(cid:104)x, x(cid:105)Rn:1. However, it is possible for In fact, the points x satisfying (cid:104)x, x(cid:105)Rn:1 < 0 are of particular relevance to hy- perbolic geometry as the n-dimensional hyperboloid Hn is comprised of just such points: Hn = {x ∈ Rn:1 (cid:104)x, x(cid:105)Rn:1 = −1, x0 > 0}. The first condition defines a hyperbola of two sheets, and the second one selects the top sheet. Shortest paths (geodesics) between points on the model are given by the hyperbola formed by the intersection of Hn and the two dimensional plane containing the origin and both of the points. The distance (along the geodesic) between two points xu, xv ∈ Hn is given by dHn (xu, xv) = arccosh(−(cid:104)xu, xv(cid:105)Rn:1) and is analogous to the length of the great circle connecting two points in spherical geometry3. We also define the tangent space of a point x ∈ Hn as TxHn = {x(cid:48) ∈ Rn:1 (cid:104)x, x(cid:48)(cid:105)Rn:1 = 0}. 3 The proof for the distance formula is given in [19] HEAT: Hyperbolic Embedding of Attributed Networks 5 We see that TxHn is the collection of all points in Rn:1 that are orthogonal to p. It can be shown (in [19]) that (cid:104)x(cid:48), x(cid:48)(cid:105)Rn:1 > 0 ∀x(cid:48) ∈ TxHn ∀x ∈ Hn. In other words, the tangent space of the hyperboloid is positive definite (with respect to the Minkowski bilinear form) for all points on the hyperboloid. This property actually defines Hn (equipped with the Minkowski bilinear form) as a Riemannian manifold [19]. Furthermore, we obtain a positive norm for any vector x(cid:48) ∈ TxHn, allowing us to preform gradient descent. 3 Method 3.1 Problem Definition We consider a network of N nodes given by the set V with V = N . We use E to denote the set of all interactions between the nodes in our network. E = {(u, v)} ⊆ V × V . We use the matrix W ∈ RN×N to encode the weights of these interactions, where Wuv is the weight of the interaction between actor u and actor v. We have that Wu,v > 0 ⇐⇒ (u, v) ∈ E. If the network is unweighted then Wu,v = 1 for all (u, v) ∈ E. Furthermore, the matrix X ∈ RN×d describes the attributes of each nodes in the network. These attributes may be discrete or continuous. We consider the problem of representing a graph given as G = (V, E, W, X) as set of low-dimensional vectors in the n-dimensional hyperboloid {xv ∈ Hn v ∈ V }, with n << min(N, d). The described problem is unsupervised. 3.2 HEAT Overview Our proposed HEAT consists of two main components: 1. A novel random walk algorithm, which samples the network to obtain train- ing samples that can capture both topological and attributional similarity. 2. A learning algorithm that learns hyperboloid embeddings from the training samples. The following subsections describe the detailed algorithms. 3.3 Sample the network using random walks with teleport Following previous works [10,18], we modify the random-walk procedure to ob- tain training samples that capture both topological and attributional similarity. The basic idea is that, in additional to standard random walks which capture topological similarity, we use attributional similarity to 'teleport' the random walker to the nodes with similar attributes. To this end, we define the attributional similarity Y as cosine similarity of the attribute vectors of the nodes. We assign a value of 0 similarity to any negative values. That is Yuv = max u Xv XuXv , 0 (cid:18) X T (cid:19) 6 David McDonald and Shan He with · the Euclidean norm. We use cosine similarity because it can handle high dimensional data well without making a strong assumption about the data. We can change the cosine similarity to a more sophisticated and problem dependant measure of pairwise node attributional similarity. We then additionally define ¯W and ¯Y to be the row normalized versions of W and Y respectively, such that each row sums to 1 (zero rows remain as zero rows). Therefore, each row in ¯W and ¯Y is a probability distribution. In particular, the entry ¯Wu,v encodes the probability of moving from node u to v based on the strength of the topological link between u and v, and ¯Yu,v likewise encodes the teleport probability based on attributional similarity. Beginning from a source node s in the network, we perform a walk with a fixed length l through the network. Each step in the walk from one node to the next is a stochastic process based on both topological structure and similarity of attributes. We define 0 ≤ α ≤ 1 to be a parameter that controls the trade-off between a topological step and an attribute step in the walk. Formally, we use i to denote the ith node in the walk (x0 = s), and for each step for i = 1, 2, ..., l we sample πi ∼ U (0, 1) and determine the ith node as follows: (cid:40) Wuv Yuv P (xi = v xi−1 = u) = if πi < α, otherwise. We follow previous works like node2vec [10] and Deepwalk [18], by considering nodes that appear close together in the same walk to be "context pairs". For a source-context pair (u, v), we aim to maximise the probability of observing v, given u, P (v u). To build the set of source-context pairs, D, we scan across all walks with a sliding window and add pairs of nodes that appear within the window. We call this window size the "context-size" and it is a user defined parameter (hereafter denoted c) that controls the size of a local neighbourhood of a node. Previous works show that increasing context size typically improves performance, at some computational cost [10]. Negative Sampling We define the probability of two nodes sharing a connec- tion to be function of their distance in the embedding space. Nodes separated by a small distances share a high degree of similarity and should, therefore have a high probability of connection. Similarly, nodes very far apart in the embedding space should have a low probability of connection. We make the common simpli- fying assumption that a source node and neighbourhood node have a symmetric effect over each other in feature space (ie: P (v u) = P (u v) for all u, v ∈ V ) [10]. To this end, we define the symmetric function P (v u) := −d2Hn (xu, xv) to be the un-normalized probability of observing a link between nodes source node u and context node v. We square the distance because this gives us stable HEAT: Hyperbolic Embedding of Attributed Networks 7 gradients [8]. We normalize the probability thusly: (cid:16) P (v u)/2σ2(cid:17) (cid:16) P (v(cid:48) u)/2σ2(cid:17) P (v u) := Z(u) := exp Z(u) exp 1 (cid:88) v(cid:48)∈V However, computing the gradient of the partition function Z(u) involves a sum- mation over all nodes v ∈ V , which for large networks, is prohibitively compu- tationally expensive [10]. Following previous works, we overcome this limitation through negative sampling. We define the set of negative samples for u, Γ (u), as the set of v for we we observe no relation with u: Γ (u) := {v (u, v) (cid:54)∈ D} There is no guarantee that the size of these sets is the same for all u, so we further define Sm(u, v) := {xi ∼ Γ (u) i = 1, 2, ..., m} ∪ {v} to be a random sample with replacement of size m from the set of negative sam- ples of u, according to a noise distribution Pn. Following [10], we use Pn = U 3 4 the unigram distribution raised to the 3 4 power. This means that the probability that a node is selected as a negative sample is proportional to its occurrence probability, that we define to be the number of times that it appeared over all the random walks for the network. 3.4 Hyperboloid Embedding Learning We now formulate a loss function L that encourages maximising the probability of observing all positive sample pairs P (v u) ∀(u, v) ∈ D and minimising the probability of observing all other pairs. To this end, we define the loss function L for an embedding Θ = {xu ∈ Hn u ∈ V } to be the the mean of negative log-likelihood of observing all the source-context pairs in D, against the negative sample noise: (cid:88) (cid:34) exp(cid:0)−d2Hn (xu, xv)/2σ2(cid:1) (cid:80) v(cid:48)∈Sm(u,v) exp (−d2Hn (xu, xv(cid:48)) /2σ2) (cid:35) L(Θ) = − 1 D log (u,v)∈D We observe that minimising L involves maximising P (v u) ∀(u, v) ∈ D as required. This encourages source-context pairs to be close together in the em- bedding space, and u to be embedded far from the noise nodes v(cid:48) [15]. 3 For x, x(cid:48) ∈ Hn lim x(cid:48)→x with lim x→−1 ∂x arccosh2(−x) → 2. Contrast this (cid:104)x, x(cid:48)(cid:105)Rn:1 → −1 and lim x→−1 ∂x arccosh(−x) → ∞ 8 David McDonald and Shan He Because we use hyperboloid model for complex network embedding, unlike previous works [15,8] that use the Poincar´e ball model and approximate gra- dients, our gradient computation is simple and exact [22]. We follow the three step procedure in [22] to compute gradients x(cid:48) u. The procedure is based on the fact that the loss function L is defined over the whole ambient Minkowski space Rn:1, which is further defined over Hn ⊂ Rn:1. Therefore, for a given point on the hyperboloid xu ∈ Hn, we can compute the gradient of L with respect to xu, denoted ∇Hn Hn. Then to perform gradient descent optimization, we move xu along −∇Hn Hn. Finally we map x(cid:48) back to Hn using an exponential mapping. L by a small amount η to x(cid:48) L ∈ Txu u ∈ Txu xu xu u L, we first compute the gradient with respect to Specifically, to compute ∇Hn the Minkowski ambient space Rn:1 as ∇Rn:1 xu L = − ∂L ∂x0 , ∂L ∂x1 , ..., ∂L ∂xn Let ouv := − arccosh(−(cid:104)xu, xv(cid:105)Rn:1)2/2σ2. Then (cid:12)(cid:12)(cid:12)(cid:12)xu (cid:33) (cid:12)(cid:12)(cid:12)(cid:12)xu xu (cid:32) (cid:88) (cid:12)(cid:12)(cid:12)(cid:12)xu (cid:88) ∇Rn:1 xu L = 1 D v (u,v)∈D v(cid:48)∈Sm(u,v) (δvv(cid:48) − P (v u)) · ∇Rn:1 xu and ∇Rn:1 xu ouv = (cid:113)(cid:104)xu, xv(cid:105)2Rn:1 − 1 arccosh (−(cid:104)xu, xv(cid:105)Rn:1) σ2 · xv ouv(cid:48) (1) (2) where δvv(cid:48) is the Kronecker delta function. We then use the vector projection formula from Euclidean geometry (replacing the dot product with the Minkowski inner product) to compute the projection of the gradient with the ambient to its component in the tangent space: L = ∇Rn:1 L + (cid:104)xu,∇Rn:1 L(cid:105)Rn:1 · xu ∇Hn (3) xu xu xu Having computed the gradient component in the tangent space of xu, we define the exponential map to take a vector x(cid:48) Hn to its corresponding point on the hyperboloid: u ∈ Txu Expxu (x(cid:48) u) = cosh(x(cid:48) u) · xu + sinh(x(cid:48) (4) Where · denotes the norm with respect to the Minkowski space: x := (cid:112)(cid:104)x, x(cid:105)Rn:1. This is analogous to the exponential map in spherical geometry u) · x(cid:48) ux(cid:48) u with maps points from the tangent space of a point on the sphere back to the sphere itself. Finally, the three-step procedure for computing u(cid:48) with learning rate η is: 1. Calculate ambient gradient ∇Rn:1 2. Project ∇Rn:1 L to tangent ∇Hn 3. Set xu = Expxu L(cid:1) (eq. (4)). (cid:0)−η∇Hn L (eqs. (1) and (2)), L (eq. (3)), xu xu xu xu HEAT: Hyperbolic Embedding of Attributed Networks 9 4 Experimental Setup & Results 4.1 Datasets We evaluate HEAT on two citation networks (obtained from [5]) and one PPI network (obtained from [11]). We select the PPI network to evaluate performance on a different network type than citation. For the PPI network, [11] use posi- tional gene sets, motif gene sets and immunological signatures as features and gene ontology sets as labels. Each connected component represents a different tissue sample. For our experiments, we select the largest connected component. Features for all networks were scaled to have a mean of 0 and a standard devia- tion of 1. Table 1 shows the network statistics of these three networks. Table 1. Network statistics. Key: N is the number of nodes, E is the number of edges, d is the dimension of node features, y is the number of classes. Network N E Cora ML 2995 8416 2879 7 Citeseer 4230 5358 2701 6 d y PPI 3480 54806 50 121 4.2 Parameter Settings Table 2 shows the parameter settings used for the following experiments. For comparison, we used the open-source implementation of the algorithm described by [15]. We used default parameters to train their embeddings. We fix α = 0.2 for all experiments. Note that HEAT requires less numbers of epochs (emax) and negative samples (m). However, as a trade-off, HEAT introduces four more control parameters, but as shown in Section 4.5, the performance of HEAT is robust to the setting of those parameters. Table 2. Parameter settings used for experiments. Key: η is the learning rate, emax is the number of epochs, m is the number of negative samples for each positive sample, b is the batch size, c is the context size, s is the number of random walks, l is the random walk length, eb is the number of epochs with reduced learning rate at beginning of training, σ is the Gaussian width, α is the topology/attribute trade off parameter. η emax m b c s l eb σ α HEAT .3 N&K 1. 1500 50 50 - 10 50 3 10 80 - 1. .2 - - 20 - 5 - 10 David McDonald and Shan He 4.3 Network Reconstruction & Link Prediction We use network reconstruction experiment to evaluate the embedding's capacity of reflecting the original data. After training our model to convergence upon the complete information, we compute distances (according the hyperbolic distance) in the embedding space between all pairs of nodes according to both models. We assign the true edges in the network positive labels and all other pairs as negatives. We then rank node pairs by their distance in increasing order and, with a sliding threshold, compute the area under the receiver operating characteristic (AUROC) curve. Following common practice, to evaluate the embeddings' link prediction abil- ity, we randomly select 15% of the edges in the network and remove them [5]. We randomly select also an equal number of non-edges in the network. We then train each model on the incomplete network and rank the pairs of nodes based on distance. We then use the removed edges as the positives and the selected non-edges as negatives and, again, compute AUROC. Table 3. AUROC scores (to 3 d.p.) for network reconstruction (Recons.) and link prediction (LP) tasks with embedding dimensions of 5, 10, 25, and 50. All scores are averaged over 30 random starting seeds. When α = 0.2 the embedding consider node attributes, while α = 0 indicates the embedding only consider network topology. Bold scores indicate the best score for each dimension for each task. Dim N&K α = 0.0 α = 0.2 N&K α = 0.0 α = 0.2 N&K α = 0.0 α = 0.2 Cora ML Citeseer PPI 5 10 25 50 Recons. 0.993 0.995 0.992 0.999 0.999 0.998 0.937 0.931 LP 0.889 0.921 0.960 0.756 0.907 0.958 0.910 0.898 0.922 0.887 Recons. 0.994 0.996 0.997 1.000 0.999 LP 0.893 0.929 0.968 0.760 0.916 0.963 0.913 0.909 0.999 0.940 0.950 0.944 0.901 Recons. 0.994 0.997 0.998 1.000 0.999 0.999 0.941 0.961 0.959 LP 0.894 0.930 0.972 0.764 0.919 0.967 0.914 0.915 0.908 Recons. 0.994 0.997 0.998 1.000 0.999 0.999 0.941 0.963 0.963 LP 0.893 0.931 0.973 0.761 0.919 0.969 0.914 0.916 0.911 Table 3 shows that, with node attributes, i.e., when α = 0.2, HEAT outper- formed N&K on all three networks for the link prediction task. Interestingly, without node attributes, i.e., when α = 0, HEAT still outper- formed N&K on the network reconstruction task and link prediction task on the Cora ML network as well as the link prediction task for the Citeseer network and both reconstruction and link prediction for higher dimensional embeddings of the PPI network. Such results indicate that HEAT can embed networks with- out node attributes in hyperbolic space accurately. We can also see that, for the Cora ML and Citeseer networks, incorporating node attributes, i.e., when α = 0.2, improved the AUROC scores significantly HEAT: Hyperbolic Embedding of Attributed Networks 11 for the link prediction task. However, for PPI, incorporating node attributes deteriorates the AUROC score, which suggests that for PPI, the homophily hypothesis, i.e., similar nodes in terms of attributes are more likely to interact, might not hold. We also observe that, for all three networks, increasing the hyperbolic embedding dimensions only provides marginal improvement. 4.4 Node Classification We make the assumption that nodes are likely to connect to nodes of the same label, and will also display similar attributes. We train an embedding using complete topology and attribute information. Note that the embedding is unsu- pervised, as it is performed with no knowledge of the ground truth labels. We project the resulting embedding to the Klein model of hyperbolic space, which preserves straight lines [17]. We use an out-of-the-box logistic regression model with the Klein embedding of the network as input features. We record micro- F1 and macro-F1 scores by varying the labelled percentage of nodes from 2% to 10%. For the PPI network, each protein has multiple labels from the Gene Ontology. To evaluate our model in the multi-label case, we adopt a one-vs-all setting, where we train a separate regression model for each label. Fig. 1. Plots of micro-averaged F1 scores against labelled percentage for the Cora ML network. Each score is averaged over 30 random starting seeds. For HEAT, when α = 0.2 the embedding considers node attributes, while α = 0 indicates the embedding only considers network topology. We present only the results for an embedding dimension of 10 for clarity, but obtain similar results for all dimensions. Figures 1 to 3 show that, for node classification, when considering node at- tributes, i.e., α = 0.2, HEAT outperformed N&K on all three networks. We also note that, even without node attributes, HEAT obtained better results 12 David McDonald and Shan He Fig. 2. Plots of micro-averaged F1 scores against labelled percentage for the Citeseer network. Each score is averaged over 30 random starting seeds. For HEAT, when α = 0.2 the embedding considers node attributes, while α = 0 indicates the embedding only considers network topology. We present only the results for an embedding dimension of 10 for clarity, but obtain similar results for all dimensions. Fig. 3. Plots of micro-averaged F1 scores against labelled percentage for the PPI net- work. Each score is averaged over 30 random starting seeds. For HEAT, when α = 0.2 the embedding considers node attributes, while α = 0 indicates the embedding only considers network topology. We present only the results for an embedding dimension of 10 for clarity, but obtain similar results for all dimensions. on Cora ML and PPI networks than N&K. In contrast, for Citeseer network, without node attributes, HEAT performed much worse than N&K in terms of HEAT: Hyperbolic Embedding of Attributed Networks 13 micro-average F1 score. These results suggest that, without node attributes, the performance of network embedding algorithms such as N&K and HEAT depends on the specific network topology. 4.5 Sensitivity of control parameters We carried out preliminary experiments to evaluate HEAT's robustness to the setting of the control parameters. Our results indicated that the most sensitive parameter is α, which controls the trade off between topology and attributes. We therefore run HEAT over a range of values α ∈ {0.0, 0.05, 0.1, 0.2, 0.5, 0.8, 1.0}. Figure 4 plot AUROC scores for link prediction and the micro averaged F1 scores for node classification obtained from a 10 dimensional embedding with 10% labelled samples for training the logistic regression model. From the two figures, we can see the performance of HEAT on the two tasks on the three networks is robust a wide range values of α. 5 Conclusion This paper presents HEAT to fill the gap of embedding attributed networks in hyperbolic space. We have designed a random walk algorithm to obtain the training samples that capture both network topological and attributional sim- ilarity. We have also derived an algorithm that learns hyperboloid embeddings from the training samples. Our results on three networks show that, by including attributes, HEAT can improve the quality of a learned hyperbolic embedding in a number of downstream machine learning tasks. We also observed that HEAT is competitive compared with N&K, a state of the art hyperbolic embedding algorithm, on unattributed networks, with smaller numbers of iterations and training samples. HEAT provides a general hyperbolic embedding method for both unattributed and attributed networks, which opens the door to hyperbolic manifold learning on a wide ranges of networks. 6 Code Availability Open source software is freely available at https://github.com/DavidMcDonald1993/ heat. References 1. Alanis-Lobato, G., Mier, P., Andrade-Navarro, M.A.: Efficient embedding of com- plex networks to hyperbolic space via their laplacian. Scientific Reports 6 (2016) 2. Alanis-Lobato, G., Mier, P., Andrade-Navarro, M.A.: Manifold learning and max- imum likelihood estimation for hyperbolic network embedding. Applied Network Science 1(1), 10 (2016) 14 David McDonald and Shan He Fig. 4. Plots to show the effect of alpha on two downstream machine learning tasks. We fix the embedding dimension to 10 for these experiments. For node classification, we train the logistic regression model with 10% labelled samples. 3. Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35(8), 1798 -- 1828 (2013) 4. Bogun´a, M., Papadopoulos, F., Krioukov, D.: Sustaining the internet with hyper- bolic mapping. Nature communications 1, 62 (2010) 5. Bojchevski, A., Gunnemann, S.: Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking. arXiv preprint arXiv:1707.03815 (2017) 6. Chamberlain, B.P., Clough, J., Deisenroth, M.P.: Neural embeddings of graphs in hyperbolic space. arXiv:1705.10359 (2017) 7. Clough, J.R., Evans, T.S.: Embedding graphs in lorentzian spacetime. PloS one 12(11), e0187301 (2017) HEAT: Hyperbolic Embedding of Attributed Networks 15 8. De Sa, C., Gu, A., R´e, C., Sala, F.: Representation tradeoffs for hyperbolic em- beddings. arXiv:1804.03329 (2018) 9. Gibert, J., Valveny, E., Bunke, H.: Graph embedding in vector spaces by node attribute statistics. Pattern Recognition 45(9), 3072 -- 3083 (2012) 10. Grover, A., Leskovec, J.: node2vec: Scalable feature learning for networks. In: ACM SIGKDD. pp. 855 -- 864. ACM (2016) 11. Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. In: Advances in Neural Information Processing Systems. pp. 1024 -- 1034 (2017) 12. Krioukov, D., Papadopoulos, F., Kitsak, M., Vahdat, A., Bogun´a, M.: Hyperbolic geometry of complex networks. Physical Review E 82(3), 036106 (2010) 13. Krioukov, D., Papadopoulos, F., Vahdat, A., Bogun´a, M.: Curvature and temper- ature of complex networks. Physical Review E 80(3), 035101 (2009) 14. Liao, L., He, X., Zhang, H., Chua, T.S.: Attributed social network embedding. IEEE Transactions on Knowledge and Data Engineering 30(12), 2257 -- 2270 (2018) 15. Nickel, M., Kiela, D.: Poincar´e embeddings for learning hierarchical representa- tions. In: Advances in neural information processing systems. pp. 6338 -- 6347 (2017) 16. Papadopoulos, F., Kitsak, M., Serrano, M., Bogun´a, M., Krioukov, D.: Popularity versus similarity in growing networks. Nature 489 (2011) 17. Papadopoulos, F., Psomas, C., Krioukov, D.: Network mapping by replaying hy- perbolic growth. IEEE/ACM Transactions on Networking (TON) 23(1), 198 -- 211 (2015) 18. Perozzi, B., Al-Rfou, R., Skiena, S.: Deepwalk: Online learning of social represen- tations. In: ACM SIGKDD. pp. 701 -- 710. ACM (2014) 19. Reynolds, W.F.: Hyperbolic geometry on a hyperboloid. The American mathemat- ical monthly 100(5), 442 -- 455 (1993) 20. Sarkar, R.: Low distortion delaunay embedding of trees in hyperbolic plane. In: International Symposium on Graph Drawing. pp. 355 -- 366. Springer (2011) 21. Thomas, J.M.e.a.: Machine learning meets complex networks via coalescent em- bedding in the hyperbolic space. Nature Communitcations 1615 (2017) 22. Wilson, B., Leimeister, M.: Gradient descent in hyperbolic space. arXiv:1805.08207 (2018)
1511.07487	3	1511	2016-02-25T13:30:25	Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads	[ "cs.SI" ]	As breaking news unfolds people increasingly rely on social media to stay abreast of the latest updates. The use of social media in such situations comes with the caveat that new information being released piecemeal may encourage rumours, many of which remain unverified long after their point of release. Little is known, however, about the dynamics of the life cycle of a social media rumour. In this paper we present a methodology that has enabled us to collect, identify and annotate a dataset of 330 rumour threads (4,842 tweets) associated with 9 newsworthy events. We analyse this dataset to understand how users spread, support, or deny rumours that are later proven true or false, by distinguishing two levels of status in a rumour life cycle i.e., before and after its veracity status is resolved. The identification of rumours associated with each event, as well as the tweet that resolved each rumour as true or false, was performed by a team of journalists who tracked the events in real time. Our study shows that rumours that are ultimately proven true tend to be resolved faster than those that turn out to be false. Whilst one can readily see users denying rumours once they have been debunked, users appear to be less capable of distinguishing true from false rumours when their veracity remains in question. In fact, we show that the prevalent tendency for users is to support every unverified rumour. We also analyse the role of different types of users, finding that highly reputable users such as news organisations endeavour to post well-grounded statements, which appear to be certain and accompanied by evidence. Nevertheless, these often prove to be unverified pieces of information that give rise to false rumours. Our study reinforces the need for developing robust machine learning techniques that can provide assistance for assessing the veracity of rumours.	cs.SI	cs	Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads Arkaitz Zubiaga1, Maria Liakata1, Rob Procter1, Geraldine Wong Sak Hoi2, Peter Tolmie1 1 University of Warwick, Gibbet Hill Road, CV4 7AL Coventry, UK 2 swissinfo.ch, Bern, Switzerland [email protected] Abstract As breaking news unfolds people increasingly rely on social media to stay abreast of the latest updates. The use of social media in such situations comes with the caveat that new information being released piecemeal may encourage rumours, many of which remain unverified long after their point of release. Little is known, however, about the dynamics of the life cycle of a social media rumour. In this paper we present a methodology that has enabled us to collect, identify and annotate a dataset of 330 rumour threads (4,842 tweets) associated with 9 newsworthy events. We analyse this dataset to understand how users spread, support, or deny rumours that are later proven true or false, by distinguishing two levels of status in a rumour life cycle i.e., before and after its veracity status is resolved. The identification of rumours associated with each event, as well as the tweet that resolved each rumour as true or false, was performed by a team of journalists who tracked the events in real time. Our study shows that rumours that are ultimately proven true tend to be resolved faster than those that turn out to be false. Whilst one can readily see users denying rumours once they have been debunked, users appear to be less capable of distinguishing true from false rumours when their veracity remains in question. In fact, we show that the prevalent tendency for users is to support every unverified rumour. We also analyse the role of different types of users, finding that highly reputable users such as news organisations endeavour to post well-grounded statements, which appear to be certain and accompanied by evidence. Nevertheless, these often prove to be unverified pieces of information that give rise to false rumours. Our study reinforces the need for developing robust machine learning techniques that can provide assistance in real time for assessing the veracity of rumours. The findings of our study provide useful insights for achieving this aim. Introduction Social media have increasingly gained popularity in recent years, enabling people not only to keep in touch with family and friends, but also to stay abreast of ongoing events and breaking news as they unfold. The potential for spreading information quickly through a large community of users is one of the most 1 valuable characteristics of social media platforms. Social media, being open to everyone, enable not only news organisations and journalists to post news stories, but also ordinary citizens to report from their own perspectives and experiences. This broadens the scope and diversity of information that one can get from social media and sometimes may even lead to stories breaking before they appear in mainstream media outlets [Kwak et al., 2010]. While this often leads to having access to more comprehensive information, it also comes with caveats, one of which is the need to sift through the different information sources to assess their accuracy [Silverman, 2013]. The spread of misinformation is especially important in the context of breaking news, where new pieces of information are released piecemeal, often starting off as unverified information in the form of a rumour. These rumours then spread to large numbers of users, influencing perception and understanding of events, despite being unverified. Social media rumours that are later proven false can have harmful consequences both for individuals and for society [Zubiaga and Ji, 2014]. For instance, a rumour in 2013 about the White House having been bombed, injuring Barack Obama, which was tweeted from AP's Twitter account by hackers, spooked stock markets in the US1. A major event that was similarly riddled with consequential rumours was Hurricane Sandy, which hit the East Coast of the US in 2012. Part of the city of New York suffered from power outages and many people had to rely on the Internet accessed through their mobile phones for information. To prevent major incidents, the US Federal Emergency Management Agency had to set up a web page specifically for rumour control2. While rumours in social media are a concern, little work has been done so far to understand how they propagate. In this work we aim to help rectify this by examining in some detail rumours generated on Twitter within the context of nine different newsworthy events. First, we introduce our methodology to collect, identify, annotate and enable the analysis of social rumours. Then, we look at how rumours are spawned, and later confirmed or debunked; our study looks at conversations around rumours in social media, exploring how social media users respond to rumours both before and after the veracity of a rumour is resolved. Our study provides insight into rumour diffusion, support and denial in social media, helping both those who gather news from social media in determining accuracy of information and the development of machine learning systems that can provide assistance in real-time for assessing the veracity of rumours [Derczynski and Bontcheva, 2014]. Background One of the main challenges when studying rumours is to come up with a sound definition of the concept. In contrast to recent research that considers a rumour as constituting information that is ultimately deemed false, here we emphasise 1http://www.bbc.co.uk/news/world-us-canada-21508660 2https://twitter.com/fema/status/264800761119113216 2 our adherence to the widely accepted definition of rumours as "unverified and instrumentally relevant information statements in circulation" [DiFonzo and Bordia, 2007]. This unverified information may turn out to be true, or partly or entirely false. Adapting the existing definition to our context of breaking news stories, we define a rumour as a "circulating story of questionable veracity, which is apparently credible but hard to verify, and produces sufficient skepticism and/or anxiety so as to motivate finding out the actual truth" [Zubiaga et al., 2015b]. Rumours and related phenomena have now been studied from many different perspectives [Donovan, 2007] ranging from psychological studies [Rosnow and Foster, 2005] to computational analyses [Qazvinian et al., 2011]. Traditionally, it has been very difficult to study people's reactions to rumours, given that this would involve real-time collection of reaction as rumours unfold. To overcome this obstacle, Allport undertook early investigations [Allport and Postman, 1946, Allport and Postman, 1947] in the context of wartime rumours. He posited the importance of studying rumours, emphasising that "newsworthy events are likely to breed rumors" and that "the amount of rumor in circulation will vary with the importance of the subject to the individuals involved times the ambiguity of the evidence pertaining to the topic at issue". This led him to set forth a motivational question which is yet to be answered: "Can rumors be scientifically understood and controlled?" [Allport and Postman, 1946]. His 1947 experiment [Allport and Postman, 1947] reveals an interesting fact about rumour circulation and belief. He looked at how US President Franklin D. Roosevelt allayed rumours about losses at the 1941 Pearl Harbor bombing. The study showed that before the President made his address, 69% of a group of undergraduate students believed that losses at Pearl Harbor were greater than officially stated; but five days later, the President having spoken in the meantime, only 46% of an equivalent group of students believed this statement to be true. This study revealed the importance of an official announcement by a reputable person in shaping society's perception of the accuracy of a rumour. Early research focused on different objectives. Some work has looked at the factors that determine the diffusion of a rumour, including, for instance, the influence of the believability of a rumour on its subsequent circulation, where believability refers to the extent to which a rumour is likely to be perceived as truthful. Early research by Prasad [Prasad, 1935] and Sinha [Sinha, 1952] posited that believability was not a factor affecting rumour mongering in the context of natural disasters. More recently, however, Jaeger et al. [Jaeger et al., 1980] found that rumours were passed on more frequently when the believability level was high. Moreover, Jaeger et al. [Jaeger et al., 1980] and Scanlon [Scanlon, 1977] found the importance of a rumour as perceived by recipients to be a factor that determines whether or not it is spread, the least important rumours being spread more. Others have attempted to categorise different types of rumours. Knapp [Knapp, 1944] introduced a taxonomy of three types of rumours: (1) 'pipe- dream' rumours, as rumours that lead to wishful thinking, (2) 'bogy' rumours, as those that increase anxiety or fear, and (3) 'wedge-driving' rumours, as those that generate hatred. 3 The widespread adoption of the Internet gave rise to a new phase in the study of rumour in naturalistic settings [Bordia, 1996] and has taken on particular importance with the advent of social media, which not only provides powerful new tools for sharing information but also facilitates data collection from large numbers of participants. For instance, Takayasu et al. [Takayasu et al., 2015] used social media to study the diffusion of a rumour in the context of the 2011 Japan Earthquake, which stated that rain in the aftermath might include harmful chemical substances and led to people being warned to carry an umbrella. The authors looked at retweets of early tweets reporting the rumour, as well as later tweets reporting that it was false. While their study showed that the appearance of later correction tweets diminished the diffusion of tweets reporting the false rumour, the analysis was limited to a single rumour and does not provide sufficient insight into understanding the nature of rumours in social media. Their case study, however, does show an example of a rumour with important consequences for society, as citizens were following the latest updates with respect to the earthquake in order to stay safe. Some researchers have looked at how social media users support or deny rumours in breaking news stories but results are, as yet, inconclusive. In some cases it has been suggested that Twitter does well in debunking inaccurate information thanks to self-correcting properties of crowdsourcing as users share opinions, conjectures, and evidence. For example, Castillo et al. [Castillo et al., 2013] found that the ratio between tweets supporting and debunking false rumours was 1:1 (one supporting tweet per debunking tweet) in the case of a 2010 earthquake in Chile. Procter et al. [Procter et al., 2013b] came to similar conclusions in their analysis of false rumours during the 2011 riots in England, but they noted that any self-correction can be slow to take effect. In contrast, in their study of the 2013 Boston Marathon bombings, Starbird et al. [Starbird et al., 2014] found that Twitter users did not do so well in distinguishing between the truth and hoaxes. Examining three different rumours, they found the equivalent ratio to be 44:1, 18:1 and 5:1 in favour of tweets supporting false rumours. In a similar way, but expanding on the findings of earlier work, we aim to look at rumours associated with a broader set of newsworthy events, differentiating also the support and denial that comes before or after a rumour is resolved as being true or false. Friggeri et al. [Friggeri et al., 2014] and Hannak et al. [Hannak et al., 2014] used a different approach to look at how people respond to false information by looking at tweets that link to web pages that debunk certain pieces of information. While their studies are of a broader nature and cover rumours within different contexts, they only looked at false rumours that had been debunked on the fact-checking website Snopes3. Moreover, they did not look into conversational characteristics around rumours, beyond looking at users who respond with a link to Snopes. Going beyond studying the nature of rumours to gain insight into how they are spread and supported by recipients, some researchers have also worked on the 3http://www.snopes.com/ 4 development of automated techniques to categorise the type of support provided in each tweet towards the underlying rumour [Lukasik et al., 2015, Qazvinian et al., 2011]. The dataset introduced in the current work provides a suitable resource to expand research in this direction. Motivation and Contribution In the absence of research that takes a broad look at the diffusion, support and denial of rumours in social media, we set out to develop a methodology that would enable us to collect, annotate and analyse a large collection of rumours that we could then analyse for how people react to them. In contrast to previous approaches, our methodology enables gathering collections of rumours that are not necessarily known a priori, so that we can collect a diverse set of rumours with different levels of popularity. We introduce a scheme for the annotation of social media, which we apply to a collection of 330 conversation threads discussing different rumours, associated with 9 different newsworthy events. Using this dataset, we analyse rumour diffusion, support and denial in social media, looking at how they evolve from the point of being released to eventual resolution. Our study shows interesting differences between true and false rumours, one of them being that the former are typically resolved much faster than the latter. We find that social media users generally show a tendency towards supporting rumours whose veracity is yet to be resolved, which questions the validity of using the aggregated judgments of users for determining the veracity of a rumour. Users do reasonably well in distinguishing true and false rumours after they are corroborated or debunked, but aggregated judgments while a rumour is still unverified leave much to be desired. This emphasises the need for developing machine learning techniques that can provide assistance in assessing rumour veracity in real-time. The main contributions of our study include: • The development of an annotation scheme for analysing conversations around rumours in social media. This annotation scheme has been devel- oped to capture observable conversational features of how people engage with information of questionable veracity. • The application of this annotation scheme to a collection of 330 threads associated with 9 different newsworthy events. We tested and validated an annotation methodology both with an expert panel as well as through crowdsourcing. We came up with a crowdsourcing methodology for the creation of the annotated dataset described and analysed here. • The performance of a quantitative analysis of rumours in social media, looking at how they are spread, as well as how users support or deny them. To the best of our knowledge, our work is the first to perform such analysis on a broad dataset involving multiple newsworthy events and rumourous stories, and the first to consider the change in veracity status (unverified, true, false) at different points of the life cycle of a rumour. 5 Materials and Methods The way the data is collected as well as the manual annotation process constitute an important part of our methodology, allowing us to analyse conversational features of social media. Data Collection Previous research has focused on collecting data for rumours known to have been circulating a priori [Qazvinian et al., 2011, Procter et al., 2013b, Starbird et al., 2014] rather than a more general approach to the collection of new rumours. By gathering rumours known a priori, one can, for instance, search for tweets with specific keywords, e.g. 'London Eye fire', to retrieve tweets associated with the rumour that the London Eye had been set on fire in the context of the 2011 England riots. However, this approach will fail to identify rumours associated with events for which specific keywords have not been previously defined. In this study, our goal has been to extend rumour collection to encompass different types of rumourous stories that are not necessarily known a priori. We did this by collecting tweets from the streaming API relating to newsworthy events that could potentially prompt the initiation and propagation of rumours. Selected rumours were then captured in the form of conversation threads. We performed the collection of rumourous conversations in three steps. First, we collected candidate rumourous stories (signaled by highly retweeted tweets associated with newsworthy current events) from Twitter. Second, journalists in the research team selected from these candidate rumours those that met the rumour criteria [Zubiaga et al., 2015a] and identified the tweets that actually introduced them. Finally, we collected associated conversations for these rumour-introducing tweets and annotated them using an annotation scheme specifically crafted for this purpose. Identification of Rumourous Stories and Conversations Data collection. We use Twitter's streaming API to collect tweets in two different situations: (1) breaking news that are likely to spark multiple rumours and (2) specific rumours that are identified a priori. We collected tweets for nine different events, which include five cases of breaking news: • Ferguson unrest: citizens of Ferguson in Michigan, USA, protested after the fatal shooting of an 18-year-old African American, Michael Brown, by a white police officer on August 9, 2014. • Ottawa shooting: shootings occurred on Ottawa's Parliament Hill in Canada, resulting in the death of a Canadian soldier on October 22, 2014. • Sydney siege: a gunman held hostage ten customers and eight employees of a Lindt chocolate caf´e located at Martin Place in Sydney, Australia, on December 15, 2014. 6 • Charlie Hebdo shooting: two brothers forced their way into the offices of the French satirical weekly newspaper Charlie Hebdo in Paris, killing 11 people and wounding 11 more, on January 7, 2015. • Germanwings plane crash: a passenger plane from Barcelona to Dusseldorf crashed in the French Alps on March 24, 2015, killing all passengers and crew. The plane was ultimately found to have been deliberately crashed by the co-pilot of the plane. And four specific rumours (known a priori): • Prince to play in Toronto: a rumour started circulating on November 3, 2014 that the singer Prince would play a secret show in Toronto that night. Some people even queued at the venue to attend the concert, but the rumour was later proven false. • Gurlitt collection: a rumour in November 2014 that the Bern Museum of Fine Arts was going to accept a collection of modernist masterpieces kept by the son of a Nazi-era art dealer. The museum did end up accepting the collection, confirming the rumours. • Putin missing: numerous rumours emerged in March 2015 when the Russian president Vladimir Putin did not appear in public for 10 days. He spoke on the 11th day, denying all rumours that he had been ill or was dead. • Michael Essien contracted Ebola: a post by a Twitter user on October 12, 2014 stated that the AC Milan footballer Michael Essien had contracted Ebola. The report was later denied by the footballer and thus exposed as a hoax. Collection through the streaming API was launched straight after the jour- nalists identified a newsworthy event likely to give rise to rumours. As soon as the journalists informed us about a newsworthy event, we set up the data collection process for that event, tracking the main hashtags and keywords pertaining to the event as a whole. Twitter's public stream was accessed through the 'statuses/filter' endpoint4, making use of the 'track' parameter to specify keywords. Note that while launching the collection slightly after the start of the event means that we may have missed the very early tweets, we kept collecting subsequent retweets of those early tweets, making it much more likely that we would retrieve the most retweeted tweets from the very first minutes. This is also consistent with our methodology as we rely on highly retweeted tweets to identify candidate rumours. Data sampling. Given the large volume of the datasets, we sampled rumour candidates by picking tweets that sparked a high number of retweets. This is justified by our definition of rumours, which specifies that a statement has to 4https://dev.twitter.com/streaming/reference/post/statuses/filter 7 generate substantial interest and be widely circulated in order to be considered as a rumour. Depending on the event, the retweet threshold used ranged from two for specific rumours (Prince, Gurlitt, Putin, Michael Essien) to up to 100 for larger-scale events (Ferguson, Ottawa, Sydney, Charlie Hebdo). Although the crash of the Germanwings airplane in the French Alps was a significant event that generated Twitter conversations over the course of several days, a lower threshold of 25 retweets was chosen in order to capture an adequate number of tweets in German. The sampled subsets of tweets were visualised in a timeline with links to the conversation thread sparked by each of the initial (source) tweets. Next, journalists were asked to identify rumourous source tweets and non-rumourous source tweets. Journalists often need to clarify information they need to use and thus have experience separating corroborated facts from rumours. Additionally, the journalists continually tracked the nine events, so they were knowledgeable about the stories and rumours associated with them. Data annotation. We developed an annotation tool to facilitate the journalists' work of annotating tweets as rumours vs non-rumours. The visualisation of source tweets in the form of a timeline, accompanied by the context added by the conversation thread they spark, made it easier to keep track of posts as the event unfolds. While data collection was concurrent with events unfolding, the annotation process was performed a posteriori, once a consensus had emerged about the facts relating to the event in question. This enabled the journalists to be more knowledgeable about the entire story, as well as any rumours that had emerged and how they had been resolved. It also encouraged careful annotation that encompassed a broad set of rumours; the journalists could go through the whole timeline of tweets as we presented them and perform the annotations. The annotation work performed by the journalists was twofold: (a) annotation of each source tweet as being a rumour or not and (b) grouping of source tweets into stories; the latter helped us group different rumourous tweets (and associated conversation threads) as being part of the same story, as is the case with the following three tweets which are all associated with the same rumour (i.e., associated with the story "Putin is facing a palace coup"): • Is Putin facing palace coup? [link] • Can Putin's Absence Indicate A Palace Coup In Moscow? #europe #russia [link] • RUMORS about #Putin started with a stroke, then he's sick, and now this [link] What's next... #Russia #cdnpoli #dirtydeeds Once rumours were grouped into stories, the journalists also annotated whether each of the stories was ultimately proven to be false, confirmed as true, or remained unverified. For cases where the story was either true or false, they also marked within the timeline of the story the tweet that, according to their judgement, was decisive for establishing the veracity of the rumour. We refer to the latter as the resolving tweet. Note that the resolving tweet is not always available for true and false stories, as such information may sometimes 8 occur outside of our data collection time frame, e.g. weeks later or outside the Twittersphere. Fig 1 shows a diagram with the structure of rumour data produced with the annotation work performed by the journalists. Our datasets consist of rumour stories, represented by squares, which can be one of true (green), false (red), or unverified (orange). Each of the rumour stories has a number of rumour threads associated with it, which we represent as black lines that form a timeline where threads are sorted by time. When a story is true or false, the journalists also picked, within the story's timeline, one tweet as the resolving tweet. Note that the resolving tweets cannot always be found within the Twitter timeline, as in example story #5. Figure 1: Diagram showing the structure resulting from the journalists' annota- tion work. Rumour stories, represented by squares, can be one of true (green), false (red), or unverified (orange). Each of the rumour stories has a number of rumour threads associated with it (black lines). When a story is true or false, the journalists also picked, where available, one tweet as the resolving tweet within the story's timeline. Outcome of the rumour annotation task. Table 1 summarises the outcome of the annotation work, showing the number of source tweets annotated as rumours and non-rumours for each of the nine events, as well as the number of rumourous stories, where each story corresponds to a group of rumourous tweets. 9 Event name Sydney Siege Ottawa Shooting Charlie Hebdo Germanwings Ferguson Prince to play in Toronto Gurlitt Putin missing Essien has Ebola TOTAL Rumour stories Annotated threads Rumour threads Non-rumour threads 61 51 61 19 42 6 3 6 1 250 1321 901 2169 1022 1183 241 386 266 18 7507 535 475 474 332 291 237 190 143 18 2695 786 426 1695 690 892 4 196 123 0 4812 Table 1: Outcome of the annotation of rumours. For the purposes of our study, we focus on the 2,695 tweets annotated as rumourous. These rumourous tweets have been annotated in three different languages: 2,460 in English, 198 in German, and 37 in French. Complementing the Dataset with Conversations. We used the above dataset of rumourous source tweets to launch collection of the conversational threads they initiated. As a native feature of Twitter, users can reply to one another. Hence, we look for all the replies to the 2,695 rumourous source tweets for the nine events in the dataset. While Twitter does not provide an API endpoint to retrieve conversations sparked by tweets, it is possible to collect them by scraping tweets through the web client interface. We developed a script that enabled us to collect and store complete conversations for all the rumourous source tweets5. The script scrapes the Twitter web client interface, collecting the responses that appear beneath the source tweet. Once the tweets below the source tweet have been scraped, the script performs the following two steps to make sure that the whole conversation started by the source tweet is retrieved: (1) The script checks if there are more pages with responses (since Twitter pages the responses), and (2) the script then retrieves, recursively, the replies to all those replying tweets, which enables retrieval of nested interactions. This process gives us the tweet IDs of all the tweets replying to the source tweet, i.e. the conversational thread, from which we can form the whole tree. To collect all the metadata for those tweets, we then access the Twitter API using the tweet IDs; specifically, we use the 'statuses/lookup' endpoint6, which provides all the metadata for up to 100 tweets at once. The collection of conversations for the 2,695 rumourous source tweets con- tained 34,849 tweets, 2,695 being source tweets and 32,154 being responses. Annotation of Rumourous Conversations Our ultimate goal is to have rumourous conversations annotated according to the support and evidence provided by tweets responding to the rumour denoted by 5The conversation collection script is available at https://github.com/azubiaga/ pheme-twitter-conversation-collection 6https://dev.twitter.com/rest/reference/get/statuses/lookup 10 Figure 2: Example of a conversation generated by a rumourous tweet. the source tweet (response tweets). We developed an annotation scheme suitable for capturing conversation properties of the Twitter threads in terms of such interactions and used it to obtain an annotated corpus using crowdsourcing. In designing the annotation scheme we considered existing schemes for factu- ality and aspects of conversation analysis, such as turn-taking and corresponding reactions. The development of the annotation scheme has been motivated by previous research that has evolved conversation analytic techniques in order to arrive at an appropriate methodology for the sociological analysis of microblogs, especially Twitter [Tolmie et al., 2015]. This research identified a number of core organisational properties within Twitter conversations that have formed the basis of the approach reported here for identifying potentially rumourous conversation threads, i.e.: that it takes more than one conversational turn for a rumour to be identified; that Twitter conversations have a sequential order; that they involve topic management; and that the production of a response is shaped by its accountable characteristics. The development of the annotation scheme was carried out in two iterations: first by testing it with a four-person expert panel of social media researchers (three PhD students, one postdoctoral researcher) and second by running preliminary tests on a crowdsourcing platform. The annotation scheme resulting from this iterative process is shown in Fig 3. It assumes that there are two types of tweets in a conversation: (1) the source tweet that initiates a rumour, and (2) response tweets that respond within the thread 11 started by the source tweet, including both direct and nested responses. The annotation scheme consists of three main dimensions which express the mode of interaction: (1) support or response type, (2) certainty and (3) evidentiality. The annotation of source tweets and response tweets presents two variants in the case of the first component; support applies to source tweets whereas response types applies to response tweets. Each tweet in a rumourous conversation is annotated in terms of the above three dimensions, which we define in detail below: Source Tweet Support (+) Supporting (+) Denying (+) Underspecified Certainty Evidentiality Replying Tweet Response Type (x2 for deep replies) (+) Certain (+) Somewhat certain (+) Uncertain (+) Underspecified (+) Agreed (+) Disagreed (+) Appeal for more information (+) Comment (+) First-hand experience (+) URL pointing to evidence (+) Quotation of person / organisation (+) Attachment of picture (+) Quotation of unverifiable source (+) Employment of reasoning (+) No evidence Figure 3: Annotation scheme for rumourous social media conversations. 1. Support and Response Type: • Support: Support is only annotated for source tweets. It defines if the message in a source tweet is conveyed as a statement that supports or denies the content of the statement. It is hence different from the rumour's truth value and reflects the view of the author of the source tweets towards the rumour's veracity. Support takes the following values: (1) Supporting, when the author of the source tweet supports the content of the statement, (2) Denying, when denying it or (3) Underspecified, when the author's view is unclear. This feature is related to the "Polarity" feature in the factuality scheme by Saur´ı et al. [Saur´ı and Pustejovsky, 2009]. • Response Type: Response Type is used to designate the support of response tweets towards a source tweet that introduces a rumourous story. Some replies can be very helpful in determining the veracity of a rumour, and thus we annotate Response Type with one of the following four values: (1) Agreed, when the author of the reply supports the statement they are replying to, (2) Disagreed, when the author of the reply disagrees with the statement they are replying to, (3) Appeal for 12 more information, when the author of the reply asks for additional evidence in relation to the statement they are responding to, or (4) Comment, when the author of the reply makes their own comment without a clear contribution to assessing the veracity of either the tweet they are responding to or the source tweet. The inclusion of the Response Type dimension in the annotation scheme follows Procter et al. [Procter et al., 2013b], who originally introduced these four types of responses for rumours. However, unlike [Procter et al., 2013b] we additionally consider the annotation of Response Type for nested replies, i.e. tweets that are not directly replying to the source tweet. In this case Response Type is annotated for two different aspects: (i) the type of response with respect to the rumour in the source tweet and (ii) the type of response towards its parent tweet, i.e. the tweet it is directly replying to. This double annotation allows us to better analyse the way conversations flow and how opinions evolve with respect to veracity. It is worth noting that the response type is not necessarily transitive, and the aggregation of pairwise agreements and disagreements with previous tweets does not necessarily match with the agreement with the source. For the purposes of our analysis here, we simplify a tweet's response type as its support with respect to the rumour; we infer this from the tweet's support with respect to the source, which can be one of these four cases: (1) if the source tweet supports a rumour and the response tweet agrees, we consider that the tweet supports the rumour, (2) if the source tweet supports a rumour and the response tweet disagrees, we consider that the tweet denies the rumour, (3) if the source tweet denies a rumour and the response tweet agrees, we consider that the tweet denies the rumour, and (4) if the source tweet denies a rumour and the response tweet disagrees, we consider that the tweet supports the rumour. Figure 4 shows an example of the double annotation, and how we determine the support towards the rumour. Figure 4: Example of annotation of rumour type, as well as how we determine support towards the rumour. 2. Certainty: Certainty measures the degree of confidence expressed by the 13 author of a tweet when posting a statement in the context of a rumour. It applies to both source tweets and response tweets. The author can express different degrees of certainty when posting a tweet, from being 100% certain, to considering it as a dubious or unlikely occurrence. The value annotated for either Support or Response Type has no effect on the annotation of Certainty, and thus it is coded regardless of the statement supporting or denying the rumour. The values for Certainty include: (1) Certain, when the author is fully confident or the author is not showing any kind of doubt, (2) Somewhat certain, when they are not fully confident and (3) Uncertain, when the author is clearly unsure. This corresponds to the Modality feature of Saur´ı et al. [Saur´ı and Pustejovsky, 2009]. 3. Evidentiality: Evidentiality determines the type of evidence (if any) provided by an author of a tweet and applies to both source tweets and response tweets. It is important to note that the evidence provided has to be directly related to the rumour being discussed in the conversation and any other kind of evidence that is irrelevant in that context is not annotated here. Evidentiality can have the following values: (1) First-hand experience, when the author of the tweet claims to have witnessed events associated with the rumour (2) Attachment of a URL pointing to evidence, (3) Quotation of a person or organisation, when an accessible source is being quoted as a source of evidence, (4) Attachment of a picture, (5) Quotation of an unverifiable source, when the source being mentioned is not accessible, such as "my friend said that...", (6) Employment of reasoning, when the author explains the reasoning behind their view and (7) No evidence, when none of the other types of evidence is given in the tweet. Contrary to the dimensions described above, more than one value can be picked for Evidentiality, with the exception of "No evidence" which excludes the selection of any other value. Hence, we cater for the fact that a tweet can provide more than one type of evidence, e.g. quoting a news organisation while also attaching a picture that provides evidence. All three of the above dimensions are annotated for each tweet in a conver- sation thread, with one exception; response tweets whose Response Type has been annotated as Comment, do not need to be annotated for Certainty and Evidentiality. Since the annotation of all 2,695 conversations we collected was not viable due to time and budget constraints, we randomly sampled a subset of the data. The sampled dataset includes 330 rumourous conversations, 297 of which are in English and 33 in German. This amounts to 4,842 tweets overall, 4,560 in English and 282 in German. The conversations are distributed across the nine events as follows: Charlie Hebdo (74), Sydney siege (71), Ottawa shooting (58), German- wings crash (53), Ferguson unrest (46), Putin missing (13), Prince/Toronto (12), Ebola/Essien (2), and Gurlitt (1). These 330 conversations were categorised into 140 different stories. Out of these 330 conversations, 159 are true, 68 are false and 103 remained unverified. 14 132 of the conversations were part of one of the 41 stories that have a resolving tweet annotated. The annotation of the sampled rumourous conversation threads we collected was performed through crowdsourcing in order to ensure its timely comple- tion [Procter et al., 2013a]. For crowdsourcing, we used the CrowdFlower platform7, because of its flexibility, which allowed us to refine the annotation work and specify prerequisites for annotators. Further details on how we split the annotation into individual tasks and the settings chosen for the crowdsourcing tasks can be found in [Zubiaga et al., 2015a], where we validated the suitability of the approach, including both the annotation scheme, as well as the crowdsourcing jobs, on a small subset of rumours. The annotation of all 330 conversation threads consisted of 68,247 judgments on 4,842 tweets performed by 233 different annotators. We combined all judg- ments for each tweet-dimension pair through majority vote. These tweets are annotated for the three features described in our annotation scheme: support or response type (which is annotated twice for deep replies), certainty, and evidentiality. A comparative analysis between the three features shows that they are independent from each other, with low pairwise Pearson correlation values: -0.31 between support and certainty, -0.15 between support and evidentiality, and -0.04 between certainty and evidentiality. To quantify the difficulty of the task, we measured the inter-annotator agreement values by comparing each individual judgment against the majority vote. Overall, the annotators achieved an agreement rate of 62.3%, which is distributed differently across different tweet types and dimensions. Table 2 shows how the agreement rates are distributed for source and response tweets when annotating for Support, Certainty and Evi- dentiality. Our findings show that annotators found it easier to annotate source tweets, as they are less ambiguous and require less context for understanding. The agreements are somewhat lower for response tweets. When we compare the different dimensions, we observe that Support is the easiest to annotate for source tweets, but very similar to Certainty overall. Evidentiality is slightly more difficult to annotate for both source and response tweets, presumably because of the large number of different values that the annotators can choose from. Source tweets Replies Overall Support Certainty Evidence 81.1% 62.2% 63.7% 74.9% 58.3% 60.8% 68.8% 59.8% 61.1% Table 2: Inter-annotator agreement values for different features and tweet types. The present study was approved by the Warwick University Humanities & Social Sciences Research Ethics Committee (HSSREC), which confirmed that the project, "Computing Veracity Across Media, Languages and Social Net- works (PHEME)", received full ethics approval (ref 69/13-14, date: 30.05.2014), 7http://www.crowdflower.com/ 15 including approval to publish extracts from social media datasets. The resulting dataset is available for research purposes8. We believe that the methodology described above, along with the conversation collection software, should enable reproducibility of similar data collections for future events. Results We begin by investigating the diffusion of rumours in our corpus and then move on to analyse the annotations for Support, Certainty, and Evidentiality for both source and response tweets. We conclude by exploring attributes of users that participate in rumour diffusion. Rumour Timelines Figure 5 shows the timelines for the rumours collected and annotated for the nine events in our dataset. Rumours are coloured according to their veracity status at each moment, as determined by the journalists through the annotation of resolving tweets. All of the rumours start off in an unverified status (orange), while some of them are later proven true (green) or found to be false (red); others, however, remain unverified, and are featured in orange through the entire timeline. These charts visualise, in all, 103 rumours which remain unverified, 159 which were later proven, and 68 which were found to be false. Out of the 227 rumours that turned out to be either true or false, we have a resolving tweet for 129 rumours (56.8%), while for the other 98 cases, the journalists could not find a resolving tweet. Some of these may have been resolved outside of Twitter, outside of the collected data, or later in time out of our data collection and annotation time frames; these are coloured in orange. The figures show variability in the duration of the unverified status of rumours; some rumours are resolved quite quickly while others take much longer to be verified. They also highlight the large number of unverified rumours in the case of both the Ferguson unrest and the Charlie Hebdo shooting. On the other hand, more rumourous stories ended up being verified in events like the Ottawa shooting or the Sydney siege. While the distribution of veracity status varies across different events, the number of rumours is generally large but the number of rumours of unverified status is remarkable. When we look at the delay between a rumour being posted for the first time on Twitter and the rumour being resolved as being either true or false, we observe that false rumours tend to be resolved significantly later than true rumours (Wilcoxon signed-rank test [Wilcoxon, 1945]: W = 812.5, p < 2.2e−16). While the median true rumour is resolved in about 2 hours, the median false rumour takes over 14 hours to be resolved. Figure 6 illustrates this large difference in the time it takes for true and false rumours to be resolved. While the vast majority of true rumours are resolved within 5 hours of being posted for the first time, 8https://figshare.com/articles/PHEME_rumour_scheme_dataset_journalism_use_ case/2068650 16 Figure 5: Rumour timelines showing the lifecycle of rumours that start as unverified stories (orange), and are occasionally later resolved as being either true (green), or false (red). Each line represents a rumour story, while the X axis represents the timeline. many of the false rumours are not resolved within the first 10 hours; indeed for those that are eventually solved, this does not happen until 15 to 20 hours after the timestamp of the source tweet introducing the rumour. Rumour Diffusion To complement the analysis enabled by the visualisation of rumour timelines, we take a closer look at the diffusion of these rumours in the form of retweets. Figure 7 shows networks of interactions between users, the connections being coloured according to their accuracy, as we describe next. Note that eight events are shown here, given that the other event, Gurlitt, did not spark sufficient retweets to create an equivalent visualisation. Here we consider a connection between two users a,b if a user a retweets a user b's rumourous tweet. Note that 17 Ottawa shootingSydney siegeFergusonGermanwings plane crashCharlie Hebdo shootingPutin missingEssien / ebolaPrince / TorontoGurlittFigure 6: Distribution of delays (in hours) in resolving false (red) and true (green) rumours. Horizontal lines represent 25, 50, and 75 percentiles. retweets of rumourous source tweets are used here for the analysis, to explore how different rumours are spread. We colour these connections based on the accuracy of the original tweets being retweeted: • Blue (accurate retweets): the retweets of tweets that are either sup- porting true rumours, or denying false rumours. • Brown (inaccurate retweets): the retweets of tweets that are wrong, i.e., either supporting false rumours, or denying true rumours. • Orange (unverified retweets): the retweets of tweets that still have an unverified status. Hence, these retweet networks visualise the degree of diffusion generated by accurate, inaccurate, and unresolved tweets. Note that each of these networks visualises all the interactions with rumours in each of the events, without considering different points in time. In these figures, we can observe that tweets reporting unverified rumours are more widely spread; the percentages of unverified tweets range from 30.73% of the retweets in Ebola/Essien, to 100% of the retweets in the Putin missing story. Retweets of inaccurate reports are especially remarkable in the Ferguson unrest (26.78%) and the Ebola/Essien hoax (69.27%). Retweets of accurate tweets can only be observed in Prince/Toronto (23.36%), Germanwings plane crash (33.32%), Ottawa shootings (35.76%), and the Sydney siege (26.64%). Table 3 shows the percentages of retweets of each type of tweet (unverified, accurate, inaccurate) for each of the events. We further examine these retweet networks by looking at the time in which the retweets occur. We want to know the extent to which each type of rumour is 18 051015202530False RumoursTrue Rumoursllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll051015202530False RumoursTrue RumoursFigure 7: Retweet networks representing retweets of unverified source tweets (orange), accurate tweets that support true rumours or deny false rumours (blue) and inaccurate tweets that deny true rumours or support false rumours (brown). retweeted, as well as whether there is a time-effect on rumour diffusion patterns, e.g. with some rumours being retweeted more at the beginning and tailing off as time passes. Figure 8 shows the average distribution of retweets for different rumourous stories (true, false) over time. The plots are normalised to show the percentages of retweets, computed every 15 minutes, representing a ratio of the total amount of retweets for that tweet. That is, once a rumourous tweet is posted, we look at the ratio of retweets that occur over time within 15 minute windows. A high percentage of retweets at a certain point in time represents that, on average, that type of source tweet receives a high number of retweets at that stage of its lifecycle. To enable a better comparison we make a distinction between tweets that occur before and after resolving tweets (post- and pre-), tweets that are part of a true or false rumour (true, false), and tweets that support or deny a rumour (supporting, denying). Note that one chart is missing, the one corresponding to the combination pre-denying-false, as there are no such 19 Unverified Accurate Inaccurate Ebola/Essien Putin missing Charlie Hebdo Prince/Toronto Ferguson Germanwings Ottawa shootings Sydney siege 30.73% 100.00% 96.22% 67.29% 73.22% 60.24% 55.02% 71.72% 0.00% 0.00% 0.00% 23.36% 0.00% 33.32% 35.76% 26.64% 69.27% 0.00% 3.78% 9.35% 26.78% 6.44% 9.23% 1.64% Table 3: Percentages of retweets for unverified, accurate, and inaccurate tweets. instances in our annotated data. We can observe that tweets posted before resolving tweets tend to spark a large number of retweets in the very first minutes, a trend which quickly fades in less than 20 minutes. After the occurrence of resolving tweets, retweets are more evenly distributed over time, which reveals that post-tweets keep being retweeted for a longer time. Hence, early tweets before a rumour is settled are highly spread as soon as they are posted; however, even after a rumour is resolved the rumour-bearing source tweets are retweeted for a while after. Extending this analysis, Figure 9 compares the number of retweets that each type of tweet receives overall, compared with the average number of retweets that all rumours get. The point of this check is to establish if certain types of tweets spark more retweets. Given that different events and rumours garner very different numbers of retweets, here we use z-scores to normalise the retweet counts across events and rumours. The z score for a given value x is given by the formula: x − µ σ z = (1) where µ is the mean of the population, and σ is the standard deviation of the entire population under consideration. The z-score represents, therefore, the number of standard deviations that a population is above or below the global average. Interestingly, the types of tweets that spark more retweets are, by far, early tweets supporting a rumour that is still unverified ("pre-supporting") (W = 118760, p < 2.2e − 16). Among these, there is also a significant difference between true and false rumours; those that end up being true are retweeted more than the false ones (W = 6145.5, p = 0.02678). The number of retweets of tweets supporting these stories ("post-supporting"), both true and false, drops dramatically after the rumours are resolved, suggesting that the rumour's interest decreases once its veracity value is known. The tweets that are retweeted the least, however, are those denying a rumour, irrespective of whether the stories are true or false and whether or not they have been corroborated. This analysis reveals an interesting pattern in behaviour, showing that users tend to support 20 Figure 8: Retweet timelines showing the percentage of retweets that each type of tweet gets in 15 minute steps. Higher retweet percentages at the beginning represent a high interest in spreading the tweet in the very first minutes. unverified rumours (whether explicitly or implicitly), potentially due to the arousal that these early, unverified stories produce and their potential societal impact. Users do not seem to make the same effort, however, to spread false rumour debunks to let others know about false reports. While the evidence from retweets suggests that social media users are not very good at distinguishing between true and false stories when retweeting rumourous tweets, the actual tweet responses to rumourous tweets may reflect a different picture. In the next section we investigate the level of support and denial observed in the replies to rumourous tweets. Rumour Support and Denial The crowdsourced annotations, which manually categorise each of the source and response tweets according to the type of support expressed with respect to the rumour, enable us to analyse the performance of social media users in terms of support and denial of rumours. Specifically, our annotation scheme categorised each tweet as supporting a rumour, denying it, appealing for more information, or making a comment. For a simplified analysis of rumour support and denial, we omit comments, which do not contribute to resolving the veracity of a rumour. We consider supporting tweets on the one hand and, on the other, we combine denying tweets and appeals for more information into denials. Having these 21 0306090120150Post−denying−falseTime (minutes)Retweets (%)0102030405060700306090120150Post−denying−trueTime (minutes)Retweets (%)0102030405060700306090120150Post−supporting−falseTime (minutes)Retweets (%)0102030405060700306090120150Post−supporting−trueTime (minutes)Retweets (%)0102030405060700306090120150Pre−denying−trueTime (minutes)Retweets (%)0102030405060700306090120150Pre−supporting−falseTime (minutes)Retweets (%)0102030405060700306090120150Pre−supporting−trueTime (minutes)Retweets (%)010203040506070Figure 9: Z-score values for retweets of different types of tweets, representing the extent to which different types of tweets receive retweets below or above the overall average of retweets that all rumours get. two types of response tweets, i.e., supports and denials, we define the support ratio as the ratio of supporting tweets over those that are either supporting or denying a rumour, so that we can normalise the support and make it comparable across rumours and events. The equation that determines the support ratio for a rumour is defined in Equation 2. supportratio = #support #support + #denials (2) We look at the support ratio observed for both true and false rumours, distinguishing between ratios before and after the resolving tweet is posted, so that we can also explore the effect of the resolving tweet on public opinion. Figure 10 shows support ratios before and after resolving tweets for rumours that turn out to be true or false. We observe that true rumours tend to have a slightly higher ratio of support (median = 0.156) than false rumours (median = 0.067) before the resolving tweet appears, which is statistically significant (W = 1099.5, p = 0.02014). However, the fact that both median values are positive and the difference between the two populations in this case is relatively small, suggests that the aggregation of collective intelligence may not be sufficient to determine the veracity of a rumour, at least not before the resolving tweet is posted. While rumours remain unverified, the overall tendency is to support them, or to assume that they are true, irrespective of their actual truth value, as shown by the prevalence of positive support ratios in these cases. Surprisingly, the support ratio decreases after the resolving tweet for both true and false rumours, in relation to the support ratio observed before the advent of the resolving tweet. Furthermore, there is a larger difference in terms 22 Pre−supporting−falsePre−supporting−falsePost−supporting−falsePost−supporting−falsePre−denying−falsePre−denying−falsePost−denying−falsePost−denying−false−0.4−0.20.00.20.40.60.8of the support ratio between true and false rumours after the advent of the resolving tweet, with a median of 0.02083 for true rumours, and a median of -0.0723 for false rumours, although without sufficient statistical significance (W = 220, p = 0.07854). The support ratio for unverified stories is very similar to the support ratio observed for false and true rumours prior to their resolution. The support ratio for unverified stories tends to be low but slightly in favour of supporting them (median = 0.04348). This confirms the overall tendency to support unverified stories. Figure 10: Distribution of support ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles. As an additional step here, we look at the number of cases in which the amount of support exceeds denial, those in which denial beats support, and those in which they have equal weight. Figure 11 visualises on a heatmap the number of rumour stories where there are either more supporting or more denying tweets, there is a tie, or there is not a single tweet questioning it or backing it up. This again buttresses our previous observation that most people tend to support rumourous stories and few people question or deny them. In fact, the chart shows that the only case in which the amount of denial beats support is that of false stories, after the resolving tweet has been posted. Apart from looking at the support ratio, we also analyse the amount of discussion that each type of rumour (true, false, pre and post resolution) produces. We define the discussion ratio as the sum of supporting and denying tweets divided by all the tweets, hence referring to the fraction of tweets that discuss the veracity of a rumour either positively or negatively (see Equation 3). discussionratio = #support + #denials #support + #denials + #comments (3) Figure 12 compares the amount of discussion that different types of rumourous 23 −0.50.00.5False, UnresolvedFalse, Debunked−0.50.00.5True, UnverifiedTrue, Verified−0.50.00.5Unverified RumoursFigure 11: Support heatmap representing the number of rumours that spark (1) more support, (2) more denials, (3) the same number of each, or (4) neither. tweets spark before and after resolving tweets. We use z-scores again to normalise the values and make them comparable to one another for rumours corresponding to different events. Here we observe that there is more discussion after the veracity of a rumour is resolved, especially for false rumours. Unverified stories all have negative z-scores, irrespective of the stories later being proven true or false, or remaining unverified. However, positive z-scores can be observed when the resolving tweet has been made public. Interestingly, the fact that the resolving tweet has been posted produces a higher number of tweets discussing the veracity of the rumour, either positively or negatively. To further quantify the amount of support and denial produced in these situations, we look next at the support ratio. Figure 13 shows the support ratio that different types of tweets (true, false, pre and post resolution) receive. The large increase of discussion for false stories once veracity is resolved, as observed above, turns into denying messages. This suggests that users are good at denying already debunked rumours, but are not so good at handling unresolved rumours. When it comes to true rumours, users do quite well in determining that an unverified rumour is true. However, and surprisingly, the slight increase in discussion after a true rumour is verified turns into a higher degree of denial. We believe that this is due to the fact that supporting messages are no longer necessary, but messages from users remaining skeptical feature more prominently in the discussions. Rumour Certainty Certainty, as we annotated it in the context of rumours, measures the degree of confidence expressed by the author of a tweet. We look at the degree of certainty expressed by users in the context of rumours of different veracity statuses to 24 moresupportmoredenialstienothingFalse,unresolvedFalse,debunkedTrue,unverifiedTrue,verifiedUnverifiedFigure 12: Z-score values for the amount of discussion generated by different rumours, representing the extent to which different rumours spark discussion below or above the overall average. Red corresponds to false rumours, green to true rumours and yellow to unverified rumours. check how it changes and whether it increases as rumours get resolved. To compute the certainty ratio observed in each of the rumourous stories, we simplify the annotations of certainty and distinguish two groups: (1) certain, which includes annotations of tweets belonging to a rumour as 'certain', and (2) not-entirely-certain, which includes annotations of tweets as 'somewhat certain' and 'uncertain'. Given these two values, the certainty ratio is computed as: certaintyratio = #certain #certain + #not-entirely-certain (4) Figure 14 shows the distributions of certainty ratios by veracity status (true, false, unverified). Certainty appears to be stable before and after resolving tweets for true rumours; with a median of 0.575 before being resolved, and a median of 0.563 afterwards, there is no significant difference in the degree of certainty after a rumour has been verified (W = 848, p = 0.345). However, we do observe a larger difference in the case of false rumours, where the median is 0.618 before resolution, and 0.556 thereafter; however, this difference cannot be considered to be statistically significant (W = 224.5, p = 0.2384). In the rest of the cases, i.e., unverified rumours, the certainty level (median = 0.571) is very similar to true rumours eventually verified as true. 25 Pre−discussion−falsePost−discussion−falsePre−discussion−truePost−discussion−truePre−discussion−unverified−0.2−0.10.00.10.2Figure 13: Z-score values for support ratios observed in different rumours, representing the extent to which different rumours get support below or above the overall average. Rumour Evidentiality Evidentiality in the context of a rumour determines the type of evidence (if any) provided by an author of a tweet. We look next at the extent to which users provide evidence with their statements for different veracity statuses. To compute the evidentiality ratio in rumourous stories, we simplify the annotations of evidence for individual tweets within a rumourous thread as follows: (1) 'evidence', which applies to a tweet that provides any kind of evidence, and (2) 'no-evidence', for tweets that have no evidence attached. Given these two values, we compute the evidentiality ratio as: evidentialityratio = #evidence #evidence + #no-evidence (5) Figure 15 shows the distributions of evidentiality ratios by veracity status. The values for evidentiality associated with rumours that are as yet unverified is very similar, irrespective of whether the rumour is proven to be false, true, or remains unverified. However, there is a noticeable drop of evidentiality after a rumour is resolved. This is especially noticeable for false rumours. Interestingly, most tweets provide supporting evidence when a false rumour is still unverified (median = 0.8944), but this number drops dramatically once the rumour is proven 26 Pre−support−falsePost−support−falsePre−support−truePost−support−truePre−support−unverified−1.0−0.50.00.5Figure 14: Distribution of certainty ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles. false (median = 0.667); this difference is statistically significant (W = 269.5, p = 0.01382). The drop in evidentiality after rumour resolution is smaller for true rumours, whose median value is 0.853, while unverified; this drops slightly to 0.8 after the advent of the resolving tweet; however, this difference cannot be deemed statistically significant (W = 932, p = 0.07333). The significant drop in evidentiality associated with false rumours once they are debunked suggests a shift in the focus of the discussion from trying to disprove a rumour. Analysis of Users Next, we examine the role that different users play in the diffusion and support of rumours. In our rumour datasets, a total of 56,099 different users participated: 208 posted rumourous source tweets, 3,129 responded to the rumours, and 53,469 retweeted the rumours. We rank users by the amount of followers they have, which, as we argue below, can be used to provide a measure of the reputation of the user. However, the number of followers might somehow be rigged, e.g., by users who simply follow many others to attract more followers. To control for this effect, we define the follow ratio as the logarithmically scaled ratio of followers over followees (see Equation 6). log10 #f ollowers #f ollowing (6) Hence, the larger the number of followers a user has in relation to the number of users they are following, the larger the follow ratio will be. To assess the validity of the follow ratio as a reputation measure, we took a closer look at users with a high follow ratio. Within the possible follow ratio values, which ranges from -1 to 7 in our dataset, we looked at the 42 users with 27 0.00.20.40.60.81.0False, UnresolvedFalse, Debunked0.00.20.40.60.81.0True, UnverifiedTrue, Verified0.00.20.40.60.81.0Unverified RumoursFigure 15: Distribution of evidentiality ratios before and after resolving tweets for true and false rumours, as well as for rumours that remain unverified. Horizontal lines represent 25, 50, and 75 percentiles. a follow ratio of 4 or higher. Among these 42 highly reputable users, 38 are news organisations, 2 are 'parody accounts' (e.g. accounts self-evidently set up to make fun of or 'spoof' real people or organisations), 1 is an activist, and 1 is an actor. Hence, the majority of users with high follow ratios in the context of newsworthy events are news organisations, which validates the follow ratio as a reputation measure for our purposes. Figure 16 shows the statistics for the average follow ratio of users who express support, certainty and evidentiality for rumours at different veracity status levels. Here we observe that users who support rumours, irrespective of whether they are ultimately true or false, tend to have a substantially higher follow ratio compared to those users who challenge them. On the contrary, users who deny rumours seem to have a lower follow ratio, irrespective of whether the story is ultimately true or false. This divergence can actually be observed both while rumours are still unverified, as well as later, when their veracity has been resolved. When we look at the relationship between follow ratio and certainty, we see that users who tend to be certain are those with a higher follow ratio, whereas users with lower follow ratios tend to be uncertain. This seems to be the case for both true and false rumours, although uncertainty is expressed by users with a high follow ratio in the case of false rumours prior to their resolution. Looking at the types of evidence that users of different levels of reputation attach to their tweets, we can see some interesting patterns that differentiate users with low and high follow ratios. In fact, users with high follow ratios tend to quote external pieces of evidence to back up their own posts, such as a link to a news story, while users with a low follow ratio tend to either not provide any evidence at all or rely on self-assertion. Given that the majority of users in our sample with a high follow ratio are news organisations, these results do not seem very surprising. One the one 28 0.00.20.40.60.81.0False, UnresolvedFalse, Debunked0.00.20.40.60.81.0True, UnverifiedTrue, Unverified0.00.20.40.60.81.0Unverified Rumourshand, we would expect journalists to orientate to the need to display compliance with professional values in reporting stories, which is consistent with quoting sources, even if these sources themselves may not be entirely trustworthy [Cotter, 2010]. On the other hand, the pressure to produce something in the face of an over-abundance of information and the presence of numerous near-identical stories from competing news organisations means that verification of a rumour may take second place to the need to publish [Boczkowski, 2010]. Figure 16: Average follow ratios for users who express different types of support, certainty and evidentiality, along with error bars that represent the 95% confidence interval. Apart from the user follow ratio, we also considered other factors to distinguish between users, such as user age, whether or not they are verified users, or the number of times they tweet, but found no significant differences. Discussion and Conclusions The methodology we have developed for collecting, identifying, and annotating rumour threads has enabled us to analyse a number of important aspects of how people react to rumours using measures of diffusion and in the context of responses that either support or deny the veracity of the rumours. The application of this methodology to 330 rumour threads associated with 9 different newsworthy events and 4,842 tweets has helped us discover behavioural patterns that can be observed in the context of rumours. To the best of our knowledge, this is the first work to perform such a comprehensive study on a diverse set of rumours. We consider conversational aspects of rumours and also differentiate between two parts of the rumour cycle, i.e., prior to the resolution of its veracity status and after a rumour has been debunked or verified. To summarise what has come out of the study results: 29 0.00.51.01.5false−external−quotation−postfalse−external−quotation−prefalse−no−evidence−postfalse−no−evidence−prefalse−self−quotation−postfalse−self−quotation−pretrue−external−quotation−posttrue−external−quotation−pretrue−no−evidence−posttrue−no−evidence−pretrue−self−quotation−posttrue−self−quotation−preAverage follow ratioEvidentiality by follow ratio0.00.51.0false−certain−postfalse−certain−prefalse−not−entirely−certain−postfalse−not−entirely−certain−pretrue−certain−posttrue−certain−pretrue−not−entirely−certain−posttrue−not−entirely−certain−preAverage follow ratioCertainty by follow ratio0.00.51.01.5false−denying−postfalse−denying−prefalse−supporting−postfalse−supporting−pretrue−denying−posttrue−denying−pretrue−supporting−posttrue−supporting−preAverage follow ratioSupport by follow ratioTrue rumours tend to be resolved faster than false rumours. Our study shows a significant difference in terms of the time it takes before true and false rumours are resolved respectively. While most true rumours are corroborated within 2 hours of the source tweet that introduces them, false rumours tend to take about 14 hours before they are debunked. This observation aligns with common sense and buttresses earlier conjectures according to which proving that a fact is not accurate is far more difficult than proving it is true [Saunders, 1984]; providing counter-evidence that disproves a statement is indeed more challenging. Rumours in their unverified stage produce a distinctive burst in the number of retweets within the first few minutes, substantially more than rumours proven true or false. When we looked at the diffusion of rumours in the form of retweets, we observed a clear tendency among users to share rumours that are still in the first phase of their life cycle, prior to them being resolved. Our analysis shows that tweets reporting nverified rumours spark a sudden burst of retweets within the very first minutes, showing that users tend to share early, unverified reports rather than later confirmations or debunks. This might seem somewhat surprising at first, as one would expect people to be more interested in the final resolution of a rumour. However, this observation aligns with the nature of rumours, which by definition produce increased levels of anxiety, increasing, in turn, the likelihood of passing on information to others. Our analysis suggests that interest in spreading a story decreases after its veracity value is resolved, especially when it is proven false. This finding also aligns with investigations of the reportability of newsworthy information in the conversation analytic literature [Sacks, 1992]. The prevalent tendency of users is to support every unverified rumour. Interesting insights into rumours are also obtained from our analysis of discussion in the form of supporting and denying tweets. When we considered both support and denial tweets together, which we referred to as discussion, we observed that the amount of discussion increases substantially once a rumour is resolved, irrespective of whether it is corroborated or debunked. When distinguishing between support and denials, we observe that it is denial which increases after rumour resolution, and support rather drops off as a rumour is resolved, irrespective of whether the story turns out to be true or false. It is surprising that an increase in messages denying a rumour not only occurs with rumours that have been debunked, but also in the case of those confirmed to be true. This suggests that those who trust the confirmation no longer feel the need to insist, while a number of skeptical users emerge rebutting the resolution of the rumour. It is worth noting that the increase in messages denying a rumour does not usually imply that they numerically exceed supporting messages. In fact, the only case in which denying messages outnumber supporting messages is when a false rumour has been debunked; this does not occur with false rumours while they are still unverified. Overall, the prevalent tendency of users is to support every unverified rumour, potentially because of the lack of counter-evidence. While the level of certainty does not change over the course of the rumour lifecycles, users provide evidence in their tweets when 30 rumours are yet to be resolved, but less so after resolution. Apart from looking at whether tweets were supporting or denying rumours, we also looked at certainty – the level of confidence expressed by the author of a tweet within the conversation thread – and evidentiality – whether there is evidence linked to tweets within a rumourous conversation thread. We found that there is not much of a difference in terms of certainty for different veracity statuses of rumours (true, false, unverified), other than debunked rumours being associated with a slightly lower level of certainty. This is potentially due to posts by skeptics, but the finding was not statistically significant. In terms of evidentiality, there is a significant drop in its presence in tweets after a false rumour has been debunked, in comparison to tweets posted at an earlier stage, when the rumour was still unverified. Highly reputable users such as news organisations tend to sup- port rumours, irrespective of them being eventually confirmed or debunked, tweet with certainty and provide evidence within their tweets. Finally, we have also looked at different characteristics of users to find out whether different types of users react differently to rumours. Among the characteristics studied, we found significant differences when separating users by follow ratio – i.e., the scaled ratio of the number of a user's followers to the number of accounts the user follows. Our analysis shows that users with high follow ratios are more likely to: (1) support any rumour, irrespective of its truth value; (2) be certain about their statements and (3) attach evidence to their tweets by quoting an external source. On the other hand, users with low follow ratios are more likely to: (1) deny rumours, irrespective of their actual truth value; (2) be rather uncertain about their statements and (3) either provide no evidence in their tweets, or provide evidence on the basis of their own experience, opinions or observations. In our context of newsworthy events the majority of users with high follow ratios are news organisations, which we might reasonably assume are at pains to act in ways that are consistent with professional standards when it comes to news reporting and so justify their readers' trust. Hence, we observe that news organisations tend to make well-grounded postings; they usually do so with certainty and often quote an external source as evidence. We also observe that despite their evident endeavour to make well-grounded statements, news organisations may nevertheless post rumours that are eventually proven false. This finding is not necessarily inconsistent with acting to maintain professional reputation, but reflects how this has to be balanced against the pressures journalists may face to publish, even when stories and sources are unverified. Hence, quoting a source, whether verified or not, represents a minimum threshold for journalists to be seen as adhering to professional standards and so be judged to be trustworthy. This finding, in fact, buttresses the observations on digital media ethics by Ward [Ward, ], who posited that values in on-line journalism have shifted from the tradition of 'accuracy, pre-publication verification, balance, impartiality, and gate-keeping' to 'immediacy, transparency, partiality and post- publication correction'. This has brought about 'a tension between traditional journalism and on-line journalism'. 31 Future work. Our study sheds new light on human reactions to rumours and provides insights towards better understanding of how to control the spread of rumours. One of our next objectives is to use the datasets we have collected to develop machine learning techniques for automatically predicting rumour veracity at different time points and using conversational patterns to help establish the veracity of rumours. In this way, we believe it will be possible to assist journalists to check their sources more quickly and thoroughly, and thus enable them to play a more effective role in mitigating the spread and impact of false rumours. We would also like to verify whether our observation of the tendency for users with high follow ratios to behave more responsibly can generalise from news organisations to other types of users with high follow ratios. Acknowledgments This work has been supported by the PHEME FP7 project (grant No. 611233). References Allport and Postman, 1946. Allport, G. W. and Postman, L. (1946). An analysis of rumor. Public Opinion Quarterly, 10(4):501–517. Allport and Postman, 1947. Allport, G. W. and Postman, L. (1947). The psychology of rumor. Boczkowski, 2010. Boczkowski, P. J. (2010). News at work. Chicago, IL: The University of Chicago. Bordia, 1996. Bordia, P. (1996). Studying verbal interaction on the inter- net: The case of rumor transmission research. Behavior Research Methods, Instruments, & Computers, 28(2):149–151. Castillo et al., 2013. Castillo, C., Mendoza, M., and Poblete, B. (2013). Pre- dicting information credibility in time-sensitive social media. Internet Re- search, 23(5):560–588. Cotter, 2010. Cotter, C. (2010). News talk: Investigating the language of journalism. Cambridge University Press. Derczynski and Bontcheva, 2014. Derczynski, L. and Bontcheva, K. (2014). Pheme: Veracity in digital social networks. Proceedings of ISA. DiFonzo and Bordia, 2007. DiFonzo, N. and Bordia, P. (2007). Rumor, gossip and urban legends. Diogenes, 54(1):19–35. Donovan, 2007. Donovan, P. (2007). How idle is idle talk? one hundred years of rumor research. Diogenes, 54(1):59–82. 32 Friggeri et al., 2014. Friggeri, A., Adamic, L. A., Eckles, D., and Cheng, J. (2014). Rumor cascades. In Proceedings of the AAAI Conference on Weblogs and Social Media. Hannak et al., 2014. Hannak, A., Margolin, D., Keegan, B., and Weber, I. (2014). Get back! you don't know me like that: The social mediation of fact checking interventions in twitter conversations. In Proceedings of the AAAI Conference on Weblogs and Social Media. Jaeger et al., 1980. Jaeger, M. E., Anthony, S., and Rosnow, R. L. (1980). Who hears what from whom and with what effect a study of rumor. Per- sonality and Social Psychology Bulletin, 6(3):473–478. Knapp, 1944. Knapp, R. H. (1944). A psychology of rumor. Public Opinion Quarterly, 8(1):22–37. Kwak et al., 2010. Kwak, H., Lee, C., Park, H., and Moon, S. (2010). What is twitter, a social network or a news media? In Proceedings of the World Wide Web Conference, pages 591–600. ACM. Lukasik et al., 2015. Lukasik, M., Cohn, T., and Bontcheva, K. (2015). Clas- sifying tweet level judgements of rumours in social media. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Pro- cessing, pages 2590–2595, Lisbon, Portugal. Association for Computational Linguistics. Prasad, 1935. Prasad, J. (1935). The psychology of rumour: A study relating to the great indian earthquake of 1934. British Journal of Psychology. General Section, 26(1):1–15. Procter et al., 2013a. Procter, R., Housley, W., Williams, M., Edwards, A., Burnap, P., Morgan, J., Rana, O., Klein, E., Taylor, M., Voss, A., et al. (2013a). Enabling social media research through citizen social science. EC- SCW 2013 Adjunct Proceedings, 3. Procter et al., 2013b. Procter, R., Vis, F., and Voss, A. (2013b). Reading the riots on twitter: methodological innovation for the analysis of big data. International Journal of Social Research Methodology, 16(3):197–214. Qazvinian et al., 2011. Qazvinian, V., Rosengren, E., Radev, D. R., and Mei, Q. (2011). Rumor has it: Identifying misinformation in microblogs. In Proceedings of EMNLP, pages 1589–1599. Rosnow and Foster, 2005. Rosnow, R. L. and Foster, E. K. (2005). Rumor and gossip research. Psychological Science Agenda, 19(4). Sacks, 1992. Sacks, H. (1992). Lectures on conversation. Saunders, 1984. Saunders, K. W. (1984). Mythic difficulty in proving a negative, the. Seton Hall Law Review, 15:276. 33 Saur´ı and Pustejovsky, 2009. Saur´ı, R. and Pustejovsky, J. (2009). Factbank: A corpus annotated with event factuality. Language resources and evaluation, 43(3):227–268. Scanlon, 1977. Scanlon, T. J. (1977). Post-disaster rumor chains: A case study. Mass Emergencies, 2(126):22–27. Silverman, 2013. Silverman, C. (2013). Verification handbook. Sinha, 1952. Sinha, D. (1952). Behaviour in a catastrophic situation: A psychological study of reports and rumours. British Journal of Psychology. General Section, 43(3):200–209. Starbird et al., 2014. Starbird, K., Maddock, J., Orand, M., Achterman, P., and Mason, R. M. (2014). Rumors, false flags, and digital vigilantes: Misin- formation on twitter after the 2013 boston marathon bombing. In Proceedings of iConference. iSchools. Takayasu et al., 2015. Takayasu, M., Sato, K., Sano, Y., Yamada, K., Miura, W., and Takayasu, H. (2015). Rumor diffusion and convergence during the 3.11 earthquake: a twitter case study. PLoS one, 10:e0121443–e0121443. Tolmie et al., 2015. Tolmie, P., Procter, R., Rouncefield, M., Liakata, M., and Zubiaga, A. (2015). Microblog Analysis as a Programme of Work. ArXiv e-prints. Ward, . Ward, S. J. A. Digital media etchics. https://ethics.journalism. wisc.edu/resources/digital-media-ethics/. Wilcoxon, 1945. Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics bulletin, pages 80–83. Zubiaga and Ji, 2014. Zubiaga, A. and Ji, H. (2014). Tweet, but verify: epis- temic study of information verification on twitter. Social Network Analysis and Mining, 4(1):1–12. Zubiaga et al., 2015a. Zubiaga, A., Liakata, M., Procter, R., Bontcheva, K., and Tolmie, P. (2015a). Crowdsourcing the annotation of rumourous conver- sations in social media. In Proceedings of the World Wide Web Conference Companion. Web Science Track. Zubiaga et al., 2015b. Zubiaga, A., Liakata, M., Procter, R., Bontcheva, K., and Tolmie, P. (2015b). Towards detecting rumours in social media. In Proceedings of the AAAI Workshop on AI for Cities. 34
1712.05110	1	1712	2017-12-14T07:24:19	Optimality Of Community Structure In Complex Networks	[ "cs.SI", "physics.data-an", "physics.soc-ph" ]	Community detection is one of the pivotal tools for discovering the structure of complex networks. Majority of community detection methods rely on optimization of certain quality functions characterizing the proposed community structure. Perhaps, the most commonly used of those quality functions is modularity. Many heuristics are claimed to be efficient in modularity maximization, which is usually justified in relative terms through comparison of their outcomes with those provided by other known algorithms. However as all the approaches are heuristics, while the complete brute-force is not feasible, there is no known way to understand if the obtained partitioning is really the optimal one. In this article we address the modularity maximization problem from the other side --- finding an upper-bound estimate for the possible modularity values within a given network, allowing to better evaluate suggested community structures. Moreover, in some cases when then upper bound estimate meets the actually obtained modularity score, it provides a proof that the suggested community structure is indeed the optimal one. We propose an efficient algorithm for building such an upper-bound estimate and illustrate its usage on the examples of well-known classical and synthetic networks, being able to prove the optimality of the existing partitioning for some of the networks including well-known Zachary's Karate Club.	cs.SI	cs	Optimality Of Community Structure In Complex Networks Stanislav Sobolevsky a Center For Urban Science+Progress, New York University, Brooklyn, NY, USA SENSEable City Laboratory, Massachusetts Institute of Technology, Cambridge, MA, USA Institute Of Design And Urban Studies of The National Research University ITMO, Saint-Petersburg, Russia Alexander Belyi Faculty of Applied Mathematics and Computer Science, Belarusian State University, Minsk, Belarus SENSEable City Laboratory, SMART Centre, Singapore and SENSEable City Laboratory, Massachusetts Institute of Technology, Cambridge, MA, USA SENSEable City Laboratory, Massachusetts Institute of Technology, Cambridge, MA, USA (Dated: July 11, 2018) Carlo Ratti 7 1 0 2 c e D 4 1 ] I S . s c [ 1 v 0 1 1 5 0 . 2 1 7 1 : v i X r a a To whom correspondence should be addressed: [email protected] 1 Abstract Community detection is one of the pivotal tools for discovering the structure of complex networks. Majority of community detection methods rely on optimization of certain quality functions char- acterizing the proposed community structure. Perhaps, the most commonly used of those quality functions is modularity. Many heuristics are claimed to be efficient in modularity maximization, which is usually justified in relative terms through comparison of their outcomes with those pro- vided by other known algorithms. However as all the approaches are heuristics, while the complete brute-force is not feasible, there is no known way to understand if the obtained partitioning is really the optimal one. In this article we address the modularity maximization problem from the other side - finding an upper-bound estimate for the possible modularity values within a given network, allowing to better evaluate suggested community structures. Moreover, in some cases when then upper bound estimate meets the actually obtained modularity score, it provides a proof that the suggested community structure is indeed the optimal one. We propose an efficient algorithm for building such an upper-bound estimate and illustrate its usage on the examples of well-known clas- sical and synthetic networks, being able to prove the optimality of the existing partitioning for some of the networks including well-known Zachary's Karate Club. Keywords: Complex networks - Community detection - Network science 2 INTRODUCTION A number of features of the increasingly interconnected world can be now described by means of complex networks - examples include but are not limited to physical and digital infrastructure, biological interactions, as well as human mobility and communications. The complex networks play crucial role in various fields such as physics, biology, economics, social sciences, urban planning. This gives paramount importance to the methods and approaches allowing to understand the underlying structure of the complex networks, in particular their community structure. The community detection saw a wide range of applications in social science [1], biology[2], economics [3]. Some more specific applied examples include studies of human mobility and interactions with applications to regional delineation [4–12]. Over the last 15 years a big number of approaches and algorithms for community detec- tion in complex networks has been suggested. Some of them are just the straightforward heuristics such as hierarchical clustering [13] or the Girvan-Newman [14] algorithm, however the vast majority rely on optimization techniques based on the maximization of various ob- jective functions. The first and the most well-known one is modularity [15, 16] assessing the relative strength of edges and quantifying the cumulative strength of the intra-community links. A large number of modularity optimization strategies have been suggested [15–25]. A good historical overview is presented in [26]. Modularity is known to have certain shortcomings including a resolution limit[22, 27] and certain alternative objective functions deserve to be mentioned: Infomap description code length [28, 29], block model likelihood [30–35], and surprise [36]. However despite its limitations modularity remains the most commonly used approach so far. Recently the authors proposed a novel universal optimization technique for commu- nity detection "Combo" [25] capable of maximizing modularity, description code length and pretty much any other metric based on the link scoring and assessing the cumulative score of the intra-community links (we will further refer to such metrics as link-scoring-based). For modularity optimization in most cases Combo outperforms other state-of-the-art algo- rithms in terms of the quality (modularity score) of the resulting partitioning which could be achieved within a reasonable time for the network of up to tens of thousands of nodes. But whichever optimization technique is used, it is always a heuristic as the problem of finding exact modularity maximum is known to be NP-hard [37] - the total number of 3 partitioning to consider scales at the order of magnitude of mn where m is the number of communities and n is the size of the network. A heuristic is never capable of proving that the provided solution is indeed the optimal one and usually can not even tell how close to the optimal the solution is, with the rare exception of algorithms with some guaranteed per- formance estimates [38]. This way the partitioning quality is usually assessed in the relative terms by comparing the algorithm's outcome with the results of the other approaches. In the present paper we consider the problem from a different angle proposing an al- gorithm for building upper-bound estimate for the possible modularity score or any other link-scoring-based partitioning quality function. Sometimes, when such an estimate meets the value achieved by the optimization algorithm, this can provide a proof that the obtained solution is indeed the optimal one. We will provide examples of cases when such a proof becomes possible. THE MODULARITY FUNCTION AND ITS TRIVIAL UPPER BOUND If weights of network edges between each pair of nodes A,B are denoted as e(A, B) then modularity [15, 16] of the partition c(A) (a mapping assigning community number c to each node A) can be defined as Q = (cid:88) q(A, B), (1) where the quantities q(A, B) for each edge AB (call q a modularity score for an edge) are A,B,c(A)=c(B) defined as where wout(A) = (cid:80) (cid:80) e(A, B) q(A, B) = C e(A, C), win(B) = (cid:80) T − wout(A)win(B) C e(C, B), T = (cid:80) T 2 , A wout(A) = (cid:80) B win(B) = A,B e(A, B). If the network is undirected then the edge modularity scores q are symmet- rical: q(A, B) = q(B, A). However even for the directed case, the modularity scores could be effectively symmetrized assigning q(A, B) := (q(A, B) + q(B, A))/2 without any impact on the total score Q. Consider a symmetrical square matrix of the edge modularity scores q. Worth mentioning that all the considerations below do not depend on the specific way the scores q are defined, so are equally valid for any other objective function implemented through a cumulative edge scoring like (1). 4 (cid:88) Q ≤ Qmax = (cid:88) Modularity Q of the complex network is known to be a normalized measure, i.e. −1 ≤ Q ≤ 1. However can one ever reach the maximal value of Q = 1? In most cases not and a trivial estimate of q(A, B) + q(A, A) (2) A,B,q(A,B)>0 A,q(A,A)<0 often leads to Qmax < 1. But even this more precise upper bound (2) is often far from the actual values that could be ever achieved. Further we provide an algorithm for building a better estimate. THE METHOD Idea of the algorithm is based on the following consideration. Assume that the network contains three nodes A, B, C such that the modularity scores q(A, B) > 0 and q(A, C) > 0, while q(B, C) < 0. Then if we were to have both positive scores q(A, B) and q(A, C) as a part of the total modularity 1, i.e. have both links AB and AC as intra-community ones with respect to the considered community structure, all three nodes A, B, C should belong to the same community. But this will come at a price of also having a link with a negative score q(B, C) as intra-community one and would cause the corresponding negative score to be also included in 1. Otherwise at least one of the positive scores q(A, B) or q(A, C) would not be included. This way a triangle ABC produces an additional "penalty" of p(A, B, C) = 2 min{q(A, B), q(B, C),q(A, C)} to the total modularity score estimate (the penalty is doubled as each edge score q(A, B) is present in the network twice as q(B, A) = q(A, B)). More generally, if a chain A1, A2, ..., Ak is such that q(Aj, Aj+1) > 0 for all j = 1, 2, ..., k− 1, while q(A1, Ak) < 0 then such a chain produces a penalty of p(A1, A2, ..., Ak) = min{q(A1, A2), q(A2, A3), ..., q(Ak−1, Ak),q(A1, Ak)}. (3) Call those triangles or chains penalized. Based on that, if one finds a set Ω of non-overlapping (i.e. not having common edges) 5 penalized undirected triangles or longer chains, then Q ≤ Qmax −(cid:88) ω∈Ω p(ω). (4) However, there is also a way to deal with the set of overlapping triangles or chains having common edges. Before we introduce the corresponding approach let us consider a more general framework which will make the idea more intuitive and will also enable further generalization of the algorithm. PENALIZED SUBNETWORKS Start with several definitions. Recall that each edge AB of the original network is char- acterized by its modularity score q(A, B) which can be positive, negative or zero. Definition 1. Subnetwork (SN) – a subset of nodes of the original network, where each edge A,B is characterized by a score q∗(A, B) such that q∗(A, B) ≤ q(A, B) and q∗(A, B)q(A, B) ≥ 0. Definition 2. Subnetwork's penalty – a difference between the sum of all edge scores q∗ of the subnetwork and the proven upper bound for its possible partitioning modularity estimate. Subnetwork with a positive penatly is called a penalized subnetwork. Definition 3. Resolved subnetwork (RSN) – a subnetwork, best known candidate partition of which matches the proved upper bound for its possible partitioning modularity estimate (making it a proven optimal partitioning). Definition 4. Reduced subnetwork of a given penalized subnetwork S – a subnetwork of S having the same penalty as S. Definition 5. Permissible linear combination LS =(cid:80) characterized by the linear combinations q∗(A, B) = (cid:80) j λjSj of the subnetworks Sj with non-negative coefficients λj – is a subnetwork of the original network which edges AB j (A, B) of the corresponding j (A, B) from the subnetworks Sj (in case AB is not present in some of the sub- j (A, B) = 0), provided that q∗(A, B) ≤ q(A, B) and scores q∗ networks consider corresponding q∗ q∗(A, B)q(A, B) ≥ 0. j λjq∗ Then it is easy to see that the following proposition holds: Summation lemma: Consider a set of subnetworks {S1, S2, ..., Sm} with proven penal- 6 j λjpj. j λjSj for ties p1, p2, ..., pm. Then then penalty for a permissible linear combination LS =(cid:80) non-negative λj is at least(cid:80) expressed as Q =(cid:80) (cid:80) subnetwork Sj by Uj. Then for any given partitioning c modularity Q(c) of the LS can be j −pj) = Qmax− j λjQj and then Q =(cid:80) j λjQj ≤(cid:80) j λjUj =(cid:80) Indeed, denote the proven upper bound estimate for the partitioning modularity for each j λj(Qmax j λjpj, where Qmax as before denotes the estimate (2) for a corresponding subnetwork. Definition 6. Resolved network/subnetwork is a network or its subnetwork for which one was able to find a proven upper bound modularity estimate and an actual community structure with matching modularity scores. Resolving a network means constructing the proof of optimality for the known partition- ing. Summation lemma provides a framework for resolving the networks through combining proven penalties of the smaller subnetworks. OVERLAPPING TRIANGLES AND CHAINS Now return to an approach of handling a set of overlapping triangles or chains having common edges using summation rule. Clearly penalized triangles or chains are examples of penalized subnetworks. If p is the penalty then one can construct a reduced subnetwork of the same penalty with the same edges scored as p and −p depending on the sign of the original edge characteristics. Then considering any set Ω = {ωj} of overlapping penalized triangles or chains ωj, one can replace it by a set of the corresponding reduced penalized triangles or chains ωj. If a linear combination of those is a permissible linear combination (i.e. if a linear combination of penalties of all the triangles/chains each edge participates in does not exceed the absolute value of the modularity score of this edge) then summation rule provides a proven cumulative penalty for the entire network as a linear combination of the penalties of all the triangles and chains. In particular if(cid:80) j ωj is a permissible linear combination (with λj = 1) then Q ≤ Qmax −(cid:88) p(ωj). (5) This way we replace overlapping penalized triangles and chains with non-overlapping (in j 7 terms of the edge scores) reduced penalized triangles and chains. THE ALGORITHM An upper bound estimate (5) largely depends on the set Ω of triangles or chains. Introduce a greedy heuristic building a sequence of triangles and chains Ω = {ω1, ω2, ..., ωm} aiming to produce the best possible upper bound estimate (5): 1. Start from k = 3 and empty Ω. 2. Look for all the penalized chains of length k: search through the negatively scored edges AB, considering all the paths between A and B which consist of all positively scored edges, evaluating penalty for every such chain. 3. If there exists at least one such chain, add to Ω the one having highest associated penalty value p(ω); reduce the edge scores of the original network by subtracting the corresponding edge scores p(ω) or −p(ω) of the reduced chain ω; repeat from the step 2. 4. Otherwise check if there still exist any penalized chain (evaluate connected components in the graph with edges having positive residual score and see if at least one of them contains an internal edge with the negative residual score). If such a chain exists increment k and repeat from step 2. 5. Otherwise stop. The algorithm is streightforward and relatively fast, however as pretty much any greedy heuristic sometimes converges to the non-optimal solutions. Alternative for the step 3 is to select a random penalized chain at each step. Also one can select the best chain with a certain probability and a random chain otherwise. In those cases each series of iterations 2-3 for each k could be performed multiple times in order to select the best intermediate result before incrementing k. 8 THE RESULTS First we provide in Table I several examples of the networks for which we are able to prove that the obtained best modulairity partitioning is indeed the optimal one. Here and further we use Combo algorithm for modularity optimization. Network Size Acheived modularity Upper-bound modularity Ratio, % Ants 1 [39] 16 Ants 2 [39] 13 Dolphin groups [40] 13 Pony [41] 17 0.0972 0.1666 0.5900 0.1375 GAMAPOS [42] 16 0.525565 PADGM [42] 16 0.364444 KNOKM [42] 10 0.14876 KNOKI [42] 10 0.0816327 Synthetic network 50 0.6399 0.0972 0.1666 0.5900 0.1375 0.525565 0.364444 0.14876 0.0816327 0.6399 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 TABLE I: Evaluation on the real-world and synthetic net- works: examples of the networks with exact or nearly ex- act match between upper-bound estimate and the achieved result. The synthetic network was generated using LFR ap- proach [43], while real-world network taken from [44] and [42] Second we perform extensive evaluations of the algorithm on a number of networks of size 25-75 created using artificial network generator of Lancichinetti- Fortunato- Radicchi (LFR) benchmark [43] with average degree equal 4, maximum degree = 16, mixing parameter µ = 0.1 − 0.2 for the topology equal to mixing parameter for the weights, exponent for the weight distribution = 1, minus exponent for the degree sequence = 2, minus exponent for the community size distribution = 1, minimum for the community sizes = 5, maximum for the community sizes equal to number of nodes divided by 4. Results are reported in the Table II Network size Avg. ratio of achieved vs upper bound modularity, % Exact matches, % 25-34 100.0 9 100.0 35-44 45-54 55-64 65-75 All 99.96 99.98 99.91 99.82 99.93 80.0 60.0 50.0 18.18 60.78 TABLE II: Evaluation on the synthetic LFR networks. Further in the Table III we present results for the other real-world networks where exact match between the upper bound estimate and the best known modularity score has not been achieved. In most cases the gap between the two is below 5%, often (if half of those cases) within 0.1 − 1%. Network Size Achieved modularity Upper-bound modularity Ratio, % 0.425789 0.548080 0.572035 0.541007 0.62767 0.7875 0.3717 98.59 96.43 99.07 97.45 96.46 99.99 99.92 93.69 99.15 99.91 99.53 94.87 92.62 98.00 90.23 96.08 94.37 # 1 2 3 Zachary's karate club [45] Dolphins [46] Les Miserables [47] 34 62 77 0.419790 0.528519 0.566688 4 Political Books [orgnet.com] 105 0.527237 5 American College Football [14] 115 0.605445 6 World colonial ties [11] 7 World language ties [11] 238 238 0.7874 0.3714 8 Neural net of C. Elegans [48] 297 0.507642 0.541827 9 10 11 12 13 14 15 16 17 Infectious diseases [49] Synthetic network Synthetic network Jazz musicians [50] WOLFN [42] THURM [42] GAMANEG [42] PADGB [42] KAPFMM [42] 410 250 500 198 20 15 16 16 15 0.860384 0.798493 0.851837 0.444787 0.0580522 0.269513 0.274673 0.39875 0.278393 10 0.86777 0.79922 0.85584 0.46885 0.0626775 0.275023 0.3044 0.415 0.295014 TABLE III: Evaluation on the real-world networks: best and upper-bound modularity score for the extensive set of net- works. FURTHER IMPROVEMENTS OF THE ALGORITHM As we see the presented algorithm can build a plausible upper-bound estimate for the possible modularity score for many real-world and synthetic networks, often being able to resolve them (i.e. matching the modularity score achievable by the community detection algorithm (Combo), proving the optimality of such a partitioning). However, there are networks that fundamentally can not be resolved with the above approach - see Figure 1 for an example of a simple unresolvable network configuration. And despite the fact that the gap between the proven upper bound estimate and achievable modularity score might be as small as 0.1% or even less this still does not allow to prove the optimality of the partitioning which is seen as the ultimate goal of the present paper. In order to overcome the fundamental limitations of the approach based on penalized FIG. 1. Example of the network configuration for which exact upper bound on modularity can not be found by algorithm considering triangles and chains. Numbers indicate scaled edges' modularity scores. 11 254-5-3-3triangles and chains we propose a more general framework able to utilize arbitrary penalized subnetworks, including the one on the Figure 1. Scheme of the framework: p1, p2, ..., pm. max. j λjpj → Phase 1. Identify a set of arbitrary subnetworks S1, S2, ..., Sm with known penalties Phase 2. Find a best possible acceptable linear combination(cid:80) could be implemented as a linear optimization problem maximizing(cid:80) Instead of a binary (i.e. having λj = 0, 1) inclusion heuristic like we tried before, Phase 2 j λjpj → max subject to linear constrains λj ≥ 0 and the corresponding constrains on each edge score, ensuring that the resulting linear combination of the subnetworks is permissible. j λjSj, such that(cid:80) The phase 1 requires us to decide which Sj to include and how to build them. A simple although perhaps not time efficient generic solution could consider all possible subnetworks of a small size up to a certain size M . For M small enough all such possible subnetworks could be analyzed within a reasonable time. Each subnetwork can be resolved using the following algorithm, which we call partial brute-force: 1. Start with m := 0 and with the upper bound modularity estimate Q∗ := 0; 2. Increment m (m := m + 1) if m < M , otherwise - stop, marking the subnetwork as unresolved; 3. Consider all combinations of excluding exactly m edges and merging nodes connected by other positive edges into communities (those combinations where edges to merge along imply merging along some of those other edges which were intended to be ex- cluded could be discarded as they have to be already considered as the ones with smaller m); 4. For each of the combinations above assess the resulting modularity Q1 and set Q∗ := max (Q∗, Q1); 5. If sum of all positive edges but the m + 1 smallest ones is higher than Q∗, return to step 2, otherwise stop marking the subnetwork as resolved. Once the algorithm finalizes marking subnetwork as resolved, then p = Qmax − Q∗ is the proven penalty. 12 And then for each penalized subnetwork resolved with the above approach we try to reduce the weights of its edges as much as possible provided that the penalty is preserved. This is done by solving a linear programming problem minimizing the sum of the edge weights with respect to the following conditions: 1. For each partition considered by the procedure above, sum of the absolute values of edge scores contributing to penalty (positive edges between communities and negative edges within communities) can not be less than p; 2. For each set of the m + 1 edges with positive scores, sum of their scores can not be less than p. The above two procedures applied to all subnetworks up to a certain size M provide a set of reduced resolved subnetworks, completing phase 1 of the framework. Then a standard linear programming solution like a simplex method could be used to implement phase 2. Current implementation of this approach for M = 6 for example allows us to improve the earlier obtained upper bound modularity estimates, resolving a famous Zachary's karate club network [45] as well as the networks 13-17 from the Table III. This way we find exact match between the best modularity score (e.g. 0.41979 for Zachary's karate club) achieved by Combo and the upper-bound estimate we provide proving the optimality of the obtained partitioning. We believe the approach might allow resolving much broader sets of the net- works, however implementation requires further optimization so that it can efficiently handle networks of larger size within a reasonable time. CONCLUSIONS We proposed an efficient algorithm for building an upper-bound estimate for the possible partitioning modularity score of a network. The algorithm is applied to a set of well- known classical and synthetic networks, being able to prove the optimality of the existing partitioning for some of the networks including well-known Zachary's Karate Club. For the rest of the networks the constructed upper bound modularity estimates happen to be pretty close to the achieved scores of the known partitioning (the gap between the achieved score and the upper-bound estimate is often within 1% of the score). The ways of further improving the approach are provided. 13 [1] M. Planti´e and M. Crampes, in Social media retrieval (Springer, 2013), pp. 65–85. [2] R. Guimer`a and L. A. Nunes Amaral, Nature 433, 895 (2005), ISSN 0028-0836, URL http: //dx.doi.org/10.1038/nature03288. [3] C. Piccardi and L. Tajoli, Phys. Rev. E 85, 066119 (2012), URL http://link.aps.org/doi/ 10.1103/PhysRevE.85.066119. [4] C. Ratti, S. Sobolevsky, F. Calabrese, C. Andris, J. Reades, M. Martino, R. Claxton, and S. H. Strogatz, PLoS ONE 5, e14248 (2010), URL http://dx.doi.org/10.1371%2Fjournal. pone.0014248. [5] V. Blondel, G. Krings, and I. Thomas, Brussels Studies. La revue scientifique ´electronique pour les recherches sur Bruxelles/Het elektronisch wetenschappelijk tijdschrift voor onderzoek over Brussel/The e-journal for academic research on Brussels (2010). [6] S. Sobolevsky, M. Szell, R. Campari, T. Couronn´e, Z. Smoreda, and C. Ratti, PloS ONE 8, e81707 (2013). [7] A. Amini, K. Kung, C. Kang, S. Sobolevsky, and C. Ratti, EPJ Data Science 3, 6 (2014). [8] B. Hawelka, I. Sitko, E. Beinat, S. Sobolevsky, P. Kazakopoulos, and C. Ratti, Cartography and Geographic Information Science 41, 260 (2014). [9] C. Kang, S. Sobolevsky, Y. Liu, and C. Ratti, in Proceedings of the 2nd ACM SIGKDD International Workshop on Urban Computing (ACM, 2013), p. 1. [10] S. Sobolevsky, I. Sitko, R. T. Des Combes, B. Hawelka, J. M. Arias, and C. Ratti, in Big Data (BigData Congress), 2014 IEEE International Congress on (IEEE, 2014), pp. 136–143. [11] A. Belyi, I. Bojic, S. Sobolevsky, I. Sitko, B. Hawelka, L. Rudikova, A. Kurbatski, and C. Ratti, International Journal of Geographical Information Science 31, 1381 (2017). [12] S. Grauwin, M. Szell, S. Sobolevsky, P. Hovel, F. Simini, M. Vanhoof, Z. Smoreda, A.-L. Barab´asi, and C. Ratti, Scientific Reports 7 (2017). [13] T. Hastie, The elements of statistical learning : data mining, inference, and prediction : with 200 full-color illustrations (Springer, New York, 2001), ISBN 0387952845. [14] M. Girvan and M. Newman, Proc. Natl. Acad. Sci. USA 99 (12), 7821 (2002). [15] M. Newman and M. Girvan, Phys. Rev. E 69 (2), 026113 (2004). [16] M. Newman, Proceedings of the National Academy of Sciences 103, 8577 (2006). 14 [17] M. E. J. Newman, Phys. Rev. E 69, 066133 (2004), URL http://link.aps.org/doi/10. 1103/PhysRevE.69.066133. [18] A. Clauset, M. E. J. Newman, and C. Moore, Phys. Rev. E 70, 066111 (2004), URL http: //link.aps.org/doi/10.1103/PhysRevE.70.066111. [19] Y. Sun, B. Danila, K. Josi´c, and K. E. Bassler, EPL (Europhysics Letters) 86, 28004 (2009), URL http://stacks.iop.org/0295-5075/86/i=2/a=28004. [20] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, Journal of Statistical Mechan- ics: Theory and Experiment 2008, P10008 (2008). [21] R. Guimera, M. Sales-Pardo, and L. A. N. Amaral, Physical Review E 70, 025101 (2004). [22] B. H. Good, Y.-A. de Montjoye, and A. Clauset, Phys. Rev. E 81, 046106 (2010), URL http://link.aps.org/doi/10.1103/PhysRevE.81.046106. [23] J. Duch and A. Arenas, Phys. Rev. E 72, 027104 (2005), URL http://link.aps.org/doi/ 10.1103/PhysRevE.72.027104. [24] J. Lee, S. P. Gross, and J. Lee, Phys. Rev. E 85, 056702 (2012), URL http://link.aps.org/ doi/10.1103/PhysRevE.85.056702. [25] S. Sobolevsky, R. Campari, A. Belyi, and C. Ratti, Physical Review E 90, 012811 (2014). [26] S. Fortunato, Physics Report 486, 75 (2010). [27] S. Fortunato and M. Barth´el´emy, Proceedings of the National Academy of Sciences 104, 36 (2007), http://www.pnas.org/content/104/1/36.full.pdf+html, URL http://www.pnas.org/ content/104/1/36.abstract. [28] M. Rosvall and C. T. Bergstrom, Proceedings of the National Academy of Sciences 104, 7327 (2007), http://www.pnas.org/content/104/18/7327.full.pdf+html, URL http://www.pnas. org/content/104/18/7327.abstract. [29] M. Rosvall and C. Bergstrom, Proc. Natl. Acad. Sci. USA 105, 1118 (2008). [30] B. Karrer and M. E. J. Newman, Phys. Rev. E 83, 016107 (2011), URL http://link.aps. org/doi/10.1103/PhysRevE.83.016107. [31] B. Ball, B. Karrer, and M. E. J. Newman, Phys. Rev. E 84, 036103 (2011), URL http: //link.aps.org/doi/10.1103/PhysRevE.84.036103. [32] P. J. Bickel and A. Chen, Proceedings of the National Academy of Sciences 106, 21068 (2009). [33] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, Phys. Rev. Lett. 107, 065701 (2011), URL http://link.aps.org/doi/10.1103/PhysRevLett.107.065701. 15 [34] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a, Phys. Rev. E 84, 066106 (2011), URL http://link.aps.org/doi/10.1103/PhysRevE.84.066106. [35] X. Yan, C. Shalizi, J. E. Jensen, F. Krzakala, C. Moore, L. Zdeborov´a, P. Zhang, and Y. Zhu, Journal of Statistical Mechanics: Theory and Experiment 2014, P05007 (2014). [36] R. Aldecoa and I. Mar`ın, PLoS ONE 6, e24195 (2011), URL http://dx.doi.org/10.1371% 2Fjournal.pone.0024195. [37] U. Brandes, D. Delling, M. Gaertler, R. Gorke, M. Hoefer, Z. Nikoloski, and D. Wagner, arXiv preprint physics/0608255 (2006). [38] T. N. Dinh, X. Li, and M. T. Thai, in Data Mining (ICDM), 2015 IEEE International Con- ference on (IEEE, 2015), pp. 101–110. [39] B. J. Cole, Science 212, 83 (1981). [40] R. C. Connor, R. A. Smolker, and A. F. Richards, Coalitions and alliances in humans and other animals pp. 415–443 (1992). [41] T. Clutton-Brock, P. Greenwood, and R. Powell, Ethology 41, 202 (1976). [42] Ucinet iv datasets, URL http://vlado.fmf.uni-lj.si/pub/networks/data/UciNet/ UciData.htm. [43] A. Lancichinetti, S. Fortunato, and F. Radicchi, Phys. Rev. E 78 (4), 046110 (2008). [44] Linton c. freeman's datasets, URL http://moreno.ss.uci.edu/data.html. [45] W. W. Zachary, Journal of Anthropological Research 33, 452 (1977). [46] D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and S. M. Dawson, Be- havioral Ecology and Sociobiology 54, 396 (2003), URL http://dx.doi.org/10.1007/ s00265-003-0651-y. [47] D. E. Knuth, The Stanford GraphBase: a platform for combinatorial computing (Addison- Wesley, 1993), URL http://www-cs-staff.stanford.edu/~{}uno/sgb.html. [48] J. G. White, E. Southgate, J. N. Thomson, and S. Brenner, Philosophical Trans- actions of the Royal Society of London. B, Biological Sciences 314, 1 (1986), http://rstb.royalsocietypublishing.org/content/314/1165/1.full.pdf+html, URL http: //rstb.royalsocietypublishing.org/content/314/1165/1.abstract. [49] L. Isella, J. Stehl´e, A. Barrat, C. Cattuto, J.-F. Pinton, and W. Van den Broeck, Journal of theoretical biology 271, 166 (2011). 16 [50] P. M. Gleiser and L. Danon, Advances in Complex Systems 06, 565 (2003), http://www.worldscientific.com/doi/pdf/10.1142/S0219525903001067, URL http://www. worldscientific.com/doi/abs/10.1142/S0219525903001067. 17
1906.12330	1	1906	2019-06-21T03:05:17	Graph Star Net for Generalized Multi-Task Learning	[ "cs.SI", "cs.CL", "cs.LG" ]	In this work, we present graph star net (GraphStar), a novel and unified graph neural net architecture which utilizes message-passing relay and attention mechanism for multiple prediction tasks - node classification, graph classification and link prediction. GraphStar addresses many earlier challenges facing graph neural nets and achieves non-local representation without increasing the model depth or bearing heavy computational costs. We also propose a new method to tackle topic-specific sentiment analysis based on node classification and text classification as graph classification. Our work shows that 'star nodes' can learn effective graph-data representation and improve on current methods for the three tasks. Specifically, for graph classification and link prediction, GraphStar outperforms the current state-of-the-art models by 2-5% on several key benchmarks.	cs.SI	cs	Graph Star Net for Generalized Multi-Task Learning Lu Haonan AARC Huawei Technologies [email protected] Tian Ye AARC Huawei Technologies [email protected] Seth H. Huang ∗ AARC Huawei Technologies [email protected] Guo Xiuyan AARC Huawei Technologies [email protected] Abstract In this work, we present graph star net (GraphStar), a novel and unified graph neural net architecture which utilizes message-passing relay and attention mechanism for multiple prediction tasks - node classification, graph classification and link prediction. GraphStar addresses many earlier challenges facing graph neural nets and achieves non-local representation without increasing the model depth or bearing heavy computational costs. We also propose a new method to tackle topic-specific sentiment analysis based on node classification and text classification as graph classification. Our work shows that "star nodes" can learn effective graph-data representation and improve on current methods for the three tasks. Specifically, for graph classification and link prediction, GraphStar outperforms the current state-of-the-art models by 2-5% on several key benchmarks. 1 Introduction Relational data, consisted of nodes, sub-graphs and edges, has a vast range of applications from sentiment analysis, text classification to protein and enzyme interactions. Graphs have implicit, complex topological structures and relationships, and graph models capable of internalizing these structures have achieved state-of-the-art performances for many task [30]. For data representation, graphs are more universal than conventional pairwise relationships, and machine learning approaches for graph analysis mainly focus on three primary tasks: node classification, graph classification and link prediction [10, 11, 15]. Graph data models a set of subjects/ entities (nodes) and their inter-relationships (edges). In a simplest classification setting, the model attempts to predict unlabeled nodes by the surrounding labeled nodes. The graph domain is also categorized into spectral and non-spectral approaches. Spectral approaches often work with methods to represent the graph. Non-spectral approaches often conduct local-convolutions on the graph in the neighborhood [5, 19]. In order to capture the global state, a local conventional model has to increase its depth, but this may encounter an over-smoothing issue [30, 3, 9]. On the other hand, [22] tackles non-local representation to capture long-range dependencies, using fully-connected-nodes attention and aggregating all features of all positions as a weighted sum, resulting in a high computational complexity. Our model aims at capturing the global state without increasing the model depth and using fully-connected attention. The codes for essential experiments will be released at https://github.com/graph-star-team/graph_star. This site complies fully with the anonymous, double-blinded policies. ∗Corresponding Author. Lu Haonan and Tian Ye are equal contributors Preprint. Under review. Graph Attention Networks, instead of using convolutional neural networks (CNN) as in [5], proposes a graph attention layer, which compared to CNNs can specify different weights to different nodes within a neighborhood [19]. They adopt the multi-head attention mechanism [19, 1], allowing for predictions of new-node and previously unseen graphs, a transductive-to-inductive approach. However, it does not allow for deep-neighborhood representation, which lessens the model capability to understand context. [22] addresses this issue with a self-attention approach to capture global state without increasing the model depth. Star Transformer [4] changes the original fully-connected attention structure to one with a virtual message-passing relay, reducing the number of connections from quadratic to linear [4]. The design has the potential of capturing long-range dependency, aggregating both local and non-local compositionality [4]. Our work has shown improvement upon previous results while simplifying the model structure and improving the computational efficiency. In this work, we introduce Graph Star Net (GraphStar), an inductive framework which extends the information boundary through relay nodes to aggregate global information. To our best knowledge, we have not seen another architecture which utilizes the virtual relay concept in graph modeling and aggregating local and non-local information. Our main contributions lie in addressing many previous limitations and unifying the approaches in multiple tasks for node classification, graph classification and link prediction simultaneously. We also proposed a new node-classification-based, topic-specific sentiment analysis method as well as graph-classification-based, text classification method. Compared with the previous node-classification-based transductive models such as [23, 27], our model can easily be used for unseen, new document predictions. GraphStar utilizes the unique data representation via an information relay across the graph and trains the model to: (1) perform inductive tasks on previously unseen graph data; (2) aggregate both local and long-range information, making the model globally-aware, extracting high-level abstraction typically not represented in individual node features; (3) the relays serve as a hierarchical representation of the graphs and can be directly used for graph classification tasks. 2 GraphStar Model Architecture In this section, we present the star relay function and the updating process and describe the theoretical and implementation advantages. Generally, our approaches will be divided into three steps: star initialization, updating of the real nodes and the updating of the stars. In each layer of the model, the second and third steps occur in a cyclical manner layer-by-layer in the updating process, and our goal is to train a multi-purpose model to represent local and global compositionality for different prediction tasks. Our model setup primarily follows the landmark works of [19, 22] as well as [4]. Figure 1: Graph Star Architecture. Left: Corresponding to Step 2, local compositionality updates for real nodes. Node i uses the representations of its neighbors, including itself, and the star in the layer t to get its own representation in layer (t+1). Right: Corresponding to Step 3, star global compositionality updates. To update the star, the model uses the updated representations of all nodes in the layer (t+1) and its own representation in layer t to get star's representation in (t+1). This star, connected to all nodes, is a natural hierarchical representation of the whole graph. 2 The input to our model is a graph G(A, F ), where A RNb×Nb is adjacency matrix and F Rn×d is the feature matrix, and each node has d features, and Nb is the number of nodes in a graph. We −→ i Rd, describe the graph with a set of learned features in each layer H t = ( h t −→ f i Rd is the initial representation of the node i and t being the layer and F = ( number. Step 1: Initial Representation of the Star −→ f 2, ..., −→ h 2, ..., −→ f Nb ), −→ h Nb ), −→ h 1, −→ f 1, −→ F mean = mean(F 0), F Rd. We have the initial star αinit,iW init −→ f i), where V representation −→ S (0) = σ((cid:80) Let iNb αinit,i = exp(< W init Q exp(< W init Q −→ F mean, W init −→ F mean, W init −→ f i >) −→ f k >) K K (cid:80) kNb (1) < −→a , Q , W init W init −→ b > is the dot product of −→a , K , W init V follow the standard Transformer setup [18]. −→ b , σ is the non-linear activation function and Step 2: Real Node Update Here we use multi-head attention to conduct real node update [19]. Inspired by rGCN [13], we treat node-to-neighborhood under relation rR, node-to-star and node-to-self as three types of relations, with the first one corresponding to the number of relations in the graph. The node importance under relation r is controlled by attention coefficients. We define the self-self, self-star and self-neighbor under relation r as controlled by WV 0, WV S and WrV 1, respectively. The work presented in our paper is based on the special case R = 1, making it single-type relation link prediction. The node update can be represented by NHead(cid:110) σ( (cid:88) (cid:88) (cid:88)m = (cid:88) (cid:88) h(t+1) i = ijrW m(t) αm rV 1 h(t) j + αm is,r=sW m(t) V S St + αm i0,r=0W m(t) V 0 ht i), (2) where(cid:102) represents the concatenation of all multi-heads, N r jN r i rR m i is the neighbors of node i under relation rR. Nhead is the number of heads in the multi-head attention setting. The multi-head outputs are then concatenated as the representation of t + 1 layer. Here we define exp(< W m(t) Q h(t) i , W m(t) rV 1 h(t) k >)+ exp(< W m(t) Q h(t) i rR kN r i V S S(t) >) + exp(< W m(t) Q h(t) i , W m(t) , W m(t) V 0 h(t) i >), (3) and the attention coefficients can thus be represented by exp(< W m(t) Q h(t) , W m(t) rV 1 h(t) j >) exp(< W m(t) Q h(t) i , W m(t) V S S(t) >) exp(< W m(t) Q h(t) i , W m(t) V 0 h(t) i >) i (cid:80)m (cid:80)m (cid:80)m , , . αm ijr = αm is,r=s = αm i0,r=0 = It should be noted that for all relations, we adopt a parameter sharing scheme WK∗ = WV ∗. Step 3: Star Update 3 NHead(cid:110) (cid:88) Following step 1 and step 2, after we got the real node representation at layer t + 1, we can now update the star nodes via S(t+1) = σ( s,jW m(t) αm V h(t+1) j + αm s,sW m(t) V St). (4) m jNb We define (cid:88)m s (cid:88) kNb = exp(< W m Q St, W m K h(t+1) k >) + exp(< W m(t) Q St, W m K St >) , and the attention coefficients are thus αm s,j = exp(< W m Q S(t), W m K h(t+1) j (cid:80)m (cid:80)m s s >) , . exp(< W m Q S(t), W m K S(t) >) αm s,s = 3 Multi-Task Setup Node Classification For node classification tasks, following standard graph neural setup described in [19, 13] we take the GraphStar model's final layer output in (2) and apply the standard sof tmax activation function. We then minimize the cross-entropy loss conventional in classification model training tasks. Graph Classification From the output of (4), in GraphStar, star and this virtual node are connected via other real nodes through attention mechanism, naturally forming hierarchical representation of the graph, meaning S(t) is a natural representation of the graph's global state. We use the final layer output of the star for graph classification. Link Predictions Here we present the application of GraphStar in single-relation data sets. From (2) We treat the GraphStar model as an encoder for node embedding, and link prediction is a downstream task. A standard setup is to make predictions on subject, relation, object (s, r, o) as described in [13]. The decoder is a score function f (s, r, o) : Rd× R ×Rd → R. The score function for the single-relation experiments is the DistMult [26]: f (s, r, o) = eT Following [26, 17, 13], we train the model with negative sampling, and for a given positive example, we sample one negative sample through a randomized method to the subject or the object, with cross-entropy loss function. s Rreo. 4 Experimental Evaluation We demonstrate the results for node classification, graph classification and link prediction tasks. We also developed several graph data sets based on standard text classification and sentiment analysis dataset IMDB. 4.1 Datasets Node Classification Tasks. For node classification, GraphStar is tested on three popular transductive learning benchmarks - Cora, Citeseer and PubMed, and one inductive learning benchmark - PPI. Topic-Specific Sentiment Analysis. Here, we propose a new method to tackle sentiment analysis based on node classification and use IMDB-binary dataset as an example. This dataset was originally 4 not a graph task; it is usually treated as a natural language processing task (NLP). It is a binary sentiment analysis consisting of 50,000 reviews labeled as either positive or negative. Our innovation is to turn the pure NLP task into a graph data task based on document embedding. First, for IMBD, we use a pre-trained large Bert model to document encoding [1], and we treat every film review as a node in a graph. We then link the nodes (film reviews) which belong to the same topic and create a graph dataset. This approach is highly generalizable to most topic-specific classification tasks. Table 1: Summary of Classification Datasets for Node Classification and Link Predictions.The notation O+:O- describes the ratio of positive and negative edges. More details regarding these datasets can be found in [14, 21] Graph Classification task. For graph classification tasks, we use Enzymes, D&D, Proteins and Mutag as our evaluation benchmarks. The description of the data is summarized in Table 2. Text Classification Tasks. For text classification, we use R8, R52, MR, 20 News Group and Ohsumed as our benchmarks. Different from previous text graph-based models[23, 27], here we conduct the text classification as graph classification instead of node classifications. In a real life setting, a trained model should be inductive, which can be used to directly make predictions on new texts or documents. In the GraphStar setup, every document corresponds to one graph, and in the graph, every node is a word. If two words appear in the same sliding window, then these two words are linked via an edge. The same words with different context are treated as different nodes. In this way, although the initial word share the same feature, the outputs embedding of the words will be different. For 20 News Group, due to the text length, maximum being over 15000, same words are treated as the same node, and we truncate the text length at 3072. We divide the original training data into 90% of training data and 10% of validation data, leaving the test data out of the training cycles. For our experiments, we report the test data results. The summary of the data is in Table 3. Table 2: Graph Classification Datasets Table 3: Summary of Text Classification Datasets for Graph Classification 5 Link Prediction. For link prediction, following [16], we use Cora, Citeseer, Pubmed as key bench- marks, and the data description is in Table 1. 4.2 Results The experimental results are summarized in Table 4, 5, 6, 7 and 8 for transductive node classification, inductive node classification, graph classification, graph text classification and link prediction. All reported results are based on 10 runs. We adopt the Adam optimizer, and the margins of error are provided in the parenthesis. Node Classification For the transductive tasks, all tasks share the same parameters and are not optimized individually for different datasets, and the only variations are the initial learning rates and L2 regularization settings. We use the attention coefficient dropout rate = 0.2. The number of GraphStar layers = 2, the number of attention head per layer Nhead = 8, hidden unit size per head is 16, and the hidden unit's dropout rate is 0.7. For the training, we keep the best model achieving the highest validation accuracy given the patience setting of 50 epochs. The initialization of the star nodes is based on the mean of raw features of the nodes. During training, for Cora, Citeseer and Pubmed, we apply the L2 regularization with ρ = 0.002, ρ = 0.004, and ρ = 0.0001, and the initial learning rates are set to be 0.001, 0.001 and 0.005, respectively. The initial learning rates for PPI and IMDB are 0.0002 and 0.0001, respectively. The L2 setting is ρ = 0 for both PPI and IMDB. For both data sets, the attention coefficient and the hidden unit dropout rates are both 0.2 with 3 layers. The number of hidden units per head is 128 for PPI and 256 for IMDB. For transductive tasks, the GraphStar classification accuracy is measured against the results in [2, 19, 23, 16, 29]. For inductive task the reported results are in Table 4, based on [2, 5, 19, 25]; the GraphStar model achieves better performance than the current state-of-the-art models for the PPI dataset. Table 4: Results of transductive and inductive node classification experiments. Left: Results of transductive experiments on Cora, Citeseer and Pubmed datasets for multiple landmark model architectures [2, 19, 23, 16, 29]. Right: Results of inductive experiment in terms of micro-average F1 scores on the PPI dataset [2, 5, 19, 25] . Table 5: Results of inductive IMDB node classification experiments with reported results based on [6, 12, 24] The IMDB sentiment analysis is reported in Table 5. The GraphStar is shown to outperform the current state-of-the-art results after only 130 seconds of training time on V100 GPU. Graph Classification / Graph-based Text Classification. For graph classification, we report the current state-of-the-art results in Table 6. For text classification, the current state-of-the-art results are reported and compared against our model[20, 28, 8] and reported in Table 7. 6 For Enzymes, Proteins, D&D and Mutag classification tasks, we use the attention coefficient dropout rate and the hidden unit dropout rate of 0.2. The number of GraphStar layers = 3, the number of attention head Nhead = 4. The learning rates are 0.0005 for all datasets. The L2 regularization is 0.0001, 0.001, 0.001, 0.0 and the hidden unit sizes per head are 16, 128, 16 and 16 for Enzymes, Proteins, D&D and Mutag, respectively. For all datasets, we perform 10-fold cross-validation for model evaluation, and the resulting accuracy and margins are based on the 10 runs. For the text classification tasks, we use the attention coefficient dropout rate = 0.3. The number of GraphStar layers = 3, the number of attention head Nhead = 4, hidden unit size per head is 128, and the hidden unit's dropout rate is 0.3. For the training, we keep the modes with the best validation accuracy given the patience setting of 20 epochs. During training, we apply the learning rate of 0.01 and L2 regularization with ρ = 0.002 for all datasets. Table 6: Results of Graph Classification tasks. Left: Science datasets. Right: Text classification datasets. Table 7: Results of graph text classification tasks. Link Prediction Task. The evaluation metric for link prediction task is the average of the AUC and AP scores. The number of attention head Nhead = 4, the number of layer is 3, initial learning rate is 0.0002, and the L2 regularization setting is 0.0005. The same setting is applied to all datasets for link prediction except the hidden size per layer and dropout. The attention coefficient dropout is 0 for all datasets, and the hidden layer dropout rate is 0.2 for Cora and 0 for Citeseer and Pubmed. The hidden unit size per head is 512 for Cora and Citeseer, and 128 for Pubmed. For the training, we keep the model achieving the highest average of the AUC and AP scores on the validation set, given the patience setting of 300 epochs. For link prediction, the [7, 16] state-of-the-art results are reported. Link prediction performance is reported as the combined average of AUC and AP scores. Table 8: Results of link prediction experiments. 4.3 Analysis of Results In this section, we study the inductive node classification training size effect and the stability of graph classification. For PPI dataset, we gradually reduce the original training set graph size and give them to validation and test sets. The result is reported in Table 9. 7 The result indicates that for PPI, GraphStar is particular effective such that, even if we reduce the training size drastically, the training accuracy can still achieve reasonable performance. We also observe the same effect in the IMDB dataset such that, by reducing the original training size of 25000 to 1997 and treating the 23003 as extra validation set, the model still maintains the accuracy of 94.0% on the test set. Table 9: Result of training effectiveness of GraphStar and training size reduction on PPI dataset Figure 2: Stability for test accuracy in 10-fold cross-validation training in graph classification. Left: Result for Enzymes dataset. Right: Result for Protein dataset Additionally, we also study the stability of 10-fold cross validation process for graph classification and present it in Figure 2. The process for Enzyme and Protein data sets are presented here, and as the graph indicates, the experiments are conducted on the 10-fold validation across epochs. All reported accuracy are based on experiment setups described above. The figure indicates that for graph classification, the test results can have high fluctuation for those datasets, and our best results, compared with the current state-of-the-art models, have considerably higher accuracy. 5 Conclusion We have presented graph star net (GraphStar) which can be used in three general graph tasks (node classification, graph classification and link prediction). GraphStar, without incurring heavy computation costs, is able to map the global state effectively. Lastly, we also propose novel approaches to conduct topic-specific sentiment analysis and text classification with graph models. GraphStar has achieved state-of-the-art performance across all tasks. 8 References [1] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. [2] Hongyang Gao, Zhengyang Wang, and Shuiwang Ji. Large-scale learnable graph convolutional networks. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1416 -- 1424. ACM, 2018. [3] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1263 -- 1272. JMLR. org, 2017. [4] Qipeng Guo, Xipeng Qiu, Pengfei Liu, Yunfan Shao, Xiangyang Xue, and Zheng Zhang. Star-transformer. arXiv preprint arXiv:1902.09113, 2019. [5] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1024 -- 1034, 2017. [6] Rie Johnson and Tong Zhang. Supervised and semi-supervised text categorization using lstm for region embeddings. arXiv preprint arXiv:1602.02373, 2016. [7] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. [8] Jia Li, Yu Rong, Hong Cheng, Helen Meng, Wenbing Huang, and Junzhou Huang. Semi- supervised graph classification: A hierarchical graph perspective. 2019. [9] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. [10] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5115 -- 5124, 2017. [11] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social repre- sentations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701 -- 710. ACM, 2014. [12] Devendra Singh Sachan, Manzil Zaheer, and Ruslan Salakhutdinov. Revisiting lstm networks for semi-supervised text classification via mixed objective function. 2018. [13] Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In European Semantic Web Conference, pages 593 -- 607. Springer, 2018. [14] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi- Rad. Collective classification in network data. AI magazine, 29(3):93 -- 93, 2008. [15] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large- scale information network embedding. In Proceedings of the 24th international conference on world wide web, pages 1067 -- 1077. International World Wide Web Conferences Steering Committee, 2015. [16] Phi Vu Tran. Multi-task graph autoencoders. CoRR, abs/1811.02798, 2018. [17] Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex embeddings for simple link prediction. In International Conference on Machine Learning, pages 2071 -- 2080, 2016. [18] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998 -- 6008, 2017. [19] Petar Velickovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. [20] Saurabh Verma and Zhi-Li Zhang. Graph capsule convolutional neural networks. arXiv preprint arXiv:1805.08090, 2018. 9 [21] Daixin Wang, Peng Cui, and Wenwu Zhu. Structural deep network embedding. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1225 -- 1234. ACM, 2016. [22] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7794 -- 7803, 2018. [23] Felix Wu, Tianyi Zhang, Amauri H. Souza Jr., Christopher Fifty, Tao Yu, and Kilian Q. Wein- berger. Simplifying graph convolutional networks. CoRR, abs/1902.07153, 2019. [24] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised Data Augmentation. arXiv e-prints, page arXiv:1904.12848, Apr 2019. [25] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. arXiv preprint arXiv:1806.03536, 2018. [26] Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. arXiv preprint arXiv:1412.6575, 2014. [27] Liang Yao, Chengsheng Mao, and Yuan Luo. Graph convolutional networks for text classifica- tion. arXiv preprint arXiv:1809.05679, 2018. [28] Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hierarchical graph representation learning with differentiable pooling. In Advances in Neural Information Processing Systems, pages 4800 -- 4810, 2018. [29] Jiani Zhang, Xingjian Shi, Junyuan Xie, Hao Ma, Irwin King, and Dit-Yan Yeung. Gaan: Gated attention networks for learning on large and spatiotemporal graphs. arXiv preprint arXiv:1803.07294, 2018. [30] Jie Zhou, Ganqu Cui, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, and Maosong Sun. Graph neural networks: A review of methods and applications. arXiv preprint arXiv:1812.08434, 2018. 10
1802.03984	3	1802	2019-06-10T06:56:25	SPINE: Structural Identity Preserved Inductive Network Embedding	[ "cs.SI" ]	Recent advances in the field of network embedding have shown that low-dimensional network representation is playing a critical role in network analysis. Most existing network embedding methods encode the local proximity of a node, such as the first- and second-order proximities. While being efficient, these methods are short of leveraging the global structural information between nodes distant from each other. In addition, most existing methods learn embeddings on one single fixed network, and thus cannot be generalized to unseen nodes or networks without retraining. In this paper we present SPINE, a method that can jointly capture the local proximity and proximities at any distance, while being inductive to efficiently deal with unseen nodes or networks. Extensive experimental results on benchmark datasets demonstrate the superiority of the proposed framework over the state of the art.	cs.SI	cs	SPINE: Structural Identity Preserved Inductive Network Embedding Junliang Guo , Linli Xu∗ and Jingchang Liu Anhui Province Key Laboratory of Big Data Analysis and Application, School of Computer Science and Technology, University of Science and Technology of China [email protected], [email protected], [email protected] 9 1 0 2 n u J 0 1 ] I S . s c [ 3 v 4 8 9 3 0 . 2 0 8 1 : v i X r a Abstract Recent advances in the field of network embedding have shown that low-dimensional network repre- sentation is playing a critical role in network anal- ysis. Most existing network embedding methods encode the local proximity of a node, such as the first- and second-order proximities. While being efficient, these methods are short of leveraging the global structural information between nodes distant from each other. In addition, most existing meth- ods learn embeddings on one single fixed network, and thus cannot be generalized to unseen nodes or networks without retraining. In this paper we present SPINE, a method that can jointly capture the local proximity and proximities at any distance, while being inductive to efficiently deal with un- seen nodes or networks. Extensive experimental results on benchmark datasets demonstrate the su- periority of the proposed framework over the state of the art. 1 Introduction Network embedding has been successfully applied in a wide variety of network-based machine learning tasks, such as node classification, link prediction, and community detec- tion, etc [Cai et al., 2017; Kipf and Welling, 2016]. Different to the primitive network representation, which suffers from overwhelming high dimensionality and sparsity, network em- bedding aims to learn low-dimensional continuous latent rep- resentations of nodes on a network while preserving the struc- ture and the inherent properties of the network, which can then be exploited effectively in downstream tasks. Most existing network embedding methods approximate local proximities via random walks or specific objective func- tions, followed by various machine learning algorithms with specific objective functions to learn embeddings [Perozzi et al., 2014; Grover and Leskovec, 2016]. Specifically, the local proximity of a node is approximated with the routine of learn- ing the embedding vector of a node by predicting its neigh- borhood, inspired by the word embedding principle [Mikolov ∗Corresponding Author Figure 1: An example of structural identities in the information dif- fusion process. Red arrows indicate information diffusions between nodes (e.g., retweeting in Twitter), and nodes in the same circle are in the same community. et al., 2013a; Mikolov et al., 2013b] which learns the embed- ding vector of a word by predicting its context. However, there still exist some potential issues that need further concerns. On the one hand, local proximity preserved methods generally do not model nodes far from each other in practice. Meanwhile, in real-world network mining tasks, nodes that are far apart but close in structural identity, or in other words, take similar roles, should perform similarly on specific tasks. Figure 1 shows an example of how nodes with different roles perform in the information diffusion pro- cesses on social networks. Nodes with different colors in- dicate different roles in social networks, i.e., structural hole spanners (red nodes), opinion leaders (yellow nodes) and or- dinary users (blue nodes) respectively [Lou and Tang, 2013; Yang et al., 2015b]. Intuitively, nodes with same roles be- have similarly even with a large distance (yellow nodes in Figure 1), which is the property that should be preserved in the embedding space. In the meantime, the local proximity of a node is also crucial in network embedding. For example in Figure 1, nodes in the same community should be clustered tightly in the embedding space. Therefore, a desirable net- work embedding method should preserve the local proximity and the global structural identity of a node simultaneously to represent the node precisely. Unfortunately, most existing methods fail to consider the local and global structural infor- mation at the same time. In principle, it is challenging to interactively integrate the two kinds of information to obtain comprehensive embeddings rather than a trivial linear combi- nation. On the other hand, most existing network embedding ap- proaches are transductive. To be specific, embeddings are learned on a fixed network, and cannot be directly applied NetworkDiffusionsCommunitiesto new joined nodes or other networks. In contrast, induc- tive methods which are able to generalize to unseen nodes or totally new networks are extensively required in real-world applications, e.g., social recommendation for new users, clas- sification of protein functions in various protein-protein inter- action graphs, etc. Unfortunately, traditional network embed- ding principles such as random walk and matrix factorization are impracticable for unseen nodes or networks, which makes inductive network embedding much more challenging than transductive problems. In this paper, we propose SPINE, an inductive network embedding framework which jointly preserves local proximi- ties and structural identities of nodes. We show that structural similarities between node pairs can be represented by a high- order proximity of the network known as Rooted PageRank (RPR) [Liben-Nowell and Kleinberg, 2007], and by assign- ing each node a structural feature vector based on RPR, we can encode structural proximities between nodes by measur- ing the similarities of their structural features. To construct an inductive framework, we learn an embedding generator rather than directly optimizing a unique embedding for each node, through which local proximities are integrated. To fur- ther encode structural identities, we propose a biased Skip- Gram Negative Sampling (SGNS) approach with a novel pos- itive sampling strategy guided by structural similarities be- tween nodes. Furthermore, the objective function of SPINE is carefully designed to enhance the structural information con- tained in the embedding generator. wi,j(cid:80)V 2 SPINE In this section, we propose Structural Identity Preserved In- ductive Network Embedding (SPINE), a novel inductive ap- proach for unsupervised network embedding, which consists of three components: structural feature generation, embed- ding generation and biased SGNS optimization. Problem Definition. Given an undirected network G = {V, E, F}, in which a set of nodes V are connected by a set of edges E, and F ∈ RV ×f is the content matrix of nodes. The adjacency matrix is A where Ai,j = wi,j is the edge weight between node vi and vj, and we denote the corresponding transition matrix as P , where Pi,j = represents the transition probability between node vi and vj. Our goal is to learn E ∈ RV ×d, where d is a small number of latent dimensions. These low-dimensional representations should well preserve the structural properties of G, including local proximities and structural identities, which can be evaluated with downstream tasks such as node classification. 2.1 Rooted PageRank and Structural Identity We start with theoretical preliminaries of our structural fea- ture generation algorithm. Here we introduce a well-known high-order proximity of a network named Rooted PageR- ank (RPR) [Liben-Nowell and Kleinberg, 2007], defined as SRPR = (1 − βRPR)(I − βRPRP )−1, where βRPR ∈ (0, 1) is the probability of the current node randomly walking to a neighbor rather than jumping back to the start node. The (i, j)-th entry of SRPR is the probability that a random walk from node vi will stop at vj in the steady state, which can be k=1 wi,k in a recursive manner as: i + (1 − βRPR)gi used as an indicator of the node-to-node proximity. There- fore, we can use the i-th row of SRPR, denoted as SRPR , to represent the global structural information of node vi. We can further rewrite SRPR i i SRPR i = βRPRP SRPR (1) where gi is the index vector of node vi whose i-th element is 1 while others are 0. Next, we are going to verify that SRPR is able to represent the structural identity of node vi. We first define the complete structural property [Batagelj et al., 1992] of a node as: Definition 1. A node property t : V → R is complete struc- tural if for any automorphism ϕ of every node vi ∈ V , it always satisfies: i t(vi) = t(ϕ(vi)) Examples of complete structural properties include t(vi) = degree of node vi, t(vi) = number of nodes at distance k from vi (k-hop neighbors), t(vi) = the centrality of vi, etc. [Batagelj et al., 1992]. Then the following theorem can be directly derived from [Batagelj et al., 1992]: Theorem 1. Given a structural description of node vi defined by a set of complete structural node properties as: Ti = [t1(vi), t2(vi),··· , tn(vi)] where Ti is an n dimensional vector, and n > 0 is the number of chosen properties. Let d(·,·) denote standard dissimilar- ities between vectors, and vi ≡ vj indicate node vi and vj have equal structural identity, then for ∀vi, vj ∈ V : vi ≡ vj ⇐⇒ d(Ti, Tj) = 0 i From Theorem 1, we can conclude that the structural iden- tities between nodes can be measured through properly de- signed structural vectors. Next we are going to show that SRPR actually represents a complete structural property. Following the examples in Definition 1, the centrality of vi can be regarded as a complete node property of vi. There are multiple ways to measure the centrality of a node, while the original PageRank [Brin and Page, 2012] is exactly a vari- ant of eigenvector centrality. At each iteration, the PageRank value πi of node vi is updated as in [Langville and Meyer, 2011]: and M is the Google matrix defined as: πT i = πT i M , i M = βP + (1 − β)gi1T (2) where gi = ( 1V ,··· , 1V )V and 1 = (1,··· , 1)V . Ac- cording to Equation (1), as a variant of PageRank, the only difference between SRPR and the original PageRank is the choice of gi. Therefore, the target matrix of Rooted PageR- ank can be written as: M RPR (3) where I is the identity matrix. Thus SRPR is the leading left hand eigenvector of M RPR )T = (SRPR is also a variant of eigenvector centrality, thus can be further regarded as a com- plete structural property to represent the structural identity of vi, i.e., Ti = SRPR = βRPRP + (1 − βRPR)I . As a consequence, SRPR i satisfies: (SRPR , i.e., SRPR )T M RPR . i i i i i i i i Algorithm 1 Rooted random walk sampling Input: the graph G, the present node vi, the continuation probability βRPR ∈ (0, 1), hyper-parameters k, m, l Ps = RootedRandomWalk(G, vi, l, βRPR) for vj in Ps do Output: the structural feature vector Ti of vi 1: Initialize a counter Ci ∈ RV 2: repeat 3: 4: 5: 6: 7: until m times 8: Ci ← Sort Ci in a descending order 9: Ti ← The first k elements of Ci 10: return Ti ← Ti/sum(Ti) Ci[j] ← Ci[j] + 1 end for i 2.2 Structural Feature Generation In this section we first state our motivation of choosing SRPR as the structural identity instead of others, based on which the structural feature generation method is introduced. i To construct an inductive method, we expect the length of the structural description of a node is independent of the total number of nodes V and fixed at k. In this paper, we use the top k values of SRPR as the structural description of vi. Compared with other high-order proximities or node proper- ties (e.g., the original PageRank), RPR captures the global structural information of the network, while being tailored to encode the local structural information of root nodes ac- cording to the definition, and thus can better represent the importance of neighbors at various distances to the present node [Haveliwala, 2002]. Moreover, it has been theoretically and empirically proven [Litvak et al., 2007] that in a network with power-law degrees, the RPR values also follow a power- law distribution. Therefore, the largest k RPR values are able to represent the structural information of a node with a suit- ably chosen k. When calculating SRPR , considering the inductive prereq- uisite, the transition matrix P which encodes the structure of the network may be unavailable. Alternatively, we approxi- mate SRPR from the local structure around vi through a Monte Carlo approximation. The complete procedure is summarized in Algorithm 1, where m and l indicate the number of repeats and the length per random walk respectively, and k controls the length of the structural feature vector. The largest k values of SRPR are taken as the structural description of vi, denoted as Ti in the rest of this paper. 2.3 Embedding Generation To construct an inductive network embedding framework, we generate embeddings instead of directly optimizing the em- bedding matrix, through which the structural information and the content information of networks can be jointly incorpo- rated. As stated above, the k values in Ti indicate k largest struc- tural proximities between vi and nodes co-occurring with vi in rooted random walks. We denote the content matrix of the i ∈ Rk×f , which is constructed corresponding k nodes as F k row by row according to the same order of RPR values in i i i i , we Ti. Given the structural features Ti and node content F k propose an embedding generation method for vi. Specifically, we first employ a multilayer percep- tron (MLP) to map nodes from the content space to the em- bedding space, then compute a linear combination of the k vectors with respect to the corresponding weights in Ti. For- mally, denote the dimensionality of embeddings as d, the weight matrix of the MLP is WM ∈ Rf×d, then the embed- ding generation process can be written as: k(cid:88) ei = σ( Ti,jF k i,jWM), (4) j=1 i,j ∈ Rf is the j-th row of F k where F k i , σ is the non-linear activation function, and ei ∈ Rd is the embedding vector of node vi. 2.4 Biased SGNS The Skip-Gram Negative Sampling (SGNS) model is widely used in representation learning, which is based on the princi- ple of learning the embedding vector of a word/node by pre- dicting its neighbors. More formally, given an embedding vector ei, SGNS is minimizing the objective function as: J(ei ep) = −log(σ(eT i ep)) − K · Evn∼Pn(vp)log(σ(−eT i en)), (5) where σ(·) is the sigmoid function and K controls the neg- ative sampling number. vp and vn are positive and negative nodes of vi respectively, while ep and en are the correspond- ing embedding vectors. Technically, negative nodes are sam- pled from a distribution Pn, and for most network embedding methods, positive nodes are defined as nodes that co-occur with vi in a fixed-size window in random walk sequences. In SPINE, to encourage the similarity of embeddings to jointly encode the similarity in terms of structural identities and local proximities simultaneously, we design a novel bi- ased positive sampling strategy based on the structural fea- tures generated from Algorithm 1. The complete procedure is illustrated in Algorithm 2. Specifically, we define a struc- tural rate α ∈ (0, 1) to control the ratio of structural sampling and local proximity sampling. With probability α, a positive sample of vi is sampled according to the similarities between their structural features (starting from line 2). Otherwise, the positive node is sampled from nodes that co-occur near vi on trivial random walks (starting from line 9), which is prepro- cessed and stored in Li. The similarity metric in line 4 can be chosen from Euclidean distance, cosine similarity, Dynamic Time Warping (DTW) [Salvador and Chan, 2007], etc. In our experiments, we use DTW which is designed to compare ordered sequences as the similarity metric. The structural sampling paradigm alleviates the limitation of distance. In practice, it is redundant to compute the struc- tural similarity between vi and all the other nodes, since nodes with completely different local structures are nearly impossible to be sampled as positive pairs through structural sampling. Intuitively, nodes with similar degrees are likely to have similar local structures. Based on this intuition, we reduce the redundancy by only considering nodes which have similar degrees with the present node vi. Specifically, given Algorithm 2 Biased positive sampling Input: the structural feature matrix T , the present node vi, a node list Li, the structural rate α Output: vp, which is a positive sample of vi 1: Initialize an empty list Ps = [] 2: if random(0, 1) < α then for j = 1 to V , j (cid:54)= i do 3: t ← Compute the similarity between Ti and Tj 4: Ps ← Append t to Ps 5: 6: 7: 8: 9: else 10: 11: end if 12: return vp end for Ps ← Normalize Ps to [0, 1] vp ← Sample a node according to Ps vp ← Randomly choose a node from Li an ordered list of node degrees, we choose the candidates for structural sampling by taking O(log V ) nodes in each side from the location of vi. As a consequence, the time com- plexity of structural sampling for each node is reduced from O(V ) to O(log V ). 2.5 Learning and Optimization We introduce the biased SGNS based objective function of our framework in this section. We propose two types of em- beddings for each node vi ∈ V , i.e., a content generated em- bedding ei as defined in Equation (4), and a structure based embedding si which is the i-th row of an auxiliary embed- ding matrix WS ∈ RV ×d. In real-world network-based datasets, the content of nodes is likely to be extremely sparse and weaken the structural information incorporated during the generation process of ei. Therefore we employ a direct interaction between ei and si to strengthen the structural in- formation contained in the learned embeddings. Formally, given the present node vi and its positive sample vj, which is sampled according to Algorithm 2, the pairwise objective function of SPINE can be written as: F (vi, vj) = λ1 · J(ei ej) + λ2 · J(si sj) + (1 − λ1 − λ2) · [J(ei sj) + J(si ej)] (6) where λ1 and λ2 control weights of different parts, and J(··) denotes the pairwise SGNS between two embeddings defined in Equation (5). Intuitively, the generator and the auxiliary embedding matrix should be well trained through their single loss as well as obtaining each other's information through the interaction loss, where λ1 and λ2 determine which one is the primary part. The structure based embeddings WS and the parameters WM of the MLP constitute all the parameters to be learned in our framework. The final objective of our framework is: V (cid:88) i(cid:54)=j min WS,WM .F (vi, vj) (7) By optimizing the above objective function, the embed- ding generator is supposed to contain the content information Method node2vec struc2vec n+s n+s+f SDNE HOPE Graphsage GAT SPINE SPINE-p Citeseer 47.2 41.1 42.7 57.0 45.2 46.1 52.6 72.5† ± 0.7 72.6 ± 0.4 73.8 ± 0.2 Cora 69.8 64.2 68.3 73.7 68.7 67.2 79.8† 83.0† ± 0.7 83.7 ± 0.4 82.2 ± 0.7 Pubmed 70.3 60.7 67.1 67.3 69.1 69.4 75.2 79.0† ± 0.3 78.5 ± 0.3 82.2 ± 0.3 Table 1: Accuracy of transductive node classification (in percent- age). "†" indicates that the results are directly copied from their papers and other results are provided by ourselves. and the structural information as well as local proximities and structural identities simultaneously. Therefore, during infer- ence, we drop the auxiliary embedding matrix WS and only keep the trained embedding generator. In the sequel, embed- dings of unseen nodes can be generated by first construct- ing structural features via Algorithm 1 and then following the paradigm described in Section 2.3. 3 Experiments 3.1 Experimental Setup We test the proposed model on four benchmark datasets to measure its performance on real-world tasks, and one small scale social network to validate the structural identity pre- served in the learned embeddings. For the node classification task, we test our method on Citation Networks [Yang et al., 2016], where nodes and edges represent papers and citations respectively. To test the performance of SPINE while gener- alizing across networks, we further include PPI [Stark et al., 2006], which consists of multiple networks corresponding to different human tissues. To measure the structural identity preserved in embeddings, we test SPINE on a subset of Face- book dataset [Leskovec and Krevl, 2014], denoted as FB-686, in which nodes and links represent users and their connec- tions, and each user is described by a binary vector. As for the baselines, we consider unsupervised net- work embedding methods including node2vec [Grover and Leskovec, 2016], struc2vec [Ribeiro et al., 2017] and their variants. Considering that node2vec and struc2vec are de- signed to preserve the local proximity and the structural iden- tity respectively, we concatenate the corresponding learned embeddings to form a new baseline, denoted as n+s, to il- lustrate the superiority of SPINE over the linear combina- tion of the two proximities. In addition, n+s+f denotes the content-incorporated variant of n+s. We also compare with SDNE [Wang et al., 2016] and HOPE with RPR matrix [Ou et al., 2016] to test the performance of our inductive Rooted PageRank approximation. On the other hand, in addition to transductive methods, we also consider the unsupervised vari- ant of Graphsage [Hamilton et al., 2017], an inductive net- work embedding method which jointly leverages structural and content information. We also report the performance of the state-of-the-art supervised inductive node classification method GAT [Velickovi´c et al., 2017]. Random and raw fea- ture results are also included as baselines in this setting. Citeseer Cora Pubmed Methods Random RawFeats Graphsage SPINE Random RawFeats Graphsage SPINE Random RawFeats Graphsage SPINE 20% 40% 60% 80% 17.7 19.5 63.9 57.7 41.4 58.5 68.7 75.4 20.1 18.8 59.6 66.6 73.1 48.6 77.9 86.7 38.9 38.5 72.9 75.7 76.4 79.9 85.7 78.8 16.7 60.3 47.8 71.5 19.1 64.6 58.8 82.1 39.3 74.6 78.2 83.0 20.4 62.2 53.9 72.1 22.0 64.7 66.4 84.1 39.8 75.4 79.4 83.7 Table 2: Accuracy of inductive node classification w.r.t node re- moval rate (in percentage). Figure 3: Node classification results on PPI. The left vertical axis indicates the micro-F1 score while the right indicates the macro-F1 score. Both are in percentage. tion, SPINE is also more compact in space complexity, as the parameter scale of SPINE during inference is O(f d), com- pared to O(f d + d2) and O((f + KC)d) for Graphsage and GAT with two layers respectively, where K is the number of attention heads and C is the number of classes. Results in the inductive setting are reported in Table 2 and Figure 3. Note that in this setting we use all the remaining nodes as training data during classification, thus the results are generally larger than that under the transductive setting. One can observe that SPINE outperforms all the baselines, in- dicating the generalization capability of the embedding gen- erator learned by optimizing our carefully designed objective function. In addition, with the increasing node removal rate which leads to greater loss of local proximities, Graphsage can perform worse than raw features, indicating the limita- tion of the methods that only preserve local proximities. In contrast, SPINE alleviates the sparsity of local information by incorporating structural identities. 3.3 Structural Identity We proceed to investigate the structural identity on the FB- 686 dataset here. We consider the transductive setting here, and the results under inductive setting can be found in the sup- plementary material. Specifically, for the original network, we construct a mirror network and relabel the nodes, and consider the union of two networks as the input. As a con- sequence, node pairs between original nodes and their mir- ror nodes are obviously structurally equivalent, thus should be projected close in the embeddings space. We then evalu- ate the Euclidean distance distribution between embeddings of the mirror node pairs and all the node pairs connected by Figure 2: Comparison of running time on Cora, with training batches of size 512 and inference on the full test set (1000 nodes). For our method, we use SPINE and SPINE-p to indicate the variants with WS randomly initialized and pretrained with node2vec respectively. To make predictions based on the em- beddings learned by unsupervised models, we use one-vs-rest logistic regression as the downstream classifier. For all the methods, the dimensionality of embeddings is set to 2001. 3.2 Node Classification We first evaluate the performance of SPINE on node classi- fication, a common network mining task. Specifically, we conduct the experiments in both transductive and inductive settings. For the transductive setting, we use the same scheme of training/test partition provided by Yang et al. [2016]. As for the inductive setting, on citation networks, we randomly remove 20%, 40%, 60% and 80% nodes and the correspond- ing edges, these nodes are then treated as test nodes with the remaining network as the training data. Meanwhile on the PPI network, we follow the same dataset splitting strategy as in [Hamilton et al., 2017], i.e., 20 networks for training, 2 for validation and 2 for testing, where the validation and testing networks remain unseen during training. For both settings we repeat the process 10 times and report the mean score. Results in the transductive setting are reported in Table 1. We can observe that SPINE outperforms all unsupervised em- bedding methods, and performs comparably with the state-of- the-art supervised framework GAT. In addition, n+s performs worse than node2vec, which implies that a simple linear com- bination of local proximity preserved and structural identity preserved embeddings is incapable of generating a meaning- ful representation that effectively integrates the two compo- nents. The superiority of SPINE over SDNE and HOPE in- dicates the efficacy of the inductive RPR approximation al- gorithm as well as the joint consideration of local proximity and structural identity. SPINE also outperforms the content- augmented variant n+s+f, which shows that the content ag- gregation method we propose can better consolidate the con- tent and structure information. Furthermore, the comparison between the two variants of SPINE indicates that while the basic model of SPINE already achieves a competitive per- formance, we can further enhance the model with initializa- tions of WS that are well pretrained by focusing only on local proximity, which also justifies the effectiveness of the paradigm of interactive integration proposed in Section 2.5. As for the comparison on training and test runtime, results are shown in Figure 2. Obviously, SPINE is more efficient in time complexity, especially in the inference stage. In addi- 1Code is avaliable at https://github.com/lemmonation/spine SPINEGraphsageGAT10−310−210−1100Time (Seconds)Training (per batch)Inference (full test set)RandomFeatsGraphsageSPINE363840424446485052micro-F1 scoreNode classification on PPI22242628303234macro-F1 scoremicro-F1 scoremacro-F1 scoreSPINE can robustly distinguish and preserve structural iden- tity as well as local proximity even with severe perturbations. 4 Related Work Most network embedding methods consider to preserve lo- cal proximity between nodes with frameworks based on ran- dom walk [Perozzi et al., 2014; Grover and Leskovec, 2016], skip-gram [Tang et al., 2015; Cao et al., 2015], matrix fac- torization [Yang et al., 2015a; Guo et al., 2017] and deep learning [Wang et al., 2016; Gao and Huang, 2018] respec- tively. However, it is worth noting that few of the exist- ing works consider the structural identity between nodes, and fail to handle proximities between nodes at distances. Struc2vec [Ribeiro et al., 2017] preserves the structural iden- tity by constructing a multi-layer complete graph and exe- cute random walk on it. HOPE [Ou et al., 2016] captures structural identity through factorizing a global Rooted PageR- ank matrix. However, while preserving the structural identity, they ignore the basic local proximities of nodes, which limits its applicability on real-world network mining tasks. Similar problems also occur in two recent methods [Tu et al., 2018; Zhang et al., 2018]. SDNE [Wang et al., 2016], a deep learn- ing based method, is only able to take the first- and second- order proximities into account. Furthermore, most of the methods mentioned above are transductive. Inductive meth- ods [Hamilton et al., 2017; Velickovi´c et al., 2017] tackles this challenge by recursively training a set of aggregators for each node to integrate its neighbors' content as the embed- ding of the current node in every iteration. As nodes at the k-th iteration contain the structural information from their neighbors within k hops, they cannot deal with nodes at ar- bitrary distances unless with sufficient iterations, which is costly for real-world tasks. 5 Conclusion In this paper, we propose SPINE, a network embedding ap- proach which is able to jointly preserve structural identities and local proximities of nodes while being generalized to un- seen nodes or networks. We assign a structural feature vec- tor to each node based on Rooted PageRank, and we learn an embedding generator leveraging the structural features of each node to incorporate the structural and content informa- tion of nearby nodes. In addition, we propose a biased SGNS algorithm with a novel positive sampling procedure, based on which a carefully designed objective function is proposed to enhance the structural information contained in the embed- ding generator. Extensive experiments demonstrate the supe- riority of SPINE over the state-of-art baselines on both trans- ductive and inductive tasks. In future work, we are interested in introducing structural identity to other network-based tasks such as social recommendation. Acknowledgements This research was supported by the National Natural Science Foundation of China (No. 61673364, No. 91746301) and the Fundamental Research Funds for the Central Universities (WK2150110008). Figure 4: Euclidean distance distribution between mirrored node pairs and connected node pairs on the FB-686 dataset. Figure 5: Euclidean distance distribution between mirrored node pairs and connected node pairs on the FB-686 datasets with vary- ing degrees of edge removal and changing content. edges, denoted as Pm and Pa respectively. Intuitively, if em- beddings successfully preserve structural identities of nodes, E[Pm] should be much smaller than E[Pa]. Results of SPINE and struc2vec with respect to the two distributions are shown in Figure 4. Obviously, compared to struc2vec, embeddings learned by SPINE yield smaller dis- tances between both mirrored node pairs and connected node pairs, indicating the structural identity and local proximity are jointly preserved better. In addition, the ratio between E[Pa] and E[Pm] is 13.40 and 5.72 for SPINE and struc2vec respec- tively, which means SPINE distinguishes the two proximities more clearly. Further to test the robustness of SPINE to edge removal and changing content, we randomly sample two new net- works from the original FB-686 network. Specifically, we preserve each edge in the original network with probability s, and randomly exchange a 1's location with another 0's lo- cation in each node's content vector. Consequently, from the view of structure, the probability for an original edge con- tained both in the two generated networks is s2, and smaller s indicates less structure correlation between the two gener- ated networks. From the view of content, mirrored nodes are nearly impossible to have identical content due to the spar- sity of content vectors. As can be observed in Figure 5, the ratio between E[Pa] and E[Pm] is not significantly affected by the degree of structure perturbation s, which indicates that 012345678Euclidean Distance (d)0.00.20.40.60.81.0P[D ≥ d]all - SPINEmirrored - SPINEall - struc2vecmirrored - struc2vecall - averagemirrored - average0123456Euclidean Distance (d)0.00.20.40.60.81.0P[D ≥ d]s=1.0 alls=1.0 mirroreds=0.8 alls=0.8 mirroreds=0.6 alls=0.6 mirroreds=0.4 alls=0.4 mirroreds=0.2 alls=0.2 mirroredReferences [Batagelj et al., 1992] Vladimir Batagelj, Anuska Ferligoj, and Patrick Doreian. Direct and indirect methods for struc- tural equivalence. Social networks, 14(1):63 -- 90, 1992. [Brin and Page, 2012] Sergey Brin and Lawrence Page. Reprint of: The anatomy of a large-scale hypertextual web search engine. Computer networks, 56(18):3825 -- 3833, 2012. [Cai et al., 2017] Hongyun Cai, Vincent W Zheng, and Kevin Chen-Chuan Chang. A comprehensive survey of graph embedding: Problems, techniques and applications. arXiv preprint arXiv:1709.07604, 2017. [Cao et al., 2015] Shaosheng Cao, Wei Lu, and Qiongkai Xu. Grarep: Learning graph representations with global In Proceedings of the 24th ACM structural information. International on Conference on Information and Knowl- edge Management. ACM, 2015. [Gao and Huang, 2018] Hongchang Gao and Heng Huang. Deep attributed network embedding. In IJCAI, 2018. [Grover and Leskovec, 2016] Aditya Grover Jure node2vec: Scalable feature learning for Leskovec. networks. In KDD, pages 855 -- 864. ACM, 2016. and [Guo et al., 2017] Junliang Guo, Linli Xu, Xunpeng Huang, and Enhong Chen. Enhancing network embedding with auxiliary information: An explicit matrix factorization perspective. arXiv preprint arXiv:1711.04094, 2017. [Hamilton et al., 2017] Will Hamilton, Zhitao Ying, and Jure Inductive representation learning on large In Advances in Neural Information Processing Leskovec. graphs. Systems, pages 1024 -- 1034, 2017. [Haveliwala, 2002] Taher H Haveliwala. Topic-sensitive pagerank. In WWW, pages 517 -- 526. ACM, 2002. [Kipf and Welling, 2016] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. [Langville and Meyer, 2011] Amy N Langville and Carl D Meyer. Google's PageRank and beyond: The science of search engine rankings. Princeton University Press, 2011. and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data, June 2014. [Leskovec and Krevl, 2014] Jure Leskovec [Liben-Nowell and Kleinberg, 2007] David Liben-Nowell and Jon Kleinberg. The link-prediction problem for social journal of the Association for Information networks. Science and Technology, 58(7):1019 -- 1031, 2007. [Litvak et al., 2007] Nelly Litvak, Werner RW Scheinhardt, and Yana Volkovich. In-degree and pagerank: why do they follow similar power laws? Internet mathematics, 2007. [Lou and Tang, 2013] Tiancheng Lou and Jie Tang. Mining structural hole spanners through information diffusion in social networks. In WWW, pages 825 -- 836. ACM, 2013. [Mikolov et al., 2013a] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. preprint arXiv:1301.3781, 2013. [Mikolov et al., 2013b] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed repre- sentations of words and phrases and their compositional- ity. In Advances in neural information processing systems, pages 3111 -- 3119, 2013. [Ou et al., 2016] Mingdong Ou, Peng Cui, Jian Pei, Ziwei Zhang, and Wenwu Zhu. Asymmetric transitivity preserv- ing graph embedding. In KDD, pages 1105 -- 1114. ACM, 2016. [Perozzi et al., 2014] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social rep- resentations. In KDD, pages 701 -- 710. ACM, 2014. [Ribeiro et al., 2017] Leonardo FR Ribeiro, Pedro HP Saverese, and Daniel R Figueiredo. struc2vec: Learning In KDD, node representations from structural identity. pages 385 -- 394. ACM, 2017. [Salvador and Chan, 2007] Stan Salvador and Philip Chan. Toward accurate dynamic time warping in linear time and space. Intelligent Data Analysis, 11(5):561 -- 580, 2007. [Stark et al., 2006] Chris Stark, Bobby-Joe Breitkreutz, Teresa Reguly, Lorrie Boucher, Ashton Breitkreutz, and Mike Tyers. Biogrid: a general repository for interaction datasets. Nucleic acids research, 34(suppl 1):D535 -- D539, 2006. [Tang et al., 2015] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large- In WWW, pages scale information network embedding. 1067 -- 1077, 2015. [Tu et al., 2018] Ke Tu, Peng Cui, Xiao Wang, Philip S Yu, and Wenwu Zhu. Deep recursive network embedding with In KDD, pages 2357 -- 2366. ACM, regular equivalence. 2018. [Velickovi´c et al., 2017] Petar Velickovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li`o, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. [Wang et al., 2016] Daixin Wang, Peng Cui, and Wenwu Zhu. Structural deep network embedding. In KDD, pages 1225 -- 1234. ACM, 2016. [Yang et al., 2015a] Cheng Yang, Zhiyuan Liu, Deli Zhao, Maosong Sun, and Edward Y Chang. Network represen- tation learning with rich text information. In IJCAI, 2015. [Yang et al., 2015b] Yang Yang, Jie Tang, Cane Wing-ki Le- ung, Yizhou Sun, Qicong Chen, Juanzi Li, and Qiang Yang. Rain: Social role-aware information diffusion. In AAAI, 2015. [Yang et al., 2016] Zhilin Yang, William W Cohen, and Rus- lan Salakhutdinov. Revisiting semi-supervised learning with graph embeddings. preprint arXiv:1603.08861, 2016. [Zhang et al., 2018] Ziwei Zhang, Peng Cui, Xiao Wang, Jian Pei, Xuanrong Yao, and Wenwu Zhu. Arbitrary-order proximity preserved network embedding. In KDD, pages 2778 -- 2786. ACM, 2018. Algorithm 3 RootedRandomWalk Input: the graph G, the present node vi, the continuation probabil- Randomly select a node vj from the neighbors of vcur Append vj to the end of Ps ity βRPR ∈ (0, 1), the walk length l vcur ← the last element of Ps if random(0, 1) < βRPR then Output: a rooted random walk sequence Ps 1: Initialize a list Ps = [vi] 2: for j = 1 to l − 1 do 3: 4: 5: 6: 7: 8: end if 9: 10: end for 11: return Ps Append vi to the end of Ps else Dataset Citeseer Cora Pubmed PPI FB-686 # Classes # Nodes 6 7 3 121 - 3327 2708 19717 56944 168 # Edges 4732 5429 44338 818716 3312 # Features 3703 1433 500 50 63 Table 3: Dataset statistics A Rooted Random Walk Please refer to Algorithm 3 for details of the Monte Carlo approximation of rooted random walk. B Dataset Details We test the proposed model on four benchmark datasets. As most existing methods are transductive, we adopt three static datasets and one across network dataset to measure the trans- ductive and inductive performance of SPINE respectively. The statistics of datasets are summarized in Table 3. Among them, the static datasets are Citation Networks: Citeseer contains 3312 publications of 6 different classes and 4732 edges between them. Nodes represent papers and edges represent citations. Each paper is described by a one- hot vector of 3703 unique words. Cora includes 2708 publications of machine learning with 7 different classes. Similar to Citeseer, Cora contains 5429 citation links between them. And each paper is described by a one-hot vector of 1433 unique words. Pubmed consists of 19717 scientific publications pertain- ing to diabetes classified into one of three classes. It contains 44338 citation links and each document is described by a TFIDF weighted word vector from a dictionary of 500 unique words. To test the performance of SPINE while generalizing across networks, we further introduce the PPI dataset: PPI contains various protein-protein interactions, where each graph corresponds to a different human tissue. Nodes represent proteins with 121 different cellular functions from gene ontology as their labels. Figure 6: Classification results with different α on subPPI. The left and right vertical axes indicates micro-F1 and macro-F1 score re- spectively. Both are in percentage. FB-686 is a subset of Facebook dataset, which consists of 168 users with 3312 links between them. Each user is de- scribed by a 0-1 vector of 63 features. C Hyperparameter Settings We keep the following settings for all tasks and datasets: ra- tios λ1 and λ2 are set to 0.4 and 0.2 respectively, while the structural rate α and the restart rate βRPR of rooted random walk are both set to 0.5. The learning rate of the Adam op- timizer is set to 0.001. For methods that leverage random walks, we set the number of repeats for each node to 10, the length of random walks to 40 and the window size to 5 for a fair comparison. D Parameter Study In order to evaluate the influence of parameters on the perfor- mance of SPINE, we conduct experiments on a subset of PPI datasets, denoted as subPPI, which contains 3 training net- works and 1 test network. We first investigate the effect of the structural rate α, then test the robustness of SPINE with respect to varying λ1 and λ2 values, both on node classifica- tion. D.1 Effect of α We vary the value of the structural rate α from 0 to 1 with an interval of 0.2, and report the results in Figure 6. As expected, with increasing α values, the classification performance in- creases first and then decreases when α gets too large. The results show that only considering the structural identity or local proximity will impair the quality of learned embeddings in real-world tasks. Therefore, the structural rate α is crucial for enhancing the quality of learned embeddings. D.2 Effects of λ1 and λ2 To investigate the influence of ratios λ1 and λ2, we vary the value of λ1 from 0.2 to 1.0 and the value of λ2 from 0.2 to 0.6, both with an interval of 0.2. Results of node classifica- tion on subPPI are reported in Table 4, from which we can conclude that the interaction between the generator and WS successfully enhances the structure information carried in the generator as expected, as λ1 = 1.0 (no interaction) achieves the worst performance. 0.00.20.40.60.81.046.046.547.047.548.0micro-F1 scoreNode classification on subPPI30.531.031.532.032.533.0macro-F1 scoremicro-F1 scoremacro-F1 scoreFigure 7: Case study for nodes with few neighbors. The top half are the current nodes and their local structures, and the bottom half are the nodes with highest structural similarities to the current node, as well as their corresponding local structures. Different colors indicate different classes. Both GraphSAGE and node2vec misclassify the current nodes. In (a) we present the current node with two neighbors, while the detected nodes have similar local and class information. Same situation occurs in (b). Table 4: Results of node classification on subPPI with varying λ1 and λ2 values (in percentage) λ1 0.2 0.2 0.2 0.4 0.4 0.6 1.0 λ2 0.2 0.4 0.6 0.2 0.4 0.2 0.0 micro-F1 macro-F1 47.1 47.0 46.6 47.2 47.8 46.7 45.4 32.0 31.4 30.5 32.1 30.5 31.2 30.4 Figure 8: Case study for nodes with many neighbors. Following same settings in Figure 7. E Case Study To intuitively analyse the impact of jointly considering lo- cal proximity and structural identity on classification perfor- mance, we conduct case studies on the Cora dataset, in the transductive setting. Specifically, we select several represen- tative case from nodes that GraphSAGE and node2vec both misclassify. From our observation, the integration of local proximity and structural identity can significantly benefit the classification performance on two types of nodes: nodes with few or lots of neighbors. We first illustrate cases of nodes with few neighbors in Fig- ure 7. These nodes are similar to the ordinary users we dis- cussed in Introduction. (blue nodes in Figure 1), which are also the majority type of nodes in real-world network dataset. Therefore the classification performance on these nodes has a heavy impact on the overall results. However, normally these nodes have few neighbors, which implies limited local prox- imity to utilize, inhibiting the methods that only consider lo- cal proximity. In Cora, there are 38.3% nodes with no more than 2 neighbors, and this percentage rises to 50.7% among the nodes misclassified by node2vec or GraphSAGE, indicat- ing it is harder for local proximity methods to deal with ordi- nary nodes than nodes with more neighbors. SPINE handles this issue by introducing structural identi- ties. As illustrated in Figure 7, we list the local structure of top 3 or 5 nodes with highest structural similarity (computed by line 4, Algorithm 3) to the current node. Obviously they have similar local structures, e.g., both have similar numbers of 1-hop and 2-hop neighbors, and more importantly, most of these nodes have the same label with the current node. SPINE successfully detects these nodes as expected. Thus, instead of singly leveraging the current node's local information, we are able to leverage much more structural and local information around these selected nodes. Finally, embeddings with higher quality are learned and better classification performance are achieved. Another case of nodes that SPINE can better deal with is those with lots of neighbors but in different classes. These nodes act as bridges between different academic fields or communities, also know as the structural hole spanners as discussed in Introduction (red nodes in Figure 1). There- fore, these nodes tend to have many neighbors from different classes, making it hard to judge their class only depending on the local information, as illustrated in Figure 8. Again, we alleviate this problem by jointly considering lo- cal proximity and structural identity. We also list the 3 most similar nodes for each case. From Figure 8, we can observe that the nodes we select not only have similar local structures to the current node, but also have many neighbors in the same class to the current node at the same time. As a benefit of the extra local structure information, better performance is achieved in classifying this kind of nodes, comparing to the local proximity methods. Current Node1-hop neighbors2-hop neighborsMost similar nodes(a)(b)Dataset Citeseer Cora Pubmed Pearson (p-value) Spearman (p-value) 0.72 (0.0) 0.77 (0.0) 0.78 (0.0) 0.74 (0.0) 0.79 (0.0) 0.84 (0.0) to the generalization ability to identify similar structural iden- tities in unseen networks. The two distributions learned from Graphsage are practically identical, justifying our intuition and the necessity of preserving structural identities. Table 5: Pearson and Spearman coefficients between structural dis- tance and Euclidean distance for connected node pairs on citation networks. Figure 9: Euclidean distance distribution between inductively trained embeddings of mirrored node pairs and connected node pairs on the perturbed FB-686 with s = 0.2. F Experiments on Structural Identity F.1 Real World Tasks We verify that embeddings learned by SPINE also preserve structural identities in real-world tasks. We compute the cor- relation between the structural distance (or similarity) defined in line 4, Algorithm 3 and the Euclidean distance in the em- bedding space for all the connected node pairs. The values of correlation measured by Pearson and Spearman coefficients are listed in Table 5, which indicates that there indeed exists a strong correlation between the two distances, validating that SPINE successfully preserves the defined structural similarity in the embedding space. Inductive Setting F.2 To verify whether SPINE can capture the structural identity across networks, we generate four new networks G1, G2, G3 and G4, from the original FB-686 network with s = 0.2. The combination of G1 and G2 is considered as training networks, while G3 and G4 constitute the test data. After training, the embeddings of nodes in G3 and G4 are inferred, on which we compute the distance distributions between embeddings of connected node pairs and mirrored node pairs. Intuitively, as G1 and G2 are separated, baselines which only consider local proximities are not able to capture the structural similarity be- tween networks. Results are shown in Figure 9. Although the structural correlation between training and testing networks is small given s = 0.2, the two distance distributions learned by SPINE are still strikingly different, indicating that SPINE can learn high-level representations of structural identities from training networks rather than just storing them, which leads 012345Euclidean Distance (d)0.00.20.40.60.81.0P[D d]all - SPINEmirrored - SPINEall - Graphsagemirrored - Graphsageall - averagemirrored - average
1607.02714	1	1607	2016-07-10T08:41:53	Privacy Leakage through Innocent Content Sharing in Online Social Networks	[ "cs.SI", "cs.CY" ]	The increased popularity and ubiquitous availability of online social networks and globalised Internet access have affected the way in which people share content. The information that users willingly disclose on these platforms can be used for various purposes, from building consumer models for advertising, to inferring personal, potentially invasive, information. In this work, we use Twitter, Instagram and Foursquare data to convey the idea that the content shared by users, especially when aggregated across platforms, can potentially disclose more information than was originally intended. We perform two case studies: First, we perform user de-anonymization by mimicking the scenario of finding the identity of a user making anonymous posts within a group of users. Empirical evaluation on a sample of real-world social network profiles suggests that cross-platform aggregation introduces significant performance gains in user identification. In the second task, we show that it is possible to infer physical location visits of a user on the basis of shared Twitter and Instagram content. We present an informativeness scoring function which estimates the relevance and novelty of a shared piece of information with respect to an inference task. This measure is validated using an active learning framework which chooses the most informative content at each given point in time. Based on a large-scale data sample, we show that by doing this, we can attain an improved inference performance. In some cases this performance exceeds even the use of the user's full timeline.	cs.SI	cs	Privacy Leakage through Innocent Content Sharing in Online Social Networks Maria Han Veiga Inst. of Computational Science University of Zurich, Switzerland [email protected] 6 1 0 2 l u J 0 1 ] I S . s c [ 1 v 4 1 7 2 0 . 7 0 6 1 : v i X r a ABSTRACT The increased popularity and ubiquitous availability of on- line social networks and globalised Internet access have af- fected the way in which people share content. The informa- tion that users willingly disclose on these platforms can be used for various purposes, from building consumer models for advertising, to inferring personal, potentially invasive, information. In this work, we use Twitter, Instagram and Foursquare data to convey the idea that the content shared by users, especially when aggregated across platforms, can potentially disclose more information than was originally intended. We perform two case studies: First, we perform user de- anonymization by mimicking the scenario of finding the iden- tity of a user making anonymous posts within a group of users. Empirical evaluation on a sample of real-world social network profiles suggests that cross-platform aggregation in- troduces significant performance gains in user identification. In the second task, we show that it is possible to infer physical location visits of a user on the basis of shared Twit- ter and Instagram content. We present an informativeness scoring function which estimates the relevance and novelty of a shared piece of information with respect to an inference task. This measure is validated using an active learning framework which chooses the most informative content at each given point in time. Based on a large-scale data sam- ple, we show that by doing this, we can attain an improved inference performance. In some cases this performance ex- ceeds even the use of the user's full timeline. Keywords Privacy, Social networks, Information value 1. INTRODUCTION User privacy is a topic that has increasingly gained trac- tion with the rise of online social networks (OSN). These platforms allow users to communicate, connect with peers and share content. Originally, OSNs mainly focused on Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full cita- tion on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- publish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. SIGIR PIR '16 Pisa, Italy c(cid:13) 2016 ACM. ISBN 123-4567-24-567/08/06. . . $15.00 DOI: 10.475/123 4 Carsten Eickhoff Dept. of Computer Science ETH Zurich, Switzerland [email protected] these core aspects, but nowadays the term also includes plat- forms which are primarily user-centric, allowing members to broadcast personal thoughts and content. In 2010, [13] find OSNs among the most frequently visited Web sites for a large population of users. Due to their prevalence and abundance in personal content, OSNs lend themselves to the study of human behavior at scale [14]. Recent successful initial public offerings (IPO) and high market valuations underline the monetary value of OSNs. However, the relation between the number of registered users, their online activity, and these valuations is not entirely clear. It has been shown in several studies that user char- acteristics, such as personality traits [12] or future route intentions [15], can be reliably inferred from corresponding OSN profiles. Although the general value of personal data is widely accepted, there have not been many studies which assign a tangible value to OSN profiles. As a consequence, both for users as well as platform providers, the value of in- formation remains a vague notion, at best. This situation is detrimental both to users who cannot be expected to make informed decisions about privacy controls, as long as they do not know the value and potential risk of disclosing a given information item, as well as platform and service providers who blindly buy and sell user data in bulk instead of sav- ing resources by concentrating on select relevant portions of information. In this paper, we aim to draw attention to accidental pri- vacy leakage through content sharing in online social net- works and make a first step toward describing a formal met- ric of task-specific informativeness of pieces of shared con- tent. Our empirical study relies on three popular OSN plat- forms: (1) Twitter, a microblogging platform whose main content comes in tweets, posts limited to 140 characters which can contain text, media (video or images), links to external Web sites, references to other users and hashtags (terms starting with the # symbol, which are used to mark keywords or topics in a tweet). (2) Instagram, a photo shar- ing platform. Its main content are visual in nature along with optional textual descriptors. (3) Foursquare, a location service platform concentrating on the notion of check-ins. Check-ins correspond to real-world venues that the user has visited. In addition to the venue name, more information such as location and venue categories are available. Our investigation is driven by the following research ques- tions: 1. How well can we uniquely identify a user based on matching a set of unseen posts to a user's online foot- print and is there a benefit in modelling user identities across more than one OSN? 2. For the same user, is the information posted in one OSN indicative of the information contained in another OSN? 3. Can we quantify the amount of new information that a piece of shared content carries, with respect to a concrete inference task? In particular, to address RQ 1, we mimic the following scenario: let us have a collection of users' online footprints and a set of q anonymous posts. We test whether we are able to correctly find the author of the anonymous posts based on seeing part of their online footprint. As for RQ 2 and 3, we consider that a user may uninten- tionally expose personal information through seemingly in- nocuous shared content. For example, when a user shares a venue check-in, it is easy to infer which venue category was visited. However, a post which does not mention a place explicitly might still contain information about a potential behaviour or visit intention. Consider two tweets from the same user: "Lol should start heading to the gym #fitness" and "What a great sunny day! ". It is clear that the first tweet contains more information about the user's intention to visit a venue type than the second. To this end, we devise an informativeness metric for shared content. The metric ex- plicitly models the item's relevancy towards a given inference task as well as its novelty in comparison to the previously seen timeline. Such a score can serve as an indication of the amount of novel information disclosure associated with an information item and can, in the future help both service providers as well as privacy advocates in making informed decisions. This paper makes three novel contributions beyond the state of the art: 1. We formulate a scoring function to quantify the infor- mation value of shared content from the perspective of improving the performance of a concrete inference task. 2. We present a practical way of exploiting the theoretical model by integration into an active learning scheme that results in enhanced user model learning rates. 3. In both practical settings, we put particular emphasis on cross-platform models of user identity The remainder of this paper is structured as follows: Sec- tion 2 gives an overview of related work dedicated to user modelling as well as privacy protection in Web scenarios. On the basis of a parallel corpus of OSN profiles belonging to the same natural person, Section 4 discusses unique user identification methods employing intra-platform as well as cross-platform information. Section 5 formally describes an informativeness score for shared OSN content and applies it to the task of predicting user traits manifested on one plat- form based on the user's activity on other OSNs. Finally, Section 6 concludes with a brief discussion of our main find- ings as well as an outlook on ongoing and future extensions to this work. 2. RELATED WORK There is a wealth of work dedicated to privacy protection in Web information systems such as search engines or social networks. A number of early studies investigate the com- mon privacy concerns of information system users [17, 18, 16, 9], finding that general concerns are abundant among Internet users but remain vague and imprecise. Many users are aware of the information collection and behavioral profil- ing activities undertaken by service providers as well as the wide range of data-driven inference efforts that have been presented by the academic and industrial communities [15, 12]. In spite of this knowledge, however, even technology- affine users cannot reliably quantify the exact risks entailed by careless information disclosure. Privacy concerns become especially prevalent in mobile computing environments [11]. De Montjoye et al. [6] show that as few as four hourly GPS samples are enough to uniquely identify 95% of all individuals in a 500k-user phone log. We encounter an even greater potential for privacy hazards in settings that go beyond raw positional traces, joining them with topical information, e.g., in Web search queries [19] or contextual advertising [1]. To counter such de-anonymization and tracking efforts, various strategies have been proposed. Dwork's concept of differential privacy [7] considers adding -noise to aggregate queries that prevents singling out individual contributions to the overall aggregate. Similarly, Carpineto and Romano [4] rely on the notion of k-anonymity, ensuring that no query should return less than k individual records. In the domain of personal information, these approaches may not go far enough since certain, frequent, characteristics that would neither be detected under k-anonymity nor differential pri- vacy could cause severe privacy hazards. This paper, in spirit, follows the reasoning of Howard et al. [10] by measuring not just the amount of information contained in a given message, but also with respect to an inference task which can be economically relevant. In this way it attributes an economic dimension to messages, which can be an interesting measure both for industry players as well as for the message's original author. On the basis of a number of concrete classification tasks, this paper aims to close the gap between the rich body of work on empirical analysis of privacy hazards on the one hand and the large range of available privacy protection measures on the other. We argue that only by understanding the concrete implica- tions of information disclosure (e.g., in the form of the value of a piece of information) can users be expected to make ed- ucated decisions about the appropriate protection measures they are willing to take. 3. DATASET As our research questions are concerned with the rela- tionship between parallel user profiles of the same natural person across different OSNs, we rely on the methodology described in [8] to assemble our dataset. We obtain a collec- tion of 618 distinct users who cross-post content from cor- responding profiles in multiple social networks, totalling 1.1 million tweets, 18000 Instagram posts and 99000 Foursquare check-ins. 4. USER DE-ANONYMIZATION Our first use case is concerned with, given a number of anonymous social media posts q and a collection of users U, finding the particular user u ∈ U that authored the posts. Inspired by general text matching strategies [2], each user's known previous posts are described in the form of a unigram language model Mu and the likelihood of said user having authored the anonymous text q corresponds to p(Muq). Us- ing Bayes' law, one can write: p(uq) = p(qu)p(u) p(q) And to select the most likely user: arg max u∈U p(qu) (1) (2) To simplify the expression further, we assume that p(q) is constant for all users and treat p(u) as uniform across all u ∈ U. Thus, we find the most likely user by estimating p(qu), the probability of posts q being generated by the language model derived from u's available timeline. Figure 1: F-1 score curve of classifiers for conditions 1-2 varying sample size and training data size These timelines are projected into a n-dimentional TF- IDF weighted vector space. To preserve the natural way in which users write, no further vocabulary pre-processing (such as lemmatization or exclusion of less common words) was applied. Based on this representation, we estimate p(qu) as the product across all terms t in the vocabulary: p(tMu) (3) p(qu) ∝ (cid:89) t∈V As described in Section 3, the dataset contains the on- line footprint of the same user on Twitter, Instagram and Foursquare. To mimic the described task, textual data from one OSN is used as the source of anonymous posts and the textual data from the remaining two OSNs is used to gen- erate the user language model p(Muq). To generate the training data, randomly sampled sections of varying length from the training source are used to gener- ate pairs of the form (Mu, user). For the test set, we remove any form of user mentions to mimic an anonymous post. Our experiments investigate a number of combinations of (training, test) data sources: 1. (Twitter, Twitter) 2. (Twitter + Foursquare + Instagram, Twitter) 3. (Twitter, Foursquare) 4. (Twitter + Instagram, Foursquare) 5. (Twitter, Instagram) 6. (Twitter + Foursquare, Instagram) For the first two cases, we split the Twitter timeline into separate training and test portions. However, the more in- teresting conditions are 3-6, as the source of anonymous posts does not come from the data source used to gener- ate the language model. Additionally, we vary the amount of available profiling in- formation by successively revealing larger parts of the train- ing data. Furthermore, we also study the impact of changing the size of the set of anonymous posts q. Multiple training sources are combined (Conditions 2, 4, and 6) in the following way: for a fixed amount of available Figure 2: F-1 score curve of classifiers for conditions 3-4 varying sample size and training data size profiling information, 20% of it is made up from Instagram or Foursquare data (or 40% if these are put in together) and the rest from Twitter, due to the relative abundance of Twitter data. The performance of our classifier is given by accuracy of predicting the correct user who generated the anonymous posts. The results are averaged across 10 randomization runs. In Figures 1-3, the micro averaged F-1 score curves for the classifiers built under different conditions are shown. In Table 1, the results for Conditions 1-2 can be found. We note that the usage of additional OSNs does not improve the de-anonymization performance. This is not too surprising as the source of the anonymous posts come from Twitter. The results for Conditions 3-4 can be found in Table 2. We remark that the classifier's performance does not improve with the addition of Instagram data as users can cross-post check-ins on Twitter and users can check-in into a venue multiple times. The more interesting results can be found in Table 3, which presents the results for Conditions 5-6. We note that in this case, as training and test data come from different sources, there is some improvement in the de-anonymization performance when we include Instagram data as an extra training source. With respect to RQ 1, we note that it is possible to match Table 3: User de-anonymization for different sizes of training and test data, using Instagram data as a source for anonymous posts Train Source Posts Seen 1 Twitter Twitter Foursquare 50 100 200 500 1000 50 100 200 500 1000 17.44 19.07 20.77 20.75 21.00 18.05 18.28 21.70 22.55 21.87 # anonymous posts 5 35.12 40.17 41.71 41.27 40.15 35.91 38.44 44.81 44.02 41.85 15 49.22 55.51 58.06 58.34 55.62 51.20 55.01 60.90 60.40 60.42 20 52.32 59.82 60.07 61.74 60.34 53.17 57.12 64.50 63.78 63.26 5. INFORMATION VALUATION Let us again start from an OSN user base U, in which each user u is defined by the set of his associated timelines u}K {M k k=1. Further, let S be an OSN such that we can de- fine the set of all posts made by u in S as his timeline, M S u . We treat the timeline as a long consecutive piece of text in which each post constitutes a sentence. We use information from the timeline to estimate the probability of a user man- ifesting a certain property A. This probability is denoted by p(AMu), where A denotes "u shows Property A" and Mu is the user's timeline. Due to our definition of timelines, the same method can be used for full timelines or subsets of posts. Regardless of the chosen scope, we now project the timeline into an n-dimensional TF-IDF weighted vector space that allows us to train a classifier CA, estimating the final p(AMu). 5.1 Measuring informativeness Our objective is to find a function which quantifies the information carried in a post. On the one hand, we are interested in capturing the relevance of a post with respect to a certain inference task, on the other hand, in order to avoid redundancy or attributing a high score to already seen information, we are interested in capturing the novelty of some content with respect to what is already known. In a spirit similar to [5], we model the information content in two ways: • Relevance ρ of the post with respect to an inference task or a set of tasks; • Novelty ν of the post with respect to the user's previ- ously posted content. We form our informativeness score as a convex combina- tion between these two quantities, thus introducing a mix- ture parameter λ ∈ [0, 1]. Now, for each newly authored post m, we can define an informativeness function I : Rn × Rn × C → R+ as follows: I(m, Mu,C) = λ ν(m, Mu) + (1 − λ) ρ(m,C) (4) 5.1.1 Relevance Measuring the relevance of shared content can be intu- itively thought of as determining which piece of shared con- tent contains features that are important for the classifier's Figure 3: F-1 score curve of classifiers for conditions 5-6 varying sample size and training data size Table 1: User de-anonymization for different sizes of training and test data, using Twitter as a source for anonymous posts Train Source Posts Seen 1 Twitter Twitter Instagram Foursquare 50 100 200 500 1000 50 100 200 500 1000 18.21 19.19 18.45 19.59 20.24 17.56 18.38 18.78 18.16 19.90 # anonymous posts 5 47.80 48.85 48.68 49.32 49.83 46.17 47.88 48.35 50.84 49.39 15 69.90 71.41 71.57 73.23 73.34 69.82 71.24 72.23 72.97 72.81 20 74.22 76.96 77.03 77.96 78.23 73.24 75.82 77.69 77.87 76.66 Table 2: User de-anonymization for different sizes of training and test data, using Foursquare data as a source for anonymous posts Train Source Posts Seen 1 Twitter Twitter Instagram 50 100 200 500 1000 50 100 200 500 1000 32.43 33.17 34.51 35.81 36.91 32.68 33.16 36.56 37.58 35.00 # anonymous posts 5 54.97 58.35 60.85 62.41 61.58 55.72 58.43 60.34 62.24 62.57 15 61.33 63.24 67.40 67.50 67.88 61.87 64.14 66.84 69.17 66.92 20 62.34 63.10 67.89 68.94 67.06 62.80 66.01 69.02 69.04 67.95 user profiles across OSNs based only on their textual data, achieving a maximum accuracy of 77.96% when using only 500 posts (the equivalent of 20% of the average length of the Twitter timelines at our disposition). Furthermore, the usage of multiple OSNs as training data source seems to im- prove the classifiers' performance when the source of anony- mous posts and training data are distinct, suggesting there is a consistency in user language and vocabulary across the chosen OSNs. We also observe that, in general, the more anonymous posts are available, the better the performance of the designed classifier becomes. (cid:88) (cid:88) decision. A popular choice of such a function describing fea- ture importance is the Gini Importance (Ig). For a feature θ, the Gini Importance for a classifier C is defined as: IgC(θ) = ∆iθ(τ, T ). (5) T τ Where τ is a node, T a decision tree and ∆i(τ ) the de- crease in Gini Impurity. The Gini Importance indicates how often a particular feature θ was selected for a split, and how large its overall discriminative value was for a particular clas- sification problem. We estimate the overall relevance ρ of a post by summing up the importance scores across features contained in the post (cid:126)m ∈ Rn: ρ( (cid:126)m,C) = n(cid:88) Igc(mi) i=0 5.1.2 Novelty For a fixed u ∈ U and OSN S, let (cid:126)m1, (cid:126)m2 ∈ Rn be the vector representation of shared contents m1, m2 ∈ M S u , the user's timeline. Informally, the function novelty : Rn × Rn → R+ should have the following properties: 1. (cid:126)m2 should have low novelty if it is contained in (cid:126)m1. 2. (cid:126)m2 should have low novelty if it is similar to (cid:126)m1. 3. (cid:126)m2 should have high novelty if it is distinct from (cid:126)m1. Let (cid:126)m1 be denoted as (m1,1, ..., m1,n). The proposed func- tion to measure novelty is the following: Let ν: Rn × Rn → R+, be a non-symmetric function defined by: n(cid:80) ν( (cid:126)m1, (cid:126)m2) = i=1 exp(−α(mi,1 + m2,i − 1)) 1[m2,i (cid:54)= 0] (6) n(cid:80) i=1 For each word, the novelty function decays with the num- ber of times that the word appears. By regulating α we can control how many times we have to observe a word to not consider it novel anymore. 5.2 Experimental Setup As a concrete example of the general data-driven label prediction problem introduced previously, we turn towards the task of predicting whether a person will visit a particular type of location (e.g., an Italian restaurant or a golf course) based on their social network timeline(s). These timelines are projected into a TF-IDF-weighted vector space. The vocabulary is curated by: removing all links and user men- tions, stop words, words which occur less than 5 times and, when possible, word lemmatisation using WordNet. For the prediction task, we use the AdaBoost algorithm [3] with de- cision trees as weak learners as our classifier since they gener- ally work well without refined parameter tuning. The classi- fier's performance is evaluated under 10-fold cross-validation. We begin by training one binary classifier per venue type that decides whether or not a user's timeline suggests they are likely to visit that type of location. For every test user u, we initiate the procedure by randomly sampling a single post from their timeline Sk and create a truncated timeline. Then, at each iteration, we sample a constant number d of Figure 4: Percentage of users visiting different venue types posts from the timeline, add them to the truncated timeline and make a prediction using this iteratively updated input vector. The procedure is iterated until user u has no more posts left (or until the truncated timeline reaches a fixed amount of posts). We obtain an ordered sequence of predic- tions: (y0, y1, ..., ynend ), where nend represents the number of iterations. When d = 1, i.e., we add one post at a time, we simulate the situation in which an existing user profile is updated over time as new content is being shared. We aggregate results taking into account the varying time- line length across users in the following way: the maximum timeline length nlmax is calculated. Then, for each user whose timeline is shorter than nlmax , the last prediction ynend is repeated to generate a sequence of predictions of length nlmax . As a performance baseline, we randomly sam- ple posts to be added. In the active method, instead, we select the posts which are most informative according to our metric presented in the previous section. Foursquare offers a hierarchical taxonomy of places that display very different relative popularity. Figure 4 plots the percentage of our user population that has visited each of the more than 500 categories. We can note that the distribution is heavily skewed. While virtually all users have, at some point, visited places that fall into broad categories such as [Arts & Entertainment] or [Food ], others are so specific that they remain almost empty (e.g., [South Tyrolean Restau- rants] or [Hunting Supplies]). For the purpose of venue pre- diction, we are forced to make a subselection of categories that are neither so broad that the prediction task would be- come trivial nor so specific that the classifier would not find sufficiently many training examples. For this reason, we fo- cus on those categories that were visited by 25% to 35% of the population, giving us a set of 37 venues (highlighted in yellow in the figure). 5.3 Results 5.3.1 Cross OSN Each user's timeline is initialized with a single post to which we iteratively add additional, randomly sampled, posts to form an updated user model. Figure 5, shows the classifier performance as a function of the number of posts available to the user model. From this overview, we note three recurring slope patterns. Some venue-specific classifiers quickly reach their optimal performance after as few as 750 posts have been observed (top figure), for others, significantly more it- erations are required (center figure), and lastly, for some particular venues, the classification accuracy hardly bene- fits at all from using more posts (bottom figure). We refer to these three situations as quick-to-learn, slow-to-learn and hard-to-learn venues, respectively. Figure 6 gives a complete overview of the relative frequency of mentions of the chosen venue categories and their affiliation to the three slope types. The general tendency seems to be that frequently mentioned venues tend to be quicker to learn than rare ones, while hard-to-learn venues appear to be randomly spread across the observation frequency range. 5.3.2 Active resource selection After having confirmed the intuitive assumption that (within the limits of our three slope types) more data results in more accurate predictions, we now proceed to describing an active selection scenario in which we expand the user model by the most informative posts according to our metric rather than random ones. To this end, we fix the novelty parameter α at 0.5, meaning that after a word appears 5 times, its nov- elty becomes negligible. Table 4 highlights this method's performance at different settings of λ. 50 posts are actively selected for this experiment and we note that our selection scheme biased towards informativeness delivers significantly better F1 performance1 than the random selection baseline at all parameter settings. The overall best performance was obtained at a setting of λ = 0.1, where the score mixture is dominated by feature importance while still taking into account novelty. Table 4: Average classifier performance across all 37 venue types, for different λ. λ baseline 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 F1-score Precision Recall 8.99 40.18 41.97 41.98 41.48 41.47 41.18 41.56 40.69 40.19 39.00 12.12 15.36 42.10 44.19 44.16 43.63 43.66 43.23 43.80 43.55 42.78 42.21 18.95 52.78 44.21 46.64 46.57 46.00 46.09 45.50 46.31 46.85 45.72 45.99 43.38 Let us return to our previously introduced categorization of learning curve slope types. Table 5 shows the influence of λ on the performance of the three slope categories. We ob- serve clearly diverging tendencies between quick and slow- to-learn venues. While quick-to-learn venues benefit from low novelty contributions, their more slowly evolving coun- terparts benefit from novelty-biased informativeness scores. Examples can be seen in Figure 7. Again, hard-to-learn venue types do not show any noticeable response to differ- ent choices of λ, as long as the relevance component is not fully turned off. Furthermore, for some particular venues, the classifier at- tained a better performance when using only 50 actively selected posts than when using the full timeline of the users. Some examples can be found in Table 6. Regarding RQ 2, we show again that there is consistency in terms of content shared across OSNs, in particular, we show that is is possible to predict venue type visits based on what is shared on Twitter and Instagram. Furthermore, 1Average F1 score for all classifiers is computed using the average of precision and recall across all classifiers. Figure 5: Venue-dependent classifier learning rates across training iterations. Figure 6: Average percentage of venue type mentions per twitter timeline, when user has visited the venue type with category assignment of quick, slow or hard to learn for venue types. Table 5: Average F1-scores across slope types. λ Quick 47.81 0.0 50.16 0.2 0.4 49.83 49.59 0.6 46.97 0.8 1.0 17.87 Slow 38.51 40.40 40.10 40.18 41.19 9.70 Hard 36.70 38.43 37.23 38.48 36.97 30.48 with respect to RQ 3, we show that using our designed met- ric, we can find the posts which are most relevant to predict venue type visits. In particular, using the active learning framework with our information measure of content in posts as selection criteria, we observe overall quicker learning rates and in some cases, we can use a significantly reduced the number of posts to attain a classifier performance which is comparable to using the full timeline of the user. 6. CONCLUSION In this paper, we studied privacy hazards pertaining to cross-platform social network usage. Individually innocuous posts can lead to leakage of critical information when aggre- gated along or across a user's OSN profiles. We quantify this effect in two experiments: (1) uniquely identifying users in an anonymous pool and (2) predicting user properties man- ifested on one OSN platform based on content from other parallel profiles. In the user de-anonymization task, we note that it is pos- sible to match user profiles across OSNs based only on their textual data, with as little as 10% to 20% of the user's full timeline. Furthermore, the inclusion of multiple OSNs as training data sources has been shown to improve the classi- fiers' performance when the source of anonymous posts and training data are distinct. This suggests that there is a con- sistency in user language and vocabulary across OSNs. In the information valuation task, we propose a general- purpose metric of textual informativeness in order to model the value of shared information items both for service providers (predictive power) as well as the user (potential privacy haz- Figure 7: Example of venue types which benefit from novelty component. Top figure: venue type Brewery. Bottom figure: Resort Table 6: Classifier performance using a truncated timeline versus a full timeline. Sushi Rest. Cocktail Bar Gastropub Brewery Nightclub Full timeline Truncated + Random Selection Truncated + Active Selection 52.86 11.49 61.72 31.48 4.97 55.84 38.53 5.36 54.79 50.88 21.51 66.15 43.40 20.11 54.23 ards). We show experimentally that the metric reflects the relative importance of posts with respect to the inference task being performed. When actively selecting a subset of posts per user, this method was always able to beat a ran- dom selection baseline. While choosing posts according to their relevance seems to lead to better performance in gen- eral, we noted that only for some venues there was a notice- able benefit in including a strong novelty component in the information scoring function. This work focused on showing the privacy hazards that arise from sharing content which seems uninformative or harmless. In the future, we are excited to extend this line of work by a dedicated investigation of information valua- tion scores on the user side (e.g., of an OSN) as it would greatly help people understand their own digital footprint and enable them to recognize moments of critical informa- tion disclosure. Furthermore, part of this work focused on proposing a metric for information valuation with respect to an inference task. We are interested to extend this line of work by an in-vivo study of monetary efficiency of advertis- ers as a consequence of introducing an informativeness-aware resource selection scheme in their real-time bidding (RTB) pipelines. 7. REFERENCES [1] Ashish Agarwal, Kartik Hosanagar, and Michael D Smith. Location, location, location: An analysis of profitability of position in online advertising markets. Journal of marketing research, 48(6):1057–1073, 2011. [2] Adam Berger. Statistical Machine Learning for Information Retrieval. PhD thesis, Pittsburgh, PA, USA, 2001. AAI3168516. [3] Leo Breiman. Arcing classifier (with discussion and a rejoinder by the author). Ann. Statist., 26:801–849, 1998. [4] Claudio Carpineto and Giovanni Romano. Semantic search log k-anonymization with generalized k-cores of query concept graph. In Advances in Information Retrieval, pages 110–121. Springer, 2013. [5] Charles L.A. Clarke, Maheedhar Kolla, Gordon V. Cormack, Olga Vechtomova, Azin Ashkan, Stefan Buttcher, and Ian MacKinnon. Novelty and diversity in information retrieval evaluation. In Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '08, pages 659–666, New York, NY, USA, 2008. ACM. [6] Yves-Alexandre de Montjoye, C´esar A Hidalgo, Michel Verleysen, and Vincent D Blondel. Unique in the crowd: The privacy bounds of human mobility. Scientific reports, 3, 2013. [7] Cynthia Dwork. Differential privacy. In Encyclopedia of Cryptography and Security, pages 338–340. Springer, 2011. [8] Maria Han Veiga and Carsten Eickhoff. A cross-platform collection of social network profiles. In Proceedings of the 39th SIGIR Conference on Research and Development in Information Retrieval. ACM, 2016. [9] Weiyin Hong and James YL Thong. Internet privacy concerns: An integrated conceptualization and four empirical studies. MIS Quarterly, 37(1):275–298, 2013. [10] Ronald Howard et al. Information value theory. Systems Science and Cybernetics, IEEE Transactions on, 2(1):22–26, 1966. [11] Mark J Keith, Samuel C Thompson, Joanne Hale, Paul Benjamin Lowry, and Chapman Greer. Information disclosure on mobile devices: Re-examining privacy calculus with actual user behavior. International journal of human-computer studies, 71(12):1163–1173, 2013. [12] Michal Kosinski, David Stillwell, and Thore Graepel. Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences of the United States of America (PNAS), 110(15):5802–5805, 2013. [13] Ravi Kumar and Andrew Tomkins. A characterization of online browsing behavior. In Proceedings of the 19th International Conference on World Wide Web, WWW '10, pages 561–570. ACM, 2010. [14] David Lazer, Alex Pentland, and et al. Computational social science. Science, 323(5915):721–723, 2009. [15] Wen Li, Carsten Eickhoff, and Arjen P. de Vries. Want a coffee?: predicting users' trails. In William R. Hersh, Jamie Callan, Yoelle Maarek, and Mark Sanderson, editors, SIGIR, pages 1171–1172. ACM, 2012. [16] Naresh K Malhotra, Sung S Kim, and James Agarwal. Internet users' information privacy concerns (iuipc): The construct, the scale, and a causal model. Information Systems Research, 15(4):336–355, 2004. [17] H Jeff Smith, Sandra J Milberg, and Sandra J Burke. Information privacy: measuring individuals' concerns about organizational practices. MIS quarterly, pages 167–196, 1996. [18] Kathy A Stewart and Albert H Segars. An empirical examination of the concern for information privacy instrument. Information Systems Research, 13(1):36–49, 2002. [19] Robert West, Ryen W White, and Eric Horvitz. Here and there: Goals, activities, and predictions about location from geotagged queries. In Proceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval, pages 817–820. ACM, 2013.
1604.08171	1	1604	2016-04-27T18:35:55	Adaptive Influence Maximization in Social Networks: Why Commit when You can Adapt?	[ "cs.SI" ]	Most previous work on influence maximization in social networks is limited to the non-adaptive setting in which the marketer is supposed to select all of the seed users, to give free samples or discounts to, up front. A disadvantage of this setting is that the marketer is forced to select all the seeds based solely on a diffusion model. If some of the selected seeds do not perform well, there is no opportunity to course-correct. A more practical setting is the adaptive setting in which the marketer initially selects a batch of users and observes how well seeding those users leads to a diffusion of product adoptions. Based on this market feedback, she formulates a policy for choosing the remaining seeds. In this paper, we study adaptive offline strategies for two problems: (a) MAXSPREAD -- given a budget on number of seeds and a time horizon, maximize the spread of influence and (b) MINTSS -- given a time horizon and an expected number of target users to be influenced, minimize the number of seeds that will be required. In particular, we present theoretical bounds and empirical results for an adaptive strategy and quantify its practical benefit over the non-adaptive strategy. We evaluate adaptive and non-adaptive policies on three real data sets. We conclude that while benefit of going adaptive for the MAXSPREAD problem is modest, adaptive policies lead to significant savings for the MINTSS problem.	cs.SI	cs	Adaptive Influence Maximization in Social Networks: Why Commit when You can Adapt? Sharan Vaswani University of British Columbia Vancouver, Canada [email protected] 6 1 0 2 r p A 7 2 ] I S . s c [ 1 v 1 7 1 8 0 . 4 0 6 1 : v i X r a ABSTRACT Most previous work on influence maximization in social net- works is limited to the non-adaptive setting in which the marketer is supposed to select all of the seed users, to give free samples or discounts to, up front. A disadvantage of this setting is that the marketer is forced to select all the seeds based solely on a diffusion model. If some of the se- lected seeds do not perform well, there is no opportunity to course-correct. A more practical setting is the adaptive set- ting in which the marketer initially selects a batch of users and observes how well seeding those users leads to a diffusion of product adoptions. Based on this market feedback, she formulates a policy for choosing the remaining seeds. In this paper, we study adaptive offline strategies for two problems: (a) MaxSpread – given a budget on number of seeds and a time horizon, maximize the spread of influence and (b) MinTss – given a time horizon and an expected number of target users to be influenced, minimize the number of seeds that will be required. In particular, we present theoretical bounds and empirical results for an adaptive strategy and quantify its practical benefit over the non-adaptive strategy. We evaluate adaptive and non-adaptive policies on three real data sets. We conclude that while benefit of going adaptive for the MaxSpread problem is modest, adaptive policies lead to significant savings for the MinTss problem. 1. INTRODUCTION Recently, there has been tremendous interest in the study of influence propagation in social and information networks, motivated by applications such as the study of spread of infections and innovations, viral marketing, and feed rank- ing to name a few (e.g., see [15, 8, 20, 21]). A prototypi- cal problem that has received much attention in the litera- ture is influence maximization (MaxSpread): given a di- rected network G, with edge weights denoting probabilities of influence between nodes, find k nodes, such that activat- ing them in the beginning leads to the maximum expected spread, i.e., expected number of activated nodes as according Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00. Laks V.S. Lakshmanan University of British Columbia Vancouver, Canada [email protected] to a given diffusion model. For the viral marketing applica- tion, nodes may model users, activation may correspond to product adoption, and seed users are given free or price dis- counted samples of the product, with the aim of achieving the maximum expected number of product adoptions. Kempe, Kleinberg and Tardos [17] formalized this as a discrete optimization problem and studied several discrete- time diffusion models including independent cascade and linear threshold (details in Section 2). They showed that MaxSpread under these models is NP-hard but the ex- pected spread function satisfies the nice properties of mono- tonicity and submodularity. Exploiting these properties, they showed that a simple greedy algorithm, which repeat- edly adds the seed with the largest marginal gain, i.e., in- crease in expected spread, until the budget k is reached, achieves a (1 − 1/e)-approximation to the optimum. There has been an explosion of research activity around this prob- lem, including development of scalable heuristics, alterna- tive diffusion models, and scalable approximation algorithms (e.g., see [5] [23] [18] [13] [12] [22]). For space limitations, we refer the reader to [3] for a more detailed survey. The majority of the work in influence maximization has confined itself to a non-adaptive setting where, in viral mar- keting terms, the marketer must commit to choosing all the k seeds up front. This means that the choice of every single seed is driven completely by the diffusion model used for capturing the propagation phenomena. In practice, it may happen that the actual spread resulting from the seeds cho- sen may fall short of the expected spread predicted by the diffusion model. Recent work by Goyal et al. [12] shows that most diffusion models tend to over-predict the actual spread. Thus, committing to the choice of all k seeds in one shot can result in a sub-optimal performance in actuality. A more re- alistic setting is one where the marketer chooses a subset of seeds and activates them. She monitors how their activa- tion spreads through the network and observes the actual spread thus far. She can then take into account this market feedback in making subsequent seed selections. We call this setting an adaptive setting, as choices of subsequent seeds are adapted to observations made so far about the actual spread achieved by previous selections. Hence, the adaptive setting introduces a policy π which specifies which node(s) to seed at a given time. It is very intuitive that adaptive seed selection should lead to a higher actual spread com- pared to non-adaptive seed selection, since it benefits from market feedback and tailors seed selections accordingly. Adaptive seed selection raises several major challenges. For instance, in the adaptive setting, in practice, there is a finite time horizon T within which the marketer wishes to conduct her viral marketing campaign. Suppose k is the seed budget of a marketer. The marketer must then con- sider the following questions. How many seeds to select at a given time, that is, what is the batch size? Which nodes should be selected in each intervention ? How long should she wait between seeding successive batches (interventions)? If T is sufficiently long, it seems intuitive that selecting one seed at a time and waiting until the diffusion completes, before choosing the next seed, should lead to the maximum spread. The reason is that we do not commit any more seeds than necessary to start a fresh diffusion and every seed se- lection takes full advantage of market feedback. We refer to the above case as unbounded time horizon. Another natural question is, what if the time horizon T is not long enough to allow many small batches to be chosen and/or diffusions to be observed in full. In this case, which we call bounded time horizon, the marketer has to choose a strategy in which the budget k is spent within the time horizon T and every seed selection benefits from as much feedback as possible. Instead of maximizing the spread, the marketer may have a certain expected spread as the target that she wants to achieve. This target may be derived from the desired sales volume for the product. A natural problem is to find the minimum number of seeds needed to achieve the target. This problem, called minimum targeted seed selection (MINTSS for short), has been studied in the non-adaptive setting [11], where it was shown that the classic greedy algorithm leads to a bi-criteria approximation to the optimal solution. An interesting question is whether an adaptive strategy for seed selection can significantly cut down on the number of seeds needed to reach a given target spread. Adaptive MaxSpread has been studied recently in [14, 10, 14], and adaptive MINTSS has been studied in [7]. While a more detailed comparison with these papers appears in Section 2, here are the key differences with our work. The market feedback model assumed by these papers is that when a node is activated (seeded), a subset of the out-edges from the node become active or "live" while others stay in- active or become "dead". This amounts to saying we get to observe which active nodes succeeded in activating which other nodes. We refer to this as "edge level" feedback. Edge level feedback assumption is unrealistic, since in practice, we may only know which other nodes activated as a result of choosing certain seeds, rather than who succeeded in acti- vating whom. The experiments conducted in these papers (if at all) are on small toy networks with 1000 nodes and they do not clarify the practical benefits of going adaptive for real large networks. All previous studies are confined to the setting of unbounded time horizon, which means the horizon is long enough for the diffusion started by each batch to complete. In practice, the horizon may be bounded and not leave enough time for successive diffusions to complete. The theo- retical results in these papers bound the performance of the greedy adaptive policy compared to the optimal adaptive policy. Notice that the optimal (adaptive) policy cannot be computed in polynomial time. The only practical options for both non-adaptive and adaptive settings are greedy approx- imations (possibly with techniques for scaling up to large datasets). Thus, a real question of practical interest is what do we gain by going adaptive, i.e., what is the gain in pe- formance of the greedy approximation algorithm when it is made adaptive? In contrast, we studyMaxSpread and MINTSS under both unbounded and bounded time horizon and quantify the benefits of going adaptive with reference to the greedy approximation algorithm, as opposed to the optimal algorithm which is not practical. Furthermore, we propose a novel node level feedback model which we use for adaptive seed selection. Node level feedback is in line what is really observable in practice: which users became active as a result of seeding the last batch of users? and make the following contributions. In this paper, we address the aforementioned questions • We define the problems of adaptive influence maxi- mization under bounded or unbounded time horizon and minimum adaptive targeted seed selection by gen- eralizing their non-adaptive counterparts (Section 3). • We propose a novel node level feedback model for cap- turing market feedback for adaptive seed selection and show that as long as the time horizon is unbounded, i.e., long enough to allow diffusions to complete, the spread function for our node level feedback model is adaptive monotone and adaptive submodular (Section 4). • We establish a bound on the spread achieved by a greedy adaptive strategy for seed selection compared to both an optimal adaptive strategy and a greedy non- adaptive strategy. The former shows that the greedy algorithm continues to provide a guaranteed approxi- mation, while the latter formally establishes the bene- fits of practical adaptive strategies over practical non- adaptive ones (Section 4). • We establish a similar bound on the number of seeds required by the greedy adaptive policy compared to the greedy non-adaptive one, in order to meet a given target expected spread and establish the practical ad- vantage of going adaptive (Section 4). • For the unbounded horizon, we scale up the classic greedy adaptive policy by leveraging the recent state- of-the-art non-adaptive randomized algorithm based on reverse reachable sets and adapt it to the adaptive setting to achieve superior performance (Section 5). • We argue that the expected spread function is com- putationally hard to optimize f or the case of bounded time horizon and propose an alternative algorithm based on sequential model-based optimization (Section 5). • We conduct a comprehensive set of experiments on 3 real datasets to measure the performance of our algo- rithms and their advantages over non-adaptive poli- cies. We report our results (Section 6). Related work is discussed in Section 2. We summarize the paper and present directions for future work in Section 7. 2. RELATED WORK Non-adaptive: Two classical models of diffusion that have been extensively studied are Independent Cascade (IC) and the Linear Threshold (LT) [17]. Both are discrete-time models. The expected spread function under both these models is monotone and submodular. A real-valued set function f : 2U → R is monotone if f (S) ≤ f (S(cid:48)),∀S ⊂ S(cid:48) ⊆ U . It is submodular if ∀S ⊂ S(cid:48) ⊂ U and x ∈ U \ S(cid:48), f (S(cid:48)∪{x})−f (S(cid:48)) ≤ f (S∪{x})−f (S), i.e., the marginal gain (increase in the objective function) by adding an element to a set cannot increase as the set grows. While MaxSpread under both IC and LT models is NP-hard, a simple greedy algorithm [19] provides a (1− 1/e)-approximation to the op- timal solution. The greedy algorithm involves successively selecting the node with the highest marginal gain. Since computing the expected spread of a given set (and hence marginal gain) is #P-hard for both IC and LT models [6, 4], Kempe et al. [17] advocated using MCMC simulations to estimate marginal gains. Using MCMC estimation of the marginal gain, the greedy algorithm yields a (1 − 1/e − )- approximation to the optimum, where > 0 is the error in marginal gain estimation. Tang et al. [22] propose a time complexity) randomized greedy near-optimal (w.r.t. (1 − 1/e − )-approximation algorithm for MaxSpread. It uses the concept of random reverse reachable (RR) sets to achieve this. We briefly review their procedure in Section 5. Here, we note that it is currently the state of the art for MaxSpread and has been shown to scale to a billion-edge network [22]. Adaptive: Adaptive influence maximization has been pro- posed previously in [10, 14, 7]. Golovin and Krause [10] extend the definitions of submodularity and monotonicity to the adaptive setting. In the adaptive setting, batches of nodes are seeded at different intervals. When a batch is seeded, an actual diffusion (called realization in [10]) un- folds as per the classical IC model. The next batch is cho- sen based on the previously observed cascade. An objective function is adaptive monotone and adaptive submodular if the marginal gain of every element is non-negative and non- increasing in every possible realization, as the size of the set (alternatively length of the policy) increases. We wish to choose a policy that maximizes such an objective function in the adaptive setting. As before, the greedy policy con- sists of selecting the node with the maximum marginal gain. Golovin and Krause [10] derive average case bounds on the performance of greedy adaptive policies. They also prove bounds on the greedy adaptive policy for adaptive submod- ular functions under matroid constraints [9]. They assume an edge level feedback mechanism with the IC model and show that the expected spread is adaptive monotone and adaptive submodular, guaranteeing an approximation algo- rithm. Guillory et al. [14] study the problem of submodular set cover in the adaptive setting in which the objective is to minimize the total number of sets required to cover a cer- tain target set and prove worst case bounds for the greedy adaptive policy. They briefly describe how their framework can be used for influence maximization in a social network with hidden information (e.g., hidden preferences of users). In this paper, we consider the more traditional influence maximization problem and assume that users do not have any hidden preferences. We establish average case guar- antees similar to [10]. Finally, [7] addresses the adaptive MinTss problem and shows that under certain conditions, the batch-greedy adaptive policy, in which the seeds are cho- sen in batches in a greedy manner, is competitive not only against the sequential greedy policy (choosing one seed at a time) but also against the optimal adaptive policy. As ex- plained in the introduction, the key difference between these papers and our work is that unlike them, we adopt a more realistic node level feedback, establish bounds relating adap- tive greedy policy with the non-adaptive greedy algorithm, thus answering the question, in practice what does one gain by going adaptive. This question is not answered by simply comparing adaptive greedy with the optimal policy which anyway could not be used in a real network owing to its intractability. 3. PROBLEM DEFINITION We consider a directed social network G = (V, E), with V = n and E = m with the edge weights giving the influ- ence probabilities between two users. In this paper, we as- sume the independent cascade (IC) diffusion model. In this model, time proceeds in discrete steps. At time t = 0, the seed nodes are active. Each active user gets one chance to in- fluence/activate her neighbor in the next time step. This ac- tivation attempt succeeds with the corresponding influence probability between the two users. An edge along which an activation attempt succeeded is said to be "live" whereas the other edges are said to be "dead". This leads to 2E possible worlds of the network. One of these possible worlds is the true world, which reflects the reality of which activation at- tempts succeeded/failed and of the whole cascade/diffusion that occurred starting from the batches of seeds chosen at various times. Clearly, we don't know the true world, al- though as described below, in an adaptive setting, it may be revealed partially from successive seed selections. We consider influence maximization in the adaptive set- ting, where we are given a seed budget k and a time horizon T . The state of a network at time t is used to determine the seed(s) which will be selected at that time. The precise definition of state depends on the type of feedback model. Golovin and Krause [10] consider an edge level feedback, where they assume after a batch is seeded, the precise sta- tus (live/dead?) of every edge is observable. We will instead adopt the node level feedback model, whereby we assume the status (active/inactive?) of every node in the network is ob- servable. A network state is a mapping ψt : V → {0, 1}. For convenience, ψt can be treated as the set of active nodes at time t. At t = 0, the nodes seeded at t = 0 are the only active nodes. Let G(cid:96) denote the true world. It is easy to observe which nodes become active: e.g., which users buy a product or share a particular story on their Facebook page. These are precisely the nodes reachable from seeds in G(cid:96) within a given time horizon T . However, we cannot observe which edges of the network G are actually present (i.e., are live) in G(cid:96): e.g., we can't trace which friend of a particular user influenced her to buy the product. Hence node level feedback is a more realistic feedback model. Notice that it makes weaker assumptions about what is observable, than edge level feedback. By a policy, we mean a seed selection strategy, i.e., a mapping π : ψt → 2V , where ψt is the network state at current time t. Thus, a policy specifies which nodes to seed next, given the current network state. It is possible to have π(ψt) = {}. This means the policy chooses to not seed any nodes and just "wait". The policy thus transforms a network from its current state ψt to the new state ψt+1 := ψt∪π(ψt). The network state can change from seeding as well as from the spreading of influence from current active nodes under the diffusion model. We use f (π) to denote the spread of a policy π within a given time horizon T in the true world (the world G(cid:96) which is realized in reality). The seeds selected by an adaptive policy depend on the true world whereas for the non-adaptive case, the selected seeds are independent of the true world. We care about the performance of both strate- gies in the true world. Since we don't know the true world, we quantify the average gain obtained by going adaptive, and do so by generating a number of candidate true worlds, finding the spread f (π) in each of these worlds and taking the average. We denote the average performance of a policy π as favg(π). σ refers to the expected spread without any feedback from the network whereas favg is the average true spread given a feedback model. Note that σ is equal to favg for a non-adaptive policy since it does not use any feedback. We use πk to denote a policy constrained to select k seeds. When the policy π seeds a node, i.e., when π(ψt) (cid:54)= ∅, we refer to it as an intervention. Since the diffusion model we use is discrete, the diffusion process completes in a maximum of D time-steps where D is the length of the longest simple path in the network. If the last intervention occurred at time t and there are no subsequent interventions, then we have ψt+D+1 = ψt+D. Since we have a model of the diffusion process, we can simulate reality (i.e., generate or sample a set of candidate true worlds) and gauge policies by measuring their perfor- mance against the set of candidate true worlds. Thus, pol- icy design is based on maximizing the average performance over the set of candidate worlds. These worlds constitute a training set in standard machine learning. The true world (the test sample in this analogy) will be generated from the same diffusion process, i.e., it will be drawn from the same probability distribution. Since the policy is known before we actually implement it in the real world, these are offline policies, as opposed to online policies, which are determined dynamically depending on the state of the network and the time left. In this paper, we focus on offline policies. We next formalize the two problems studied in this paper. We assume C is a set of candidate true worlds, chosen by some oracle, which will be used for computing favg. Problem 1 (MaxSpread). Given a directed probabilistic network G = (V, E), seed budget k, and time horizon T , find the optimal policy πopt,k = arg max{favg(πk)}. Notice that in the calculation of spread f (and hence favg), the diffusion is restricted to a length of at most T , where T is the given time horizon. When T ≥ kD, we say the time horizon is unbounded. Otherwise, it is bounded. We consider MaxSpread under both bounded and unbounded horizons. MinTss in the non-adaptive case aims to find a seed set of the smallest size such that the resulting expected spread is above a given threshold Q. The best known result states there is a bi-criteria approximation: given β > 0, the greedy β (cid:101)), where algorithm yields a seed set of size ≥ OP T (1+ln(cid:100) Q OP T is the optimal seed set size and the expected spread of the greedy selection is ≥ Q − β. We generalize it to the adaptive setting. Let cavg(π) denote the average number of seeds chosen by policy π across the true worlds C on which it is tested. Problem 2 (MinTss). Given a directed probabilistic net- work G = (V, E), time horizon T , and spread threshold Q, find the policy πopt that leads to the smallest seed set, i.e., πopt = arg min{cavg(π) f (π) ≥ Q}. Notice that policy π may end up using different numbers of seeds in different choices of true worlds. cavg(π) is the average number of seeds chosen by π across those worlds. Finally, notice that we require that the spread achieved by πopt must be ≥ Q in every true world chosen and not on an average across candidate true worlds. This is motivated by practical considerations: if the policy underperforms (i.e., f < Q) in some worlds and the true world is among them, the marketer won't be happy! 4. THEORETICAL RESULTS Recall, we say time horizon is unbounded if H ≥ kD, where k is the seed budget, D is the length of the longest path of G, and H is the time horizon. We consider un- bounded time horizon up to Section 4.2. Bounded time hori- zon is addressed in subsection 4.3. Our first result is that our spread function based on node level feedback is adap- tive monotone and adaptive submodular, thus affording an efficient approximation algorithm. Theorem 1. For unbounded time horizon, if the diffusion process is allowed to complete after every intervention, node level feedback is equivalent to edge level feedback w.r.t. marginal gain computation and therefore the expected spread function is adaptive submodular and adaptive monotone under node level feedback. Proof. We will show that node level feedback is equivalent to edge level feedback from the perspective of marginal gain computation. In [10], the authors show that the expected spread function under edge level feedback is adaptive mono- tone and adaptive submodular. The above theorem will fol- low from this. Specifically, we prove that (a) for every edge level feedback based network state, there is a corresponding state based on node level feedback, which preserves marginal gains of nodes, and (b) vice versa. Given edge level feedback, we clearly know which nodes are active. These are precisely nodes reachable from the seeds via live edge paths in the revealed network. In the rest of the proof, we show that for each node level feedback state, there is a corresponding edge level feedback state that preserves marginal gains. Let S0 be the set of seeds chosen at time t = 0. Given node level feedback, we can infer the corresponding edge level feedback based network state using the following rules. Consider an edge from an active node u to node v. Notice that the status of an edge leaving an inactive node is unknown in either feedback model. Rule 1: If node u is active, v is inactive, and there is an edge from u to v, then infer that edge (u, v) is dead. Rule 2: If nodes u and v are both active and u is the only in-neighbor of v, then conclude that the edge (u, v) is live. Rule 3: If nodes u and v are both active and u has more than one in-neighbor, arbitrarily set the status of the edge (u, v) to be live or dead. We now show that the way edge status is chosen to be live or dead in Rule 3 plays no role in determining the marginal gains of nodes. We make the observation that if the diffusion process is allowed to complete after each intervention, the only extra information about the network that is observed using edge level feedback over node level feedback is the status of edges between 2 active nodes. Given that the node u is active, we need to calculate the marginal gain of every other node in the network for the next intervention. Next we show that the status of edges between 2 active nodes does not matter in the marginal gain computation for any node. For the rest of the argument, we consider the both u and v are active and that v has multiple active in-neighbours, i.e., the case that is addressed by Rule 3. Consider an arbitrary node w the marginal gain of which we need to calculate. There maybe multiple paths from w to a node reachable from w. These paths can be classified into those which in- volve the edge (u, v) and ones which don't. The marginal gain computation involving the latter paths is independent of the status of the edge (u, v). Since the diffusion pro- cess is allowed to complete, all nodes which can be reached (in the "true" possible world) from w through (u, v) have already been activated. Hence paths going through (u, v) do not contribute to the marginal gain for w. Thus, the status of the edge (u, v) does not matter. Since w is any arbitrary node, we can conclude that the marginal gain of every node remains the same under states based on both feedback models. Adaptive monotonicity and submodular- ity are properties of marginal gains. Since marginal gains are preserved between edge level and node level feedback, it follows that these properties carry over to our node level feedback model. 4.1 MaxSpread There are four types of policies – the greedy non-adaptive policy (abbreviated GNA), the greedy adaptive policy (GA), the optimal non-adaptive policy (ONA) and the optimal adaptive policy (OA). We use πGA,k to denote the greedy adaptive policy constrained to select k seeds and σ(πGA,k) to refer to the expected spread for this policy. While previous results bound the performance of greedy (adaptive) policies in relation to optimal adaptive policies, they do not shed light on practically implementable policies under either set- ting. These previous results do not quite answer the question "What do we gain in practice by going adaptive?" since both the optimal non-adaptive or optimal adaptive policies are in- tractable. We establish relations between two key practical (and hence implementable!) kinds of policies – the greedy non-adaptive policy and the greedy adaptive policy – for both MaxSpread and MinTss. These relations quantify the average "adaptivity gain", i.e., the average benefit one can obtain by going adaptive. We first restate Theorem 7 from [7]. This theorem gives a relation between the spreads obtained using a batch greedy adaptive policy which is constrained to select seeds in batches of size b and the optimal adaptive policy. Fact 1. If σ(πGA,lb) is the average spread obtained by using a greedy batch policy with a batchsize b and σ(πOA,mb) is the spread using an optimal sequential policy (the optimal policy if we are able to select one seed per intervention) constrained to selecting a number of seeds divisible by the batchsize b, then σ(πGA,lb) > (1 − e −l αγm )σ(πOA,mb) (1) where α is the multiplicative error in calculating the marginal gains. gamma is a constant and equal to ( e e−1 )2. Proposition 1. Let the horizon be unbounded. Let πGA,nGA be the greedy batch policy that select nGA seeds overall in batches of size bGA, and let πOA,nOA be the optimal adaptive policy that selects nOA seeds overall in batches of size bOA. Then σ(πGA,nGA ) ≥ σ(πOA,nOA ) (2) where α ≥ 1 is the multiplicative error in calculating the marginal gains and γ = ( e (cid:7) (cid:6) nGA αγ(cid:6) nOA (cid:7)(cid:19)(cid:35) 1 − exp e−1 )2 is a constant. (cid:18) bGA bOA (cid:34) − Proof. Fact 1 gives a relation between the spreads obtained by a batch greedy adaptive policy constrained to select lb seeds and the optimal adaptive policy constrained to select mb seeds. Both these policies are constrained to select seeds in batches of size b. The relation is in terms of the number of batches used by the policies. Let l and m be the number of batches for the greedy and optimal policies respectively. We make the following observations. First, the two policies can be constrained to select seeds in different batchsizes, bGA and bOA respectively. Next, the number of seeds selected by the policies need not be divisible by the batchsizes. We can follow a similar proof procedure as Theorem 7 in [7] and replace l by (cid:100) nGA (cid:101) and m by (cid:100) nOA (cid:101). bGA bOA Theorem 2. Let πGN A,k be a greedy non-adaptive policy, πGA,k and πOA,k be the greedy and optimal adaptive policies respectively with batch-sizes equal to one i.e. the adaptive policies are sequential. All policies are constrained to select k seeds. Then we have the following relations: −1/αγ)σ(πOA,k) σ(πGA,k) ≥ (1 − e (3) σ(πGN A,k) ≥ (1 − 1 e − )2σ(πOA,k) (4) Proof. Proposition 1 gives us bounds on the ratio of the spread achieved by batch-greedy adaptive policy and that achieved by the optimal adaptive policy. We set nOA = k and bGA = bOA = 1 and obtain equation 3 of the theorem. Theorem 2 of [1] states that for a submodular monotone function, there exists a non-adaptive policy which obtains (1 − 1/e − ) fraction of the value of the optimal adaptive policy. In our context, this implies that the spread due to an optimal non-adaptive policy constrained to select nON A seeds is within a (1 − e−nON A/nOA − ) factor of the spread of an optimal adaptive policy selecting nOA seeds. More precisely, σ(πON A,nON A ) ≥ (1 − e −nON A/nOA − )σ(πOA,nOA ) (5) The classical result from Nemhauser [19] states that the greedy non-adaptive algorithm obtains a (1−1/e−) fraction of the value of the optimal non-adaptive algorithm, where is the additive error made in the marginal gain computation. Moreover if the greedy non-adaptive policy is constrained to select nGN A seeds and the optimal non-adaptive policy se- lects nON A seeds we have the following: σ(πGN A,nGN A ) ≥ (1 − e −nGN A/nON A − )σ(πON A,nON A ) (6) Combining equations 5 and 6, we obtain the following result σ(πGN A,nGN A ) ≥ (1−e −nGN A/nOA−)(1−e (7) Setting nGN A = nGA = nOA = k, we obtain equation 4 of the theorem. −nGN A/nOA−)σ(πOA,nOA ) Discussion: To clarify what this theorem implies, lets assume that we can estimate the marginal gains perfectly. Let's set = 0 and α = 1. We thus obtain the following re- lations: σ(πGA,k) ≥ (1 − e−1/γ)σ(πOA,k) and σ(πGN A,k) ≥ (1 − 1 e )2σ(πOA,k). These two factors are almost equal (in fact non-adaptive is slightly better) and in the case of per- fect marginal estimation, there is not much gain in going adaptive. This intuition is confirmed by our experiments in section ??. 4.2 MinTss Given that it takes the optimal adaptive policy nOA seeds to achieve a spread of Q, we seek to find the number of seeds that it will take the greedy adaptive and traditional greedy non-adaptive policy to achieve the same spread. Since the non-adaptive policy can be guaranteed to achieve the tar- get spread only in expectation, we allow it to have a small shortfall βON A. In addition, we allow both the greedy poli- cies to have a small shortfall against their optimal variants. We formalize these notions in the following theorem. Theorem 3. Let the target spread to be achieved by the optimal adaptive policy be Q. Let the allowable shortfall for the optimal non-adaptive policy over the optimal adaptive policy be βON A. Let βGA and βGN A be the shortfall for the greedy adaptive and non-adaptive policies over their optimal variants. Let the number of seeds required by the four policies - OA, ONA, GA and GNA be nOA, nON A, nGA and nGN A. Then we have the following relations nGA ≤ nOA(αγ ln(Q/βGA)) (cid:19) (cid:18) Q βON A − Q ln (cid:18) (cid:18) nGN A ≤ nOA ln nGN A ≤ nOA ln Figure 1: Theoretical comparison of adaptive and non-adaptive strategies in the network. In networks where this length is small [?], most diffusions complete within a short time. This is also helped by the fact that in practice, influence probabilities are small. However, if we are given a very short time hori- zon, the diffusion process may not complete. In this case, seed selection is forced to be based on observations of in- complete diffusions. We show that the spread function in this case is no longer adaptive submodular. Theorem 4. The spread function with the IC diffusion model is not adaptive submodular if the diffusion process after each intervention is not allowed to complete. Q − βON A βGN A − (Q − βON A) (9) (cid:19) (cid:18) (cid:19) Q βGA − βGN A − Q Q − βGA + βGN A ln βGN A − (Q − βGA + βGN A) (8) (cid:19) (10) Proof. If in proposition 1, we set bGA = bOA = 1 , σ(πOA,nOA ) = Q and σ(πGA,nGA ) = Q − βGA, after some algebraic manip- ulation we can obtain equation 8 of the theorem. Setting σ(πON A,nON A ) = Q−βON A, σ(πOA,nOA ) = Q in equation 5, we obtain the intermediate relation 11. Q (11) Setting σ(πON A,nON A ) = Q − βON A and σ(πGN A,nGN A ) = Q − βON A − βGN A, we obtain the following relation. Q − βON A nON A ≤ nOA ln( ) nGN A ≤ nON A ln( Q − βON A βGN A − (Q − βON A) ) (12) We constrain the spreads for the greedy adaptive and greedy non-adaptive policies to be the same. Hence, Q−βGA = Q− βON A − βGN A. Hence βON A = βGA − βGN A. By combining equations 11 and 12 and substituting βON A as βGA − βGN A, we obtain equation 10 of the theorem. Discussion: To understand the implications of this the- orem, set α = 1, = 0. Let the βGN A = 2 and βGA = 1, thus allowing for a shortfall of only 2 nodes in the spread. We obtain the following relations: nGA ≤ nOAγ ln(Q/2) and nGN A ≤ nOA ln(Q) ln((Q − 1)/2). Figure 1 shows the growth of these functions with Q. We can see that as Q increases, the ratio nGA grows much slower nOA than nGN A . Hence, for the MinTssproblem, there is clearly nOA an advantage on going adaptive. This is confirmed by our experiments in section ??. 4.3 Bounded Time Horizon In discrete diffusion models (e.g., IC), each time-step rep- resents one hop in the graph, so the time needed for a diffu- sion to complete is bounded by D, the longest simple path Figure 2: Counterexample to show that the spread is not adaptive submodular under incomplete diffusion Proof. We give a counterexample. Consider the network shown in Figure 2 and the true world, where the edge (u, v) is live and (v, w) is dead. Let H = 2, k = 2. Suppose at t = 0, we choose the seed set S = {u}, so the next intervention must be made at time t = 1. Based on the true world, we observe that nodes u and v are active at time t = 1. Hence we infer the edge (u, v) to be live. We do not know the status of edge (v, w). Even though w is reachable in the network G, there is incomplete information in the realization revealed at t = 1 to decide if the node w is active or not, since the observed diffusion is incomplete. Thus, the expected spreads w.r.t. the realization above are as follows: σ(S) = 2+(1−p) and σ(S ∪ {w}) = 3. Let S(cid:48) = {u, v}. Then σ(S(cid:48)) = 2 and σ(S(cid:48) ∪ {w}) = 3. This is because w is one hop away from v ∈ S(cid:48) and the realization tells us that w is not active. Thus, we have σ(S ∪{w})− σ(S) < σ(S(cid:48) ∪{w})− σ(S(cid:48)). This was to be shown. What are our options, given that the spread under bounded time horizon is in general not adaptive submodular? The- orem 24 in [10] shows that if a function is not adaptive submodular, no polynomial algorithm can approximate the optimal expected spread within any reasonable factor. Thus, we may continue to use adaptive greedy policy, but without any guarantees in general. In our experiments ??, we use a novel Sequential Model Based Optimization (SMBO) ap- proach for finding a reasonably good policy when the time horizon is bounded. 5. ALGORITHMS To obtain a greedy adaptive policy, we need to repeatedly select nodes with the maximum marginal gain at every inter- vention. This implies that we need to run the greedy influ- ence maximization algorithm to compute the marginal gain over the entire network multiple times. Fortunately, this can be done efficiently by exploiting the recent work [22] which describes a near-optimal and efficient greedy algorithm – Two-phase Influence Maximization (TIM) for non-adaptive influence maximization. We first review TIM and describe the modifications we made to it for the adaptive case. 5.1 Two phase Influence Maximization Overview of TIM: Given a budget of k seeds, a network with m edges and n nodes and an appropriate diffusion model such as IC, TIM obtains a (1− 1/e− ) fraction of the optimal spread in the non-adaptive case, incurring a near- optimal runtime complexity of O(k +l)(n+m)logn/2. TIM operates by generating a large number of random Reverse Reachable (RR) sets. An RR set is defined for a particular node v and a possible world W of the network. It consists of the set of nodes that can reach the node v in the possi- ble world W . Given enough number (see [22] for an explicit bound) of RR sets, the nodes which cover a large number of RR sets are chosen as the seed nodes: the node u which appears in the maximum number of RR sets is chosen as the first seed. Once a node u is selected as a seed, we re- move all the RR sets containing u and the next seed is the node which covers the maximum of the remaining RR sets and so on until a seed set S with k nodes is chosen. Tang et al. [22] show that this simple strategy is enough to guarantee a (1 − 1/e − )-approximation factor in near optimal time. Adaptive TIM: In a greedy adaptive policy, we need to select seed nodes in every intervention. After each interven- tion, a certain number of nodes are influenced and become active. These already active nodes should not be selected as seeds. To ensure this, we eliminate all RR sets covered by any of these active nodes. If the number of nodes which became active is large, it brings the number of remaining RR sets below the required bound, which in turn can invali- date the theoretical guarantees of TIM, as the marginal gain of seeds selected in the next intervention may not be esti- mated accurately. Hence after each intervention, we need to re-generate the RR sets to effectively select seeds for the next intervention. To avoid this expensive repeated RR set generation, we instead eliminate all active nodes from the original network, by making all the incoming and outgoing edges have a zero probability, and generating the required number of RR sets for the new modified network. This guar- antees that the resulting RR sets do not contain the already active nodes. This is equivalent to running the greedy non- adaptive algorithm multiple times on modified networks and results in retaining preserves the theoretical guarantees of TIM. For the unbounded time horizon, the optimal policy consists of selecting one seed per intervention and letting the diffusion complete. For the IC model, the diffusion can take a maximum of D time steps where D is the lenght of the longest simple path in the network. 5.2 Sequential Model Based Optimization In the case of bounded time horizon (i.e., T < kD), as discussed at the end of Section 4.3, there is no straightfor- ward strategy to find or approximate the optimal policy. The policy depends on the precise values of time horizon T and properties of the network. For MaxSpread the two ex- treme cases are the non-adaptive policy and the completely sequential greedy policy. The non-adaptive policy does not take any feedback into account and is hence suboptimal. For a sequential policy, the inter-intervention time T /k will be less than D. Hence the completely sequential policy will re- sult in incomplete diffusions and from 4 will be suboptimal. A similar reasoning applies for MinTss For both problems, we are either forced to seed more than one node per inter- vention or wait for less than D time-steps between interven- tions, or both. We split the problem of finding the optimal policy into two parts - finding the intervention times and the number of nodes to be seeded at each intervention and which nodes need to be seeded at each intervention. Us- ing the logic in 4, we solve the latter problem by using the adaptive TIM algorithm described above. For the for- mer problem, we resort to a heuristic approach since the expected spread function we need to optimize does not have any nice algebraic properties w.r.t. time. In order to find the best offline policy, we need to calculate favg for each candidate policy. Calculating favg across all the candidate possible worlds is expensive. Thus we need to maximize an expensive function without any nice mathematical proper- ties. Hence we resort to a bayesian optimization technique known as sequential model based optimization (SMBO)[16]. The SMBO approach narrows down on the promising config- urations (in our case, policies) to optimize a certain function. It iterates between fitting models and using them to make choices about which configurations to investigate. We now show the above problems can be encoded for solv- ing these problems using SMBO. Consider MaxSpread We have a maximum of k interventions. Some of these inter- ventions may seed multiple nodes whereas other might not seed any. There are another k− 1 variables corresponding to the inter-intervention times. Since the number of variables is 2k− 1, SMBO techniques will slow down as k increases. It is also non-trivial to add the constraint that the sum of seeds across all interventions will add to k. Since this leads to an unmanageable number of variables for large k, we introduce a parameter p which we refer to as the policy complexity. Essentially, p encodes the number of degrees of freedom a policy can have. For every i < p, we have a variable si which is the number of nodes to be seeded at a particular interven- tion. We have also have a variable ti which encodes waiting time before making the next intervention. For example, if p = 2 and s1 = 2, t1 = 5, s2 = 3, t2 = 7 we initially seed 2 nodes, wait for 5 time-steps, then seed 3 nodes, wait for 7 time-steps before the next intervention. In the next inter- vention, we repeat the above procedure, until we run out of time, i.e., reach T or get too close (within s1 or s2) to the budget of k seeds. In the latter case, the last intervention just consists of using the remaining seeds. We use the same strategy to encode policies for MinTssIn this case, however, we stop if the time reaches T or if ≥ Q nodes become ac- tive. Since we have a manageable number of parameters, we can easily use SMBO techniques to optimize over these parameters. The objective function for the first problem is to maximize the spread. The constraint is covered by the encoding. For the second problem, the objective function is to minimize the seeds to achieve a spread of Q. This can be modelled by introducing penalty parameters λ1 and λ2. The function can be written as, minimize g(x) + λ1(Q − f (x)) + λ2(f (x) − Q) (13) where x is the parameter vector, g(x) is the number of seeds, f (x) is the spread, Q is the target spread. The parameter λ1 penalizes not achieving the target spread whereas λ2 pe- nalizes over-shooting the target spread. λ1 encodes the hard constraint whereas λ2 is used to direct the search. 6. EXPERIMENTS 6.1 Datasets We run all our experiments on 3 real datasets – the author collaboration network NetHEPT (15k nodes and 62k edges), the trust network Epinions (75k nodes and 500k edges) and Flixster. On NetHEPT and Epinions where real influence probabilities are not available, we set the probability of an edge into a node v to 1/indegree(v), following the popular approach [5, 23, 4]. We use the Flixster network under the topic-aware independent cascade model of diffusion [2] for which the authors learned the probabilities using Ex- pectation Maximization. Their processed network has 29k nodes and 10 topics. We choose the topic which results in the maximum number of non-zero edge weights. The result- ing sub-network of Flixster consists of 29k nodes and 200k edges. 6.2 Experimental Setup As mentioned earlier, we consider only the IC model of diffusion. We compare between greedy non-adaptive, greedy sequential adaptive and the batch-greedy adaptive policies. Since the actual true world is not known, we sample each edge in the network according to its influence probability and generate multiple true worlds. Since we are interested in the performance of a policy on an average, we randomly generate generate 100 true worlds and average our results across them. For either problem, the seeds selected by the non-adaptive policy is based on expected case calculations and remain the same irrespective of the true world. Only the performance of the policy is affected by the true possible world. Also note that for MinTss in some true worlds the spread of the non-adaptive policy might be less than the target Q. The shortfall can be modelled by the factor β introduced in Section 4. 6.3 Sequential Model Based Optimization We use Sequential Model-Based Optimization for General Algorithm Configuration (SMAC) [16]. SMAC is the state of the art tool used for automatic algorithm configuration for hard problems including SAT and Integer Linear Pro- gramming. Based on the performance of a configuration on certain kinds of benchmark instances characterized by prob- lem specific properties, SMAC creates a random forest model and uses it to choose promising candidate configurations to be evaluated. SMAC uses the model's predictive distribu- tion to compute its expected positive improvement over the the incumbent (current best solution). This approach au- tomatically trades exploitation for exploration. SMAC can easily handle both numerical and categorical parameters. For our case, we need to optimize an expensive black- box function (as defined in the previous section) over 2p configurations where p is the policy complexity. Because the function is hard to evaluate a simple brute-force grid search over the parameter space is not feasible. SMAC is implicitly able to leverage the structure present in the problem and come up with promising solutions to the problem. The benchmark instances consist of seeds for the random process generating 10 true worlds at a time. Hence, the evaluation of each configuration on each instance involves running the algorithm 10 times. We use a training set of 1000 such instances and a separate test set of 50 instances to evaluate the policies found by SMAC. We restrict the number of function evaluations SMAC can make to 500 and set the tuner timeout (the maximum time that can be spent by SMAC in building the random forest model and deciding which configuration to evaluate next) is set to 100 seconds. 6.4 MaxSpread 6.4.1 Unbounded time horizon For MaxSpread we vary the number of seeds k over {1, 10, 20, 50, 100}. For the unbounded horizon, we compute the spread obtained using the greedy non-adaptive and the greedy adaptive sequential policies. We set = 0.1. Figure 3: NetHEPT: Average Spread vs Number of seeds Figure 4: Flixster: Average Spread vs Number of seeds Figures 3 and 4 show the average spread favg across 100 possible true worlds as the number of seeds is varied in the given range. We quantify the the effect of adaptivity by the ratio favg (πGA) favg (πGN A) , which we call the average adaptivity gain. We see that the average adaptivity gain remains constant as the number of seeds are varied. We obtain similar results even with higher (100 to 500) values of k. This finding is consistent with the observations made in Section 4. For the adaptive greedy sequential strategy in which we select one seed at a time, we generate RR sets for k = 1 and regenerate the RR sets between each pair of interventions. The run-time graphs are shown in Figures 5 and 6. As can be seen, although this method scales linearly with the num- ber of seeds, it is much slower than the non-adaptive case and will prohibitive for larger datasets. Instead we can gen- erate a large number of RR sets upfront and use these sets to Figure 5: NetHEPT: Runtime vs Number of seeds Figure 8: Epinions: Runtime vs Number of seeds Figure 6: Flixster: Runtime vs Number of seeds select seeds for the first few interventions. The RR sets are regenerated as soon as the change in the number of active nodes becomes greater than a certain threshold (the regen- eration threshold θ). The intuition is that if the number of active nodes has not increased much in a few interven- tions, the number of RR sets does not decline significantly and they still represent information about the state of the network well. We call this optimization trick lazy RR(LR) set regeneration to contrast it with the full RR(FR) set re- generation. We observe that because of submodularity, the frequency of RR set (re)generation decreases as the number of seeds (and time) increases. For our experiments, we em- pirically set θ equal to 10. Higher values of θ lead to lower runtimes but to a smaller average adaptivity gain. 6.4.2 Bounded time horizon For the bounded time horizon, the policy will be forced to group sets of seeds together to form a batch. From Fact 1, we know that the average spread for such a policy will be lower and hence the average adaptivity gain will further de- crease. To verify this, we conduct an experiment on the NetHEPT dataset in which we decrease the time horizon T from a large value (corresponding to unbounded time hori- zon) to low values of the order of the length of the longest path in the network. We vary the policy complexity p to be 1 or 2 in this case. We aim to find the best configuration by varying the batch-size in the range 1 to 100 and the inter- intervention time between 1 and the D of the network. Since the difference between the spreads for the non-adaptive pol- icy vs. the greedy adaptive sequential policy is so small, for the bounded time horizon, SMAC is unable to find a unique optimal policy. Different runs of SMAC yield different poli- cies for the same number of seeds, sometimes converging to the non-adaptive policy even for reasonably large time hori- zons! A higher configuration time for SMAC might lead to stable results or alternatively we might need to encode the problem differently. We leave this for future work. 6.5 MinTss 6.5.1 Unbounded time horizon For all 3 datasets, for the unbounded time horizon, we compare the greedy non-adaptive and greedy adaptive poli- cies with different batch sizes in the range {1, 10, 50, 100}. Because a large number of seeds may be required to satu- rate a certain fraction of the network, we use the lazy RR set generation approximation explained above and set to 0.2. Figures 9, 10 show the comparison between the non-adaptive and various adaptive greedy policies for the NetHEPT and Flixster datasets. Epinions shows a similar trend. Figure 7: Epinions: Average Spread vs Number of seeds We use this strategy to find the spread for both NetHEPT and Flixster. As can be seen from the runtime graphs and average spread graphs, this strategy does not decrease the spread much but leads to significant computational savings. After verifying this strategy, we use it to compare the 2 policies on the larger Epinions dataset, where the same trend is observed – see Figures 7 and 8. The average adaptivity gain is small even for the greedy adaptive sequential policy in case of unbounded time horizon. Figure 9: NetHEPT: Number of seeds required vs Target fraction 6.5.2 Bounded time horizon T ShortFall (β) Number of seeds Objective function Policy(s,t) 10 709 200 7290 (100,6) 50 174.98 177.33 1927.1 (28,8) 100 10.54 171.11 276.51 (20,11) 1000 0 168 168 (3,12) Table 1: Policies of p = 1 recovered by SMAC for varying time horizons(T ) for Flixster with Q = 5800 We now consider the important question, how good is the effect of adaptivity for a bounded time horizon for the MinTss problem. For this, we vary the time horizon T from 10 to 1000 and the policy complexity p is set to either 1 or 2. We use the Flixster dataset and fix the target fraction of nodes to 0.2. As in the previous problem, we aim to find the best configuration by varying the batch size in the range 1-100 and the inter-intervention time between 1 and the D of the network. Since each configuration run involves solving MinTss 500 times, to save computation time we use a relatively high = 0.5. We verified that similar results hold for smaller values of . The optimal policy returned by SMAC is evaluated on a different set of instances (possible true worlds) averaging the results over 50 such instances. Table 1 shows the results for this experiment. For both p = 1, 2, as the time horizon increases, the shortfall goes to zero and the objective function is just the number of seeds required. We see that even for a low time horizon, SMAC is able to find a policy for which the number of seeds is close to the policy (which uses 163 seeds) for an unbounded time horizon. It is still much better than the non-adaptive version of the policy which uses a large number of seeds even for unbounded time horizon. As T increases, in the policy found by SMAC, the number of seeds/interventions decreases and inter-intervention time increases. In fact, for T = 1000 the p = 1, the policy found by SMAC seeded 3 nodes per intervention and had a inter-intervention time equal to 12 (which is greater than D of the graph). For ex- tremely small T , the policy found by SMAC had 100 nodes per intervention and a very short inter-intervention time of 3. We observe similar behaviour even for p = 2 and with the NetHEPT dataset as well. Note that as long as T > D, the non-adaptive version will require the same number of seeds it needs for the unbounded horizon case. This shows us the benefit of adaptivity even when the time horizon is severely constrained. These experiments show the effective- ness of SMAC in finding reasonably good policies for any time horizon for MinTss. 7. CONCLUSION We studied adaptive influence maximization in social net- works and focused on the MaxSpread and MinTss prob- lems. We considered both the unbounded and bounded time horizon. For unbounded horizon, we derived theoreti- cal bounds on the performance of the greedy adaptive over greedy non-adaptive policies, thus quantifying the practi- cal benefit of going adaptive. We studied how the adap- tivity gain is affected by batch-size and number of seeds for MaxSpread and by target spread for MinTss. From our experiments on real networks, we conclude that while the benefit of going adaptive is modest for the MaxSpread Figure 10: Flixster: Number of seeds required vs Target fraction As can be seen, the non-adaptive policy is competitive for smaller number of target nodes. But as the target fraction increases, the adaptive policies are better able to exploit the market feedback mechanism and lead to large savings in the number of seeds. This again agrees with our theoreti- cal results which showed that the adaptivity gain increases as the number of target nodes increases. As the size of the network increases, the estimated spread calculation in the non-adaptive case is averaged across greater number of true worlds and hence becomes less efficient. We observed that in many cases, the final true spread for the non-adaptive policy either overshoots the target spread or misses the tar- get spread by a large amount. We conclude that adaptive policies are particularly useful if we want to influence a sig- nificant fraction of the network. We give some intuition for the difference in the adaptivity gains for the two problems. For adaptive policies, the rate of increase in the expected spread is fast in the beginning before the effects of submodularity take over. Hence adap- tive policies require fewer seeds than non-adaptive to reach a comparable target spread. However, once submodularity kicks in, the additional seeds added contribute relatively lit- tle to the spread. Hence for MinTss, where the objective is to reach a target spread with minimum seeds, the adaptivity gain is higher. However for MaxSpread, even though the adaptive policy reaches a high spread with fewer seeds, the remaining seeds in the budget don't add much more to the true spread. Figure 11: Flixster: Runtime vs Target fraction We also plot the runtime graph for the Flixster dataset. The non-adaptive time dominates because it needs to choose a larger number of seeds. Since the batch-greedy policies select batches of seeds and consider feedback less often, they have a lower running time which decreases as the batch-size increases. Figure 11 shows the runtime variation for Flixster. Results on other datasets show a similar trend. 2011. [13] A. Goyal, W. Lu, and L. V. Lakshmanan. Simpath: An efficient algorithm for influence maximization under the linear threshold model. In Data Mining (ICDM), 2011 IEEE 11th International Conference on, pages 211–220. IEEE, 2011. [14] A. Guillory and J. Bilmes. Interactive submodular set cover. In ICML, pages 415–422, 2010. [15] H. W. Hethcote. The mathematics of infectious diseases. SIAM review, 42(4):599–653, 2000. [16] F. Hutter, H. H. Hoos, and K. Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pages 507–523. Springer, 2011. [17] D. Kempe, J. Kleinberg, and ´E. Tardos. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 137–146. ACM, 2003. [18] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance. Cost-effective outbreak detection in networks. In Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 420–429. ACM, 2007. [19] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of approximations for maximizing submodular set functions. Mathematical Programming, 14(1):265–294, 1978. [20] J. J. Samper, P. A. Castillo, L. Araujo, J. Merelo, O. Cordon, and F. Tricas. Nectarss, an intelligent rss feed reader. Journal of Network and Computer Applications, 31(4):793–806, 2008. [21] X. Song, Y. Chi, K. Hino, and B. L. Tseng. Information flow modeling based on diffusion rate for prediction and ranking. In Proceedings of the 16th international conference on World Wide Web, pages 191–200. ACM, 2007. [22] Y. Tang, X. Xiao, and S. Yanchen. Influence maximization: Near-optimal time complexity meets practical efficiency. 2014. [23] C. Wang, W. Chen, and Y. Wang. Scalable influence maximization for independent cascade model in large-scale social networks. Data Mining and Knowledge Discovery, 25(3):545–576, 2012. problem, adaptive policies lead to significant savings (i.e., gain) for the MinTss problem. For bounded time horizon, we argued that finding the optimal policy is hard and used sequential model based optimization (SMBO) techniques to find a good policy for both the problems. Several interesting directions for future work remain. Ex- tending our framework to the LT model and also to con- tinuous time models is interesting. We believe that with continuous time models and the use of queries to infer the state of the network, adaptive influence maximization will bring the theory much closer to the practical needs of a real viral marketer. 8. REFERENCES [1] A. Asadpour, H. Nazerzadeh, and A. Saberi. Maximizing stochastic monotone submodular functions. arXiv preprint arXiv:0908.2788, 2009. [2] N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influence propagation models. Knowledge and information systems, 37(3):555–584, 2013. [3] W. Chen, L. V. Lakshmanan, and C. Castillo. Information and influence propagation in social networks. Synthesis Lectures on Data Management, 5(4):1–177, 2013. [4] W. Chen, C. Wang, and Y. Wang. Scalable influence maximization for prevalent viral marketing in large-scale social networks. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1029–1038. ACM, 2010. [5] W. Chen, Y. Wang, and S. Yang. Efficient influence maximization in social networks. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 199–208. ACM, 2009. [6] W. Chen, Y. Yuan, and L. Zhang. Scalable influence maximization in social networks under the linear ı A Areshold model. In Proc. 2010 IEEE Int. Conf. on Data Mining, pages 88a A¸S–97, 2010. [7] Y. Chen and A. Krause. Near-optimal batch mode active learning and adaptive submodular optimization. In Proceedings of The 30th International Conference on Machine Learning, pages 160–168, 2013. [8] P. Domingos and M. Richardson. Mining the network value of customers. In Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, pages 57–66. ACM, 2001. [9] D. Golovin and A. Krause. Adaptive Submodular Optimization under Matroid Constraints. Computing Research Repository, abs/1101.4, 2011. [10] D. Golovin and A. Krause. Adaptive submodularity: Theory and applications in active learning and stochastic optimization. Journal of Artificial Intelligence Research, 42(1):427–486, 2011. [11] A. Goyal, F. Bonchi, L. Lakshmanan, and S. Venkatasubramanian. On minimizing budget and time in influence propagation over social networks. Social Network Analysis and Mining, 3(2):179–192, 2013. [12] A. Goyal, F. Bonchi, and L. V. Lakshmanan. A data-based approach to social influence maximization. Proceedings of the VLDB Endowment, 5(1):73–84,
1804.07251	1	1804	2018-03-05T15:00:25	The Algorithm of Accumulated Mutual Influence of The Vertices in Semantic Networks	[ "cs.SI", "cs.CY" ]	In this article the algorithm for calculating a mutual influence of the vertices in cognitive maps is introduced. It has been shown, that in the proposed algorithm, there is no problem in comparing with a widely used method - the impulse method, as the proposed algorithm always gives a result regardless of whether impulse process, which corresponds to the weighted directed graph, is a stable or not. Also the result of calculation according to the proposed algorithm does not depend on the initial impulse, and vice versa the initial values of the weights of the vertices influence on the result of calculation. Unlike the impulse method, the proposed algorithm for calculating a mutual influence of the vertices does not violate the scale invariance after increasing the elements of the adjacent matrix, which corresponds to the cognitive map, in the same value. Also in this article the advantages of the proposed algorithm on numerous examples of analysis of cognitive maps are presented.	cs.SI	cs	The Algorithm of Accumulated Mutual Influence of The Vertices in Semantic Networks Oleh O. Dmytrenko, Dmitry V. Lande In this article the algorithm for calculating a mutual influence of the vertices in cognitive maps is introduced. It has been shown, that in the proposed algorithm, there is no problem in comparing with a widely used method – the impulse method, as the proposed algorithm always gives a result regardless of whether impulse process, which corresponds to the weighted directed graph, is a stable or not. Also the result of calculation according to the proposed algorithm does not depend on the initial impulse, and vice versa the initial values of the weights of the vertices influence on the result of calculation. Unlike the impulse method, the proposed algorithm for calculating a mutual influence of the vertices does not violate the scale invariance after increasing the elements of the adjacent matrix, which corresponds to the cognitive map, in the same value. Also in this article the advantages of the proposed algorithm on numerous examples of analysis of cognitive maps are presented. Keywords: Semantic Networks, Cognitive Map, Directed Graph, Influence Matrix, Accumulated Influence Purpose: to overcome disadvantages of the impulse method by means of a proposed in this article new algorithm for calculating mutual influence of the vertices in cognitive maps. 1. Introduction A semantic network is a network that represents semantic relations between concepts. This is often used as a form of knowledge representation. It is a directed or undirected graph consisting of vertices, which represent concepts, and edges, which represent semantic relations between concepts [1]. A semantic network is used when one has knowledge that is best understood as a set of concepts that are related to one another. Most semantic networks are cognitively based. Cognitive mapping is one of a new direction in modern decision theory [2]. A cognitive map is a directed graph, the edges of which (sometimes the vertices) have the weight. Cognitive map, like any graph, is defined by the adjacency matrix – the weights of the edges, which connects to that comprises elements corresponding vertices . Some concepts corresponding to the vertices of a cognitive map and cause-effect (casual) relations between concepts correspond to the edges (relations). 1 Wijw1,2,...,nuuuIntroduction of the weights characterizing the force of influence reveals the main focus of the cognitive approach development for the situations analysis. For well- structured situations with quantitate parameters the weights are used, and the value of influence for various paths are summarized. For fuzzy cognitive mapping analysis is well known approach, which was proposed by B. Kosko [3,4], who also introduced the term FCM (FCM – fuzzy cognitive maps). Depending on a concrete type of tasks that are solved, different modifications of FCM are considered. The basis of FCM analysis methods is the operations of fuzzy mathematics [5]. In Kosko model an influence is calculated by the following way. An through the path indirect influence action (An indirect effect) of vertex on vertex that directs from vertex to vertex is defined as , where path is a set of edges of the path ; is weight of the edge of the , which value is defined by the terms of linguistic variables. is defined as The general influence of vertex on vertex , where the maximum is taken from all ways that follow from vertex , and defines the weakest link of the path As a result, to vertex . defines the strongest influence among the indirect influences . Also for fuzzy cognitive mapping analysis the Delphi Method is used. For example, in the work [6] the Delphi Method is used for determining the relationships between factors, which effect the planning process on the purpose of building strategic information systems. Impulse method One of the methods of cognitive mapping analysis is the impulse method, which was introduced in 1970th [7,8]. According to the impulse method, each vertex takes a . The weight of the edge has a value positive value ( leads to the ) if increasing the weight of the vertex in a discrete moment of time increasing of the weight of the vertex . And vice versa, the weight of the edge has a negative value ( ) if decreasing of the weight of the vertex leads to decreasing of the weight of the vertex . if the vertices and are not related. The value of each vertices changes in a discrete moment of time according to the equation: 2 (1) pIijPij(,)()minpklklEPIw()EPPklw(,)klP(,)ijTij(,)(,)maxijpPijTIijpIP(,)ijTpI()ivt0,1,2,...t0ijwiuju0ijwiuju0ijwiuju()ivt0,1,2,...t1(1)()(,)()niiijjjvtvtwuuptwhere is an amount of the vertices in the graph. The impulse is defined by the following equation: , . At the initial moment : and are defined. Disadvantages of impulse method Despite widespread use of the impulse method, it has some disadvantages. The main and most important disadvantage of the impulse method is a diverging of the calculation result in case when an impulse process, which corresponds to the weighted directed graph, is not stable. If all non-zero characteristic constants of the weighted directed graph defined by the adjacency matrix are different and absolute value does not exceed the one, then the directed graph is impulse stable for all simple impulse processes. Otherwise, the weighted directed graph is unstable within the meaning of impulse for some simple impulse process [7]. This means that the vertex will be found, which gets the initial impulse. The impulse in some (maybe another) vertex will become infinitely large. In other words, in the impulse method the value is not defined when . That is why methods of stabilization exist [9]. Also, in [10] a critical analysis of the main methods of cognitive maps researching has been carried out and a number of drawbacks and contradictions that arise in applying the impulse method are given: 1) Diverging 2) The result of calculation - with , in step, when series in (1) diverge. depends on according to (1) on the initial values to . 3) Initial value the expression for does not at all affect the dependence as item). from (comes into 4) Increasing the values of each elements of matrix only change the value component of vector distribution. In some cases it leads to the numeric series diverging (1). in the same value, does not , but changes their rank of The algorithm for calculating mutual influence of vertices In this article a new approach for cognitive maps investigation – the algorithm for calculating a mutual influence of the vertices in cognitive maps is proposed. The idea of this algorithm is that all vertices of a weighted directed graph are considered in is calculated (where pairs and the value of influence on vertex of vertex ). As a result the influence matrix consisting of elements is obtained. For calculating the value of general influence of vertex on vertex the next steps has been taken: 1) all possible simple ways from the vertex to the vertex are built. The block 3 n()()(1)jiiptvtvt0t0t(0)jp(0)ivW()ivtt()ivtt()ivt(0)jp(0)iv()ivtt()ivtW()vtijziuju,1, 2, ...,ijnZijzijziujuiujudiagram of a recursive algorithm used is presented on the fig.1 At the input the pair of the vertices are given, where is a vertex the impact of which is determined (the initial vertex) and is the vertex under the influence (the final vertex). I. A step in the opposite direction to the edge from the final vertex into the initial vertex is taken. The vertex is added to the path. The last added vertex is the current. II. "Conditional test 1" means that the verification is performed. "Conditional test 1" means the verification of whether a vertex directly reachable from the current vertex (path of length 1) but not included into the path exists. If it is true then "Conditional test 2" is made. Otherwise, the path that is built is a deadlock, and the last added vertex is deleted from the path. III. At the stage "Conditional test 2" a check whether a vertex directly reachable from the current vertex (path of length 1) coinciding with the initial vertex is performed. If it does not coincide then the vertex is added to the path and the next step is performed from this vertex (the vertex becomes the current). The next is step II. Otherwise, if the vertex directly reachable from the current vertex (path of length 1) coincides with the initial vertex then this vertex is added to the path. The found path is displayed on the screen in a reverse order. The next, the previous added vertex is deleted from the path and a move to step II is performed. 2) at every path the influence from the vertex on the vertex is calculated, considering the weights of the edges: the impulse from the vertex the chain from according to the rules a) – d) [11]: to is distributed in a) If the vertex has a positive influence on the vertex , and has a negative . As a on influence on , then the vertex increases the negative influence of result the vertex has a positive influence on . b) If the vertex decreases a negative influence of the vertex on , then the vertex has a positive influence on . c) In this case the has a positive influence on , increasing a positive influence of the vertex on . d) 4 (,)ijuuiujujuiujuiuiuiujuiuiujuikjuuuiukukujuiukujuiujuikjuuuiukujuiujuikjuuuiujukujuikjuuuIn this case, the vertex has a negative influence on the vertex and has a on positive influence on , then the vertex decreases the positive influence of As a result the vertex has a negative influence on . Fig. 1 Flowchart of an algorithm for building all possible simple ways between a pair of vertices of the cognitive map For calculating the partial influence on the final vertex accumulated form the vertex on the simple path, it is necessary to calculate the general influence on the vertex , which is accumulated from all vertices that are involved in the - way (taking into account the rules a)–d) ); then to subtract from the influence on the vertex , which is accumulated from all vertices involved into the -way, 5 iukujuiukujuiujujuiukkijzjuktqkkijzkijzjuktqkwhich does not involve the initial vertex . and are calculated iteratively according to the formulas: where is a sequence of the vertices involved in a -way ( , (2) (3) ); , and (where is a number of vertices that are involved in the -way). The initial conditions: , , where , , ( is a dimension of the cognitive map). The general influence on the vertex accumulated from the vertex is the sum of the partial influences calculated as the delta between the (2) and the (3) on all simple ways from the vertex into the vertex where is a number of all simple ways from the vertex to the vertex . If the vertex is not reachable from the vertex , then it means that . The -normalizing on every step makes opportunities to avoid a large value of the coefficient and adequately display the influence of a previous vertex. The coefficient , is the function of a stochastic distribution for . , where an exponential distribution with the parametric variable that equals to The proposed algorithm has an exponential complexity is a number of vertices of a cognitive map. It means that after the increase of a number of the cognitive map vertices, a number of iterations will increase exponentially, that will lead to the increase of a time complexity. A process of building of all simple ways between two vertices has a great contribution in a computational complexity, because in each of the vertices it is necessary to choice the next vertex not added to path. Also for decreasing a number of iterations (as consequently for decreasing of the main load on computing complexity), the initial step is proposed, which main idea is: a) to build the reachable matrix b) every element for the initial matrix ; of the initial matrix to multiply on the corresponding 6 0iqukijzkijz1()(1)1sign(())(,)kijkkkkijijttztztztwqq1()(1)1sign(())(,)kijkkkkijijrrzrzrzrwqqktqk0iqu1mjqu0,1,...,2tm1,...,2rmmk(0)0kijz(1)0kijzmaxijw0,1,...,in0,1,...,jnnijzjuiuiuju1()skkijijijkzzzsiujujuiu0ijz2()1xxe0x22()nOenAWijwWelement of the matrix . A new obtained sparse matrix possible simple ways between vertex gives an opportunity to find beforehand all (step 1) to define whether the and vertex vertex is reachable from the vertex . Accordingly if the way from vertex to vertex does not exist (in a sparse matrix ), then the element of a matrix is equal to zero ( ), which means that the vertex does not influence the vertex . Advantages and examples of the algorithm work for calculating mutual influence of vertices One of the advantages of the proposed algorithm, unlike the impulse method, is of vertices and for the limitedness of the algorithm result for any finite number any values of the weighted directed graph. For the coefficient , which characterizes the accumulative influence, is limited . As the result of any finite values and of the value is limited , the values and corresponding to (2) and (3) are limited: accordingly and . Then the difference that characterizes the value of influence of vertex on vertex for a -way, is also limited: . As a result the sum is limited for (where is a number of all simple ways from the vertex to the vertex ). Since and are finite, then is limited because and are limited. It means that for any finite number result of vertices as a result of limitedness (because of the is limited), the exists and limited by a number value. For example, for complete graph with a dimension , a number of a simple ways from the vertex to the vertex is: In this case the result of calculating – is limited: 7 . ijaAWiujujuiuiuju0ijwijzZ0ijziuju,ijznijw0x()x0()1xrt1()x01()2x()kijzr()kijzt0()2kijzr0()2kijztkkijijzziujuk22kkijijzz1()skkijijkzz12()2skkijijkszzs0ssiujus1()skkijijkzz1()skkijijkzzsn()x,ijznsiuju11111...(2)!(2)!2!3!(3)!(2)!snennn,ijz,2*(2)!2*(2)!ijenzenThe graph of a simple impulse process defined by the adjacent matrix: (4) and corresponded to the weighted directed graph (fig. 2) is shoved on the fig. 3. Fig. 2 Weighted directed graph All non-zero characteristic constants of the adjacency matrix , which describes the weighted directed graph, are different and equal to ; ; , and their absolute value do not exceed the one, therefore the directed graph defined by (4) is impulse stable for all simple impulse processes. The directed graph defined by (4) is impulse stable for all simple impulse processes, therefore among all characteristic constants the value that exceeds the one does not exist. For gets an unit impulse, on the 61th iterative step the impulse example, if the vertex method converge completely (fig. 3). 8 00,3910,12100001=00011000WW0,80,4(13)i0,4(13)i1uFig. 3 Graph of a simple impulse process (the influence of the vertex , when gets The general influence of each vertex and their ranking for the matrix (4) are a unit impulse) presented in the tab. 1. Table 1 Vertex (No) 3 2 4 1 4,899 4,346 3,098 2,098 The influence matrix obtained as a result of the algorithm use for calculating a mutual influence of the vertices is (5) The general influence of each vertices for the influence matrix is defined according to the rule: , 9 (6) 1u1usmInfpInf00,3910,1210,7570,86500,0131=0,8650,2450110,3380,1050ZamInfZ1niamijjInfzwhere is a number of vertices. The general influence of each vertex and their ranking for the matrix (5), according to (6), are presented in the tab. 2. Table 2 Vertex (No) 3 2 4 1 2,11 1,878 1,443 1,269 Tab. 1 and tab. 2 are demonstrating that the ranking of vertices for the impulse method and for the algorithm for calculating a mutual influence of the vertices does not change. Fig. 4 Weighted directed graph for unstable impulse process All non-zero constant characteristics of the adjacency matrix: , (7) which describes the impulse process (fig .4), are not different and exceed the one ( ). ; ; Then the directed graph defined by (7) is unstable within the meaning of impulse for some simple impulse process. In another words, there is the vertex which obtains a unit impulse, that in some (maybe another) vertex, the impulse becomes an indefinitely large. In this case the impulse method is diverged and the value is not defined. As a result, the algorithm for calculating a mutual influence of the vertices in cognitive maps is used. The influence matrix for the matrix (7) is 10 namInfamInf01100001=00011000W1,261,261,26v (8) The general influence of each vertex and their ranking for the matrix (8) are presented in the tab. 3. Table 3 Vertex (No) 1 4 3 2 3,729 2,729 2,492 1,976 Fig. 5 Weighted directed graph The adjacency matrix for the weighted directed graph (fig. 5) looks as (9) All non-zero constant characteristics of the adjacency matrix (9), which describes the weighted directed graph (fig. 5) are equal: ; ; . 11 0111,7290,86500,1111=0,8650,6280110,8650,8650ZamInfamInf00000028=39052010W5,923,483,48Therefore, the directed graph defined by (9) is unstable within the meaning of impulse for some simple impulse process. The influence matrix for the adjacency matrix (9) is (10) The general influence are presented in the tab. 4. Table 4 of each vertex and their ranking for the matrix (10) Vertex (No) 3 2 4 1 22,31 12,17 5,39 0 In the work [12] the weighted directed graph is considered (fig. 6). This graph is built for the analysis of the problem of solid waste removal from cities (the example is taken from "Maruyama M. The Second Cybernetics: Deviation-Amplifying Mutual Causal Processed. – Amer. Scientist. – 51. – 1963. – P. 164-179"). Fig. 6 Weighted directed graph of the analysis of the problem of solid waste removal from cities: 1 – Amount of garbage per unit area; 2 – Number of inhabitants in the city; 3 – Improving living conditions in the city; 4 – Migration to the city; 5 – Number of treatment facilities; 6 – Number of diseases; 7 – Bacteriological contamination per unit area For the weighted graph presented on the fig. 6, the adjacency matrix is 12 00001,20701,1699,794=1,3939011,9172,5981,7910ZamInfamInf (11) The directed graph defined by the matrix (11) is unstable within the meaning of impulse for some simple impulse process because absolute values of all non-zero characteristic constants of the adjacency matrix (11) are equal: ; ; ; ; ; ; . The influence matrix for the matrix (11) is (12) The general influence are presented in the tab. 5. Table 5 of each vertex and their ranking for the matrix (12) Vertex (No) 5 3 6 2 4 7 1 5,841 5,274 4,552 4,487 2,947 2,468 1,999 On the fig. 7 an example of the weighted graph of a cognitive model of electricity consumption is considered (the example is taken from "Roberts F.S. Signed 13 000000110100000001100=0100001000001101000000000010W1,191,091,090,840,841,1416e4,7916e00,1110,0050,0090,0090,86511010,8650,8650,0050,7530,341,1660110,9750,791=0,86510,86500,1110,00020,1070,6881,4920,6880,65401,8650,4550,86510,8650,6280,62800,5670,1110,8640,1110,19Z0,1910amInfamInfDiagraphs and the Growing Demand for Energy, Environment and Planning. – 1971. – P. 395-410") For the weighted graph presented on the fig. 7, the adjacency matrix is (13) Fig. 7 Example of the weighted graph of a cognitive model of electricity consumption: 1 – Power capacity; 2 – Cost of electricity; 3 – Power consumption; 4 – State of the environment; 5 – The population; 6 – Number of workplaces; 7 – Number of enterprises The directed graph defined by matrix (13) is unstable within the meaning of impulse for some simple impulse process because absolute values of all non-zero ; characteristic constants of the adjacency matrix (13) are equal: ; ; ; ; . The influence matrix for the matrix (13) is 14 010000100100001101000=0000100001000000001000010010W1,4291,1251,1250,7430,7439,97 (14) The general influence are presented in the tab. 6. Table 6 of each vertex and their ranking for the matrix (14) Vertex (No) 1 3 7 2 5 6 4 5,888 5,7 5,585 4,042 3,821 2,208 2,097 Also the advantage of the algorithm for calculating a mutual influence of the vertices in cognitive maps is that after increasing the values of each elements of matrix in the same value the vertices ranking does not change according to the degree of their influence. It means that the proposed algorithm does not violate the scale invariance after increasing of elements of the matrix . The general in influence increasing every vertices in expressions (3) and (4) as a multiplier. . As the result of -normalizing, for each of the vertices is scaled up proportionally in after enters into the For example, in the work [12] the cognitive map of the interplay of a model of sanitary condition (fig. 8) is considered. 15 01,1121,7340,7390,4390,86510,865010,8650,6250,060,62811,865010,860,1110,865=0,1110,1160,865010,00020,0050,8650,97610,86500,0050,1110,1110,1160,8650,111100,0050,8690,9811,110,8690,7531Z0amInfamInfWWamInfFig. 8 The cognitive map of mutual influence of a sanitary condition model: 1 – Population of the city; 2 – Migration to the city; 3 – Modernization; 4 – Garbage dump; 5 – Sanitary condition; 6 – Morbidity; 7 – Bacteria The adjacency matrix of the weighted directed graph presented on fig. 8 is (15) The weighted directed graph defined by the adjacency matrix (15) is impulse stable for all simple impulse processes. All non-zero characteristic constants of the . weighted directed graph are equal: of each vertex and According to the impulse method the general influence ; ; ; ; ; their ranking for the matrix (15) are presented in the tab. 7. Table 7 Vertex (No) 3 5 1 4 7 6 2 5,44 4,06 3,38 2,21 1,79 1,27 0,4 16 000,60,90000,100000000,7000,900=0000000,9000000,90,90,3000000000000,80W0,6860,6860,6340,6250,6259,2217epInfpInfThe influence matrix (15) is (16) The general influence are presented in the tab. 8. of each vertex and their ranking for the matrix (16) A vertices ranking (tab. 8) of the algorithm for calculating a mutual influence of the vertices implementation is similar by implication to a ranking as a result the impulse method use (tab. 7). The discrepancy of the ranking is a result of the algorithm for calculating a mutual influence of the vertices specification. Table 8 Vertex (No) 5 3 1 4 7 6 2 3,816 3,599 3,371 1,683 1,65 1,598 0,493 As a result of every element of an adjacency matrix , in case of the impulse method implementation, a scale invariance violation is observed, on the other hand, the algorithm for calculating a mutual influence of the vertices saves a vertices ranking in comparison to the ranking of the vertices of the matrix increase by . (17) 17 00,5150,60,90,6620,0080,6840,100,120,1790,0550,0020,03680,1640,700,2280,90,8670,74=0,040,0160,01500,0210,9260,90,4450,2320,330,49501,590,7220,30,1690,2920,4380,21700,1830,2490,1190Z,1310,1970,150,80amInfamInfWW2001,21,80000,200000001,4001,800=0000001,8000001,81,80,6000000000001,60WFor example, for the adjacency matrix (17) obtained as result of every elements of the adjacency matrix doubling (15), the vertices ranking in comparison to the vertices of matrix ranking (15) remains the same (tab. 9), on the other hand in this case the impulse method is not applicable because the directed graph defined by the adjacency matrix (17) is unstable within the meaning of impulse for some simple impulse ; process (an absolute values of all non-zero characteristic constants are equal: ; ; ; . ; Vertex (№) for for 5 3 1 4 7 6 2 3,816 3,599 3,371 1,683 1,65 1,598 0,493 7,632 7,198 6,742 3,336 3,299 3,197 0,986 The results of the algorithm for calculating a mutual influence of the vertices for each of the vertices is implementation also show that the general influence scaled up proportionally in after increasing every vertex in . For example, for the adjacency matrix (15) after increasing it in 0,01; 0,1; 10 and is scaled up proportionally in ) the general influence та 100 ( , , 0,01; 0,1; 10 and 100 (tab. 10). Table 10 Vertex (№) 5 3 1 4 7 6 2 for for for 3,816 3,599 3,371 1,683 1,65 1,598 0,493 0,038 0,036 0,034 0,017 0,0165 0,016 0,005 0,382 0,36 0,337 0,168 0,165 0,16 0,049 for 38,161 35,99 33,71 16,831 16,496 15,982 4,928 for 381,612 359,903 337,1 168,308 164,96 159,827 49,276 Conclusion Proposed algorithm for calculating mutual influence of the vertices in cognitive maps permits to overcome the disadvantages of the impulse method. For example, for any finite number of vertices and for any values of the adjacency matrix describing the weighted directed graph the algorithm for calculating a mutual influence of the vertices gives a limited result , unlike the impulse 18 1,371,371,271,251,251,8416eamInfWamInf2WamInf0,01W0,1W10W100WamInfamInfWamInf0,01WamInf0,1WamInf10WamInf100W,ijzmethod. In the algorithm the initial values and influence the dependency of the values and (then and ). and according to the (2) and (3) do not depend on the initial impulse. Also after the values increase of each elements of matrix the ranking of a vertices does not in for change according to the degree of their influence and the general influence each of the vertices is scaled up in proportionally. The algorithm for calculating a mutual influence of the vertices proposed in this paper has one disadvantage connected with a search of all simple ways between all pair of vertices of the cognitive map. The search algorithm has an exponential complexity . However, sparse matrices implementation permits to employ the algorithm for calculating a mutual influence of the vertices for calculating the cognitive maps of large dimension analysis. References 1. Sowa, J.F. Semantic Networks, in Encyclopedia of Artificial Intelligence, S.C. Shapiro (Ed.) NewYork: Wiley and Sons (1987). 2. Axelrod R. The Structure of Decision: Cognitive Maps of Political Elites. – Princeton: Princeton University Press, 1976. – 404 p. 3. Kosko B. Fuzzy Cognitive Maps // International Journal of Man-Machine Studies, 1986. – №24. – pp. 65–75. 4. Kosko B., Fuzzy Thinking. – Hyperion, 1993. 5. Carlsson C., Fuller R. Fuzzy reasoning in decision making and optimization. – Physica, 2012. 6. Nalchigar, Soroosh; Nasserzadeh, S. M. R.; Akhgar, Babak. Simulating strategic information systems planning process using fuzzy cognitive map. International Journal of Business Information Systems, 2011, 8.3: 286-306. 7. Roberts F S. Discrete Mathematical Models with Applications to Social, Biological, and Environmental Problems / Fred Roberts. – New Jersey: Rutgers University, Prentice-Hall Inc., 1976. 8. Згуровский М.З., Романенко В.Д., Милявский Ю.Л. Принципы и методы управления импульсными процесса-ми в когнитивных картах сложных систем. Часть 1 // Проблемы управления и інформатики, 2016. – №2. – С. 21–29 (in Russian). 9. Zgurovsky M.Z., Romanenko V.D., Milyavsky Y.L. Advances in Dynamical Systems and Control. Springer International Publishing. – Springer International Publishing, 2016. – pp. 363–374. 10. Snarskii A.A., Zorinets D.I., Lande D.V., Levchenko A.V. K-method of cognitive mapping analysis // arXiv Preprint, 2016. – arXiv:1605.08243 19 (0)kijz(0)kijzt()kijzt()kijzt(0)0kijz(1)0kijz()kijzt()kijztWsmInf2()nOe11. Dmytrenko O.O., Lande D.V. The Method of accumulated influence for cognitive maps analysis // Winter InfoCom 2017: Materials of the V International Scientific and Practical Conference on Information Systems and Technologies, Kiev, 2017. – pp. 47–50 (in Ukrainian). 12. Таран Т.А., Зубов. Д.А. Искусственный интеллект. Теория и приложения. – Луганск: Восточноукраинский национальный ун-т им. Владимира Даля, 2006. – 242 с. (in Russian). 20
1602.08033	2	1602	2016-08-24T21:30:33	Modeling Precursors for Event Forecasting via Nested Multi-Instance Learning	[ "cs.SI" ]	Forecasting events like civil unrest movements, disease outbreaks, financial market movements and government elections from open source indicators such as news feeds and social media streams is an important and challenging problem. From the perspective of human analysts and policy makers, forecasting algorithms need to provide supporting evidence and identify the causes related to the event of interest. We develop a novel multiple instance learning based approach that jointly tackles the problem of identifying evidence-based precursors and forecasts events into the future. Specifically, given a collection of streaming news articles from multiple sources we develop a nested multiple instance learning approach to forecast significant societal events across three countries in Latin America. Our algorithm is able to identify news articles considered as precursors for a protest. Our empirical evaluation shows the strengths of our proposed approaches in filtering candidate precursors, forecasting the occurrence of events with a lead time and predicting the characteristics of different events in comparison to several other formulations. We demonstrate through case studies the effectiveness of our proposed model in filtering the candidate precursors for inspection by a human analyst.	cs.SI	cs	Modeling Precursors for Event Forecasting via Nested Multi-Instance Learning Yue Ning1, Sathappan Muthiah1, Huzefa Rangwala2, Naren Ramakrishnan1 1Discovery Analytics Center, Virginia Tech, Arlington, VA 22203 2Department of Computer Science, George Mason University, Fairfax, VA 22030 Abstract Forecasting events like civil unrest movements, disease outbreaks, financial market movements and government elections from open source indicators such as news feeds and social media streams is an important and challenging problem. From the perspective of human analysts and policy makers, forecasting algorithms need to provide supporting evidence and identify the causes related to the event of interest. We develop a novel multiple instance learning based approach that jointly tackles the problem of identifying evidence-based precursors and forecasts events into the future. Specifically, given a collection of streaming news articles from multiple sources we develop a nested multiple instance learning approach to forecast significant societal events across three countries in Latin America. Our algorithm is able to identify news articles considered as precursors for a protest. Our empirical evaluation shows the strengths of our proposed approaches in filtering candidate precursors, forecasting the occurrence of events with a lead time and predicting the characteristics of different events in comparison to several other formulations. We demonstrate through case studies the effectiveness of our proposed model in filtering the candidate precursors for inspection by a human analyst. 1 Introduction Forecasting societal uprisings, civil unrest movements, and terror threats is an important and challenging problem. Open source data sources (e.g., social media and news feeds) have been known to serve as surrogates in forecasting a broad class of events, e.g., disease outbreaks [1], election outcomes [18, 22], stock market movements [6] and protests [19]. While many of these works focus on predictive performance, there is a critical need to develop methods that also yield insight by identifying precursors to events of interest. This paper focuses on the problem of identifying precursors (evidence) for forecasting significant so- cietal events. Modeling and identifying the precursors for a given protest is useful information for the human analyst and policy makers as it discerns the underlying reasons behind a civil unrest movement. Specifically, the objective of this study is to forecast protest across different cities in three Latin American countries (Argentina, Brazil and Mexico). 6000 news outlets are tracked daily across these countries with the goal of forecasting the occurrence of a protest with atleast one day of lead time. From the news feeds, we also aim to identify the specific news articles that can be considered as precursors for the targer event. We formulate the precursor identification and forecasting problem within a novel multiple instance learning algorithm (MIL). Multiple instance learning algorithms [3, 28] are a class of supervised learning techniques that have labels for a group of instances, but not for individual instances. We make a similar mapping for the group of collected news articles to have no individual class label associated with every news article but the group of news articles are attached with a label indicating the occurrence of a protest. We extend the standard MIL algorithm by introducing a nested structure, where we group news articles published in a given day at the first level and then group the collection of individual days at the second level. This nested MIL approach allows for modeling the sequential constraints between the news articles 1 (grouped by days) published on different days and also provides a probabilistic estimate for every news article and the collection of news article. This estimate is signficant because it indicates for a given news article the probability of it signaling a protest event. Note, in our datasets we do not have any training labels to indicate the protest indicator per news article. Figure 1 shows an example of precursors detected by our model. On the right of the timeline, is a news report about a protest event in Argentina. The connected dots denotes the generated probability of this day being positive. From this example, we find that within 10 days before the event, there are multiple similar reported events that were selected as highly probable leading indicators or evidence. Most of planned societal events are a consequence of several factors that affect the different entities within communities and their relationships with each other (or the government) over time. In this specific example, the leading precursor some days before the protest was an article commenting on standards of living in Argentina and rising poverty levels. The International Court of Justice also provided a verdict on the debt crisis. All these factors led to the final protest where general population demanding work opportunities. Figure 1: Precursors story line for a protest event in Argentina. The x-axis is the timeline. The dots above with numbers are the probabilities for each day that the model generated for the target event. Each precursor document is titled in the timeline. The main contributions of this study are summarized as follows: 1. A novel nested framework of multi-instance learning for event forecasting and precursor mining. We formulate event forecasting and precursor mining for multiple cities in a country as a multi-instance learning problem with a nested structure. By estimating a prediction score for each instance in the history data, we automatically detect significant precursors for different events. 2. Harness temporal constraints in multi-instance learning. We explore different penalty func- tion and regularizations where we employ the temporal information in our dataset under assumption that most events of interest are follow-up reports of other events that happened before, and most planned events are developing over time. 3. Modeling for various event categories in multiple geo-locations. We extend the nested MIL formulation for general purpose multi-class classification to determine necessary attributes of events in terms of their underlying population. 4. Application and evaluation with comprehensive experiments. We evaluate the proposed methods using news data collected from July 2012 to December 2014 in three countries of Latin America: Argentina, Mexico, and Brazil. For comparison, we implement other multi-instance al- gorithms, and validate the effectiveness and efficiency of the proposed approach. We also perform qualitative and quantitative analysis on the precursors inferred by our model. 2 0.5680.4480.6060.6210.6380.6460.7722014-08-07 Protesters demand for labor; Millions living below the poverty line.2014-07-29: The Economic Convocation Argentina organized a demonstration declare entering a technical default(0.57)2014-07-28: Develop agenda for the migrant perspective (0.68)PovertyDebt CrisisGovernment2014-08-01: Argentina debt crisisto International Court of Justice(0.71)2014-07-31: Unresolved debtsin Argentina (0.74)2014-07-30: Standard of Poverty(0.79)2014-08-05: By Lear layoffs, workers blocked highway (0.82)2014-08-03: Barenboim government: Imagination to power (0.77)The rest of this paper is organized as follows. We discuss related work in Section 2. Section 3 introduces the problem setup and our proposed model based on Multi-Instance Learning is presented in Section 4. This section is then followed by experiments and evaluations on real world datasets presented in Section 5 and Section 6 . Finally, we concludes with a summary of the research in Section 7. 2 Related Work Event Detection and Forecasting. Event detection and forecasting from online open source datasets has been an active area of research in the past decade. Both supervised and unsupervised machine learning techniques have been developed to tackle different challenges. Linear regression models use simple features to predict the occurrence time of future events [4, 6, 10, 18]. Advanced techniques use a combination of sophisticated features such as topic related keywords, as input to support vector machines, LASSO and multi-task learning approaches [23, 20]. Ramakrishnan et al. [19] designed a framework (EMBERS) for predicting civil unrest events in different locations by using a wide combination of models with heterogeneous input sources ranging from social media to satellite images. Zhao et al. [27] combine multi-task learning and dynamic features from social networks for spatial-temporal event forecasting. Generative models have also been used in [26] to jointly model the temporal evolution in semantics and geographical burstiness within social media content. Laxman et al. [13] designed a generative model for categorical event prediction in event streams using frequent episodes. However, few existing approaches provide evidence and interpretive analysis as support for event forecasting. Identifying Precursors. Identifying precursors for significant events is an interesting topic and has been used extensively for interpretive narrative generation and in storytelling algorithms [11]. Rong et al. [21] developed a combinational mixed Poisson process (CMPP) model to learn social, external and intrinsic influence in social networks. Multiple Instance Learning. In the multiple instance learning (MIL) paradigm, we are given labels for sets of instances commonly referred as bags or groups. However, individual instance-level labels are unknown or missing. The bag-level labels are assumed to be an association function (e.g., OR, average) of the unknown instance level labels. One approach to MIL adapts support vector machines (SVMs) by: (i) modifying the maximum margin formulation to discriminate between bags rather than individual instances [3], and (ii) developing kernel functions that operate directly on bags [9]. Other multiple instance learning approaches and various applications are found in a detailed survey [2]. Specifically, the generalized MIL [25] formulation assumes the presence of multiple concepts and a bag is classified as positive if there exists instances from every concept. Relevant to our work, besides predicting bag labels, Liu et al. [15] seek to identify the key instances within the positively-labeled bags using nearest neighbor techniques. Recent work [12] has focused on instance-level predictions from group labels (GICF) and allowed for the application of general aggregation functions with applications to detecting sentiments associated with sentences within reviews. The methods proposed in this paper can be viewed as complementary to prior work, casting the forecasting and precursor discovery problems within novel extensions of multiple instance learning. 3 Problem Formulation Given, a collection of streaming media sources (e.g., news feeds, blogs and social network streams), the objective of our study is to develop a machine learning approach to forecast the occurrence of an event of interest in the near future. Specifically, we focus on forecasting "protests" or civil unrest movements in Latin America from a daily collection of published news articles. Besides forecasting the protest, we want to identify the specific news articles from the streaming news outlets that can be considered as supporting evidence for further introspection by an intelligence analyst. We refer to these identified articles as precursors for a specific protest. Figure 2 provides an overview of our proposed approach and problem formulation. We show groups of news articles collected daily, five days prior to the specific protest event (being forecast). Within our proposed MIL-based formulation, each news article is an individual instance, the collection of news 3 Variable S = {S} S = [Xi], i ∈ {1, ..., t} Xi = {xij}, j ∈ {1, ..., ni} xij ∈ (cid:60)V ×1 Y ∈ {−1, +1} P ∈ [0, 1] Pi ∈ [0, 1] pij ∈ [0, 1] C ∈ {1, 2, ..., K} Table 1: Notations Meaning a set of n super bags in our dataset an ordered set of t "bags" in S a set of instances with ni=Xi, number of instance in a bag Xi the j-th instance in set Xi, a V-dimension vector label of super bag estimated probability for a super bag the probability of bag i in super bag to be positive the probability of an instance xij in bag Xi in super bag to be positive multi-class label of super bag articles published on a given day is a bag, and the ordered collection of bags (days) is denoted by super- bag (explained in detail later). For this study, each individual news article is represented by a distributed representation for text derived using deep learning framework and text embeddings [14]. Figure 2 shows that for certain days within the collection we attempt to identify news articles (highlighted) that are considered as precursors from the entire collection of input news articles used for forecasting the occurrence of a specific target. 3.1 Formal Definition and Notations For a given protest event e occurring on day t + k, we assume that for each day before the event we are tracking a multitude of news sources. We represent the collection of ni news articles published on a given day i by Xi = {xi,1 . . . xi,ni}, where the j-th news article is represented by xij. The ordered collection of news articles for the protest event up to day t can be represented as a super-bag, S1:t = {X1, . . . ,Xt}. The occurrence of the protest event at time t + k is denoted by Yt+k ∈ {−1, +1} where 1 denotes a protest and −1, otherwise. The forecasting problem can be formulated as learning a mathematical function f (S1:t) → Yt+k that maps the input, ordered collection of news articles extracted per day to a protest indicator k days in the future from the day t. To identify the news articles considered as precursors (evidence), we want to estimate a probability for each news article on any given day that signifies the occurrence of a given protest. For a news article xij, we denote this estimated probability value by pij. As such, given the collection of news articles we identify the precursor set as the ones with pij greater than a fixed threshold τ . We represent this precursor set of documents as a subset of the original super-bag, given by {xij ∈ S1:t pij > τ}. As a secondary objective, we want to forecast the occurrence of an event with a long lead time i.e., large values of k. Table 1 captures the notations and definitions used in this study. 4 Methods We first provide our intuition behind formulating the precursor discovery and forecasting problem within a novel extension of multiple instance learning algorithm. Parallel to the standard multiple instance learning algorithms we have a group of news articles (bags) with labels available only for the entire bag (i.e., leading to a protest); and one of the objectives is to train a classifier to predict the bag-level label. In addition to predicting the group-level labels, we also care about predicting the labels for individual news articles (instances) since, they signify the precursor. Various MIL formulations extend this approach to estimate the key instances within a bag or provide instance-level labels. However, our problem setting has a two-level grouping structure with sequential constraints i.e., we capture news articles per day (bags) and group the days to form a super-bag with labels only available at the super-bag level. As such, we propose nested multiple instance learning formulations for predicting the super bag level labels (forecast), and then estimate the bag-level and instance-level probabilities for identifying association of the bag and 4 Figure 2: Overview of Proposed Approach for Forecasting and Precursor Discovery. instance with the event, respectively. We developed various extensions of our proposed approach, that tie the different sequential and group constraints. 4.1 Nested MIL model (nMIL ) We model the instance level probability estimates pij for a news article j on day i to associate with a targeted event e with a logistic function. This probability estimates indicate how related the specific instance is to the target event, e . Higher the probability value, the more related the document is to the target event and most probably represents a precursor that contains information about causes of the target event. pij = σ(wT xij) = 1 1 + e−wT xij . (1) Here, w denotes the learned weight vector for our model. The probability for a day (or bag) is then modeled as the average of probability estimates of all instances in a day [12]. Hence, for each bag: Pi = A(Xi, w) = 1 ni pij, (2) where A is an aggregation function. We then model the probability of a super-bag S (associated with an event e ) being positive as the average of the probability of all t bags within the super bag to be positive (related to the target event). Thus: ni(cid:88) j t(cid:88) i P = A(S, w) = 1 t Pi (3) For a given super bag S, as all the t bags within it are temporally ordered, the probability estimates for a given bag (day) is assumed to be similar to its immediate predecessor. This consistency in consecutive bag probabilities is modeled by minimizing the following cross-bag cost as below: g(Xi,Xi−1) = (Pi − Pi−1)2 (4) Finally, given a set of true labels Y for the super bags, we can train our model by minimizing the following cost function w.r.t to w: 5 Eventday t-5day t-4day t-3day t-2day t-1day tP(t-5)Ptdocumentsdeep features…………estimateYtP(t-4)P(t-3)P(t-2)P(t-1)(cid:88) S∈S β n J(w) = t(cid:88) i=1 1 t g(Xi,Xi−1, w) (5) (cid:88) ni(cid:88) S∈S; Xi,Xi−1∈S 1 n t(cid:88) 1 t 1 ni f (S, Y, w) + (cid:88) + 1 n S∈S;Xi∈S xij∈Xi i=1 j=1 h(xij, w) + λR(w) Here, • f (S, Y, w) = −I(Y = 1)logPi − I(Y = −1)(log(1 − P )) is the negative log-likelihood function that penalizes the difference between prediction and the true label for super bag S where I(·) is the indicator function. • g(Xi,Xi−1, w) is the cross-bag cost defined in Equation. 4 • h(xij, w) = max(0, m0 − sgn(pij − p0)wT xij) represents the instance level cost. Here, sgn is the sign function; m0 is a crucial margin parameter used to separate the positive and negative instances from the hyper line in the feature space; p0 is a threshold parameter to determine positiveness of instance. • R(w) is the regularization function. • β, λ are constants that control the trade-offs between the loss function and regularization function. 4.1.1 Cross-bag Similarity (nMIL∆) The cross-bag similarity g(, ) in the above equation does not allow for sudden changes in the day-level probabilities caused due to newer events happening on the current day. We update the cost function across days (bags) (Equation 4) as follows: g(Xi,Xi−1) = ∆(Xi,Xi−1)(Pi − Pi−1)2 (6) The objective function above allows for label information to spread over the manifold in the feature-space. As such, we compute ∆(, ) as the pairwise cosine similarity between the news articles in Xi and Xi−1. Since, we do not have ground truth labels for the bag level (day) we make this consistency assumption that estimated probabilities for consecutive days should be similar if the news articles have similarity in the feature space as well. This model is referred by nMIL∆and allows for sudden changes in how events unfold. 4.2 Sequential Model (nMILΩ) The basic nMIL models assume that there exists a single weight vector across all the days (bags) within a super bag. To model the sequential characteristics of the articles published across consecutive days, we extend this formulation by learning individual weight vectors for each of the historical days. Assuming t days within a super bag S we learn a weight vector for each individual day represented as Ω = [w1, . . . , wt]; where wj is the weight vector learned for day j. In this setting, the individual weight vectors are still learned together in a joint fashion as the Multiple-Task Learning approaches [7]. However, the probability of a news article j on day i will be given by pij = σ(wT i xij). The is formulation is called nMILΩand given by: 6 g(Xi,Xi−1, wi) (7) J(Ω) = β n (cid:124) + (cid:88) S∈S; Xi,Xi−1∈S 1 t t(cid:88) (cid:123)(cid:122) i=1 (cid:88) S∈S 1 n (cid:124) empirical loss f (S, Ω, Y ) (cid:123)(cid:122) (cid:88) 1 n + (cid:124) (cid:125) t(cid:88) (cid:123)(cid:122) i=1 1 ni 1 t ni(cid:88) j=1 unsupervised loss S∈S;Xi∈S xij∈Xi (cid:125) sequential loss h(xij, wi) + λR(Ω) (cid:125) Just like the Multi-Task learning algorithms, the regularization term R((Ω)) can be modified to capture the various relationship-based constraints. However, in this study we ignore these specialized approaches focusing only on the MIL paradigm. 4.3 Multiclass Classification We also extend our developed nMIL formulations to solve general purpose multiclass classification prob- lems rather than binary classification problems. Within our domain, each labeled event is manually attached with two attributes: event type and event population. Event type provides information about the nature of the event. Event population indicates the size/community of people who participated in the protest event. For the multiclass classification problems, we train one-versus-rest classifiers for each of the classes learning a separate weight vector per class. When classifying a super bag to a specific event type/population we first forecast the binary protest indicator label for a super bag. Next, we apply the multi-class classi- fication only on the predicted positive examples. 4.4 Optimization We perform online stochastic gradient decent optimization to solve our cost function and test our model on new data to predict super bag label. For every iteration in our algorithm, we randomly choose a super-bag (S, Y ) from the training dataset S by picking an index r ∈ {1, . . . , n} using a standard uniform distribution. Then we optimize an approximation based on the sampled super-bag by: J(w; S) = βf + 1 t gi + 1 t t(cid:88) i t(cid:88) ni(cid:88) 1 ni i j The gradient of the approximate function is given by: hij + λR(w) (8) (9) pij(1 − pij)xij ∂J(w; S) ∂w t(cid:88) i 1 t 1 ni ∇tJ(w) = − Y − P P (1 − P ) t(cid:88) t(cid:88) t(cid:88) − 1 t 1 t + i i − 1 t 1 ni i ni(cid:88) k = λw ni(cid:88) ni(cid:88) ni−1(cid:88) 1 ni 1 j ni−1 j j 7 sgn(pij − p0)xij(oij) 2(Pi − Pi−1) pij(1 − pij)xij 2(Pi − Pi−1) pvj(1 − pvj)xvj where v = i − 1, oij = I(sgn(pij − p0)wxij < m0). We update the weight vector using a varied learning rate and wt+1 = wt − ηt∇t(w) using mini-batch stochastic gradient descent. 4.5 Precursor discovery using nMIL In the nMIL model, each super-bag consists of an ordered set of bags and each bag represents the documents in one day in the city for which we are forecasting a protest event. We present in Algorithm 1 the steps to identify news articles as precursors based on their estimated probability given by pij > τ . Input: S = {(Sr, Yr)}r∈n+,M Output: {(psr, Yr)}r∈n+ for super bag (Sr, Yr) do Algorithm 1 Precursor Discovery in nMIL 1: procedure PD-nMIL 2: 3: 4: 5: 6: 7: psr = [] for t = 1,2,...,d(history days) do yt = [] for k ∈ Xt do 8: 9: 10: 11: 12: 13: ytk = σ( wxtk) if ytk > τ then yt ← (k, ytk) sort(yt) by ytk in descending order psr ← k where k in top(yt) return {(psr, Yr)}r∈n+ 5 Experiments 5.1 Datasets The experimental evaluation was performed on news documents collected from around 6000 news agencies between July 2012 to December 2014 across three countries in South America including Argentina, Brazil and Mexico. For Argentina and Mexico, the input news articles were mainly in Spanish and for Brazil, the news articles were in Portuguese. The ground truth information about protest events, called gold standard report (GSR) is exclusively provided by MITRE [19]. The GSR is a manually created list of civil unrest events that happened during the period 2012-2014. A labeled GSR event provides information about the geographical location at the city level, date, type and population of a civil unrest news report extracted from the most influential newspaper outlets within the country of interest. These GSR reports are the target events that are used for validation of our forecasting algorithm, and also used for analyzing the identified precursors. We have no ground truth available for verifying the validity of the precursors. Argentina: We collected data for Argentina from newspaper outlets including Clarin and Lanacion from the period of July 2010 to December 2014. There are multiple protest events that happened in Argentina during that time. For instance, people protested against the government and utility/electricity- providing companies because of heatwaves in Dec. 2013. Brazil: For Brazil, we obtained data from news agencies including the three leading news agencies in Brazil; O Globo, Estadao, and Jornal do Brasil from November 2012 to September 2013. During this period Brazil faced several mass public demonstrations occurred across several Brazilian cities stemming from a variety of issues ranging from transportation costs, government corruption and police brutality. These mass protests were initiated due to a local entity advocating for free public transportation. This 8 period had an unusually high social media activity and news coverage and is also known as the "Brazilian Spring" 1. Mexico: For Mexico, we tracked news agencies including the top outlets: Jornada, Reforma, Milenio from January 2013 to December 2014. Over 619 days, we noticed 71 news articles per day on average. There were more than 2000 protest events in this two-year period with major unrest movements in 2013 was led by teachers and students demanding education reform by protesting against the government. 5.2 Experimental Protocol The GSR signifies the occurrence of a protest event on a given day at a specific location. To evaluate the MIL-based forecasting and precursor discovery algorithms, for each protest event we extract all the published news articles for up to 10 days before the occurrence of the specific event. This ordered collection of per-day news documents up to the protest day are considered as positive super bags. For negative samples, we identify consecutive sets of five days within our studied time periods for the different countries when no protest was reported by the GSR. The ordered collection of per-day news documents not leading to a protest are considered as negative super bags for the nMIL approach. For any news article (i.e., an individual instance) within a positive/negative super-bag we have no label (or ground truth). As part of the precursor discovery algorithm, we estimate a probability for an individual instance to signal a protest (by showing evidence). It is important to note that the GSR linked news article for a protest is never used for training purposes. Having identified the positive and negative samples, we split our datasets into training and testing partitions and perform 3-fold cross-validation. We study the performance of forecasting models with varying lead time days and varying historical days. Lead time (l) indicates the days in advance the model makes predictions and historical days (h) is the number of days over which the news articles are extracted as input to the prediction algorithms. As an example, if l is set to 1, then the model forecasts if a protest event is planned for the next day. Setting the historical days, h to 5 denotes that we use news from five days before the current day to make the forecast. We varied l from 1 to 5 and h from 1 to 10 and trained 50 different models for the different approaches to study the characteristics of the developed approaches with varying lead time and historical days. For event forecasting, we evaluate the performance by the standard metrics including precision, recall, accuracy and F1-measure. 5.3 Comparative Approaches We compare the proposed nMIL models to the following approaches: • SVM : We use the standard support vector machine formulation [8] by collapsing the nested grouping structure and assigning the same label for each news article as it's super-bag (for training). During the prediction phase, the SVM yields the final super-bag prediction (forecast) by averaging the predicted label obtained for each of the instances. • MI-SVM [3]: The MI-SVM model extends the notion of a margin from individual patterns to bags. Notice that for a positive bag the margin is defined by the margin of the "most positive" instance, while the margin of a negative bag is defined by the "least negative" instance. In our case, we collapse the news articles from the different historical days into one bag and apply this standard MIL formulation. • Relaxed-MIL (rMILnor ) [24]: Similar to the MI-SVM baseline, we collapse the news articles into one bag. However, unlike the MI-SVM formulation the rMILnor can provide a probabilistic estimate for a given document within a bag to be positive or negative. • Modified Relaxed-MIL (rMILavg ): This approach is similar to the rMILnor , except we compute the probability of a bag being positive by taking average of estimate of each instance in the bag rather than using the Noisy-OR model discussed above. 1http://abcnews.go.com/ABC_Univision/brazilian-spring-explainer/story?id=19472387 9 • GICF [12]: This model optimizes a cost function which parameterized the whole-part relationship between groups and instances and pushes similar items across different groups to have similar labels. 5.4 Feature Description In practice, finding good feature representations to model the news articles is not a trivial problem. Tra- ditionally the bag of words representation allows for easy interpretation but also needs pre-processing and feature selection. Several researchers have developed efficient and effective neural network representations for language recently [5, 16, 17]. Specifically, we learn deep features for documents by taking advantage of the existing doc2vec model. For each document, we generate a 300 dimension vector for training with con- textual window size of 10 in an unsupervised version. We compared the performance of deep features with traditional TF-IDF features but the results showed little difference. Thus, we only report the evaluation of models with deep features. 6 Results and Discussion In this section, we evaluate the performance of the proposed models. Firstly, we evaluate the effectiveness and efficiency of the methods on real data in comparison with baseline methods on multiple configurations of forecasting tasks. Then, we study and analyze the quality of precursors with respect to quantitative and qualitative measures. Multi-class forecasting evaluation is also provided for one of the countries. At last, we perform a sensitivity analysis of performance regarding parameters in the proposed model. Table 2: Event forecasting performance comparison based Accuracy (Acc) and F-1 score w.r.t to state- of-the-art methods. The proposed nMIL , nMIL∆, nMILΩmethod outperform state-of-the-art methods across the three countries. Method Argentina Brazil Mexico Acc 0.611(±0.034) SVM MI-SVM 0.676(±0.026) 0.330(±0.040) rMILnor 0.644(±0.032) rMILavg 0.589(±0.058) GICF 0.709(±0.036) nMIL 0.708(±0.039) nMIL∆ 0.687(±0.038) nMILΩ F-1 0.406(±0.072) 0.659(±0.036) 0.411(±0.092) 0.584 (±0.055) 0.624(±0.048) 0.702(±0.047) 0.714(±0.034) 0.680(±0.045) Acc 0.693(±0.040) 0.693(±0.040) 0.505(±0.012) 0.509(±0.011) 0.650(±0.055) 0.723(±0.039) 0.705(±0.048) 0.713(±0.028) F-1 0.598(±0.067) 0.503(±0.087) 0.661(±0.018) 0.513(±0.064) 0.649 (±0.031) 0.686 (±0.055) 0.698(±0.045) 0.687(±0.038) Acc 0.844(±0.062) 0.880(±0.025) 0.499(±0.009) 0.785(±0.038) 0.770(±0.041) 0.898(±0.031) 0.861(±0.014) 0.871(±0.013) F-1 0.814(±0.091) 0.853(±0.040) 0.655(±0.025) 0.768(±0.064) 0.703(±0.056) 0.902(±0.030) 0.868(±0.014) 0.879(±0.014) 6.1 How well does the nMIL forecast protests? 6.1.1 Comparative Evaluation. Table 2 reports the prediction performance of the nMIL approach in comparison to other baseline ap- proaches for the task of forecasting protests. Specifically, we use set β = 3.0, λ = 0.05, m0 = 0.5 and p0 = 0.5 and report the average accuracy and F1 score along with standard deviation for predicting protests across multiple runs of varying historical days with lead time set to 1. We observe that the nMIL approaches outperform the baseline approaches across all the three countries. The rMILnor ap- proach performs poorly because the the noisy-or aggregation function associating the bag-level labels to instance-level labels forces most of the news articles within the positive bags to have probability values close to 1. However, given the large collection of news articles available per day only a subset of them will indicate signals/evidence for a protest. For Argentina, the nMIL and nMIL∆approaches outperformed the best baseline (MI-SVM ), by 7% and 8% with respect the average F1 score, respectively. Figure 3 shows the changes to F1 score for the proposed nMIL approach in comparison to SVM , MI-SVM and rMILavg for different number of historical days that are used in training with lead 10 (a) Argentina (b) Brazil (c) Mexico Figure 3: Forecasting evaluation on 3 countries with respect to F1 score for SVM , rMILnor , rMILavg and nMIL . X-axis is the number of historical days used in the training process. Y-axis shows the average F1 score of 10 runs of experiments. Table 3: F1-measure for rMILavg and nMIL models on Argentina, Brazil, and Mexico with history days from 1 to 5. Country History Days 1 rMILavg nMIL rMILavg nMIL rMILavg nMIL rMILavg nMIL rMILavg nMIL 0.719 0.745 0.659 0.664 0.674 0.649 0.656 0.676 0.669 0.626 Argentina 2 0.714 0.735 0.624 0.675 0.606 0.669 0.558 0.693 0.676 0.676 3 0.690 0.722 0.554 0.740 0.622 0.560 0.588 0.670 0.590 0.687 4 0.710 0.691 0.665 0.710 0.543 0.669 0.556 0.712 0.567 0.773 5 0.705 0.716 0.500 0.751 0.578 0.737 0.476 0.631 0.575 0.737 1 0.717 0.734 0.695 0.699 0.694 0.687 0.729 0.754 0.710 0.683 2 0.692 0.768 0.651 0.611 0.682 0.639 0.712 0.584 0.588 0.665 Brazil 3 0.696 0.721 0.573 0.738 0.620 0.674 0.720 0.736 0.616 0.657 4 0.662 0.735 0.672 0.639 0.715 0.717 0.628 0.735 0.548 0.697 5 0.680 0.717 0.565 0.721 0.622 0.742 0.621 0.725 0.570 0.735 1 0.815 0.842 0.846 0.825 0.819 0.856 0.809 0.872 0.828 0.833 2 0.803 0.868 0.875 0.889 0.787 0.903 0.822 0.888 0.845 0.937 Mexico 3 0.789 0.863 0.860 0.914 0.808 0.884 0.798 0.894 0.810 0.878 4 0.796 0.884 0.878 0.909 0.750 0.909 0.878 0.916 0.733 0.935 5 0.873 0.884 0.912 0.886 0.853 0.900 0.772 0.874 0.889 0.931 Leadtime 1 Leadtime 2 Leadtime 3 Leadtime 4 Leadtime 5 time set to 2. We trained 10 different models that use different number of historical days respectively varying from 1 to 10. These results show the methods that utilize the nested structure (nMIL∆) within the multi-instance learning paradigm, generally performed better than others. Moreover, the proposed nMIL∆models performed well consistently across different countries with different number of history days. 6.1.2 How early can the nMIL forecast? In order to study the changes of performance with and without the nested structure, we show the F1 score with varying lead times and historical days from 1 to 5 for rMILavg and nMIL models in Table 3, respectively. We observe that with larger lead time (i.e., forecasting earlier than later), the nMIL model does not necessary lose forecasting accuracy; but is sometimes even better. This can be explained by the fact that several times protests are planned a few days in advance and that civil unrest unfold as a series of actions taken by multiple participating entities over a sequence of days. As the lead time increases, F1 score for forecasting initially drops and then increases back. This behavior is also noted in prior work by Ramakrishnan et. al [19], which includes protest related data from these countries. In comparison to the nMIL model, the rMILavg approach, which collapses the sequential structure encoded within the history of days seems to perform inconsistently with increasing lead time. 6.2 Do the precursors tell a story? Quantitative Evaluation Figures 4a and 4b show the distribution of the estimated probabilities for instances within positive and negative super bags for Argentina and Mexico, respectively. The instances within the negative super bags show lower probability estimates by the proposed model and the instances within the positive super bags show higher probability estimates. For Mexico, fewer instances within the positives are assigned high probabilities indicating strength of the proposed model to identify and rank the precursors. 11 0.10.20.30.40.50.60.70.80.901234567891011F1 ScoreHistoryDaySVMMI-SVMrMILnMILΔavg00.10.20.30.40.50.60.70.801234567891011F1 ScoreHistory DaySVMMI-SVMrMILnMILΔavg0.40.50.60.70.80.9101234567891011F1 ScoreHistory DaySVMMI-SVMrMILnMILavg(a) Argentina (b) Mexico Figure 4: The estimated probabilities for negative examples (purple) and positive examples (green) for Argentina and Mexico Figure 5: Mean of relative cosine values w.r.t target events in history days for Argentina and Mexico (a) Argentina (b) Mexico Relative cosine similarity is computed as the pairwise normalized cosine similarity, scaled relative to each event. Figures 5a and 5b show the average cosine similarity value for the precursor documents (probability estimate greater than 0.7) with the target GSR documents. For Argentina, we observe that on average, the documents on day 5 have the highest semantic similarity to the target event documents (GSR). The documents on day 3 and day 10 have lower similarity compared to the target event. In order to investigate the relationship between the semantic similarity and the estimated probability by the proposed models, we compare the distribution of relative cosine similarity and relative entity hit score of the precursor documents with the target GSR documents with respect to bag of words features. Entity words in each news document are extracted by an enrichment tool for natural language processing. The relative entity hit score is calculated as the the intersection of entity set of precursor document and the target event divided by the relative minimum length of these two sets. Figures 6a and 6b show the fitted Gaussian distribution of relative cosine similarities for all documents (green lines) and precursor documents (blue lines) for Argentina and Mexico, respectively. Figures 7a and 7b show the distribution of relative entity hit score for Argentina and Mexico, respectively. These distribution figures demonstrate that the proposed model assigns higher probability to news articles with higher semantic similarity to the GSR articles representing the protests events. These results show the strength of our proposed models in identifying the precursor articles. Case Studies We present findings about the identified precursors based on the probability estimate by nMIL across three observed protests. In Figure 1, we present a protest event against government in Argentina, and the selected precursors before its occurrence with their estimated probabilities. The titles 12 (a) Argentina (b) Mexico Figure 6: The figures show the distribution of cosine similarity for all documents (green line) and the distribution for precursor documents (blue line) with probability greater than 0.7. Left is for Argentina and right is fro Mexico. (a) Argentina (b) Mexico Figure 7: The figures show the distribution of relative entity hit score for all documents (green line) and the distribution for precursor documents (blue line) with probability greater than 0.7. Left is for Argentina and right is for Mexico. of news reports as precursors are shown in the timeline. In Figures 9a and 9b, we present story lines by precursors that were discovered for two different protest events in Argentina and Mexico, respectively. Figure 9a showcases the story line about a protest event in Argentina in December 2014. In this case, the police were protesting against government for better salaries. Before this event, clashes between police and gendarmerie (military policy) had occurred leading to involvement of several policemen from different parts of the country. The text from news articles demonstrate the tense situation between the police and government in La Pampa, Argentina identified as precursors. Figure 9b shows another story line of a continuous protest event in Mexico regarding the infamous case of 43 missing students 2. The resulting outrage triggered constant protests which were identified by our proposed model. The figure shows a timeline of how the events turned violent leading up to the burning of congressional offices and depicts how different communities joined the movement. 6.3 Can nMIL forecast event populations? We also evaluated the performance of our nMIL approaches for predicting the event populations by solving a multi-class classification problem. In Table 4 we depict the weighted-average F1 score for event populations (here, with categories such as Government, Wages, Energy, Others drawn from the GSR). 2https://en.wikipedia.org/wiki/2014_Iguala_mass_kidnapping 13 Table 4: Multi-Class F1-Measure for rMILavg and nMIL models on Argentina and Mexico with historical days from 1 to 5. Argentina Mexico History Days 1 rMILavg nMIL rMILavg nMIL 0.512 0.523 0.576 0.570 2 0.512 0.552 0.526 0.583 3 0.473 0.515 0.447 0.560 4 0.417 0.485 0.547 0.615 5 Average(Variance) 0.457 0.537 0.493 0.545 0.474(1e-3) 0.524(7e-4) 0.518(3e-3) 0.575(7e-4) Due to space limitations, we only depict the performance of weighted average F1 score on event population across 1 to 5 historical days with lead time of 1. The proposed multi-class nMIL model outperforms the multi-class rMILavg model. On average, for event population, nMIL outperformed rMILavg by 10.5% and 10.6% for Argentina and Mexico, respectively. 6.4 How sensitive is nMIL to parameters? (a) β in nMIL (b) λ in nMIL Figure 8: Sensitivity analysis on β and λ. X-axis represents the varying values for the parameter and Y-axis is the test accuracy. There are three main parameters in the proposed nMIL model, which are the regularization parameter λ, weight for super bag loss β and threshold for instance level hinge loss m0. Figures 8a and 8b illustrate the performance of the proposed nMIL by varying β and λ, respectively. The test accuracy for different values of λ and β is relatively stable. 7 Conclusion and Future Directions This paper presents a novel extension of the multi-instance learning framework for event forecasting and identifying precursors for protest events. Most existing multi-instance approaches solve problems in object detection in images, drug activity prediction or identify sentimental sentences in text reviews. In contrast, we provide a novel application of MIL algorithms that require a two-level nested structure for event forecasting and precursor modeling. Specifically, we study the strengths of our developed methods on open source news datasets from three Latin American countries. Through extensive evaluation and analysis we show the strong forecasting per- formance of the proposed methods with varying lead time and historical data. We also show qualitatively via several case studies, the richness of the identified precursors for different protests across different cities. In the future, we plan to incorporate heterogeneous data sources like social media streams for event forecasting within the developed framework. We will also extend our nested multi-instance learn- 14 0 0.2 0.4 0.6 0.8 1 1.20.10.20.30.40.50.60.70.80.91Test AccuracyParameter βMexicoBrazilArgentina 0 0.2 0.4 0.6 0.8 1 1.20.010.050.10.150.20.250.30.350.40.45Test AccuracyParameter λMexicoBrazilArgentina(a) A continuous police protest in Argentina against government for better salary. In the beginning, policemen at Cordoba were requesting for better salaries. Later on, police in Catamarca were involved in clashes with gendarmerie. Three days before the target event, the government sent out troops and more and more police joined for the same purpose. One day before the event, Buenos Aires state called for a strike. (b) Protester in Mexico burned the congressional offices for justice for the missing teachers. In the beginning, students were marching for justice. Gradually, more communities such as artistic and policing community joined the event. Later on, children, youth, adults, students and teachers blocked traffic for protest. ing framework by exploring regularized multi-task learning approaches for enforcing similarity of learned parameters, while enforcing spatial and temporal constraints. 8 Acknowledgments Supported by the Intelligence Advanced Research Projects Activity (IARPA) via DoI/NBC contract number D12PC000337, the US Government is authorized to reproduce and distribute reprints of this work for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DoI/NBC, or the US Government. References [1] H. Achrekar, A. Gandhe, R. Lazarus, S.-H. Yu, and B. Liu. Predicting flu trends using twitter data. In IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pages 702–707, April 2011. 15 0.6030.8910.80.8570.8570.8560.8332013-12-11 In La Pampa there were incidents of looting and police protest 2013-12-02 31 residents of the village negotiate with Urban Highway(0.67)2013-12-03 Córdoba policemenwere quartered demanding better salaries(0.89)2013-12-06 Chaos looting in Catamarca: there were clashes between police and gendarmerie (0.92)2013-12-07 Policemaintain the claim(0.89)2013-12-10Call for a strikeby the Buenos Aires state(0.92)2013-12-09Corrientes policebowed to claim for higher wages (0.90)2013-12-09 Mar del Plata policejoined for better wages(0.89)2013-12-08 The government sent 2500 troops(0.85)0.640.520.660.650.710.620.750.780.82014-11-09: Studentstake highway(0.82)2014-11-10: Oaxaca teachersblock Ayotzinapa(0.93)2014-11-08: Ask the PGR justice for Normalistas(0.89)2014-11-07: BlockFederal Palace (0.79)2014-11-05: Studentstook radio stations in the capital(0.78)2014-11-02: The federal gov hinders the work of police(0.55)2014-11-11: CETEG teachers burn congressional offices2014-11-04: Members of the artistic communitymarched(0.76)2014-11-06: 500 peopleinvolved in the mobilization(0.81)StudentsArtistic CommunityPolicing CommunityChild, Youth, Adults2014-11-01: Studentsmarched through the streets (0.79)[2] J. Amores. Multiple instance classification: Review, taxonomy and comparative study. Artificial Intelligence, 201:81–105, 2013. [3] S. Andrews, I. Tsochantaridis, and T. Hofmann. Support vector machines for multiple-instance learning. In Advances in neural information processing systems, pages 561–568, 2002. [4] M. Arias, A. Arratia, and R. Xuriguera. Forecasting with twitter data. ACM Transactions on Intelligent Systems and Technology (TIST), 5(1):8:1–8:24, Jan. 2014. [5] Y. Bengio, R. Ducharme, P. Vincent, and C. Janvin. A neural probabilistic language model. J. Mach. Learn. Res., 3:1137–1155, Mar. 2003. [6] J. Bollen, H. Mao, and X. Zeng. Twitter mood predicts the stock market. Journal of Computational Science, 2(1):1 – 8, 2011. [7] R. Caruana. Multitask learning. Machine Learning, 28(1):41–75, July 1997. [8] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, Sept. 1995. [9] T. Gartner, P. A. Flach, A. Kowalczyk, and A. J. Smola. Multi-instance kernels. In ICML '02: Proceedings of the Nineteenth International Conference on Machine Learning, pages 179–186, San Francisco, CA, USA, 2002. Morgan Kaufmann Publishers Inc. [10] J. He, W. Shen, P. Divakaruni, L. Wynter, and R. Lawrence. Improving traffic prediction with tweet semantics. In Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence, IJCAI, pages 1387–1393, 2013. [11] M. S. Hossain, P. Butler, A. P. Boedihardjo, and N. Ramakrishnan. Storytelling in entity networks to support In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge intelligence analysts. Discovery and Data Mining, KDD, pages 1375–1383, New York, NY, USA, 2012. [12] D. Kotzias, M. Denil, N. de Freitas, and P. Smyth. From group to individual labels using deep features. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD, pages 597–606, New York, NY, USA, 2015. [13] S. Laxman, V. Tankasali, and R. W. White. Stream prediction using a generative model based on frequent episodes in event sequences. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD, pages 453–461, New York, NY, USA, 2008. [14] Q. V. Le and T. Mikolov. Distributed representations of sentences and documents. CoRR, abs/1405.4053, 2014. [15] G. Liu, J. Wu, and Z.-H. Zhou. Key instance detection in multi-instance learning. In ACML, volume 25 of JMLR Proceedings, pages 253–268. JMLR.org, 2012. [16] T. Mikolov, K. Chen, et al. Efficient estimation of word representations in vector space. CoRR, abs/1301.3781, 2013. [17] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean. Distributed representations of words and phrases and their compositionality. CoRR, abs/1310.4546, 2013. [18] B. O'Connor, R. Balasubramanyan, B. R. Routledge, and N. A. Smith. From tweets to polls: Linking text sentiment to public opinion time series. In Proceedings of the Fourth International Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 2010. [19] N. Ramakrishnan, P. Butler, S. Muthiah, and et al. "Beating the News" with EMBERS: Forecasting Civil Unrest Using Open Source Indicators. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD, pages 1799–1808, New York, NY, USA, 2014. [20] J. Ritterman, M. Osborne, and E. Klein. Using prediction markets and twitter to predict a swine flu pandemic. In Proceedings of the 1st International Workshop on Mining Social, 2009. [21] Y. Rong, H. Cheng, and Z. Mo. Why it happened: Identifying and modeling the reasons of the happening of social events. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '15, pages 1015–1024, New York, NY, USA, 2015. ACM. [22] A. Tumasjan, T. Sprenger, P. Sandner, and I. Welpe. Predicting elections with twitter: What 140 characters reveal about political sentiment. In Proceedings of the Fourth International AAAI Conference on Weblogs and Social Media, pages 178–185, 2010. 16 [23] X. Wang, M. S. Gerber, and D. E. Brown. Automatic crime prediction using events extracted from twitter posts. In Proceedings of the 5th International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction, SBP, pages 231–238, Berlin, Heidelberg, 2012. [24] X. Wang, Z. Zhu, C. Yao, and X. Bai. Relaxed multiple-instance SVM with application to object discovery. CoRR, abs/1510.01027, 2015. [25] N. Weidmann, E. Frank, and B. Pfahringer. A two-level learning method for generalized multi-instance In The European Conference on Machine Learning and Principles and Practice of Knowledge problems. Discovery in Databases, pages 468–479, 2003. [26] L. Zhao, F. Chen, C. Lu, and N. Ramakrishnan. Spatiotemporal event forecasting in social media. In Proceedings of the SIAM International Conference on Data Mining, Vancouver, BC, Canada, pages 963–971, 2015. [27] L. Zhao, Q. Sun, J. Ye, F. Chen, C.-T. Lu, and N. Ramakrishnan. Multi-task learning for spatio-temporal event forecasting. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD, pages 1503–1512, New York, NY, USA, 2015. [28] Z.-H. Zhou and J.-M. Xu. On the relation between multi-instance learning and semi-supervised learning. In Proceedings of the 24th International Conference on Machine Learning (ICML), volume 227, pages 1167–1174, 2007. 17
1408.1228	3	1408	2016-04-01T09:00:05	Location Prediction: Communities Speak Louder than Friends	[ "cs.SI", "physics.soc-ph" ]	Humans are social animals, they interact with different communities of friends to conduct different activities. The literature shows that human mobility is constrained by their social relations. In this paper, we investigate the social impact of a person's communities on his mobility, instead of all friends from his online social networks. This study can be particularly useful, as certain social behaviors are influenced by specific communities but not all friends. To achieve our goal, we first develop a measure to characterize a person's social diversity, which we term `community entropy'. Through analysis of two real-life datasets, we demonstrate that a person's mobility is influenced only by a small fraction of his communities and the influence depends on the social contexts of the communities. We then exploit machine learning techniques to predict users' future movement based on their communities' information. Extensive experiments demonstrate the prediction's effectiveness.	cs.SI	cs	Location Prediction: Communities Speak Louder than Friends Jun Pang University of Luxembourg FSTC & SnT [email protected] Yang Zhang University of Luxembourg [email protected] FSTC 6 1 0 2 r p A 1 ] I S . s c [ 3 v 8 2 2 1 . 8 0 4 1 : v i X r a ABSTRACT Humans are social animals, they interact with different com- munities of friends to conduct different activities. The lit- erature shows that human mobility is constrained by their social relations. In this paper, we investigate the social im- pact of a person's communities on his mobility, instead of all friends from his online social networks. This study can be particularly useful, as certain social behaviors are influ- enced by specific communities but not all friends. To achieve our goal, we first develop a measure to characterize a per- son's social diversity, which we term 'community entropy'. Through analysis of two real-life datasets, we demonstrate that a person's mobility is influenced only by a small frac- tion of his communities and the influence depends on the social contexts of the communities. We then exploit ma- chine learning techniques to predict users' future movement based on their communities' information. Extensive experi- ments demonstrate the prediction's effectiveness. Categories and Subject Descriptors H.2.8 [Database Management]: Database Applications- Data mining General Terms Algorithms, theory, experiments Keywords Human mobility, social networks, network communities 1. INTRODUCTION Humans are social animals, everyone is a part of the so- ciety and gets influences from it. For example, our daily behaviors, such as what types of music we listen to, where we have lunch on weekdays and what activities we conduct on weekends, are largely dependent on our social relations. Normally, we categorize our social relations into different groups, i.e., social communities, using different criteria and considerations. By definition, a community is a social unit of any size that shares common values.1 Typical commu- nities include family, close friends, colleagues, etc. In daily life, humans are engaged in various social environments, and they interact with different communities depending on the environments. For our specific behaviors, social influences, in most of cases, are not from all our friends but from certain communities. For example, we listen to similar types of mu- sic as our close friends, but not as our parents; we have lunch together with our colleagues on weekdays, but not with our college friends living in another city; on weekends we spend more time with family, but not with our colleagues. Location-based social network services (LBSNs) have been booming during the past five years. Nowadays, it is com- mon for a user to attach his location when he publishes a photo or a status using his online social network account. Moreover, users may just share their locations, called check- in, to tell their friends where they are or to engage in social games as in Foursquare. Since these large amount of lo- cation and social relation data become available, studying human mobility and its connection with social relationships becomes quantitatively achievable (e.g., [16, 10, 9, 34, 15, 7, 8]). Understanding human mobility can lead to compelling applications including location recommendation [44, 41, 43, 14, 20], urban planning [42], immigration patterns [5], etc. Previous works, including [1, 6, 10, 33], show that human mobility is influenced by social factors. However, there is one common shortcoming: they all treat friends of users equally. Similar to other social behaviors, in most cases mobility is influenced by specific communities but not all friends. For example, the aforementioned colleagues can influence the place a user goes for lunch but probably have nothing to do with his weekend plans. Meanwhile, where a user visits on weekends largely depends on his friends or family, but not his colleagues. Therefore, the impact on a user's mobility should be considered from the perspectives of communities instead of all friends. In a broader view, community is arguably the most useful resolution to study social networks [39]. In this paper, we aim to study the im- Contributions. pact from communities on a user's mobility and predict his locations based on his community information. First, we partition each users' friends into communities and propose a notion namely community entropy to quantify a user's social diversity. Second, we analyze communities' in- fluences on users' mobility and our main conclusions include: COSN'15, November 2–3, 2015, Palo Alto, California, USA. . 1http://en.wikipedia.org/wiki/Community (1) communities' influences on users' mobility are stronger than their friends'; (2) each user is only influenced by a small number of his communities; and (3) such influence is typi- cally constrained by temporal and spatial contexts. Third, we predict users' locations using their community informa- tion. Experimental results on two real-life datasets with millions of location data show that the community-based predictor achieves a strong performance. Organization. After the introduction, we present a few preliminaries and our datasets in Section 2. Then we de- scribe the community detection process and propose the no- tion of community entropy in Section 3. The relationship between users and their communities on mobility is ana- lyzed in Section 4. Based on our analysis, we propose a location predictor with features linked to community infor- mation and present experimental results in Section 5. We discuss related work in Section 6 and conclude our paper with some future work in Section 7. 2. PRELIMINARIES We summarize the notations in Section 2.1 and describe the datasets that we use throughout the paper in Section 2.2. 2.1 Notations All users are contained in the set U while a single user is denoted by u. We use the set f (u) to represent u's friends. A community of a user u is a subset of his friends denoted by c and c ⊆ f (u). Meanwhile, C (u) represents all the com- munities of u, i.e., C (u) is a set of sets of u's communities. Every friend of a user is assigned into one of the user's com- munities, the union of all his communities is the set of all his friends. In this work, we only consider non-overlapping com- munities, namely c ∩ c(cid:48) = ∅ for c, c(cid:48) ∈ C (u). However, this assumption is not crucial to our approach and our results can be extended for overlapping communities as well. A check-in of u is denoted by a tuple (cid:104)u, t, (cid:96)(cid:105), where t represents the time and (cid:96) is the location that corresponds to a pair of latitude and longitude. We use CI (u) to represent all the check-ins of u. Without ambiguity, we use location and check-in interchangeably in the following discussion. 2.2 The datasets We exploit two types of social network datasets for this work. The first one is collected by the authors of [10] from Gowalla – a popular LBSN service back in 2011. The dataset was collected from February 2009 to October 2010 and it contains 6,442,892 check-ins. Besides location information, the dataset also includes the corresponding social data which contains around 1.9 million users and 9.5 million edges. Due to the large data sparsity, we mainly focus on the check-in data in two cities in US, including New York (NY (G)) and San Francisco (SF (G)). They are among the areas with most check-ins in the dataset. In addition, when performing mobility analysis and location prediction, we only focus on users who have conducted at least 100 check-ins in each city and we term these users as active users. The second dataset is collected from Twitter from Decem- ber 2014 to April 2015 by the authors of this paper. Again, we focus on the data in New York (NY (T)) and San Fran- cisco (SF (T)) and treat all the geo-tagged tweets (tweets labeled with geographical coordinates) as users' check-ins. We exploit Twitter's Streaming API2 to collect all the geo- tagged tweets. Each check-in is organized as a 4-tuple. (cid:104)uid , time, latitude, longitude(cid:105) Figure 1 depicts a sample of check-ins in New York. To col- lect the social relationships among users, we adopt Twitter's REST API3 to query each user's followers and followees. Two users are considered friends if they follow each other mutually. Similar to the Gowalla dataset, we only focus on active users (users with more than 100 check-ins) in the Twitter dataset. Moreover, we also filter out the users who have more than 2,000 check-ins since most of them are public ac- counts such as @NewYorkCP which publishes 16,681 check- ins at the exact same location. Table 1 summarizes the two datasets. The Twitter dataset is available upon request. 3. COMMUNITIES We first show how to detect communities in social net- works in Section 3.1 and then propose a new notion to char- acterize users' social diversity in Section 3.2. 3.1 Community detection in social networks Community detection in networks (or graphs) has been extensively studied for the past decade (e.g., see [25, 31, 2, 18, 32, 22, 38, 37, 21, 39, 23]). It has important applica- tions in many fields, including physics, biology, sociology as well as computer science. The principle behind community detection is to partition nodes of a large graph into groups following certain metrics on the graph structure [18]. In the context of social networks, besides the social graph, each user is also affiliated with attributes. These information can also be used to detect communities (e.g., see [22, 38, 23]). For example, people who graduate from the same university can be considered as a community. Since the datasets we use only contain social graphs and no personal information are provided, we apply the algorithms that are based on infor- mation encoded in graph structure to detect communities. According to the comparative analysis [18], among all the community detection algorithms, Infomap [31] has the best performance on undirected and unweighted graphs and has been widely used in many systems [26, 29]. Therefore, we apply it in this work. Next we give a brief overview of In- fomap and describe how we use it to detect communities. The main idea of Infomap can be summarized as follows: information flow in a network can characterize the behav- ior of the whole network, which consequently reflects the structure of the network. A group of nodes among which information flows relatively fast can be considered as one community. Therefore, Infomap intends to use information flow to detect communities in a network. In the beginning, Infomap simulates information flow in a network with ran- dom walks. Then the algorithm partitions the network into communities and exploits Huffman coding to encode the net- work at two levels. At the community level, the algorithm assigns a unique code for each community based on the infor- mation flow among different communities; at the node level, the algorithm assigns a code for each node based on the in- formation flow within the community. Infomap allows the Huffman codes in different communities (node level) being 2https://dev.twitter.com/streaming/overview 3https://dev.twitter.com/rest/public # of users # of check-ins Avg.# of check-ins # of active users Avg.# of friends (active user) NY (G) SF (G) NY (T) 207,805 SF (T) 113,383 176,324 177,357 2,325,907 2,163,959 7,786 6,617 21.6 175 79.4 26.8 236 69.7 11.2 1,636 376.9 19.1 1,626 289.0 Figure 1: Check-ins in New York. Table 1: Summary of the datasets. duplicated which results in a more efficient encoding (less description length). In the end, finding a Huffman code to concisely describe the information flow while minimizing the description length is thus equivalent to discovering the net- work's community structure. In other words, the objective of Infomap is to find a partition of a network such that the code length for representing information flow among com- munities and within each community is minimized. Since it is infeasible to search all possible community partitions, Infomap further exploits a deterministic greedy search algo- rithm [11, 36] to find partitions. In our work, to detect communities of u, we first find all his friends as well as the links among them. Then, we delete u and all edges linked to him and apply Infomap algorithm to the remaining part of the graph. Figure 2 presents the detected communities of two users in the Gowalla dataset. Each community is marked with a different color. (a) User 521 (15 friends) (b) User 727 (50 friends) Figure 2: Communities of users 521 and 727. Avg.# of communities Avg. community size Gowalla Twitter 5.3 20.8 4.5 13.2 Table 2: Community summary of active users. Table 2 lists the summary of community information of all active users in the two datasets. Each active user in Gowalla has on average 4.5 communities while the value is 5.3 for the Twitter users. In addition, the average com- munity size of Twitter users is bigger than Gowalla users (20.8 vs. 13.2). This is because active users in the Twitter dataset have more friends than those in the Gowalla dataset Figure 3: Distribution of users w.r.t the number of commu- nities and distribution of communities w.r.t their size. (see Table 1), which indicates general social network ser- vices, such as Twitter, contain more users' social relation- ships than LBSN services, such as Gowalla. In spite of the differences on the average value in Table 2, community num- ber and community size in the two datasets follow a similar distribution. As we can see from Figure 3, both community number and size follow the power law: most of the users have small number of communities and most of the detected communities are small as well. 3.2 Community entropy After detecting communities, we are given a new domain of attributes on users. We are particularly interested in how diverse a user's communities are. We motivate this social di- versity through an example. Suppose that a user is engaged in many communities, such as colleagues at work, family members, college friends, chess club, basketball team, etc, then he is considered an active society member. Users of this kind are always involving in different social scenarios or environments, and his daily behaviors are largely dependent on his social relations. Although we do not have the semantics of each of our de- tected communities, such as the aforementioned colleagues at work or chess club, we can still use the information en- coded in the graph to define a user's social diversity. For instance, for a user with several communities whose sizes are more or less the same, his social diversity is for sure higher than those with only one community. To quantify the social diversity of a user, we introduce the notion of community entropy. Definition 1. For a user u, his community entropy is de- fined as coment(u) = 1 1 − α ln (cid:88) ( c∈C (u) c f (u) )α. 122126830852191921945206317543472634727347283472934730347313473226911224229238367672774375712021221132713361377139520821352372221845812446127421310018872368484349591764817994201378454221602216772398576230632308920264915802610416133672557239935399363993739938399393994039941# of Communities100101102103104Probability10-410-310-210-1100GowallaTwitterCommunity Size100101102Probability10-510-410-310-210-1100GowallaTwitterOur community entropy follows the definition of R´enyi en- tropy [30]. Here, α is called the order of diversity, it can control the impact of community size on the value which gives more flexibility to distinguish users when focusing on the sizes of their communities. In simple terms, our com- munity entropy, • when α > 1, values more on larger communities; • when α < 1, values more on smaller communities. The limit of coment(u) with α → 1 is the Shannon entropy.4 In general, if a user has many communities with sizes equally distributed, then his community entropy is high and this indicates that his social relations are highly diverse. We set α > 1 in the following discussion to limit the im- pact of small communities since a user may randomly add strangers as his friends in online social networks and these strangers normally form small communities (such as a one- user community5), which have less impact on the user's mo- bility. For example, if a user u has three communities with sizes equal to 1, 1 and 10, then his communities are not that diverse following the above intuition. When we set α less than 1, such as 0.5, we have coment(u) = 0.79 which is a high value indicating u's social circles are diverse. On the other hand, if we set α bigger than 1, such as 10, coment(u) drops to 0.20 which captures our intuition. In the following experiments, we set α = 10 when calculating users' com- munity entropies. Note that we have also set α to other numbers bigger than one and observed similar results. Fig- ure 4 shows the histogram of community entropies of all active users in two datasets. Figure 4: Distribution of community entropies of active users in Gowalla (left) and Twitter (right). 4. COMMUNITIES AND MOBILITY It has been proved that social factors play an important role on users' mobility, e.g., see [10]. For instance, one may go to lunch with his friends or go to a bar to hangout with his friends. Meanwhile, for a user, friends of his social networks (as well as in real life) are not all equal. Instead friends nor- mally belong to certain communities. When considering a user's mobility, intuitively different communities can impose different influence within certain contexts or social environ- ments. Continuing with the above example, the people the user has lunch with are normally his colleagues while the people he meets at night are his close friends. Therefore, in order to analyze the impact from a user's social relations on 4https://en.wikipedia.org/wiki/Renyi_entropy 5In our community detection algorithm, if u(cid:48) himself forms a community of u, then it indicates that u(cid:48) does not know any other friends of u. his mobility, it is reasonable to focus on social influence at the community level. In this section, we first study communities' influence on users' mobility. After that, we study the characteristics of the influential communities with the following two intuitions in mind: (1) a user's daily activities are constrained, and the number of communities he interacts with is limited; (2) communities influence a user's social behavior under differ- ent contexts. 4.1 Influential communities Figure 5a depicts a user's two communities' check-ins in Manhattan of the New York City. We can observe a quite clear separation between these two communities' check-ins: members of community 1 mainly visit Uptown and Mid- town Manhattan while community 2 focuses more on Mid- town. This indicates that different communities have their social activities at different areas. In a broader view, this shows that partitioning users' check-ins at the social network level (through community detection) can result in meaning- ful spatial clusters as well. A single community also has several favorite places. For example, community 1 in Figure 5a visits Times Square and Broadway quite often while members of community 2 like to stay close to Madison square park. A user may socialize with different communities at different places, for example, he may go to watch a basketball game with his family at the stadium and have lunch with his colleagues near his office. Therefore, to study influences on mobility from communities to a user, we need to summarize each community's frequent movement areas. To discover a community's frequent move- ment areas, we perform clustering on all locations that the community members have been to. Each cluster is then rep- resented by its central point and a community's frequent movement areas are thus represented by the centroids of all clusters. The clustering algorithm we use is the agglom- erative hierarchical clustering. We regulate that any two clusters can be aligned only if the distance between their corresponding centroids is less than 500m which is a reason- able range for human mobility. To illustrate the mobility influence from communities to users, we choose to use 'distances'. More precisely, we rep- resent the influence by the distances between a user's loca- tions and the frequent movement areas of his communities. Shorter distances imply stronger influences. For each loca- tion a user has visited, we calculate the distances between the location and all his communities' frequent movement ar- eas. Then, for each community of the user, we choose the shortest distance between the location and the community's frequent movement areas as the distance between the lo- cation and the community. The community which has the smallest distance to the location is considered as the influ- ential community of the user at this location. The distance between the influential community and the user's location is further defined as the distance between the user's location and his communities. Note that a user can have multiple influential communities and an influential community can influence a user on multiple locations. Figure 5b depicts the distribution of distances between users' locations and their communities in New York and San Francisco in the two datasets. As we can see, most of the distances are short which indicates the communities are quite close to users' locations. To illustrate that these Community Entropy01234Probability00.050.10.150.20.25Community Entropy01234Probability00.020.040.060.080.10.120.14(a) (b) (c) Figure 5: (a) A user's two communities' check-ins in Manhattan; (b) distribution of distances between users and their communities; (c) cumulative distribution function of distances between users and their communities, friends, sample friends and random users. (a) (b) (c) Figure 6: (a) Distribution of the number of influential communities; (b) distribution of influence entropies (bucketed by 0.2) (c) influence entropy vs. community entropy. short distances are not due to the limits of the city areas, for each location of a user, we pick some random users in the city, summarize their frequent movement areas through clustering and find the minimal distance between their fre- quent movement areas and the location. In Figure 5c6, the curve of cumulative distribution function (CDF) for these random users (purple) is much lower than the one for com- munities (blue). This means that these random users are farther away from the users than communities. To show that community is a meaningful level to study mobility, we also calculate distances between a user and all his friends. The curve for friends (red) in Figure 5c is lower than the one for communities as well, meaning that a user is closer to his communities than to all his friends in general. As a user's community is a subset of his friends, to illustrate that the shorter distances for communities than friends are not caused by frequent movement areas clustered from a small number of friends' check-ins, for each community of a user, we randomly sample the same number of his friends to build a "virtual" community and calculate the distances between the user and his virtual communities. The CDF curve in Figure 5c (yellow) shows that these virtual communities are even farther away from users than all friends. 6The results in Figure 5c are based on the data from two cities in both datasets. From the above analysis, we conclude that (1) communi- ties have strong influences on users' mobility and (2) com- munity is a meaningful resolution to study users' mobility. 4.2 Number of influential communities Research shows that a user's mobility is constrained geo- graphically (see [9, 10]), e.g., a user normally travels in or around the city where he lives. Meanwhile, social relations are not restricted by geographic constrains. For instance, a user's college friends as a community can spread all over the world. Now we focus on how many communities ac- tually influence a user's mobility i.e., how many influential communities a user has. Intuitively, this number should be small as each user only interacts with a limited number of communities in his daily life such as colleagues and family. We plot the distribution of the number of user's influential communities in Figure 6a. From two datasets, we can ob- serve a similar result. Most of the users are influenced only by a small number of communities and there are more users who have two influential communities than others. For ex- ample, almost 30% of users in New York have two influential communities in the Twitter dataset. Each location corresponds to an influential community. We proceed with studying how a user's influential commu- nities are distributed over his check-ins. We first propose a community 1community 2Distance (km)10-1100101102Probability100102104106Distance (km)10-1100101102CDF00.20.40.60.81CommunityFriendsSample FriendsRandom Users# of Influential Communities123456789101112Probability00.050.10.150.20.250.30.350.4NY (G)SF (G)NY (T)SF (T)Influence Entropy00.20.40.60.811.21.41.61.822.2Probability00.050.10.150.20.250.30.350.4NY (G)SF (G)NY (T)SF (T)Community Entropy0 0.20.40.60.81 1.21.4Avg. Influence Entropy00.20.40.60.811.21.4NY (G)SF (G)NY (T)SF (T)notion named influence entropy, it is defined as CI (u, c) CI (u) CI (u, c) CI (u) ln infent(u) = − (cid:88) c∈C (u) where CI (u, c) represents u's check-ins that are closest to the community c. The influence entropy is defined in the form of Shannon entropy: higher influence entropy indicates that the user's locations are close to his different communi- ties more uniformly. Figure 6b depicts the distribution of users' influence entropies. As we can see, in New York (NY (T)), around 20% of users' influence entropies are between 0 and 0.2 which means they have one dominating influen- tial community that is close to most of their locations. We also notice that there is a peak around 0.6 in all the cities. For example, if a user u's 50% check-ins corresponds to one influential community and the other 50% corresponds to an- other one, then infent(u) = 0.69 which falls into this range. This shows that around 20% of users are influenced by their two major communities at a similar level. Community entropy introduced in Section 3 is a notion for capturing a user's social diversity. We further study the relationship between community entropy and influence en- tropy. As shown in Figure 6c, more diverse a user's social relationship is, more probably his locations are distributed uniformly over his influential communities. From the above analysis, we conclude that only a small number of communities have influences on users' mobility. 4.3 Communities under contexts Influential communities are constrained by contexts. For instance, a user has lunch with his colleagues and spends time with his family near where he lives. Here, the lunch hour and the home location can be considered as social con- texts, and the two communities (colleague and family) have impact on the user's behavior under each of the context, respectively. Thus it is interesting to study whether this hypothesis holds generally. Temporal contexts. First, we focus on temporal contexts. The pair of contexts we choose are Lunch (11am–1pm) and Dinner (7pm–9pm) hours on Wednesday. For each user, we extract his check-ins during lunch and dinner time and find his influential communities w.r.t. these two contexts. We randomly choose four users and plot the distributions of their check-ins over their influential communities under these two contexts in Figure 7. As we can see, a user's com- munities behave quite differently on influencing his check-ins during lunch and dinner time. For example, the first user in New York in the Twitter dataset is only influenced by his community 3 during lunch time while communities 1 and 2 give him similar influences during dinner time. This simply reflects the fact that the people who users have lunch and dinner with are different. In addition, users' average influ- ence entropies drop as well under different temporal con- texts compared with the general case (see Table 3), this suggests that the influential communities tend to become more unique. For each user during lunch (dinner) time, we create a vec- tor where the i-th component counts the number of locations that are the closest to community i. We then exploit the co- sine similarity between a user's lunch and dinner vectors as his influence similarity. The results are listed in Table 4. Note that, we also choose other pairs of temporal contexts for analysis, such as working hours (9am–6pm) and nightlife (10pm–6am) and have similar observations. Spatial contexts. Next we study the influence of spatial contexts. In each city, we pick two disjoint regions (called Region 1 and Region 2, respectively) including Uptown and Downtown Manhattan in New York and Golden Gate Park and Berkeley in San Francisco. Then, we extract users' check-ins in these areas. By performing the same analysis as the one for temporal contexts, we observe similar results (see Figure 8, Table 3 and Table 4). Note that we choose the areas without special semantics in mind, e.g., business areas or residential areas. Influence entropy NY (G) SF (G) NY (T) SF (T) General Temporal (Lunch) Temporal (Dinner ) Spatial (Region 1 ) Spatial (Region 2 ) 0.56 0.35 0.27 0.45 0.42 0.73 0.39 0.43 0.20 0.21 0.69 0.22 0.30 0.52 0.61 0.70 0.25 0.31 0.23 0.26 Table 3: Influence entropy under different social contexts. Influence similarity NY (G) SF (G) NY (T) SF (T) Temporal Spatial 0.80 0.77 0.74 0.56 0.67 0.48 0.66 0.41 Table 4: Influence similarity w.r.t. social contexts. From the above analysis, we can conclude that community impact is constrained under spatial and temporal contexts. 5. LOCATION PREDICTION Location prediction can drive compelling applications in- cluding location recommendation and targeted advertising. On the other hand, it may also threat users' privacy [35]. Following the previous analysis, we continue to investigate whether it is possible to use community information to effec- tively predict users' locations, using machine learning tech- niques. More precisely, the question we want to answer is: given a user's community information, whether he will check in at a given place at a given time. Note that the time here is a certain hour on a certain day (Monday to Sunday). We first list all the features in the community-based loca- tion prediction model. Then, we present the baseline pre- dictors. Experimental results are described in the end. 5.1 Community-based location predictor To predict whether a user will visit a certain location, we use one of his communities' information to establish the feature vector, i.e., the influential community of the location (see Section 4). Community related features. Having chosen the com- munity, we extract its following features for prediction. • Distance between the community and the location. This is the distance between the location and the com- munity's nearest frequent movement area. • Community size. Number of users in the community. (a) NY (G) (b) SF (G) (c) NY (T) (d) SF (T) Figure 7: Distribution of influential communities on users' check-ins (temporal contexts). (a) NY (G) (b) SF (G) (c) NY (T) (d) SF (T) Figure 8: Distribution of influential communities on users' check-ins (spatial contexts). • Number of the community's frequent movement areas. • Community's total number of check-ins. • Community connectivity. This is the ratio between the number of edges in the community and the maximal number of possible edges. Time. Check-ins are related to time as well. Figure 9a (Figure 9b) plots the total number of check-ins in New York and San Francisco in a daily (weekly) scale. Since we aim to predict whether a user will check in at a place at a cer- tain time, the time-related features we consider are the total number of check-ins at the time7 and the day (i.e., Monday to Sunday) from all users. 5.2 Baseline models Sample friends. In our community-based predictor, each location corresponds to the user's nearest community. To illustrate the effectiveness of communities on predicting a user's mobility, in the first baseline model, for each location, we randomly sample the same number of friends as the com- munity and use these friends to build a "virtual community" (as in Section 4). We then replace the community related features with this virtual community's corresponding ones. The time-related features of this model are exactly the same as the ones for the community-based model. Friends. In the second baseline model, we consider a user's all friends instead of his communities. The features include the shortest distance from his friends to the location and the time-related features. User. It has been shown in [10, 6] that a user's past mobility can predict his future mobility effectively. Therefore, we also extract features from a user himself to perform prediction. The features include the following. 7We consider time at a per hour unit, thus the feature is the number of check-ins of all the users at that hour. (a) Daily (b) Weekly Figure 9: Check-in time in the datasets. • The shortest distance from a user's frequent movement areas (through hierarchical clustering with cut-off dis- tance equal to 500m) to the location.8 • The total number of check-ins during the day. • The total number of check-ins during the hour. User and community. In the last baseline model, we com- bine the features from the user's model and our community- based predictor. 8To avoid overfitting, we use half of each user's check-ins to discover his frequent movement areas and the other half are used for training and testing the model. Com1Com2Com3Com1Com2Com1Com2Com3Com1Com2Probability00.20.40.60.81Wed. LunchWed. DinnerCom1Com2Com1Com2Com1Com2Com3Com1Com2Com3Probability00.20.40.60.81Wed. LunchWed. DinnerCom1Com2Com3Com1Com2Com1Com2Com3Com1Com2Com3Probability00.20.40.60.81Wed. LunchWed. DinnerCom1Com2Com1Com2Com1Com2Com3Com1Com2Probability00.20.40.60.81Wed. LunchWed. DinnerCom1Com2Com1Com2Com3Com4Com1Com2Com1Com2Probability00.20.40.60.81Region 1Region 2Com1Com2Com1Com2Com3Com4Com1Com2Com3Com4Com5Com1Com2Probability00.20.40.60.81Region 1Region 2Com1Com2Com1Com2Com1Com2Com3Com1Com2Com3Com4Probability00.20.40.60.81Region 1Region 2Com1Com2Com3Com4Com1Com2Com3Com4Com5Com1Com2Com1Com2Probability00.20.40.60.81Region 1Region 20:006:0012:0018:000:00Probability00.0050.010.0150.020.025NY (G)SF (G)NY (T)SF (T)SuMTuWThFSaSuProbability00.0050.010.0150.02NY (G)SF (G)NY (T)SF (T)5.3 Metrics We partition the cities into 0.001×0.001 degree latitude and longitude cells, a user is said to be in a cell if he has been to any place belonging to the cell. Let TP, FP, FN and TN denote true positives, false positives, false negatives and true negatives, respectively. The metrics we adopt for evaluation include (1) Accuracy, Accuracy = TP + TN TP + FP + FN + TN ; (2) F1 score, We further calculate the Pearson's correlation coefficient9 between community entropy and our prediction results. In the Twitter dataset, the correlation coefficient for New York and San Francisco is 0.88 and 0.97 respectively,10 indicat- ing that community entropy and the prediction results are strongly correlated. This validates our intuition that a user with high social diversity is clearly influenced by his com- munities. We can conclude that community information can be explored to achieve promising location predictions, espe- cially for those users with high community entropies. F1 = 2 · Precision × Recall Precision = Precision + Recall TP TP + FP , Recall = , with TP TP + FN ; and (3) AUC (area under the ROC curve). 5.4 Experiment setup We build a classifier for each user. A classifier needs both positive and negative examples. So far we only have the positive ones, i.e., a user visits a location. To construct the negative examples, for each location a user visits, we ran- domly sample a different location (within the city) as the place that he does not visit at that moment. In this way, a balanced dataset for each user is naturally formed. As in the data analysis, we only focus on active users who have at least 100 check-ins in the city. For each user, we sort his check-ins chronologically and put his first 80% check-ins for training the model and the rest 20% for testing. The ma- chine learning classifier we exploit here is logistic regression. In all sets, we perform 10-fold cross validation. 5.5 Results Performance in general. As depicted in Figure 10, our community-based predictor's performance is promising and it outperforms two baseline models that exploit friends' in- formation. Especially for the sample friends model, the community-based model is almost 20% better among all three metrics in the Twitter dataset. By studying logis- tic model's coefficients, the most important feature is the distance between the community and the location, followed by the community connectivity and size. On the other hand, two predictions that are based on user's own information perform better than our community- based predictor. Also, the predictor combining user and community information does not improve the performance. This indicates that a user's past check-ins are the most useful information for predicting where he will be in the future which also validates the results proposed in [10, 6]. Prediction vs. community entropy. In Figure 11, we bucket community entropy by intervals of 0.2 and plot its relationship with the prediction results (AUC). As we can see, with the increase of community entropy, the AUC grows for the community-based model which means the predictor works better for users with high community entropies. For example, the AUC value increases more than 5% in San Francisco in the Gowalla dataset (community entropy from [0, 0.2) to [1.2, 1.4)). Figure 11: AUC as a function of community entropy, values of the Pearson's correlation: 0.60 (NY (G)), 0.75 (SF (G)), 0.88 (NY (T)), 0.97(SF (T)). Difference between cities. In Figure 10 and Figure 11, we observe that the prediction results are different between two cities. New York has the better performance than San Francisco in the Gowalla dataset. On the other hand, the prediction results are similar in the Twitter dataset. The reason for different performances in different cities could be due to the density of the cities (e.g., New York's population density is higher than San Francisco), or the adoption of LBSN services by users in different cities. We leave the investigation as a future work. 5.6 Other strategies to choose communities So far, we have shown that exploring community informa- tion can lead to effective location prediction. The commu- nity we choose is the one that has the closest frequent move- ment area to the target location. We would like to know if other strategies to choose community can achieve similar re- sults. We consider three strategies including choosing the community with most users (max-size), the community with highest connectivity (max-con) and random community (ran- dom). Table 5 summarizes the prediction performances in New York in the Twitter dataset. As we can see, our orig- inal strategy outperforms these three. Among these three strategies, max-con performs slightly better than the other two, but it is still relatively worse than our original strategy to choose community. This again validates our observation in Section 4 that influential communities are constrained by contexts (spatially or temporally), in other words one com- munity cannot influence every location of the user. 9Pearson's correlation coefficient is the covariance of two variables divided by the product of their standard devia- tions. 10The two values are slightly smaller for the Gowalla dataset, which is probably due to the fact that the Twitter dataset contains more information on social relations than the Gowalla dataset (see discussions in Section 3). Community Entropy00.20.40.60.811.21.4AUC0.70.750.80.850.9NY (G)SF (G)NY (T)SF (T)(a) AUC (b) Accuracy (c) F1score Figure 10: Prediction results. Community max-size max-con random AUC Accuracy F1score 0.83 0.73 0.74 0.71 0.79 0.74 0.74 0.72 0.78 0.72 0.73 0.71 Table 5: Performance of community-choosing strategies. 5.7 Comparison with the PSMM model In [10], the authors establish a mobility model (PSMM) for each user based on his past check-ins. The assumption behind this model is that a user's mobility is mainly cen- tered around two states such as home and work. Each state is modeled as a bivariate Gaussian distribution and the total mobility is then formalized into a dynamic Gaussian mixture model with time as an independent factor. The check-ins that do not fit well with the two states are considered as social check-ins and are modeled through another friends- based distribution. We implement the PSMM model and compare its performance with our community-based predic- tor. Each user's first 80% check-ins are used for training his PSMM model. For testing, besides the rest 20% check- ins, we also construct the same number of locations that the user does not go at the moment (as our classification setup). As the PSMM model's output is the exact location of the user, we consider the prediction is correct when the output location is within 1km of the real location. Table 6 shows the accuracy between our model and PSMM. In all the datasets, our community-based predictor significantly outperforms PSMM. As suggested in [33], this is probably because two states are not enough to capture a user's mo- bility in a city. Moreover, a user's check-in data is also too sparse to train a good PSMM model. We leave the further investigation as a future work. NY (G) LA (G) NY (T) SF (T) Community PSMM 0.76 0.55 0.67 0.60 0.78 0.67 0.81 0.65 Table 6: Comparison with PSMM on prediction accuracy. 6. RELATED WORK Thanks to the emerging of LBSNs, mobility as well as its connection with social relations have been intensively stud- ied [9, 34, 15]. There are mainly two directions of research going on in the area. One direction is to use the location information from LBSNs to predict friendships (see e.g. [19, 13, 12, 6, 33, 28, 40]), the other studies the impact from friendships on locations [1, 6, 10, 33, 24] which is what we focus on in the current work. Backstrom, Sun and Marlow [1] study the friendship and location using the Facebook data with user-specified home addresses. They find out that the friendship probability as a function of home distances follows a power law, i.e., most of friends tend to live closely. They also build a model to pre- dict users' home location based on their friends' home. Their model outperforms the predictor based on IP addresses. The authors of [6] use the Facebook place data to study check- in behaviors and friendships. They train a logistic model to predict users' locations. Besides that, they also inves- tigate how users respond to their friends' check-in and use the location data to predict friendships. Cho, Myers and Leskovec [10] investigate the mobility patterns based on the location data from Gowalla, Brightkite as well as data from a cellphone company. Based on their observation, they build a dynamic Gaussian mixture model for human mobility in- volving temporal, spatial and social relations features. Sadilek, Kautz and Bigham [33] propose a system for both location and friendship prediction. For location prediction, they use dynamic Bayesian networks to model friends' locations (un- supervised case) and predict a sequence of locations of users over a given period of time. McGee, Caverlee and Cheng [24] introduce the notion of social strength based on their obser- vation from the geo-tagged Twitter data and incorporate it into the model to predict users' home locations. Experi- mental results show that their model outperforms the one of [1]. Jurgens in [17] proposes a spatial label propagation algorithm to infer a user's location based on a small num- ber initial friends' locations. Techniques such as exploiting information from multiple social network platforms are inte- grated into the algorithm to further improve the prediction accuracy. The main difference between previous works and ours is the way of treating friends. We consider users' friends at a community level while most of them treat them the same (except for the paper [24] which introduce 'social strength', which is based on common features but not on communities). Moreover, our location predictor doesn't need any user's own information but his friends' to achieve a promising result, es- pecially for users' with high community entropies. Other mi- NY (G)SF (G)NY (T)SF (T)AUC00.20.40.60.81CommunitySample FriendsFriendsUserUser+CommunityNY (G)SF (G)NY (T)SF (T)Accuracy00.20.40.60.81CommunitySample FriendsFriendsUserUser+CommunityNY (G)SF (G)NY (T)SF (T)F1score00.20.40.60.81CommunitySample FriendsFriendsUserUser+Communitynor differences include the prediction target: we want to pre- dict users' certain locations in the future not their home [1, 24, 17] or a dynamic sequences of locations [33]. We focus on understanding users' mobility behavior from social network communities. The authors of [4] tackle the inverse problem, i.e., they exploit users' mobility informa- tion to detect communities. They first attach weights to the edges in a social network based on the check-in information, then the social network is modified by removing all edges with small weights. In the end, a community detection al- gorithm (louvain method[2]) is used on the modified social graph to discover communities. The experimental results show that their method is able to discover more meaningful communities, such as place-focused communities, compared to the standard community detection algorithm. More recently, Brown et al. [3] analyze mobility behav- iors of pairs of friends and groups of friends (communities). They focus on comparing the difference between individual mobility and group mobility. For example, they discover that a user is more likely to meet a friend at a place where they have not visited before; while he will choose a familiar place when meeting a group of friends. 7. CONCLUSION AND FUTURE WORK In this paper, we have studied the community impact on user's mobility. Analysis leads us to several important con- clusions: (1) communities have a stronger impact on users' mobility; (2) each user is only influenced by a small number of communities; and (3) different communities have influ- ences on mobility under different spatial and temporal con- texts. Based on these, we use machine learning techniques to predict users' future locations focusing on community in- formation. The experimental results on two types of real-life social network datasets are consistent with our analysis and show that our prediction model is very effective. The scripts for conducting the analysis and experiments as well as the Twitter dataset are available upon request.11 In the future, we plan to extend our work in several di- rections. First, we have shown in this paper that commu- nities can be exploited to achieve a promising location pre- diction. We are also interested in extending our work to other applications such as location recommendation. It is possible to redesign the cost function in matrix factoriza- tion based methods for location recommendation by taking into account community information. Second, we would like to conduct the analysis of community impact on other social behaviors such as information sharing or interests adoption. Third, in a broader point of view, our current work is ac- tually a demonstration of the communities' effect on human behaviors. As pointed by [39], community is the most mean- ingful resolution to study social network. Therefore, we also plan to investigate a user's role in his social network based on the structure of his communities. 8. REFERENCES [1] L. Backstrom, E. Sun, and C. Marlow. Find me if you can: improving geographical prediction with social and spatial proximity. In Proc. 19th International Conference on World Wide Web (WWW), pages 61–70. ACM, 2010. 11 Preliminary results of this work are reported as a poster [27]. [2] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10):P10008, 2008. [3] C. Brown, N. Lathia, C. Mascolo, A. Noulas, and V. Blondel. Group colocation behavior in technological social networks. PLoS ONE, 9(8):e105816, 2014. [4] C. Brown, V. Nicosia, S. Scellato, A. Noulas, and C. Mascolo. The importance of being placefriends: discovering location-focused online communities. In Proc. ACM Workshop on Online Social Networks (WOSN), pages 31–36. ACM, 2012. [5] S. Castles, M. J. Miller, and G. Ammendola. The Age of Migration: International Population Movements in the Modern World. Taylor & Francis, 2005. [6] J. Chang and E. Sun. Location3: How users share and respond to location-based data on social networking sites. In Proc. 5th AAAI Conference on Weblogs and Social Media (ICWSM), pages 74–80. The AAAI Press, 2011. [7] X. Chen, J. Pang, and R. Xue. Constructing and comparing user mobility profiles for location-based services. In Proc. 28th ACM Symposium on Applied Computing (SAC), pages 261–266. ACM, 2013. [8] X. Chen, J. Pang, and R. Xue. Constructing and comparing user mobility profiles. ACM Transactions on the Web, 8(4):article 21, 2014. [9] Z. Cheng, J. Caverlee, K. Lee, and D. Z. Sui. Exploring millions of footprints in location sharing services. In Proc. 5th AAAI Conference on Weblogs and Social Media (ICWSM), pages 81–88. The AAAI Press, 2011. [10] E. Cho, S. A. Myers, and J. Leskovec. Friendship and mobility: user movement in location-based social networks. In Proc. 17th ACM Conference on Knowledge Discovery and Data Mining (KDD), pages 1082–1090. ACM, 2011. [11] A. Clauset, M. E. Newman, and C. Moore. Finding community structure in very large networks. Physical Review E, 70(6):066111, 2004. [12] D. J. Crandalla, L. Backstrom, D. Cosley, S. Suri, D. Huttenlocher, and J. Kleinberg. Inferring social ties from geographic coincidences. Proceedings of the National Academy of Sciences, 107(52):22436–22441, 2010. [13] J. Cranshaw, E. Toch, J. Hone, A. Kittur, and N. Sadeh. Bridging the gap between physical location and online social networks. In Proc. 12th ACM International Conference on Ubiquitous Computing (UbiComp), pages 119–128. ACM, 2010. [14] H. Gao, J. Tang, X. Hu, and H. Liu. Exploring temporal effects for location recommendation on location-based social networks. In Proc. 7th ACM Conference on Recommender Systems (RecSys), pages 93–100. ACM, 2013. [15] H. Gao, J. Tang, and H. Liu. Exploring social-historical ties on location-based social networks. In Proc. 6th AAAI Conference on Weblogs and Social Media (ICWSM), pages 114–121. The AAAI Press, 2012. [16] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi. Understanding individual human mobility patterns. Nature, 453(7196):779–782, 2008. [17] D. Jurgens. That's what friends are for: Inferring location in online social media platforms based on social relationships. In Proc. 7th AAAI Conference on Weblogs and Social Media (ICWSM). The AAAI Press, 2013. [18] A. Lancichinetti and S. Fortunato. Community detection algorithms: a comparative analysis. CoRR, abs/0908.1062, 2010. [19] Q. Li, Y. Zheng, X. Xie, Y. Chen, W. Liu, and W.-Y. Ma. Mining user similarity based on location history. In Proc. 16th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (GIS), page 34. ACM, 2008. [20] B. Liu and H. Xiong. Point-of-interest recommendation in location based social networks with topic and location awareness. In Proc. 13th SIAM International Conference on Data Mining (SDM), pages 396–404. SIAM, 2013. [21] E. L. Martelot and C. Hankin. Fast multi-scale detection of relevant communities in large-scale networks. The Computer Journal, 56(9):1136–1150, 2013. [22] J. J. McAuley and J. Leskovec. Learning to discover social circles in ego networks. In Proc. 26th Annual Conference on Neural Information Processing Systems (NIPS), pages 548–556. NIPS, 2012. [23] J. J. McAuley and J. Leskovec. Discovering social circles in ego networks. ACM Transactions on Knowledge Discovery from Data, 8(1):article 4, 2014. [24] J. McGee, J. Caverlee, and Z. Cheng. Location prediction in social media based on tie strength. In Proc. 22nd ACM International Conference on Information & Knowledge Management (CIKM), pages 459–468. ACM, 2013. [25] M. E. Newman. Modularity and community structure in networks. Proceedings of the National Academy of Sciences, 103(23):8577–8582, 2006. [26] S. Nilizadeh, A. Kapadia, and Y.-Y. Ahn. Community-enhanced de-anonymization of online social networks. In Proc. 21st ACM Conference on Computer and Communications Security (CCS), pages 537–548. ACM, 2014. [27] J. Pang and Y. Zhang. Exploring communities for effective location prediction (poster paper). In Proc. 24th World Wide Web Conference (Companion Volume) (WWW), pages 87–88. ACM, 2015. [28] H. Pham, C. Shahabi, and Y. Liu. EBM: an entropy-based model to infer social strength from spatiotemporal data. In Proc. 2013 ACM International Conference on Management of Data (SIGMOD), pages 265–276. ACM, 2013. [29] D. Quercia, R. Schifanella, L. M. Aiello, and K. McLean. Smelly maps: the digital life of urban smellscapes. In Proc. 9th AAAI Conference on Weblogs and Social Media (ICWSM), pages 237–236. The AAAI Press, 2015. [30] A. R´eny. On measures of information and entropy. In Proc. 4th Berkeley Symposium on Mathematics, Statistics and Probability, pages 547–561, 1960. [31] M. Rosvall and C. T. Bergstrom. Maps of information flow reveal community structure in complex networks. Proceedings of the National Academy of Sciences, 105(4):1118–1123, 2008. [32] M. Rosvall and C. T. Bergstrom. Multilevel compression of random walks on networks reveals hierarchical organization in large integrated systems. PLoS ONE, 6(4):e18209, 2011. [33] A. Sadilek, H. Kautz, and J. P. Bigham. Finding your friends and following them to where you are. In Proc. 5th ACM International Conference on Web Search and Data Mining (WSDM), pages 459–468. ACM, 2012. [34] S. Scellato, A. Noulas, R. Lambiotte, and C. Mascolo. Socio-spatial properties of online location-based social networks. In Proc. 5th AAAI Conference on Weblogs and Social Media (ICWSM), pages 329–336. The AAAI Press, 2011. [35] R. Shokri, G. Theodorakopoulos, J.-Y. L. Boudec, and J.-P. Hubaux. Quantifying location privacy. In Proc. 32nd IEEE Symposium on Security and Privacy (S&P). IEEE CS, 2011. [36] K. Wakita and T. Tsurumi. Finding community structure in mega-scale social networks (extended abstract). In Proc. 16th International Conference on World Wide Web (WWW), pages 1275–1276. ACM, 2007. [37] J. Yang and J. Leskovec. Overlapping community detection at scale: a nonnegative matrix factorization approach. In Proc. 6th ACM International Conference on Web Search and Data Mining (WSDM), pages 587–596. ACM, 2013. [38] J. Yang, J. J. McAuley, and J. Leskovec. Community detection in networks with node attributes. In Proc. 13th IEEE International Conference on Data Mining (ICDM), pages 1151–1156. IEEE CS, 2013. [39] J. Yang, J. J. McAuley, and J. Leskovec. Detecting cohesive and 2-mode communities indirected and undirected networks. In Proc. 7th ACM International Conference on Web Search and Data Mining (WSDM), pages 323–332. ACM, 2014. [40] Y. Zhang and J. Pang. Distance and friendship: A distance-based model for link prediction in social networks. In Proc. 17th Asia-Pacific Web Conference (APWeb), LNCS. Springer, 2015. To appear. [41] V. W. Zheng, Y. Zheng, X. Xie, and Q. Yang. Collaborative location and activity recommendations with GPS history data. In Proc. 19th International Conference on World Wide Web (WWW), pages 1029–1038. ACM, 2010. [42] Y. Zheng, F. Liu, and H.-P. Hsieh. U-air: when urban air quality inference meets big data. In Proc. 19th ACM Conference on Knowledge Discovery and Data Mining (KDD), pages 1436–1444. ACM, 2013. [43] Y. Zheng, L. Zhang, Z. Ma, X. Xie, and W.-Y. Ma. Recommending friends and locations based on individual location history. ACM Transactions on the Web, 5(1), 2011. [44] Y. Zheng, L. Zhang, X. Xie, and W.-Y. Ma. Mining interesting locations and travel sequences from GPS trajectories. In Proc. 18th International Conference on World Wide Web (WWW), pages 791–800. ACM, 2009.
1909.02346	2	1909	2019-09-08T02:26:27	A complex network approach to political analysis: application to the Brazilian Chamber of Deputies	[ "cs.SI", "physics.soc-ph" ]	In this paper, we introduce a network-based methodology to study how clusters represented by political entities evolve over time. We constructed networks of voting data from the Brazilian Chamber of Deputies, where deputies are nodes and edges are represented by voting similarity among deputies. The Brazilian Chamber of deputies is characterized by a multi-party political system. Thus, we would expect a broad spectrum of ideas to be represented. Our results, however, revealed that plurality of ideas is not present at all: the effective number of communities representing ideas based on agreement/disagreement in propositions is about 3 over the entire studied time span. The obtained results also revealed different patterns of coalitions between distinct parties. Finally, we also found signs of early party isolation before presidential impeachment proceedings effectively started. We believe that the proposed framework could be used to complement the study of political dynamics and even applied in similar social networks where individuals are organized in a complex manner.	cs.SI	cs	A complex network approach to political analysis: application to APS/123-QED the Brazilian Chamber of Deputies Ana C. M. Brito Institute of Mathematics and Computer Science, University of Sao Paulo, Sao Carlos, SP, Brazil Filipi N. Silva Indiana University Network Science Institute, Bloomington, Indiana 47408, USA Diego R. Amancio∗ Institute of Mathematics and Computer Science, Department of Computer Science, University of Sao Paulo, Sao Carlos, SP, Brazil (Dated: September 10, 2019) Abstract In this paper, we introduce a network-based methodology to study how clusters represented by political entities evolve over time. We constructed networks of voting data from the Brazilian Chamber of Deputies, where deputies are nodes and edges are represented by voting similarity among deputies. The Brazilian Chamber of deputies is characterized by a multi-party political system. Thus, we would expect a broad spectrum of ideas to be represented. Our results, however, revealed that plurality of ideas is not present at all: the effective number of communities represent- ing ideas based on agreement/disagreement in propositions is about 3 over the entire studied time span. The obtained results also revealed different patterns of coalitions between distinct parties. Finally, we also found signs of early party isolation before presidential impeachment proceedings effectively started. We believe that the proposed framework could be used to complement the study of political dynamics and even applied in similar social networks where individuals are organized in a complex manner. ∗ [email protected] 1 I. INTRODUCTION In recent years, the availability of information in the form of open datasets together with the capabilities to store and process data have been promoting the rise of many new studies in a variety of disciplines, including Biology [1], Social Sciences [2], Linguistics [3 -- 6] and Physics [7, 8]. Many of these systems can be regarded as being too complex for traditional methodologies in which each of its components is isolated and studied individually. This includes the political dynamics of a country, which may depends on many aspects such as economic factors, culture, mass media, social media, etc. In a democratic system, these aspects are reflected (or should be) on the decisions made by people's representatives and on how they are organized (such as partisanship or alliances between parties). In this context, the relationships between politicians could be drawn indirectly from how they vote in propositions [9 -- 11], i.e., their agreement or disagreement among a set of voted proposals. Politicians with similar voting patterns can be understood as having similar views and interests, thus can be connected in a political network. Being a complex system, it makes sense to explore such a system by using methods borrowed from network science [12]. The Brazilian political environment becomes an interesting system to be studied under this approach since it is a multi-party system, thus, supposedly presenting a diverse spectrum of political ideas. Also, it is a system that has undergone many changes over the past few years. Moreover, data for decisions and political organization are openly available. In this paper, we propose a framework based on complex networks to study the evolution of the Brazilian political system in terms of how politicians vote in proposals along time. For that, we collected and processed voting data from 1991 to 2018, spanning 6 terms in the lower chamber and build vote correlation networks. Because the obtained networks were weighted and dense, we applied a filtering method to preserve the community structure. Using concepts borrowed from network science, we defined relevant political concepts, in- cluding coalition, fragmentation and isolation. We also devised a measure to quantify the effective number of groups, regardless of the number of political parties. Several interesting results were obtained in the Brazilian political scenario. We found in most of years a low correspondence between communities and topological groups, meaning that the majority of deputies tend to share votes with different parties. Most importantly, despite the large number of political parties in the Brazilian political scenario (29 parties in 2 2019), effectively only roughly 3 groups of deputies were identified. The quantification and visualization of shared coalitions and isolation were consistent with the political scenario unfolded along the considered period. We also found different behavior of coalitions: while some parties tend to be aligned to the president party, other parties are always in distinct network communities. Finally, we identified an isolation pattern that could be an early indication of the start of an impeachment process. While we focused on the analysis of the Brazilian case, we advocate that the methods are robust enough to be applied in other democratic systems and even in other social networks. The paper is structured as follows. In Section II, we review some recent works using a graph representation to study political systems. In Section III A, we detail the creation of networks from voting data. The procedure employed to prune the obtained weighted, dense networks is presented in Section III B. Section III C presents a discussion on how to effec- tively measure the quality and quantity of clusters identified in networks. In Section III D, we introduce some measurements to quantify the concepts of political coalition, isolation and fragmentation. Finally, results and future works are discussed in Sections V and VI, respectively. II. RELATED WORKS Some studies used the concept of complex networks to model some political concepts, including the concept of political coalition and cohesion [10, 11, 13, 14]. Both concepts are quantified via networks extracted from parliamentary votes. In [10], network edges are established according to the similarity of votes between par- liamentarians. Votes were considered to be chosen among three options. Deputies can vote for and against a specific proposition. In addition, they may abstain from voting. All data were extracted from the Italian parliament in 2013. A topological analysis to quantify the concepts of cohesion, polarization and government influence was created in terms of network parameters. The cohesion of political parties was defined using the following two density (cid:88) (cid:88) i∈C (cid:88) (cid:88) j∈C i∈C j(cid:54)∈C indexes: ρint(C) = ρext(C) = 1 nc(nc − 1) 1 nc(nc − 1) 3 wij, wij, (1) (2) where nc is the number of node in cluster (community) C, and wij is the weight of the edge linking nodes i and j. A given community C is considered cohesive if the internal density ρint(C) is significantly higher than the average internal density of the whole network ((cid:104)ρint(cid:105)). In a similar fashion, C may also be considered cohesive whenever ρext(C) > (cid:104)ρext(cid:105). In [10], network communities and relevant parliamentarians are identified by measuring the effect of removing and inserting nodes in distinct network communities. More specifically, the authors measured the variation in the modularity dQ [15] when parliamentarians with distinct political positions are placed in their respective opposite political communities. The degree of polarization was then defined as proportional to the absolute variation dQ resulting from this procedure. The cohesion defined in equations 1 and 2 revealed that some parties are well defined, while some parliamentarians display a voting pattern that is most similar to the patterns observed in other parties. Upon analyzing the modularity of the networks, the authors advocate that it may not be the best alternative to characterize the obtained clustered networks. Their analysis also showed that the emergence of three main clusters, where two groups represent parliamentarians voting according to the government preferences, while the remaining cluster was found to represent opposition parties. Finally, a temporal analysis revealed patterns of voting behavior along time. Such a temporal analysis was useful to identify e.g. parties do not display a consistent voting behavior. In [13], a dataset of propositions analyzed between 2011 and 2016 was used to analyze votes in the Brazilian political scenario. Nodes were defined as deputies, which are linked by the number of identical votes in the considered period. Different variations in the network construction were considered because abstentions may lead to different interpretations on whether two deputies agree when at least one of them abstain from voting. This study also defined political concepts such as polarization in social networks using correlation cluster- ing [16] and its symmetric relaxed version to quantify polarization. The analysis carried out in [13] revealed that several parties do not comply with their respective original ideas as expressed during the campaign period. In addition, this study showed that, in recent years, the group of deputies allied to the Brazilian government di- minished. Even after presidential reelection, allied deputies turned out to display a voting pattern in disagreement with government propositions. Finally, this study reports that the distance between centre- and right-wing deputies decreased in recent years. 4 The research conducted in [14] studied the network community organization of parliamen- tarians, focusing on their temporal evolution. The data were obtained from both Brazilian Chamber of Deputies and US Parliament from 2003 to 2017. Temporal networks were cre- ated without overlapping. In the network representation, parliamentarians are nodes and edge weights are established according to the vote similarity. In the Brazilian case, consid- ered two additional possibilities for each vote: (i) abstention; and (ii) obstruction. In the latter, it may also be considered as a vote against the proposition under discussion. The authors proposed a measurement that depends on the community structure of the network. The Partisan Discipline, computed for a given parliamentarian m, was defined as N(cid:88) i=1 pd(m) = 1 N I(m, pm, i), (3) where N is the total number of propositions analyzed and I(m, pm, i) = 1 iff m and the respective group (i.e. network community or party) to which he/she belongs voted equally in the i-th proposition. The application of equation 3 for the data obtained from the Brazilian Chamber of Deputies revealed that the values of Partisan Discipline are higher when groups are the identified communities, in comparison to groups being traditional parties. For the US Par- liament, the obtained values of Partisan Discipline are similar when groups and communi- ties are considered. The concept of polarization was analyzed by observing that cohesive communities emerges in a political scenario characterized by a high degree of polarization. Because members of the same community should have similar sets of neighbors, the number of shared neighbors was used as an indicative of political polarization. The polarization analysis revealed that the Brazilian Chamber of Deputies has a low degree of polarization. In addition, this study showed that deputies oftentimes do not stay in the same dense, well-defined community for long periods. Some studies used political networks to identify the ideology of political parties. While many studies focused on voting data, the study reported in [17] focused on party-switching affiliations from one election to another to generate networks. In such a network, nodes are parties and two parties are linked if there was a migration of deputies between these parties. The community structure of the obtained networks revealed a tendency of deputies to switch between parties with similar ideology. A model was also created to quantify the ideology of political parties on the left-right scale. The methodology was found to be robust as it does 5 not depend on the level of representation of parties, since only affiliation data is required to create political networks. The European Parliament was studied using networks in the work conducted in [11]. A dataset with roll-call voting data from the European Parliament was used to create a network where parliamentarians are nodes and edges are weights according to the similarity of voting patterns. Twitter [18] was also used to create a network where parliamentarians are connected if there is a "retweet" relationship between them. The main results revealed the emergence of political coalitions, characterized by cooperation of opposite political wing groups. In addition, they found a tendency of cooperation among small and large political groups. Finally, this study also reports that political alignments in Twitter are consistent with voting patterns. Unlike the above mentioned studies, we propose metrics based on shortest path lengths which can be understood as the connection strength between parliament members. We also use temporal series to show our results, enabling thus the visualization of cluster dynamics. Differently from other works, we compare the traditional structure of political parties with the actual structure observed from voting behavior. Our focus is on the characterization of groups, which encompasses the definition of political fragmentation, isolation and diversity of parties. III. METHODOLOGY The proposed framework to analyze the relationship among deputies according to their voting patterns can be summarized in the following steps: 1. Network construction: a network is created from voting patterns in a given period. Two deputies are linked if they vote in a similar fashion in several propositions. 2. Network backbone extraction: this step involves the removal of the weakest edges. This is essential to create a network that can be treated by most of the traditional community detection methods. Mostly important, the removal of the weakest edges does not affect the community structure of the obtained networks. 3. Cluster detection and analysis: community detection methods are employed to identify groups of deputies displaying similar voting patterns. 6 FIG. 1. Schematic representing the methodology in this paper. In (a), the voting data is analyzed. In (b), the pairwise similarity of votes in distinct propositions generates a weighted network of deputies. Edges weight represent the distance between two deputies, according to their voting behavior. In (c), edges are filtered so as to preserve the community structure. In (d), we identify the community structure of the network. Finally, in (e), the resulting network is analyzed by measuring e.g. distance between parties or clusters of deputies. 4. Network characterization: metrics are extracted from the network in order to quantify political concepts such as fragmentation, isolation and effective number of groups. These four steps are detailed, respectively, in subsections III A, III B, III C and III D. All involved steps are illustrated in Figure 1. A. Network construction Networks are constructed by representing deputies as nodes. Edges are established be- tween two deputies if the percentage of agreement considering all propositions in the dataset is higher than the percentage of disagreement. Mathematically, all votes established by the i-th deputy can be stored in a N -dimensional vector v(i), where N is the total number of considered propositions. The j-th element of v may store three different values. vj may take the values −1 or +1, representing respectively a vote for or against a given proposition. 7 D1D2D3DND1D2D3DND1D2D3DND1D2D3DNYESNOYESNOYESNOYESABSTAINYESNOYESNOYESNOYESYESYESNONOYESABSTAINYESNONOYESYESNOABSTAINABSTAINABSTAINNONODeputies' voting historymoving windowTimePairwise ScoreBackbone ExtractionNetwork CharacterizationCluster Detection(a)(b)(c)(d)(e)TimeTimeThus, the pairwise level of agreement between two deputies is defined as: N(cid:88) k=1 wij = 1 N v(i) k v(j) k = v(i) · v(j) N , (4) where v(i) · v(j) is the dot product between v(i) and v(j). Note that, when deputies have the same opinion about the proposition being voted, the level of agreement is increased by 1. Conversely, when views are opposite, wij is decreased by 1. Whenever at least one of the deputies absence from voting in a proposition, that proposition is not considered when computing wij. Because edges weights are given by wij, all obtained networks are undirected and weighted. As we shall show, the networks may be constructed using a particular time interval, and also considering sliding (overlapping) windows of fixed length. B. Network backbone extraction Even considering only agreement between deputies as connections, the proposed approach can still lead to densely connected networks. While such networks can still be regarded as a weighted, it is usually desirable to remove noise from data, such as weak connections that may correspond to spurious links. For that reason, we included a preprocessing step to our analysis pipeline in which we deal with the problem of keeping only important connections in a network. Such a task is known as edge pruning [19]. A simple way to accomplish edge pruning is keeping only edges with weights larger than a fixed threshold. Such a threshold can be determined, for instance, in terms of percentiles to reach a certain network density. The main problem with that approach is that it may introduce some bias toward highly connected nodes which can lead to substantial undesired effects on the network topology. A more sophisticated approach to the edge pruning task is the backbone extraction method [20]. This technique is based on the idea that when determining the importance of an edge, we must take into account how its ending points are connected to the network. More specifically, the authors proposed that the importance of a node can be determined in terms of a disparity filter, so that a value αij for an edge can be drawn from a p-value determined in terms of null model. This null model considers the probability of a node having an edge with a certain weight based on its other connections. In practice, one can 8 compute αij for a certain edge eij existing in the network as: αij = 1 − (ki − 1) (1 − x)ki−2dx, (cid:90) pij 0 where ki is the number of connections for node i and pij is a probability defined as wij(cid:80) . ik ∈ E wik pij = (5) (6) The resulting network is obtained by removing all edges with αij higher than a certain threshold α. Because the networks under analysis are undirected, we consider, for each edge eij, the lowest between αij and αji. Here, we opted no to use a fixed threshold for α. We instead used the smallest α such that each network have at least 80% of the nodes belonging to the giant component. This results in more consistent networks along time and allow us to analyze the data based on the major component. This is important in network analysis because some measurements, such as shortest path distances, are only measured in a connected component. In order to illustrate the edge pruning process, we show in Figure 2 a visualization of a network before and after the pruning process is applied. Interestingly, there is a clear separation in communities of deputies with very distinct voting patterns. The emergence of several communities is not as clear when the filtering procedure is not applied. C. Cluster detection and analysis Once the network is obtained, the clustering of nodes is necessary to understand how deputies are clustered by taking into account voting patterns. The identification of com- munities is also important to analyze if the organization of deputies in parties is consistent with the natural organization emerging from their voting behavior. Several community detection methods rely on the modularity to measure the quality of the obtained clusters in complex networks [22 -- 24]. This measurement quantifies the number of links that occur inside a community that are above the same number expected in a null model. Mathematically, the network modularity Q is given by δ(ci, cj), (7) (cid:34) (cid:88) ij Q = 1 2m (cid:35) aij − kikj 2m 9 FIG. 2. Network obtained considered all proposition voted in 2012. This network was constructed using the procedure described in (a) Section III A (before edge pruning); and (b) Section III B (after edge pruning). Each node represents a member from the Brazilian lower chamber. Two members are linked if they voted in a similar way. Colors represent nodes degree. This visualization was created using the Networks3D visualizer software [21]. where aij = 1 if nodes i and j are connected and aij = 0, otherwise. m = 1/2(cid:80) total number of edges in the network and ki =(cid:80) ij aij is the j aij. δ(ci, cj) = 1 only if nodes i and j are in the same community. Otherwise, δ(ci, cj) = 0. In this particular study, we used the Leiden method [25], which is an improved version of the Louvain algorithm [22]. While we report our results based on this Leiden method, a preliminary study revealed that the choice of community detection method has no influence on the results reported in this paper. The Louvain algorithm is based on the iterative repetition of two steps. First, each single node is considered as a single community in the network. Then, for each node, the method evaluates the gain in modularity when it is removed from its current community and placed to a neighbor community. The change of community membership is effectively performed if such a change corresponds to the highest gain in modularity among all possible changes in the current iteration. If the highest gain is negative (i.e. the change of community membership 10 (a)(b)leads to a decrease in modularity), then the node stays in its original community. This process is repeated until no gain in modularity is achieved. The efficiency of the method stems from the fact that the gain of modularity can be computed in a efficient way. The algorithm proposed in [25] guarantees that the obtained communities are connected. One particular feature that we are interested in the study of political networks is the correspondence between political parties and network communities. Such an analysis can be performed by comparing the total number of clusters obtained from these two partitioning strategies. A more refined notion group diversity (i.e. total number of groups) in a partition -- obtained via political parties or community detection -- may be derived from the concept of true diversity [26]. This quantity measures the effective number of groups in a partition by considering the size of groups. Let S = {s1, s2, . . . sR} be the size of each group in a partition comprising R groups. The computation of the true diversity requires that the sizes in S are normalized. For this reason, they are normalized to yield the following normalized j sj. The true diversity D(q) of set of normalized sizes Π = {π1, π2, . . . πR}, where πi = si/(cid:80) (cid:33)1/(1−q) the distribution Π is defined as D(q) = πq i , (8) (cid:32) R(cid:88) i=1 where q is the exponent given to the weighted generalized mean of πi's. If q = 1, the true diversity can be computed as (cid:33)1/(1−q) (cid:32) R(cid:88) i=1 D(q=1) = lim q→1 πq i = 1(cid:81)R i=1 ππi i = exp (cid:32) − R(cid:88) i=1 (cid:33) πi ln πi , (9) which corresponds to the exponential of the entropy of Π. The true diversity has been employed in other contexts to measure the effective number of species [27]. In network theory, the accessibility of nodes in complex networks is measured in terms of the true diversity and quantify the effective number of neighbors [28 -- 32]. Because we are measuring the diversity of clusters size, hereafter we refer to the true diversity as the effective number of political parties and communities. D. Network characterization In this section, we introduce the measurements that are used to analyze network obtained from voting patterns. We are specially interested in the analysis of distances among parties 11 along time. The first important measurement that can be extracted from political networks is the topological distance d(A, B) between two groups A and B, defined as the average distance among their elements: d(A, B) = 1 A × B (cid:88) (a,b)∈A×B l(a, b). (10) where l(a, b) is the shortest path length between a ∈ A and b ∈ B. When groups are political parties, the distance d(A, B) can be used to quantify the level of coalition between A and B. In this case, low values of distances correspond to strong coalitions. Another important quantity is the average distance dG between all nodes in the network. It is defined as: dG = 2 n(n − 1) (cid:88) a(cid:54)=b l(a, b). (11) Note that the distance l can not be computed directly from the network obtained as described in Section III B, because edge weights represent a similarity relationship. In order to map the similarity wij defined in equation 4 into a dissimilarity index, we used the following equation: ∆(wij) =(cid:2)2 · (1 − wij)(cid:3)1/2. (12) A more detailed description on network transformations, such as the one in equation 12, can be found in [33]. The political isolation of group A, I(A), is defined as the distance between group A and all other groups in a given division of the network. It is defined as (cid:88) X(cid:54)=A I(A) = X · d(A, X) (cid:14)(cid:88) X(cid:54)=A X, (13) which represents the average distance (as defined in equation 10) between A and all other groups. Note that the distances are weighted by the size of each group X (cid:54)= A. According to equation 13, the isolation reflects how distant nodes belonging to the group of reference are from the other nodes in the network. Interestingly, there is an analogy between the isolation of groups and the ability (complexity) of unsupervised classification [34]. In other words, isolated groups are analogous to outlier clusters in unsupervised data analysis. The concept of political fragmentation for group A, F (A), can be defined in terms of intra-group dispersion, i.e. F (A) = d(A, A). (14) 12 Therefore, a group is highly fragmented whenever the nodes in that group are distant from each other, according to the topology of the network. Note that the quantities defined in equations 13 and 14 are independent. High-values of isolation do not imply low values of fragmentation and vice-versa. IV. DATASET The voting data were obtained from the Brazilian Chamber of Deputies [35]. The dataset comprises a total of 3, 530 propositions voted between April, 1991 and Juny, 2019. The most frequent political parties appearing in this dataset are: Brazilian Democratic Movement (Movimento Democr´atico Brasileiro - MDB), Brazilian Social Democracy Party (Partido da Social Democracia Brasileira - PSDB), Workers' Party (Partido dos Trabalhadores - PT), Progressive Party (Partido Progressista - PP), Democrats (Democratas - DEM), Communist Party of Brazil (Partido Comunista do Brasil - PCdoB), Brazilian Socialist Party (Partido Socialista Brasileiro - PSB), Brazilian Labour Party (Partido Trabalhista Brasileiro - PTB), Liberal Party (Partido Liberal - PL), Democratic Labour Party (Partido Democr´atico Tra- balhista - PDT) and Social Liberal Party (Partido Social Liberal - PSL). It is important to note that a few of these parties have undergone splitting or merging processes leading to changes in their names over the past few years. However, for the sake of simplicity, we only consider their most recent names, as of 2019. The total number of propositions voted in each year is shown in Figure 3. In Figure 4, we show the distribution of parties (i.e. the total number of parliamentarians) along time. Interestingly, the representation of nontraditional parties have been increasing along time, while the size of traditional political parties (e.g. MDB and PSDB) is decreasing. A visualization of the obtained network over time is shown in Figure 2. V. RESULTS AND DISCUSSION Our main objective here is to leverage network science to better understand the dynamics of groups of deputies from a networked perspective. Using the proposed methodology, we intend to address the following general research questions, which are analyzed in the specific case of the Brazilian Chamber of Deputies. 13 FIG. 3. Yearly number of propositions analyzed by the Brazilian Chamber of Deputies between 1991 and 2019. While the number of propositions analyzed along time is not regular, more recently at least 100 propositions were analyzed in each year. FIG. 4. Number of seats of political parties as observed in the Brazilian Chamber of Deputies between 1991 and 2019. The number of deputies from nontraditional parties ("others") clearly increasing along time. Only the deputies in the largest component of the network were considered. 1. Is the organization in parties consistent with clusters formed via voting behavior? 2. Does the diversity of political parties correspond to the diversity of clusters? 3. Is it possible to observe and/or quantify political concepts such as fragmentation and approximation of political parties along time? 4. Is there any signal from network topology indicating major political transitions, such as in presidential impeachment? The response to the above questions is important to understand the dynamics of deputies relationships, both at the microscopic and mesoscopic level. Such an better understanding may point out to better ways to improve the democratic system of a country. In the same line, the extraction of information from unstructured date might assist the population to 14 take decisions on who they choose as representatives based on the voting history of deputies and parties alike. In order to address the above research questions, we divide our analysis in two parts. In Section V A, we analyze the global properties of the networks and compare the groups defined via topology and political parties. In Section V B, we analyze the network at the mesoscopic (party) level. A. Global analysis The global evolution of the obtained political networks is depicted in Figure 5, where colors represent political parties. An interactive visualization of these networks is available online [36]. The first important property that can be recovered from the networks is the global cohesiveness. This can be computed using a measurement that quantifies the distance between two nodes (deputies) in the network. If deputies vote in a similar fashion, one expects that the average distance between them is short. The obtained evolution of the global distances (dG in equation 11) is shown in Figure 6. All distances typically range in the interval 4 ≤ dG ≤ 6. Two patterns, however, can be identified. In the period between 1994 and 2002, distances are in the interval 4 ≤ dG ≤ 5. Conversely, in the period between 2003 and 2014, the average distances typically are typically higher, in the interval 5 ≤ dG ≤ 6. Interestingly these two intervals correspond to opposite parties occupying the presidency. This observation might indicate that different government ideologies may affect the way in which deputies vote. The consistency of political parties can be measured by probing whether groups defined via political parties are consistent with topological groups, which are obtained using only the connectivity of the network. Topological groups are obtained via modularity optimization and, for this reason, this method tends to cluster nodes in the same group if they are strongly connected in the group, with a few external links. The evolution of the quality of the clustering obtained via topology and political parties are shown in Figure 7. The modularity of groups obtained via topology outperforms the modularity of groups obtained via political parties. This means that there is a high tendency of deputies of different political parties to be connected. While the difference in modularity seems to be roughly constant along time, it is possible to observe higher deviations in particular periods. This is apparent 15 FIG. 5. Visualization of the time evolving network obtained from voting data. Each yearly snapshot was obtained considering only the proposals voted in that year. Colors correspond to parties. 16 199519981997199619992002200120002003200620052004200720102009200820112014201320122015201820172016PSDBPSDBPSDBPSDBDEMDEMDEMDEMMDBMDBMDBMDBPTPTPTPTPPPPPPPPOthersPRPRPRPRPCdoBPCdoBPCdoBPCdoBPSBPSBPSBPSBPTBPTBPTBPTBPLPLPLPLFIG. 6. Evolution of the average shortest path length. While the average distance ranges in the interval 4 ≤ dG ≤ 6, the networks obtained in the interval between 1994 and 2002 are typically more cohesive than the ones obtained between 2003 and 2012. e.g. in the years 2016 and 2017, which coincides with a period of post-impeachment political transition in Brazil. In both years, only PT was found to form an independent community, according to Figure 5. An opposite effect occurred in 2015, when the difference in modularity reached its minimum value. This is evident in Figure 5, because groups formed by parties are consistent with topological clusters. FIG. 7. Evolution of the modularity measurement. To compute the modularity, we considered groups obtained from the community detection method (blue curve) and clusters matching deputies' political parties (orange curve). The divergence between topological clusters and political parties can be further analyzed via Normalized Mutual Information (NMI) [37], which measures the quality of the obtained partitions (topological communities) based on a set of nodes membership labels (political parties). The evolution of the NMI index is shown in Figure 8. Overall, the correspondence 17 between political parties and topological partitions varies along time, but a major rupture seems to occurs just before the first PSDB term (1994) and just before impeachment (2016). Notably, the highest discrepancy (lowest value of NMI) occurred in 1996, which is in accor- dance with the visualization provided in Figure 5. This result confirms that, in the Brazilian political scenario, in most years the natural organization of deputies in communities has a low correspondence with party organization. FIG. 8. Normalized Mutual Information (NMI) between communities and political parties. A major transition affecting the NMI occurred just before the first PSDB presidency, from 1993 to 1994. The discrepancy between political parties and topological clusters is also evident when one study the actual and effective quantity of groups. The effective number of parties and communities was computed using the true diversity of the distribution of clusters size, as described in Section III C. The results are depicted in Figure 9. Since 2005, an increase in the total number of parties (green curve) is clear. In 2019, the Brazilian Chamber of Deputies is made up of roughly 29 political parties. When the effective number of parties is analyzed (blue curve), the number of parties decreases to 20 political parties. Surprisingly, the topology of the network revealed the existence of only about 5 clusters of deputies. In other words, even though 29 political parties exist, D(communities) (cid:39) 3.37 (see equation 9) in 2019. In other words, only 3.37 groups are effectively present in the Brazilian political scenario in 2019. 18 NMIFIG. 9. Evolution of clusters size and the diversity of clusters size. Even though the Brazilian Chamber of Deputies is represented by 29 political parties, the effective number of topological clusters (community diversity) in 2019 is only 3.37. B. Mesoscopic analysis The obtained networks can also shed light into the understanding of the political dynamic at the mesoscopic level. An interesting analysis concerns the quantification of the pairwise distance between political parties along time. This analysis is important, e.g. to quantify the level of coalision (or opposition) between parties and party isolation. Such concepts are oftentimes used in politics because abrupt or recurrent changes in these quantities may precede a relevant change in the political scenario. In Figure 10, we show the temporal evolution of the distances between PSDB and other parties. The distances are computed using equation 10. In this figure, we also show pres- idential terms in different colors, including the transition post-impeachment led by MDB presidency (2016 − 2018, green background). Interestingly, along the years of PSDB presi- dency, PT was found to be the most distant party from PSDB. This is consistent with the traditional sentiment of polarization between PT and PSDB in these years [38] [39]. Dur- ing PT presidency, we note that PT and PSDB were still distant, however, two additional parties (MDB and PP) were found to increase (and keep) a larger distance from PSDB. This is consistent with the real political scenario, since MDB and PT shared a coalition along the period characterized by PT presidency. These observations are consistent with the visualization provided in Figure 5. Another interesting pattern that can be extracted from Figure 10 concerns the evolution 19 FIG. 10. Average shortest path length between PSDB deputies and other parties. Distances are computed using equation 10. The PSDB fragmentation (i.e. d(PSDB, PSDB)) is computed using equation 14. Presidential terms are highlight in the period between 1993 and the end of 2018. of PSDB fragmentation (i.e. F (PSDB) = d(PSDB, PSDB)), as defined in equation 14. The PSDB fragmentation corresponds to the blue curve. Note that the typical distance between PSDB deputies is compatible with the distance between PSDB and DEM virtually along all considered years and, markedly, before 2002. Such a proximity between these two parties is also evident from the visualization of the network structure. A strong concordance with MDB and PP is also apparent before 2002 and after 2016. In Figure 11, we show the distances between MDB and other parties. In this case, the transition in government seems to play a prominent role in the behavior of this party. Between 1991 and 2001, the only party distant from MDB is PT. Conversely, between 2002 and 2011, PT and MDB became much closer. This is a consequence of the the political coalition shared by these parties along these years. Differently from the traditional alliance in previous years, MDB dissociated itself from PSDB and DEM between 2002 and 2016 (see also Figure 5). The rapprochement between these parties took place only in 2016, in the post-impeachment transition. The MDB fragmentation along time is roughly constant, however, apart from 2012-2015, the fragmentation is always similar to the typical distance between MDB and the closest parties. Finally, it is interesting to note that MDB and PP display a similar evolution of distances, according to Figures 11 and 12, respectively. The distance between PT and other political parties is depicted in Figure 13. It is clear from the data the PT and PSDB are continuously distant from each other since 1995. During 20 MDBChamber of deputies and presidential termsPSDBPSDBPTPTPTPTMDBinterim presidential termFIG. 11. Average shortest path length between MDB deputies and other parties. Distances are computed using equation 10. MDB fragmentation (i.e. d(MDB, MDB) is computed using equation 14. FIG. 12. Average shortest path length between PP deputies and other parties. Distances are computed using equation 10. The PP fragmentation (i.e. d(PP, PP) is computed using equation 14. Note that the distances behavior is similar with the ones found for MDB (see Figure 11). PSDB presidency, the isolation of PT is clear: no other party is closer to PT than PT itself. In the first years of PT presidency, PT became closer to other parties, including MDB and PP. Some years before a new transition in governments, however, PT became distant from all others, specially after 2015. This isolation scenario became consistent even after the presidential impeachment (2016). In Figure 14, we analyze the isolation of political parties few months before and after the impeachment of the Brazilian President, Dilma Rousseff (PT). PT and MDB terms are represented in red and green background colors, respectively. The transition between these 21 Chamber of deputies and presidential termsinterim presidential termMDBPSDBPSDBPTPTPTPTMDBMDBChamber of deputies and presidential termsPSDBPSDBPTPTPTPTMDBFIG. 13. Average shortest path length between PT deputies and other parties. Distances are computed using equation 10. The PT fragmentation (i.e. d(PT, PT)) is computed using equation 14. two presidencies is shown in gray background color. Apart from PT, all parties have similar values and patterns of isolation along time. While PT isolation was compatible with many the behavior displayed by many other parties, its isolation increased (relatively to other parties) months before impeachment. During the first months of the interim presidential term, PT isolation increased relatively to other parties. After 2017, PT turned out to be the most isolated political party. PT isolation is also clear in the visualization provided in Figure 5, specially in 2017. A sudden change in relative isolation months before impeachment could indicate the vulnerability of the party, and potentially a signal of a major change in the political scenario. VI. CONCLUSION In a democratic nation, the political organization can reflect its cultural context, eco- nomic and social interactions. The understanding of the role of deputies and parties alike is essential to improve any democratic system. Here we proposed a framework to assist the understanding of the dynamics of political parties. The relationship between deputies in the lower chamber was represented as a network, where two deputies are linked if they voted in a similar way in different propositions. Using concepts borrowed from network science, we defined politics concepts such as isolation, fragmentation and coalition, which were measured in terms of topological distances. 22 MDBChamber of deputies and presidential termsPSDBPSDBPTPTPTPTMDBinterim presidential termFIG. 14. Isolation of political parties few months before and after impeachment. The isolation of parties is computed using equation 13. PT turned out to be the most isolated party a few months before impeachment. The relative isolation became stronger after 2017. Several interesting results were found when the proposed framework was applied to an- alyze the Brazilian Chamber of Deputies. During PSDB term, PT was found to be the most distant party from PSDB and vice-versa. The coalition shared by parties was easily identified from both distance metrics and the visualization of the obtained networks. The networks displayed a modular topology, however, we found a low degree of consistency when comparing topological groups and political parties in most years. Surprisingly, we found that even though the Brazilian political scenario is highly fragmented, only a few topologi- cal groups emerge. While in 2019 the lower chamber comprised 29 distinct parties, we found that effectively there are roughly only 3 clusters of deputies, according to the diversity mea- surement. Finally, a detailed analysis in the period between 2015 and 2019 revealed that PT isolation increased a few months before the presidential impeachment of Dilma Rousseff (PT). This could be an early sign before impeachment proceedings against former president Rousseff effectively started. As future work, we intend to use additional information to analyze and possibly predict the results of votes based on the history of deputies, parties. This could be accomplished by using textual data obtained from the propositions and transcripts from deputies' discourses. While the focus of this manuscript was in identifying and quantifying relevant political quantities via network science, we advocate that this framework could also be used in other scenarios. An analogous approach could be used to analyze other types of social networks with similar characteristics, i.e. entities with pre-defined labels that interact in a complex 23 Chamber of deputies and presidential terms20192018201720162015interim presidential termPTPSLMDBmanner. This is the case of the Science of Science field [40], where authors interact via collaborations and, at the same time, they can be labeled according to diverse criteria: country, university, research field/group or others. ACKNOWLEDGMENTS DRA acknowledges financial support from Sao Paulo Research Foundation (FAPESP Grant no. 16/19069-9) and CNPq-Brazil (Grant no. 304026/2018-2). ACMB acknowledges Capes- Brazil for sponsorship. 24 [1] D. W. Huang, B. T. Sherman, and R. A. Lempicki, "Bioinformatics enrichment tools: paths toward the comprehensive functional analysis of large gene lists," Nucleic Acids Research, vol. 37, no. 1, pp. 1 -- 13, 2008. [2] A. Java, X. Song, T. Finin, and B. Tseng, "Why we twitter: understanding microblogging usage and communities," in Proceedings of the 9th WebKDD and 1st SNA-KDD 2007 workshop on Web mining and social network analysis, pp. 56 -- 65, ACM, 2007. [3] A. Acerbi, V. Lampos, P. Garnett, and R. A. Bentley, "The expression of emotions in 20th century books," PLoS ONE, vol. 8, no. 3, p. e59030, 2013. [4] Y. He and D. Zhou, "Self-training from labeled features for sentiment analysis," Information Processing & Management, vol. 47, no. 4, pp. 606 -- 616, 2011. [5] E. A. Correa Jr, A. A. Lopes, and D. R. Amancio, "Word sense disambiguation: A complex network approach," Information Sciences, vol. 442, pp. 103 -- 113, 2018. [6] J.-B. Michel, Y. K. Shen, A. P. Aiden, A. Veres, M. K. Gray, J. P. Pickett, D. Hoiberg, D. Clancy, P. Norvig, J. Orwant, et al., "Quantitative analysis of culture using millions of digitized books," Science, vol. 331, no. 6014, pp. 176 -- 182, 2011. [7] P. Baldi, P. Sadowski, and D. Whiteson, "Searching for exotic particles in high-energy physics with deep learning," Nature Communications, vol. 5, p. 4308, 2014. [8] A. D. Broido and A. Clauset, "Scale-free networks are rare," Nature Communications, vol. 10, no. 1, p. 1017, 2019. [9] J. Moody and P. J. Mucha, "Portrait of political party polarization," Network Science, vol. 1, no. 1, p. 119 -- 121, 2013. [10] C. Dal Maso, G. Pompa, M. Puliga, G. Riotta, and A. Chessa, "Voting behavior, coalitions and government strength through a complex network analysis," PLoS ONE, vol. 9, no. 12, p. e116046, 2014. [11] D. Cherepnalkoski, A. Karpf, I. Mozetic, and M. Grcar, "Cohesion and coalition formation in the european parliament: roll-call votes and twitter activities," PLoS ONE, vol. 11, no. 11, p. e0166586, 2016. [12] M. Newman, Networks: an introduction. Oxford university press, 2018. 25 [13] M. Levorato and Y. Frota, "Brazilian congress structural balance analysis," Journal of Inter- disciplinary Methodologies and Issues in Science, vol. 2, 2017. [14] C. H. G. Ferreira, B. de Sousa Matos, and J. M. Almeira, "Analyzing dynamic ideological com- munities in congressional voting networks," in International Conference on Social Informatics, pp. 257 -- 273, Springer, 2018. [15] V. De Leo, G. Santoboni, F. Cerina, M. Mureddu, L. Secchi, and A. Chessa, "Community core detection in transportation networks," Physical Review E, vol. 88, no. 4, p. 042810, 2013. [16] H. Becker, "A survey of correlation clustering," Advanced Topics in Computational Learning Theory, pp. 1 -- 10, 2005. [17] J. Faustino, H. Barbosa, E. Ribeiro, and R. Menezes, "A data-driven network approach for characterization of political parties' ideology dynamics," Applied Network Science, vol. 4, no. 1, p. 48, 2019. [18] https://twitter.com. [19] F. Zhou, S. Mahler, and H. Toivonen, "Simplification of networks by edge pruning," in Biso- ciative Knowledge Discovery, pp. 179 -- 198, Springer, 2012. [20] M. ´A. Serrano, M. Bogun´a, and A. Vespignani, "Extracting the multiscale backbone of com- plex weighted networks," Proceedings of the National Academy of Sciences, vol. 106, no. 16, pp. 6483 -- 6488, 2009. [21] F. N. Silva, D. R. Amancio, M. Bardosova, L. F. Costa, and O. N. Oliveira Jr, "Using network science and text analytics to produce surveys in a scientific topic," Journal of Informetrics, vol. 10, no. 2, pp. 487 -- 502, 2016. [22] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, "Fast unfolding of communities in large networks," Journal of Statistical Mechanics: Theory and Experiment, vol. 2008, no. 10, p. P10008, 2008. [23] H. V. Ribeiro, L. G. Alves, A. F. Martins, E. K. Lenzi, and M. Perc, "The dynamical structure of political corruption networks," Journal of Complex Networks, vol. 6, no. 6, pp. 989 -- 1003, 2018. [24] H. F. Arruda, L. F. Costa, and D. R. Amancio, "Using complex networks for text classification: Discriminating informative and imaginative documents," EPL (Europhysics Letters), vol. 113, no. 2, p. 28007, 2016. 26 [25] V. Traag, L. Waltman, and N. J. van Eck, "From louvain to leiden: guaranteeing well- connected communities," Scientific Reports, vol. 9, 2019. [26] M. O. Hill, "Diversity and evenness: a unifying notation and its consequences," Ecology, vol. 54, no. 2, pp. 427 -- 432, 1973. [27] A. Chao, C.-H. Chiu, and L. Jost, "Phylogenetic diversity measures and their decomposition: a framework based on hill numbers," Biodiversity Conservation and Phylogenetic Systematics, p. 141, 2016. [28] E. A. Correa Jr, F. N. Silva, L. d. F. Costa, and D. R. Amancio, "Patterns of authors contribution in scientific manuscripts," Journal of Informetrics, vol. 11, no. 2, pp. 498 -- 510, 2017. [29] H. F. Arruda, F. N. Silva, L. F. Costa, and D. R. Amancio, "Knowledge acquisition: A complex networks approach," Information Sciences, vol. 421, pp. 154 -- 166, 2017. [30] D. R. Amancio, "Comparing the topological properties of real and artificially generated sci- entific manuscripts," Scientometrics, vol. 105, no. 3, pp. 1763 -- 1779, 2015. [31] B. A. N. Traven¸colo and L. F. Costa, "Accessibility in complex networks," Physics Letters A, vol. 373, no. 1, pp. 89 -- 95, 2008. [32] J. V. Tohalino and D. R. Amancio, "Extractive multi-document summarization using mul- tilayer networks," Physica A: Statistical Mechanics and its Applications, vol. 503, pp. 526 -- 539, 2018. [33] C. H. Comin, T. K. Peron, F. N. Silva, D. R. Amancio, F. A. Rodrigues, and L. F. Costa, "Complex systems: features, similarity and connectivity," arXiv:1606.05400, 2016. [34] M. Z. Rodriguez, C. H. Comin, D. Casanova, O. M. Bruno, D. R. Amancio, L. F. Costa, and F. A. Rodrigues, "Clustering algorithms: A comparative approach," PLoS ONE, vol. 14, no. 1, p. e0210236, 2019. [35] https://dadosabertos.camara.leg.br. Last accessed: September, 2019. [36] https://filipinascimento.github.io/politiciannetworks. [37] P. Zhang, "Evaluating accuracy of community detection using the relative normalized mutual information," Journal of Statistical Mechanics: Theory and Experiment, vol. 2015, no. 11, p. P11006, 2015. [38] https://en.wikipedia.org/wiki/Brazilian_Social_Democracy_Party. [39] https://en.wikipedia.org/wiki/Workers%27_Party_(Brazil). 27 [40] S. Fortunato, C. T. Bergstrom, K. Borner, J. A. Evans, D. Helbing, S. Milojevi´c, A. M. Petersen, F. Radicchi, R. Sinatra, B. Uzzi, et al., "Science of science," Science, vol. 359, no. 6379, p. eaao0185, 2018. 28
1409.0205	2	1409	2015-12-29T07:55:45	Integrated structure investigation in complex networks by label propagation	[ "cs.SI", "physics.soc-ph" ]	The investigation of network structure has important significance to understand the functions of various complex networks. The communities with hierarchical and overlapping structures and the special nodes like hubs and outliers are all common structure features to the networks. Network structure investigation has attracted considerable research effort recently. However, existing studies have only partially explored the structure features. In this paper, a label propagation based integrated network structure investigation algorithm (LINSIA) is proposed. The main novelty here is that LINSIA can uncover hierarchical and overlapping communities, as well as hubs and outliers. Moreover, LINSIA can provide insight into the label propagation mechanism and propose a parameter-free solution that requires no prior knowledge. In addition, LINSIA can give out a soft-partitioning result and depict the degree of overlapping nodes belonging to each relevant community. The proposed algorithm is validated on various synthetic and real-world networks. Experimental results demonstrate that the algorithm outperforms several state-of-the-art methods.	cs.SI	cs	Integrated structure investigation in complex networks by label propagation Tao Wua,b,d,∗, Yuxiao Guoa, Leiting Chena,b,d, Yanbing Liuc aSchool of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China bInstitute of Electronic and Information Engineering in Dongguan, University of Electronic Science and Technology of China, cChongqing University of Posts and Telecommunications, Chongqing 400065, China dDigital Media Technology Key Laboratory of Sichuan Province, Chengdu 611731, China Chengdu 611731, China 5 1 0 2 c e D 9 2 ] I S . s c [ 2 v 5 0 2 0 . 9 0 4 1 : v i X r a Abstract The investigation of network structure has important significance to understand the functions of various complex networks. The communities with hierarchical and overlapping structures and the special nodes like hubs and outliers are all common structure features to the networks. Network structure investigation has attracted considerable research effort recently. However, existing studies have only partially explored the structure features. In this paper, a label propagation based integrated network structure investigation algorithm (LINSIA) is proposed. The main novelty here is that LINSIA can uncover hierarchical and overlapping communities, as well as hubs and outliers. Moreover, LINSIA can provide insight into the label propagation mechanism and propose a parameter-free solution that requires no prior knowledge. In addition, LINSIA can give out a soft-partitioning result and depict the degree of overlapping nodes belonging to each relevant community. The proposed algorithm is validated on various synthetic and real-world networks. Experimental results demonstrate that the algorithm outperforms several state-of-the-art methods. Keywords: Hierarchical and overlapping community, Hubs and outliers, Label propagation 1. Introduction Complex networks provide a way to represent complex systems of interacting objects, where nodes denote the objects and edges describe the interactions between them. For example, in ecological systems, nodes represent lives and edges represent dependencies, and in protein association networks, nodes represent proteins and edges represent physical interactions. Common sense dictates that the network structure is characterized by at least three common features: communities, hubs and outliers. Network communities are often correspond to groups of nodes that share a common property, role or function. The nodes within one community are more likely to interact to each other than to the rest of the network. Community structure allows for understanding the network functions that cannot be studied when considering only individual object or the entire system. Moreover, hubs are the nodes that bridge multiple communities, and hubs identification is crucial to network exploration. For example, hubs in epidemiology network could be the nodes that spread epidemics across groups, and immunizing the hubs could help to prevent the spread of epidemics. In addition, the identification of outliers that are marginally connected with the community members is also crucial to networks, as they can be used for abnormal nodes detection in complex networks. The network structure investigation is an important aspect of complex networks research. In the issue of network structure investigation, early work mainly focused on the community detection in naive networks (non-hierarchical and non-overlapping structures). Many classical methods have been proposed to detect the community structure of complex networks, including division methods [1,2], agglomerative methods [3,4], matrix related methods (non-negative matrix factorization [5], spectral method [6]), model-based methods (label propagation methods [7,8], mixture models [9], stochastic block models [10,11]), etc. In a word, the traditional detection problem about the community structure in naive networks is likely to be properly solved in the last decade. However, the problem of community detection becomes much harder for the networks with complex communities where hierarchical and overlapping structures emerge at the same time and tangle with each other. In real-world networks, communities are always nested, and the networks exhibit hierarchical community structure, such as the organization of a large company. Moreover, to use a social metaphor, common sense goes that people can belong to different social communities, depending on their friends, professions, hobbies, etc. In network terms, each node ∗Corresponding author Email address: [email protected] (Tao Wu) Preprint submitted to Elsevier August 21, 2018 can be shared between communities, forming overlapping communities. Thus, the networks demand methods that are able to detect complex community with hierarchical and overlapping structures. So later work paid close attention to the community detection in hierarchical and overlapping cases, which generally includes two main genres: hierarchical community detection [12-14] and overlapping community detection [15-22], although several researchers made a pioneering attempt on hierarchical and overlapping community detection [23-25]. Based on the above discussion, most of the existing methods detect community by considering the hierarchical and overlapping structures separately, and there are few integrated algorithms for hierarchical and overlapping community detection. Meanwhile, in the issue of network structure investigation, there are few methods focusing on hubs and outliers identification, although hubs and outliers are important, common features of complex net- works. Thus, network structure investigation should also take hubs and outliers identification under consideration. Therefore, how to discover the hierarchical and overlapping communities, hubs and outliers comprehensively and efficiently remains a important task to date. To better capture the network structures, an intuitive idea is that one should analyze networks in a local view. Thus the computation of the community memberships and the roles of each node can be based only on local information, which can avoid the complex organizational structures with a variety of properties. In this paper, we develop a label propagation based integrated network structure investigation algorithm (LINSIA). The basic idea behind LINSIA is that there has a network topology related equilibrium state between the nodes' community label choices, and the reasonable structure division can be derived based on the label equilibrium state. LINSIA controls the label propagation process to discover multi-scale communities at different aggregation levels, and a community on a large scale corresponds to multiple underlying communities. The main contributions include: (1) LINSIA can reveal hierarchical and overlapping communities, as well as hubs and outliers; (2) LINSIA can provide insight into the label propagation mechanism and propose a parameter-free solution that requires no prior knowledge; (3) LINSIA can give out a soft-partitioning result and depict the degree of overlapping nodes belonging to each relevant community. Experimental results demonstrate that the algorithm outperforms several state-of-the-art methods. The rest of the paper is organized as follows. Section 2 introduces some influence measures. Section 3 proposes a new label propagation strategy. Section 4 presents the LINSIA algorithm. Our algorithm is tested on a diverse set of networks in Section 5. Discussion and Conclusion are given in Section 6. 2. ENCoreness based influence measures Before we proceed, it is worthwhile to introduce some necessary definitions for LINSIA. The node influence represents node importance in the full network, and the label influence measures the popularity of each community label in label voting process. The node influence and the label influence are all regulated by an adaptive variable α. In order to compute node importance accurately, node influence should consider node's global role and its local topology information comprehensively. Thus, it exploits extended neighborhood coreness (ENCoreness) centrality [26] to measure node's global importance and includes node degree to capture node's local topology. Definition 2.1 (Node Influence). Let N I(i) indicates the influence of node i, and it is defined as N I(i) = EN Coreness(i) + X j∈N (i) EN Coreness(j) degree(j)α ∗ wi,j (1) where EN Coreness(i) is the ENCoreness centrality of node i, degree(j) is the degree of node j, wi,j is the weight degree(j)α is called of edge ei,j, N (i) is the neighbor set of node i, and α is an adaptive real variable. The term normalization factor, which is used to consider the different influences of nodes with diverse degrees. 1 For the purpose of label selection, we define node i's candidate label set LCi as the union of neighbor nodes' label sets and let Lsetj denote the label set of node j. Then, we let neighbors vote for each community label in the candidate label set LCi and use the voting support rate as the label influence. The detailed definition is presented as follows. Definition 2.2 (Label Influence). Let LIseti is the label influence set of the labels in the candidate label set LCi, and it is defined as LIseti = {LI p i LI p i = X j∈N (i),q∈Lsetj δ(p, q) ∗ N I(j) degree(j)α ∗ LI q j P m∈Lsetj LI m j ∗ wi,j , p ∈ LCi} (2) where LI p i Other notations mean the same thing as the notations in Eq. (1). The term is the influence of label p on node i, function δ(p, q) takes value 1 if p is identical to q and 0 otherwise. is the label influence ratio LI q j P m∈Lsetj LIm j 2 of label q to the all labels of node j, which denotes the degree of node j belonging to the community with label q. For the iterative updating of the label influence, all nodes are assigned a unique label initially, and they have the same initial label influence. Why do LINSIA needs the variable α? Variable α balances the cohesion inside the community and the compe- tition between communities, and it is related to the equilibrium state between nodes' community label choices. If setting a small value to variable α, the node influence and label influence of the core nodes will be much greater than that of the periphery nodes (because the ENCoreness values of the core nodes are always bigger than that of the periphery nodes). In this way, the nodes connected with a few core nodes and many periphery nodes will have same community labels as the core nodes rather than the periphery nodes. The way things are going, perhaps all nodes tend to a common label choice. Reversely, if the value of variable α is large, all nodes may tend to a unique label choice. So variable α should be determined based on the network topology by self-adaption. In LINSIA, we assume that α equals to 1.0 initially, and we compute the total sum of node influence firstly. Then we get the cumulative sum of node influence sequentially in descending order of k-shell value [27] until the cumulative sum equals to half of the total. The ratio r of the number of nodes corresponding to the cumulative sum to the total number of nodes can be viewed as an indicator of imbalance degree of influence distributed among network nodes. To the networks with completely balanced influence distribution, the ratio r takes value of 0.5. To the networks with extremely imbalanced influence distribution, we assume that at least top ten percent of nodes are needed to have half of the total influence. That is say, we assume the ratio r takes value of 0.1 in extremely imbalanced networks. So the median value of the ratio is 0.3. Therefore, the value of the variable α is depend on a function F (0.3, 1.0, r). Thus the variable α is self-adaptive and generally takes values over an interval of real number (0.4, 1.6) in networks with different influence distribution. Then LINSIA computes the normalize factor value and regulates the label selection behavior to make the label updating converge to the equilibrium state. The detailed definition is presented as follows. Definition 2.3 (Adaptive Variable). The adaptive variable α is defined as α = 1.0 + (0.3 − ′ N N )1/3 (3) where N is the total number of network nodes, and N sum in the above discussion. ′ is the number of nodes conresponding to the cumulative 3. Label propagation strategy 3.1. Label selection and hubs According to Eq. (2), in non-overlapping networks, we select the label with the greatest label influence in LIseti as node i's community label. In overlapping networks, we amplify each label influence in LIseti of node i, and select the label whose influence has the same order of magnitude with the greatest influence in LIseti as one of its labels, defined by Eq. (4). Lseti = {p log (max(LIseti))3 − 1 < log (LI p i )3 < log (max(LIseti))3, f orLI p i ∈ LIseti} (4) If node i's label set Lseti has more than one label, we define the node as hub node. The formal definition of hubs is presented as follows Hubs = {ilen(Lseti) ≥ 1, f or i ∈ V } (5) where len(Lseti) is the number of the labels of node i, and V is the node set of the network. 3.2. Label propagation strategy and outliers As nodes' label updating depend more on the core nodes' community labels than that of the periphery nodes, accurate label selection of the core nodes is crucial to efficient network structure analysis, which requires updating node labels in ascending order of node influence. Otherwise, inaccurate labels of the core nodes will propagate to the full network and disturb the other nodes' label selection. Moreover, outliers are the nodes that have negligible connection with communities, and they might not participate in any community. Namely, only the nodes with small node degree and independent community label are meaningful to outliers detection. According to the common sense, the nodes with these features are generally the periphery nodes. Therefore, the labels of the periphery nodes with low node influence should be updated preferentially to strive for the independent community label and the equilibrium 3 state between the label choices of the core nodes and the periphery nodes. Thus, we sort nodes according to their node influence values and update labels in a particular order, i.e., from a low influence to a high influence. Based on the above analysis, we can find that outliers detection will benefit greatly from updating labels in ascending order. The formal definition of outliers is presented as follows Outliers = {ilen(Lseti) == 1 & degree(i) ≤ 2 & φ(labeli, Inlseti) == 1, f or i ∈ V } (6) where len(Lseti) is the number of the labels of node i, labeli is the community label of node i, Inlseti is the set of initial labels of node i's neighbors and its own, and φ(labeli, Inlseti) equals to 1, if labeli is identical to one of the labels in Inlseti. 3.3. Participation intensity In overlapping networks, to represent the correlation degree of overlapping node to each community, the com- munity detection result and the label influence are used to define participation intensity set P IntensitySeti = {P I q i P I q i = LI q i P m∈Lseti LI m i , q ∈ Lseti} (7) where P I q i is the intensity of node i participating in the community with label q. 4. Description of the algorithm 4.1. Algorithm LINSIA In real-world networks, communities at one level are always nested into communities at an upper level, and the networks exhibit hierarchical community structure. Based on the label selection mechanism and the label propagation strategy defined in section 3, LINSIA constructs a weighted bottom-up super-network structure to detect the hierarchical relationship between the communities. The first step is to detect communities in primitive network and build a weighted super-network by replacing the communities with super-nodes. If the community information is known, it builds a weighted super-network based on the communities directly. The second step is to find non-overlapping communities in the super-network, and mapping the communities to the primitive network to merge the communities detected in the first step. LINSIA executes the last two steps iteratively until the community structure no longer changes. The construction process is illustrated by Fig. 1(a). In the super-network, the edge's weight is defined by wi,j = X m∈comi,n∈comj am,n/len(Cm) · len(Cn) Ni,j (8) where comi is the node set of the community corresponding to the super-node i, am,n is the element of the primitive network's adjacency matrix, len(Cm) is the number of communities to which node m related, and Ni,j is the number of the nodes in comi and comj that have edge connecting the communities. With the super-network structure, LINSIA can find multi-scale communities at different aggregation levels. Here a large-scale community at a high level corresponds to multiple small-scale communities at a a low level, which reflects the hierarchical relationship between the communities in real-world networks. LINSIA returns the hierarchical community structure and the best community division with the greatest modularity Q [4] or extended modularity EQ [24]. Based on the results, LINSIA gets hubs and outliers according to Eq. (5) and Eq. (6). The process of LINSIA is illustrated in Fig. 1 and the pseudo-code is presented in Algorithm 1. 4.2. Initial label influence and adaptive variable In the iterative updating of label influence, all node labels have the same initial influence. Generally, because the label updating is based on the label influence ratio, the initialized values would not influence the label selection result. Fig. 2(a) plots the finding number of communities with different initial label influence from 1 to 5 on two synthetic networks with different real community number. From this plot, we can see that LINSIA allows yielding stable community partitioning with the initial label influence on different values. So LINSIA is not sensitive to the initial label influence. Next, we study the influence of the variable α on the number of communities, and plots the result in Fig. 2(b) with α ranging from 0.5 to 1.5 on a synthetic network with community information. From this plot, we can see that the Eq. (3) gives a pretty well estimation to the value of the variable α corresponding to the real community number. So the definition in Eq. (3) is effective for LINSIA. 4 Fig. 1. The illustration of LINSIA. (a) The construction process of the bottom-up super-network structure. (b) The hierarchical community structure of the network. (c) The final network structure division, which includes 4 communities, 3 hubs and 2 outliers. 18 14 15 12 s e i t i n u m m o c f o r e b m u N 9 6 3 0 1.0 1.5 2.0 3.0 Initial label influence 2.5 3.5 4.0 12 10 s e i t i n u m m o c f o r e b m u N 8 6 4 2 0 real community number adaptive value got by Eq.(3) 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 α (a) Initial label influence vs. community number. (b) Adaptive variable α vs. community number. Fig. 2. The sensitivity of initial label influence and adaptive variable α on community detection. 4.3. Computational complexity The total complexity of LINSIA is the sum of the complexity of community detection at all aggregation levels and the complexity of hubs and outlier identification. The complexity of community detection at one aggregation level is O(n) + O(n · log(n)) + O(t · m), where n and m are the number of the nodes and the edges in the network respectively, O(n) is the complexity of computing node influence, O(n · log(n)) is the complexity of node ordering, O(t · m) is the complexity of updating community labels, and t is the number of iterations of label propagation process. If there are h layers in the bottom-up super-network structure, the total complexity of community detection is the sum of the complexities at the aggregation levels corresponding to the h layers. In addition, the complexity of hubs and outlier identification is O(n). Experiments show that t and h are usually small numbers and the number of nodes and edges reduces rapidly as the number of the layers increases. 5 Algorithm 1 LINSIA algorithm Input: Network G = (V, E), V = {v1, v2, ..., vn}. Output: Community division set Dic = {CS1, CS2, ...}, the best community division CSbest = {C1, C2, ...}, par- ticipation intensity P IntensitySeti, and the set of hubs and outliers. 1: t ← 0, and assigning a unique label to each node. 2: t ← t + 1, and find initial community division CSt. (a) Compute N I(i) using (1). Arrange nodes in ascending order of node influence, and store the nodes in vector N List. (b) For each node vi ∈ N List, update its labels according to (2) and (4). (c) Repeat (b) until node labels do not change any more. Then, get initial community division CSt = {C1, C2, ...} according to node labels and Dic[t] ← CSt. 3: Find community division set Dic = {CS1, CS2, ...}. (a) Build weighted super-network SNt based on CSt using (8), then detect non-overlapping community detection on SNt and get community division CCSt in super-network. (b) t ← t+1, and project CCSt into primitive network to get the merged community division CSt = {C1, C2, ...}, then Dic[t] ← CSt. (c) Iterate the last two steps until the community division no longer changes. 4: Evaluate the divisions in Dic = {CS1, CS2, ...}. Then copy the division CSi with the greatest measure value to CSbest, and copy the underlying division with the second-greatest measure value to CSsubc. 5: Find the set of hubs and outliers based on CSbest using (5) and (6), and compute P IntensitySeti using (7). 6: Return CSbest, CSsubc, P IntensitySeti, and the set of hubs and outliers. 5. Experimental results In this section, to demonstrate the benefits of the proposed algorithm LINSIA, we apply it on synthetic networks and real-world networks. Selection of comparison methods. To evaluate the performance of LINSIA, we compare it to several representatives of community detection algorithms. LPA [7] and NIBLPA [8] are label propagation based algorithms that use network structure alone as their guide and require neither optimization of a pre-defined objective function nor prior information about communities. Newman Fast Algorithm (NF) [3] is a popular community detection algorithm, which constructs a "den- drogram" by repeatedly joining communities together in pairs based on the greatest increase (or smallest decrease) in modularity measure and cuts through the dendrogram at different levels to give divisions of networks. Louvain [12] is a well-known modularity based algorithm. It allows for hierarchical community detection and has low time complexity. OCSBM [22] is an overlapping community detection method based on a principled statistical approach using generative networks models. EAGLE [24] deals with a set maximal cliques and adopts an agglomerative framework. It aims to detect both the hierarchical and overlapping structures of complex community together. To evaluate the performance of LINSIA, we compare LINSIA with LPA and NIBLPA to test its capability of community detection in naive networks (non-overlapping and non-hierarchical structures). Then we compare LINSIA with NF and Louvain for hierarchical community detection, and we compare LINSIA with OCSBM for overlapping community detection. Finally, we compare LINSIA with EAGLE in the detection of hierarchical and overlapping communities. For all experiments, we set the initial label influence to 1.0 for LINSIA as the default value, and we set α to 0.5 for NIBLPA. Evaluation metrics. To extensively compare different community detection algorithms with respect to effec- tiveness, we evaluate the division results in two ways. (a) Networks with class label. If community information is already known, the normalized mutual information NMI [28] and ENMI [29] are used to measure the performance of LINSIA for disjoint community detection and overlapping community detection respectively. NMI is proposed to provide a indication of the shared information between a pair of clusterings and ENMI is the extension of NMI for the similarity quantification of the set of true clusters and the set of detected clusters by taking overlapping nodes under consideration. (b) Networks without class label. Since the ground-truths are unknown, the modularity Q [4] and extended modularity EQ [24] are used to evaluate the quality of the disjoint communities and the overlapping communi- ties. Modularity Q is proposed to quantify the community structure in a network corresponds to a statistically surprising arrangement of edges and EQ is the extension of Q for the goodness evaluation of overlapped community decomposition. 6 Table 1. Parameters of synthetic networks. Table 2. Analysis of synthetic networks. Table 3. Real-world networks. 5.1. Synthetic networks Fig. 3. Hubs and outliers identification test. In this section, we generate several synthetic networks based on LFR benchmark networks [30] to evaluate the performance of LINSIA. There are many parameters to control the generated networks, including the number of nodes n, the average node degree k, the mixing ratio µ (each vertex shares a fraction µ of its edges with nodes in other communities), the number of overlapped nodes On, and the number of memberships of each overlapped is used to denote the mixing ratio for the communities at the high level node Om. in networks with hierarchical structures. In our experiments, LRF 1 and LRF 2 are naive synthetic networks, LRF 3 and LRF 4 are synthetic networks with hierarchical structures, LRF 5 and LRF 6 are synthetic networks with overlapping structures, and LRF 7 is a synthetic network with overlapping and hierarchical structures. The parameters for the networks are listed in Table 1. In addition, parameter µ ′ The results of the methods on the synthetic networks are shown in Table 2. We can see that LPA and NIBLPA can detect satisfactory community divisions in the naive network LRF 1 and LRF 2. Meanwhile, NF and Louvain yield acceptable results for the network LRF 3 and LRF 4. OCSBM almost achieves the perfect community divisions given the number of communities on the network LRF 5 and LRF 6. Although EAGLE allows for hierarchical and overlapping community detection, its accuracy in the network LRF 5, LRF 6 and LRF 7 is low. By contrast, LINSIA can find communities which almost exactly match the ground-truths on all the networks. So LINSIA is competitive with the other methods in the synthetic networks. In addition, we design a synthetic network to test the capability of LINSIA on hubs and outliers identification. The result is shown in Fig. 3, in which the nodes with the same community label are rendered by the same color. The result shows that the hub node (in red color) serves as a bridge between the nodes with different community label, and the outlier node (in pink color) is marginally connected with a community (in yellow color). So LINSIA can find rational hubs and outliers in the synthetic network. 5.2. Real-world networks In this section, we conduct experiments on a wide range of real-world networks shown in Table 3. In order to sufficiently evaluate LINSIA's performance, we first test it in terms of non-overlapping and non-hierarchical community detection, hierarchical community detection, overlapping community detection and overlapping and hierarchical community detection. Then we test it for hubs and outliers identification. Moreover, in addition to measuring the results with evaluation metrics, we provide the results of LINSIA on several real-world networks in visual graphics to demonstrate LINSIA's capability of community, hubs and outliers detection. (1) Non-overlapping and non-hierarchical community detection We compare our algorithm with LPA, NIBLPA for non-overlapping and non-hierarchical community detection, and Table 4 shows the results of the methods. We can see that LINSIA gets a better result than LPA and NIBLPA. 7 Table 4. The performance of different algorithms for non-overlapping and non-hierarchical community detection. Fig. 4. Result of LINSIA on Football network. Moreover, we find that LPA is a parameterized method and it gets different community divisions on the same network, which is inapplicable in practical problems. It is easy to understand that LINSIA outperforms LPA and NIBLPA since LINSIA improves the traditional method in terms of influence definition, label propagation and label selection. Below we analyze one of the networks in detail. The American College Football network is the network of games between Division IA colleges. There are 115 teams and 12 different conferences except for 8 independent teams. Fig. 4 plots the communities which are detected by LINSIA. It is interesting to note that LINSIA automatically finds 11 strong sense communities with NMI 0.7411. From this figure, we can observe that most of the teams are correctly assigned into corresponding communities. For LPA and NIBLPA, however, they are difficult to discover the natural community structure, and the NMI values are 0.6832 and 0.7027 respectively. (2) Hierarchical community detection To evaluate the performance of our algorithm on hierarchical community detection, we compare LINSIA with NF and Louvain, and the results are shown in Table 5. For the labeled networks, we present the numbers and the NMI values of the macro communities at the high level of the hierarchical community structures. For networks without class information, we compute the average Q of the communities of the hierarchical community structure. Table 5 shows that LINSIA achieves a good performance. However, Louvain results in a relatively low values of NMI, and the performance of NF is a bit worse. It is not surprise that LINSIA outperforms NF and Louvain since LINSIA can propagate local labels iteratively, and it can consider label information about the full network to achieve global optimal solution in each label selection. By contrast, NF and Louvain do not have similar mechanism to provide structure information of the full network for individual decision, which may leads to locally optimal solution. Table 5. The performance of different algorithms for hierarchical community detection. 8 Table 6. The performance of different algorithms for overlapping community detection. Fig. 5. Result of LINSIA on Polblog network. (3) Overlapping community detection We compare LINSIA with OCSBM to demonstrate the capability of LINISA on overlapping community detec- tion, and the results are shown in Table 6. As there is no overlapping ground-truth information of all these networks, we use EQ to evaluate the algorithms. Given the community number, we find that OCSBM can get high-quality community division in the labeled networks. However, the method is unsatisfactory in the networks without prior community information. We can see that LINSIA detects better results than OCSBM in most of the networks. Below we analyze the Polblog network in detail and give a graphical display. This network is derived from the links between weblogs about US politics published around the time of the 2004 presidential election. The links between weblogs were automatically extracted from a crawl of the front page of the weblog. Each blog is labeled with '0' or '1' to indicate whether they are "liberal" or "conservative". LINSIA groups these weblogs into two categories, which well represent the corresponding liberal and conservative weblogs respectively. (4) Hierarchical and overlapping community detection As shown in Table 7, we compare LINSIA with EAGLE for hierarchical and overlapping community detection. As there is no overlapping ground-truths, we compute the average EQ of the overlapping communities to measure the quality of the detected hierarchical community structures of the networks. Meanwhile, we use the number of the macro communities at the high level of the hierarchical structures as a evaluation metric. From Table 7, we can see that LINSIA shows a clear advantage over EAGLE based on the average EQ. Moreover, the community structure detected by EAGLE tends to have a large number of communities with small size. The reason is that EAGLE is a maximum cliques based method. As real-world networks are generally sparse, most of the maximum cliques detected by EAGLE only have several nodes. Dolphins network is an undirected social network representing frequent associations between 62 dolphins in a community living off Doubtful Sound, New Zealand, and the hierarchical community structure detected by LINSIA of the Dolphins network is shown in Fig. 6. Fig. 6(a) represents the overlapping communities at the high level, and Fig. 6(b) represents the overlapping communities at the low level. According to Fig. 6, we can find that the hierarchical relationship between the overlapping communities. According to [31], we can conclude that LINSIA can detect almost the real communities of the network. Furthermore, schematic network is a clique dominated network which is used for performance evaluation of EAGLE in [24]. For the performance comparison of LINSIA against EAGLE, we apply LINSIA on the network, and the overlapping communities and the hierarchical relationship between them are shown in Fig. 7. The community with blue color in Fig. 7(a) includes the community with green color and the community with blue color in Fig. 7(b). According to the network topology, we can conclude that the result is reasonable. Meanwhile, the community division found by LINSIA is identical with the result detected by EAGLE in [24]. Thus we can conclude that EAGLE performs well on clique dominated networks, and LINSIA can detect rational overlapping and hierarchical communities on multiple types of networks. 9 Table 7. The performance of different algorithms for hierarchical and overlapping community detection. (a) EQ=0.3314, C=2. (b) EQ=0.3301, C=3. Fig. 6. Result of LINSIA on Dolphins network. (a) EQ=0.4986, C=3. (b) EQ=0.4801, C=4. Fig. 7. Result of LINSIA on schematic network. (5) Hubs and outliers identification Given that there is no universal standard for hubs and outliers evaluation, the capability of LINSIA on hubs and outliers identification is evaluated based on result analysis. To evaluate LINSIA's effectiveness on hubs detection, we apply it on three real-world networks. One of the networks is the Dolphins network. The hubs of the Dolphins network detected by LINSIA are shown as red nodes in Fig. 6. Considering the network structure and community division comprehensively, we can conclude that the detected hubs are rational. Furthermore, we apply LINSIA on the Polblog network in Fig. 5 and the schematic network in Fig. 7. We can conclude that the detected hubs in red color are also reasonable. In addition, LINSIA has the advantage of soft-partitioning solution, and it can depict the degree of hubs belonging to each relevant community. For example, in the Dolphins network, the participation intensity set of the hubs in Fig. 6(a) is: {1:[(57, 0.746), (51, 0.253)], 7: [(57, 0.501), (51, 0.499)], 19: [(57, 0.5708), (51, 0.429)], 32: [(32, 0.600), (57, 0.399)], 39: [(57, 0.428), (51, 0.572)]}. Take the hub node 39 in the participation intensity set for example, the item means that node 39 belongs to the community labeled by 57 (in blue color) with degree 0.428 and belongs to the community labeled by 51 (in yellow color) with degree 0.572. LINSIA's capability on outliers identification can also be demonstrated by the results of the Polblog network in Fig. 5 and the Dolphins network in Fig. 6. In the results, the outliers in Fig. 5 (in green color) and the outliers in Fig. 6 (in pink color) are rational according to the network topology. The reason is that our algorithm proposes a 10 new label updating strategy to make the periphery nodes have opportunity to compete with the core nodes in label selection process. For example, in Fig. 6(a), firstly outlier node 60 selects node 32's label as its community label. Then node 32 acts as an overlapping node after weighting the label influence of neighbor nodes. In this way, node 60 can exist with its local label information, instead of being overwhelmed by core nodes' labels. So our algorithm can find rational outliers efficiently. In total, the experiments on all real-world networks demonstrate that LINSIA not only allows for extracting good complex communities in terms of the internal measures (Q and EQ) and external measures (NMI and ENMI), but also can identify reasonable hubs and outliers. 5.3. Running time Fig. 8. Running times comparison. (a) The running times of algorithms for non-overlapping and non-hierarchical community detection. (b) The running times of algorithms for hierarchical community detection. (c) The running times of algorithms for overlapping community detection. (d) The running times of algorithms for overlapping and hierarchical community detection. To assess and evaluate the scalability of LINSIA with respect to network size, we compare the running times of LINSIA as well as the comparison algorithms on the real-world networks listed in Table 3. Fig. 8 shows the running times of the different community detection algorithms. We can observe that LINSIA is faster than the other methods except Louvain and OCSBM. Although Louvain and OCSBM are faster than LINSIA, Louvain focuses only on hierarchical community detection, and OCSBM focuses only on overlapping community detection and requires prior community number for sufficiently good performance. By contrast, LINSIA is not only applicable to community detection, but also can identify hubs and outliers. In addition, we can find that LINSIA can process the networks with hundreds of thousands of nodes within two hours. So for the comprehensive complex network structure analysis, our algorithm LINSIA performs well in term of running time. 6. Discussion and Conclusion Network structure analysis has always been an important research topic. Though many related researches have been done, most of them focus only on one of the structural features. The problem of investigating network struc- ture comprehensively is very important, with many potential applications in real-world networks. However, little has been down in this research direction. In this paper, we have proposed an integrated network structure inves- tigation algorithm LINSIA. The proposed algorithm outperforms several representative network structure analysis algorithms. However, we are still far away from thoroughly understanding the structure of complex networks, and some more issues still remain open. In this paper, the method focuses only on structural aspect. Besides struc- ture, there are many dynamical processes on complex networks such as synchronization [38]. We hope the method and results in this paper can inspire some network structure analysis methods taking the network dynamics under consideration. 11 Acknowledgements The work was supported partially by the National Natural Science Foundation of China (Grant No. 61202255), University-Industry Cooperation Projects of Guangdong Province (Grant No. 2012A090300001) and the Pre- research Project (Grant No. 51306050102). We thank Tao Zhou and JunMing Shao for their advices. We also thank Xin Li, Yunpeng Xiao, Xingping Xian and Yuanping Zhang for enlightening discussions and careful reading of the manuscript. The authors also wish to thank the anonymous reviewers for their thorough review and highly appreciate their useful comments and suggestions. References [1] M. Girvan, M. E. J. Newman, "Community structure in social and biological networks", Proc. Natl. Acad. Sci. USA 99 (2002) 7821C7826. [2] M. Girvan, M. E. J. Newman, "Finding and evaluating community structure in networks", Phys. Rev. E 69 (2004) 026113. [3] M. E. J. Newman, "Fast algorithm for detecting community structure in networks", Phys. Rev. E 69 (2004) 066133. [4] M. E. J. Newman, "Modularity and community structure in networks", Proc. Natl. Acad. Sci. USA 103 (2006) 8577-8582. [5] J. Yang, J. Leskovec, "Overlapping community detection at scale: a nonnegative matrix factorization approach", WSDM' 13, pp. 587-596. [6] M. E. J. Newman, "Spectral methods for network community detection and graph partitioning", Phys. Rev. E 88 (2013) 042822. [7] U. N. Raghavan, R. Albert, S. Kumara, "Near linear time algorithm to detect community structures in large-scale networks", Phys. Rev. E 76 (2007) 036106. [8] Y. Xing, F.R. Meng, Y. Zhou, "A Node Influence Based Label Propagation Algorithm for Community Detection in Networks", The Sci. World J. 2014 (2014) 627581. [9] M. E. J. Newman, E. Leicht, "Mixture models and exploratory analysis in networks", Proc. Natl. Acad. Sci. USA 104 (2007) 9564-9569. [10] Brian Karrer, M. E. J. Newman, "Stochastic blockmodels and community structure in networks", Phys. Rev. E 83 (2011) 016107. [11] Zh. Q. Xu, Y. P. Ke, Y. Wang, "A Fast Inference Algorithm for Stochastic Blockmodel", ICDM'14, pp. 620-629. [12] V. D. Blondel, J. L. Guillaume, R. Lambiotte, E. Lefebvre, "Fast unfolding of communities in large networks", J. Stat. Mech. (2008) P10008. [13] J. Huang, H. Sun, J. Han, H. Deng, Y. Sun, "SHRINK: A Structural Clustering Algorithm for Detecting Hierarchical Communities in Networks", CIKM'10, pp. 219-228. [14] M. Sales-Pardo, R. Guimera, A. A. Moreira, L. A. Amaral, "Extracting the hierarchical organization of complex systems", Proc. Natl. Acad. Sci. USA 104 (2007) 15224-15229. [15] S. Gregory, "Finding overlapping communities in networks by label propagation", New J. Phys. 12 (2010) 103018. [16] G. Palla, I. Derenyi, I. Farkas, T. Vicsek, "Uncovering the overlapping community structure of complex networks in nature and society", Nature 435 (2005) 814-818. [17] J. R. Xie, B. K. Szymanski, X. M. Liu, "Slpa: Uncovering Overlapping Communities in Social Networks via a Speaker- Listener Interaction Dynamic Process", ICDMW'11, pp. 344-349. [18] D. Chen, M. Shang, Z. Lv, Y. Fu, "Detecting overlapping communities of weighted networks via a local algorithm", Physica A 389 (2010) 4177-4187. [19] Y. Z. Cui, X. Y. Wang, J. Q. Li, "Detecting overlapping communities in networks using the maximal sub-graph and the clustering coefficient", Phys. A 405 (2014) 85-91. [20] P. K. Gopalan, D. M. Blei, "Efficient discovery of overlapping communities in massive networks", Proc. Natl. Acad. Sci. USA 110 (2013) 14534-14539. [21] B. J. Sun, H. W. Shen, X. Q. Cheng, "Detecting overlapping communities in massive networks", EPL. 108 (2014) 68001. 12 [22] B. Ball, B. Karrer, M. E. J. Newman, "Efficient and principled method for detecting communities in networks", Phys. Rev. E 84 (2011) 036103. [23] A. Lancichinetti, S. Fortunato, J. Kertesz, "Detecting the overlapping and hierarchical community structure of complex networks", New J. Phys. 11 (2009) 033015. [24] H. Shen, X. Cheng, K. Cai, M.B. Hu, "Detect overlapping and hierarchical community structure in networks", Phys. A 388 (2009) 1706-1712. [25] F. Havemann, J. Glaser, M. Heinz, M. Struck, "Identifying Overlapping and Hierarchical Thematic Structures in Net- works of Scholarly Papers: A Comparison of Three Approaches", PLoS ONE 7 (2012) e33255. [26] J. Bae, S. Kim, "Identifying and ranking influential spreaders in complex networks by neighborhood coreness", Phys. A 395 (2014) 549-559. [27] M. Kitsak, L. K. Gallos, S. Havlin, F. Liljeros, L. Muchnik, H. E. Stanley, H. A. Makse, "Identification of influential spreaders in complex networks", Nature Phys. 6 (2010) 888-893. [28] A. Strehl, J. Ghosh, "Cluster ensemblesa knowledge reuse framework for combining multiple partitions", The Journal of Machine Learning Research 3 (2003) 583-617. [29] A. McDaid, D. Greene, and N. Hurley, "Normalized Mutual Information to evaluate overlapping community finding algorithms", CoRR (2011) abs/1110.2515 [30] A. Lancichinetti, S. Fortunato, F. Radicchi, "Benchmark graphs for testing community detection algorithms.", Phys. Rev. E 78 (2008) 046110. [31] W. W. Zachary, "An information flow model for conflict and fission in small groups", J. Anthro. Res. (1977) 452-473. [32] D. Lusseau, "The emergent properties of dolphin social network", Proc. Biol. Sci. 270 (2003) 186-188. [33] Lada A. Adamic, Natalie Glance, "The political blogosphere and the 2004 U.S. election: divided they blog", LinkKDD '05, pp. 36-43. [34] M. E. J. Newman, "The structure of scientific collaboration networks", Proc. Natl. Acad. Sci. USA 98 (2001) 404-409. [35] J. Leskovec, K. Lang, A. Dasgupta, M. Mahoney, "Community Structure in Large Networks: Natural Cluster Sizes and the Absence of Large Well-Defined Clusters", Internet Mathematics, 6 (2009) 29-123. [36] B. Viswanath, "On the Evolution of User Interaction in Facebook", WOSN'09, pp. 37-42. [37] J. McAuley, J. Leskovec, "Image Labeling on a Network: Using Social-Network Meta data for Image Classification", ECCV14, pp. 828-841. [38] M. Y. Zhou, Z. Zhuo, S. M. Cai, Z. Q. Fu, "Community structure revealed by phase locking", Chaos 24 (2014) 033128. 13
1705.02202	1	1705	2017-02-26T21:56:55	Structured sampling and fast reconstruction of smooth graph signals	[ "cs.SI", "cs.IT", "cs.IT" ]	This work concerns sampling of smooth signals on arbitrary graphs. We first study a structured sampling strategy for such smooth graph signals that consists of a random selection of few pre-defined groups of nodes. The number of groups to sample to stably embed the set of $k$-bandlimited signals is driven by a quantity called the \emph{group} graph cumulative coherence. For some optimised sampling distributions, we show that sampling $O(k\log(k))$ groups is always sufficient to stably embed the set of $k$-bandlimited signals but that this number can be smaller -- down to $O(\log(k))$ -- depending on the structure of the groups of nodes. Fast methods to approximate these sampling distributions are detailed. Second, we consider $k$-bandlimited signals that are nearly piecewise constant over pre-defined groups of nodes. We show that it is possible to speed up the reconstruction of such signals by reducing drastically the dimension of the vectors to reconstruct. When combined with the proposed structured sampling procedure, we prove that the method provides stable and accurate reconstruction of the original signal. Finally, we present numerical experiments that illustrate our theoretical results and, as an example, show how to combine these methods for interactive object segmentation in an image using superpixels.	cs.SI	cs	Structured sampling and fast reconstruction of smooth graph signals Gilles Puy and Patrick P´erez Abstract. This work concerns sampling of smooth signals on arbitrary graphs. We first study a structured sampling strategy for such smooth graph signals that consists of a random selection of few pre-defined groups of nodes. The number of groups to sample to stably embed the set of k-bandlimited signals is driven by a quantity called the group graph cumulative coherence. For some optimised sampling distributions, we show that sampling O(k log(k)) groups is always sufficient to stably embed the set of k-bandlimited signals but that this number can be smaller -- down to O(log(k)) -- depending on the structure of the groups of nodes. Fast methods to approximate these sampling distributions are detailed. Second, we consider k-bandlimited signals that are nearly piecewise constant over pre-defined groups of nodes. We show that it is possible to speed up the reconstruction of such signals by reducing drastically the dimension of the vectors to reconstruct. When combined with the proposed structured sampling procedure, we prove that the method provides stable and accurate reconstruction of the original signal. Finally, we present numerical experiments that illustrate our theoretical results and, as an example, show how to combine these methods for interactive object segmentation in an image using superpixels. 1. Introduction This work was initially inspired by studying developments in edit propagation for interactive image or video manipulations, where the goal is to propagate operations made by a user in some parts of the image to the entire image, e.g., propagate foreground/background scribbles for object segmentation [1 -- 5] or propagate manual color/tone modifications [3 -- 7]. First, a graph G modelling the similarities between the pixels of the image is built. Second, the user-edits specify values at some nodes (pixels) of G. Finally, the complete signal is estimated by assuming that it is smooth on G. The quality of the propagation depends on the structure of G and on the location of annotated nodes. If a part of the image is weakly connected to the rest, the user-edits do not propagate well to this region unless this region is edited directly. Therefore, highlighting beforehand which regions or groups of nodes are important to edit to ensure a good propagation would be a useful feature to facilitate user interactions. Furthermore, designing a fast reconstruction/propagation method is also important so that the user can visualise immediately the effect of his inputs. To this end, "superpixels" -- small groups of connected pixels where the image varies only little -- are sometimes used among other things to speed up computations, e.g., in [3]. We address these problems from a graph signal processing point-of-view [8]. More precisely, we view them as a sampling problem where we should select a structured set of nodes (regions to edit), "measure" the signal on this set (user-edits), and reconstruct the signal on the entire graph. We believe that the sampling strategy and the fast reconstruction method proposed here provide useful tools to optimize user's inputs and accelerate computations. As mentionned above, the user-edits are propagated to the entire image by assuming that the global signal is smooth on G. In the context of graph signal processing, smooth or k-bandlimited signal is a widely used and studied model. Several sampling methods have been designed to sample such signals. Pesenson introduced the notion of uniqueness set (of nodes) for k-bandlimited graph signals in [9, 10]. Two different k-bandlimited signals are also necessarily different when restricted to a uniqueness set. Therefore, one can sample all k-bandlimited signals on a uniqueness set. Then, Anis et al., [11, 12] and Chen et al., [13, 14] proved that one can always find a uniquess set of k nodes to sample all k-bandlimited signals. Finding this set is however computationally expensive. In [15], graph spectral proxies are used to find such a set more efficiently. Yet the combinatorial problem that needs to be solved to find such a set makes the method still difficult to use for large graphs. Other authors used the idea of random sampling to be able to handle large graphs [14, 16, 17]. Recently, Puy et al. [17] proved that there always exists a random sampling strategy for which sampling O(k log(k)) nodes G. Puy and P. P´erez are with Technicolor, 975 Avenue des Champs Blancs, 35576 Cesson-S´evign´e, France.. 1 2 G. PUY AND P. P´EREZ is sufficient to stably embed the set of bandlimited signals. They also designed a fast and scalable algorithm to estimate the optimal sampling distribution. In this paper, we study first a random sampling strategy for k-bandlimited signals where we sample few groups of nodes instead of sampling individual nodes. We introduce the concept of local group graph coherence that quantifies the importance of sampling each group. Second, in order to build a fast reconstruction technique for k-bandlimited signals, we use the intuition that a smooth graph signal is a signal that varies slowly from one node to its connected nodes. If we group few connected nodes together, we usually expect a bandlimited signal to be essentially constant on this set of nodes; as long as we do not group together too many weakly connected nodes. We propose here to use this property to accelerate the reconstruction of such k-bandlimited signals, i.e., k-bandlimited signals nearly piecewise constant over (pre-defined) groups of nodes. When combined with the proposed sampling technique, we prove that this fast method provides stable and accurate reconstructions of the signals of interest. Finally, we illustrate how to use these results for interactive object segmentation in an image, with required node groups being superpixels. 1.1. Contributions The random sampling strategy that we propose generalises the method proposed in [17]. We use here a structured sampling strategy. Let us already acknowledge here that such strategies are also studied in the field of compressed sensing [18 -- 21] and that some of our solutions are directly inspired by these works. First, in this structured sampling setting, we show that the number of groups to sample is directly linked to a quantity called the group graph cumulative coherence. This quantity generalises the concept of graph cumulative coherence introduced in [17] and characterises how much the energy of k-bandlimited can stay concentrated in each group of nodes. Second, we can then choose to sample the groups non-adaptively or to optimise the sampling distribution to minimise the number of groups to sample. With this optimised sampling distribution, our result shows that, in the worst case, sampling O(k log(k)) groups of nodes is sufficient to ensure the reconstruction of all k-bandlimited signals. As each group can contain many nodes, we might have to sample a large number of nodes. This is the potential price to pay when sampling the nodes by groups. Fortunately, a smaller number of groups -- down to O(log(k)) -- might already be sufficient if the groups are well designed. Third, we describe a method to estimate the optimal sampling distribution without the need of computing the graph Fourier matrix and which is thus able to handle large graphs. Fourth, estimating the optimal sampling distribution may still be too slow when a large number of groups are involved. We thus also present a sufficient recovery condition that involves a relaxed version of group graph cumulative coherence. The sampling distribution that minimises this relaxed coherence is fast to estimate for large graphs and large number of groups. With this sampling strategy, we prove that sampling O(k log(k)) groups is always sufficient to ensure the reconstruction of all k-bandlimited signals. This strategy is mainly interesting at small k. Finally, we propose a fast reconstruction method for k-bandlimited signals that are also nearly piecewise constant over pre-defined groups of nodes. We show that we can reduce drastically the dimension of the reconstruction problem for such signals. When the above sampling procedure is used to sample them, we prove that the proposed method provides accurate and stable recovery of this type of signals. 1.2. Applications The proposed sampling methods can have interest in several applications. For example, if one needs to develop sensors in a large scale network, it might be easier to deploy and install these sensors at nodes around few spatial locations instead of scattering them all over the network. Finding the best regions to monitor is thus important. In a social network, one might be interested in monitoring a signal defined over communities and thus should find which groups of users are the most important Structured sampling and fast reconstruction of smooth graph signals 3 to sample. individual nodes. Let us now detail such an example for semi-supervised classification. In semi-supervised learning, it may be easier to label jointly some similar nodes than The task we consider is interactive object segmentation in an image where, similarly to what is done in [22] for instance, we build a graph G that models similarities between the pixels of the image, ask the user to label some regions depending on whether they are part of the object or not, diffuse the result on the complete image by supposing that indicator vectors of the object is smooth on G. To propose the regions to label, we view this segmentation problem as a sampling problem of a smooth signal on G where the regions to label are chosen to ensure the recovery of a k-bandlimited signal. To choose the regions to label, we start by partitioning the image into superpixels, e.g., with SLIC technique [23]. We see in Fig. 1 that the superpixels divide the image into homogeneous regions. The superpixels follow the edges and most of them thus belong to either the tiger or the background but rarely both. One interest of dividing the image into superpixels is that it facilitates user interactions. It is easier to determine if a superpixel belongs to the object of interest than if a pixel belongs to the tiger, especially at the boundaries, as recently exploited for segmentation on touch-screens [24]. Using the sampling method that we developed we can propose to the user a small number of superpixels to label. Furthermore, this proposition is adapted to the structure of the graph. Another advantage of using superpixels is that the indicator function of the object, beyond being smooth on G, is also approximately piecewise constant on the superpixels. We can thus use our reconstruction method to estimate rapidly the segmentation result from the labelled superpixels. Figure 1. One image (left) and its partition into superpixels with SLIC technique [23] (right). 1.3. Notations and definitions We consider that G = {V,E, W} is an undirected weighted graph, where V is the set of n nodes, E is the set of edges, and W ∈ Rn×n is the weighted adjacency matrix with nonnegative entries. We denote the graph Laplacian by L ∈ Rn×n. We assume that L is real, symmetric, and positive semi-definite, e.g., the combinatorial graph Laplacian L := D − W, or the normalised one L := I − D−1/2WD−1/2. The matrix D ∈ Rn×n is the diagonal degree matrix and I ∈ Rn×n is the identity matrix [25]. The diagonal degree matrix D has entries Dii :=(cid:80)i(cid:54)=j Wij. We denote by U ∈ Rn×n the orthonormal eigenvectors of L and by 0 = λ1 (cid:54) . . . (cid:54) λn the ordered , where Λ := diag(λ1, . . . , λn) ∈ Rn×n. The matrix U is the real eigenvalues of L. We have L = UΛU graph Fourier basis [8]. For any signal x ∈ Rn defined on the nodes of the graph G, x = U x contains the Fourier coefficients of x ordered in increasing frequencies. This work deals with k-bandlimited signals x ∈ Rn on G, i.e., signals whose Fourier coefficients xk+1, . . . , xn are null. Let Uk be the restriction of U to its first k vectors: (cid:124) (cid:124) (1) Uk := (u1, . . . , uk) ∈ Rn×k. 4 G. PUY AND P. P´EREZ Definition 1.1 (k-bandlimited signal on G). A signal x ∈ Rn defined on the nodes of the graph G is k-bandlimited with k ∈ N \ {0} if x ∈ span(Uk), i.e., there exists η ∈ Rk such that (2) x = Ukη. the definition of k-bandlimited signals. This definition was also used in [11, 14, 17]. We assume that λk (cid:54)= λk+1 to avoid any ambiguity in Finally, for any matrix X ∈ Rn1×n2 , (cid:107)X(cid:107)2 denotes its spectral norm and (cid:107)X(cid:107)F its Frobenius norm; when n1 = n2, λmax(X) denotes its largest eigenvalue and λmin(X) its smallest eigenvalue. For any vector x ∈ Rn1 , (cid:107)x(cid:107)2 denotes the Euclidean norm of x. Depending on the context, xj may represent the jth entry of the vector x or the jth column-vector of the matrix X. The entry on the ith row and jth column of X is denoted by Xij. The identity matrix is denoted by I (its dimensions are determined by the context). We present in Fig. 2 a representation of the important variables and processes involved in this paper in order to facilitate the understanding of the different results. Figure 2. Summary of the main variables and processes involved in the presented sampling strategies. All variables are defined when needed hereafter. 2. Sampling using groups of nodes In this section, we explain our sampling strategies, starting with the definition of the groups of nodes. 2.1. Grouping the nodes We consider that the n nodes of G are divided into N different groups N1, . . . ,NN ⊆ {1, . . . , n}. The size of the (cid:96)th group is denoted N(cid:96). We suppose that these groups form a partition of {1, . . . , n}, so that each node belongs to exactly one group. We have (3) ∪N (cid:96)=1 N(cid:96) = {1, . . . , n}, and N(cid:96) ∩ N(cid:96)(cid:48) = ∅. x2Rnx2RNy2RsAAy2RmMeML2Rn⇥neL2RN⇥NAveragingSamplingp2RNStructured sampling and fast reconstruction of smooth graph signals 5 For the object segmentation application discussed in the introduction, these groups represent the superpixels. However, we do not impose the groups to be made of neighbouring nodes in the graph; they can be made of nodes "far" from each other. N(cid:96)}, we associate a matrix N((cid:96)) ∈ RN(cid:96)×n that restricts a graph For each group N(cid:96) = {n((cid:96)) 1 , . . . , n((cid:96)) signal to the nodes appearing in N(cid:96), i.e., N((cid:96)) (4) ij Note that (5) for j = n((cid:96)) i otherwise. , N((cid:96)) = I. 0 (cid:124) :=(cid:26) 1 N(cid:88)(cid:96)=1 :=(cid:40) β−1/2 N((cid:96)) n((cid:96)) The case of overlapping groups can be handled by changing the definition of N((cid:96)) to N((cid:96)) ij (6) where 1 (cid:54) βi (cid:54) N , i = 1, . . . , n, is the number of times node i appears in the different groups N1, . . . ,NN . Equation (5) also holds in this case. All results presented in Section 2 are valid for overlapping groups with this definition of N((cid:96)). 0 i for j = n((cid:96)) i otherwise, , 2.2. Sampling the groups The sampling procedure consists in selecting s groups out of the N available ones. In the application of Section 1.2, it corresponds to the selection of the superpixels to label. We select these groups at random using a sampling distribution on {1, . . . , N} represented by a vector p ∈ RN . The probability of selecting the (cid:96)th group is p(cid:96). We assume that p(cid:96) > 0 for all (cid:96) = 1, . . . , N , so that all groups may be selected with a non-zero probability. We obviously have(cid:80)N The indices Ω := {ω1, . . . , ωs} of the selected groups are obtained by drawing independently -- thus (cid:96)=1 p(cid:96) = 1. with replacements -- s indices from the set {1, . . . , N} according to p, i.e., ∀j ∈ {1, . . . , s} and ∀(cid:96) ∈ {1, . . . , N}. (7) The selected groups are Nω1 , . . . ,Nωs and the total number of selected nodes is (8) P(ωj = (cid:96)) = p(cid:96), m := Once the groups are selected, we build the sampling matrix M ∈ Rm×n that satisfies (9) and which restricts any signal to the nodes belonging to the selected groups. For a signal x ∈ Rn defined on the nodes of G, its sampled version y ∈ Rm satisfies (10) y := Mx. Our goal now is to determine what number s is enough to ensure that all k-bandlimited signals can be reconstructed from its sampled version obtained with M. To conduct this study, we need to define few more matrices. First, we associate the matrix P((cid:96)) := p−1/2 (11) to each group N(cid:96). Then, once the groups are drawn, we construct the block diagonal matrix P ∈ Rm×m (12) I ∈ RN(cid:96)×N(cid:96), P := diag(cid:16)P(ω1), . . . , P(ωs)(cid:17) . (cid:96) s(cid:88)j=1(cid:12)(cid:12)Nωj(cid:12)(cid:12) . M :=  , N(ω1) N(ωs) ... 6 G. PUY AND P. P´EREZ This matrix takes into account the probability of sampling each group and will be used to rescale 2.1 Both y for norm preservation. This matrix ensures that s−1 EΩ (cid:107)Py(cid:107)2 2 = s−1 EΩ (cid:107)PMx(cid:107)2 2 = (cid:107)x(cid:107)2 matrices P and M depend on Ω and are random. 2.3. Group graph coherence The sampling procedure in [17] is similar to the ones proposed here at the difference that the nodes are sampled individually and not by groups. It was proved there that the number of nodes to sample is driven by a quantity called the graph coherence. This quantity measures how the energy of k- bandlimited signals spreads over the nodes. Similarly, we prove here that the number of groups to sample is driven by a quantity that measures the energy of k-bandlimited signals spreads over the groups. We now introduce this quantity. The matrix N((cid:96))Uk is the matrix that restricts a k-bandlimited signal to the nodes belonging to N(cid:96). Therefore, (13) (cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13)2 = sup η∈Rk:(cid:107)η(cid:107)2=1(cid:13)(cid:13)(cid:13)N((cid:96))Ukη(cid:13)(cid:13)(cid:13)2 measures the energy on the nodes N(cid:96) of the normalised k-bandlimited signal that is most concentrated on N(cid:96). This energy varies between 0 and 1. When this energy is close to 1, there exists a k-bandlimited signal whose energy is essential concentrated on N(cid:96). This signal lives only on the nodes in N(cid:96) and does not spread elsewhere. On the contrary, when this energy is close to 0, there is no k-bandlimited signal living only on N(cid:96). The sampling distribution p is adapted to the graph and the structure of the groups if: whenever (cid:107)N((cid:96))Uk(cid:107)2 is high, p(cid:96) is high; whenever (cid:107)N((cid:96))Uk(cid:107)2 is small, p(cid:96) is small. In other words, the ratio between (cid:107)N((cid:96))Uk(cid:107)2 and p(cid:96) should be as constant as possible. This ensures that the groups where some k-bandlimited signals are concentrated are sampled with higher probability. Similarly to what was done in [17] with individual nodes, we define the group graph weighted coherence as the largest ratio between (cid:107)N((cid:96))Uk(cid:107)2 and p−1/2 Definition 2.1 (Group graph cumulative coherence). The group graph cumulative coherence of order k is . (cid:96) (14) The quantity (cid:107)N((cid:96))Uk(cid:107)2 is called the local group graph coherence. νp := max 1(cid:54)(cid:96)(cid:54)N(cid:110)p−1/2 (cid:96) (cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13)2(cid:111) . of the graph weighted coherence introduced in [17]. In the extreme case where the groups N1, . . . ,NN reduce all to singletons, we recover the definition It is easy to prove that νp is lower bounded by 1. Indeed, for any η ∈ Rk with (cid:107)η(cid:107)2 = 1, we have 1 = (cid:107)Ukη(cid:107)2 2 2 = N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Ukη(cid:13)(cid:13)(cid:13) 1(cid:54)(cid:96)(cid:54)N(cid:40)(cid:13)(cid:13)N((cid:96))Ukη(cid:13)(cid:13)2 p(cid:96) ·(cid:13)(cid:13)N((cid:96))Ukη(cid:13)(cid:13)2 N(cid:88)(cid:96)=1 (cid:41) (cid:54) max (cid:41) = ν2 1(cid:54)(cid:96)(cid:54)N(cid:40)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 p(cid:96) p(cid:96) p(cid:96) = 2 2 2 p. (15) = max 2 (cid:54) (cid:107)p(cid:107)1 · max 1(cid:54)(cid:96)(cid:54)N(cid:40)(cid:13)(cid:13)N((cid:96))Ukη(cid:13)(cid:13)2 p(cid:96) 2 (cid:41) We have νp = 1 in, for example, the degenerated case where N1 = {1, . . . , n}. 2.4. Stable embedding We now have all the tools to present our main theorem that shows that sampling O(ν2 p log(k)) groups is sufficient to stably embed the whole set of k-bandlimited signals. Hence, it is possible to reconstruct any x ∈ span(Uk) from its measurements y = Mx. 1This property is a consequence of (65), proved in Appendix A. Structured sampling and fast reconstruction of smooth graph signals 7 Theorem 2.2 (Restricted isometry property - RIP). Let M be a random subsampling matrix con- structed as in (9) using the groups N1, . . . ,NN and the sampling distribution p. For any δ, ξ ∈ (0, 1), with probability at least 1 − ξ, (16) (cid:54) 1 (1 − δ)(cid:107)x(cid:107)2 2 for all x ∈ span(Uk) provided that (17) Proof. See Appendix A. 2 2 (cid:54) (1 + δ)(cid:107)x(cid:107)2 s (cid:107)PM x(cid:107)2 p log(cid:18) 2k ξ (cid:19) . δ2 ν2 s (cid:62) 3 (cid:3) In the above theorem, we recall that s is the number of selected groups, each of them containing several nodes. We thus control the number of groups to sample and not directly the number of nodes. As the lower bound on νp is 1, sampling O(log(k)) groups might already be sufficient if the groups and the sampling distribution are well-designed. The number of groups to sample is driven by νp, which itself depends on the structure of the groups N1, . . . ,NN and on the sampling distribution p. To reduce the number of samples to measure, we might optimise the structure of the groups and the sampling distribution. For example, if we were able to construct n/L groups (L (cid:62) 1) such that (cid:107)N((cid:96))Uk(cid:107)2 ≈ (L/n)1/2 -- i.e., no k-bandlimited signals have more than 100 · (L/n)1/2 percent of its energy concentrated in each group -- then setting p(cid:96) = L/n, (cid:96) = 1, . . . , n/L, would yield ν ≈ 1. In this case, sampling one group is enough to embed the set of k-bandlimited signals. However, it is not obvious how we can construct such groups in practice and we might not even have the flexibility to modify the structure of the groups. In such a case, the only possibility to reduce the number of measurements is to optimise the sampling distribution p to minimise νp. The sampling distribution minimizing the coherence νp is the distribution p∗ ∈ RN that satisfies p∗(cid:96) := (18) and for which (19) 2 (cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 (cid:96)(cid:48)=1(cid:13)(cid:13)N((cid:96)(cid:48))Uk(cid:13)(cid:13)2 (cid:80)N 2 , for all (cid:96) ∈ {1, . . . , N}, ν2 p∗ = . 2 2 N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) > p∗(cid:96)(cid:48)−1(cid:13)(cid:13)(cid:13)N((cid:96) (cid:48))Uk(cid:13)(cid:13)(cid:13) 2 2 ν2 p(cid:48) (cid:62) p(cid:48)(cid:96)(cid:48)−1(cid:13)(cid:13)(cid:13)N((cid:96) = ν2 p∗ , 2 2 (cid:48))Uk(cid:13)(cid:13)(cid:13) (20) Indeed, let p(cid:48) (cid:54)= p∗ be another sampling distribution. As the entries of p(cid:48) and p∗ are nonnegative and sum to 1, we necessarily have p(cid:48)(cid:96)(cid:48) < p∗(cid:96)(cid:48) for some (cid:96)(cid:48) > 0. Then, where the last equality is obtained by replacing p∗(cid:96)(cid:48) with its value. Therefore, νp(cid:48) > νp∗ for any p(cid:48) (cid:54)= p∗. As similar proof can be found in, e.g., [18] where the authors derive the optimal sampling distribution for a compressive system. We notice that (21) ν2 p∗ = N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) 2 2 (cid:54) N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) 2 F = k. Hence, by using this distribution, (17) shows that sampling O(k log(k)) groups is always sufficient to sample all k-bandlimited signals. The exact number is proportional to ν2 p∗ log(k). This is not in contradiction with the fact that at least k measurements are required in total as one group contains at least one node. We also have (22) ν2 p∗ = (cid:54) N, N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) 2 2 8 G. PUY AND P. P´EREZ (cid:54) 1. Therefore, in any case, the bound never suggests to sample much more than N as (cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 2 groups, as one would expect it. We recall that the results in [17] proves that it is always sufficient to sample O(k log(k)) nodes to embed the set of k-bandlimited signals. When sampling the nodes by groups, it is the number of groups to sample that should be O(k log(k)). This can be a large number of individual nodes but this is the potential price to pay when sampling the nodes by groups. Variable density sampling [19, 26, 27] and structured sampling [18, 20, 21] are also important topics in compressed sensing. The method proposed here is closely inspired by these studies, especially by [20, 21]. Our results thus share several similarities with these works. We however benefit from a simpler signal model and take advantage of the graph structure to refine the results, propose simpler decoders to reconstruct the signal, and design efficient algorithms to estimate p∗. 2.5. A more practical result at small k's Optimising the sampling distribution reduces to estimating the spectral norm of the matrices N((cid:96))Uk. We will present in Section 4 a method avoiding the computation of Uk. However, the method might still be too slow when a large number of groups is involved as it requires an estimation of a spectral norm for each group separately. It is thus interesting to characterise the performance of the proposed method using other quantities easier to compute. In this section, we present results involving the following quantity (23) ¯νp := max 1(cid:54)(cid:96)(cid:54)N(cid:110)p−1/2 (cid:96) (cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13)F(cid:111) . The only difference between νp and ¯νp is that we substituted the Frobenius norm for the spectral norm. As (cid:107)X(cid:107)2 more pessimistic than those involving νp. We have ¯νp (cid:62) k. Indeed, (cid:54) (cid:107)X(cid:107)F for any matrix X, we have νp (cid:54) ¯νp; hence the results involving ¯νp will be (24) k = (cid:107)Uk(cid:107)2 F = N(cid:88)(cid:96)=1(cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) 2 F = p(cid:96) ·(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 p(cid:96) F N(cid:88)(cid:96)=1 (cid:54) (cid:107)p(cid:107)1 · max 1(cid:54)(cid:96)(cid:54)s(cid:40)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 p(cid:96) F (cid:41) = ¯ν2 p. As with νp, the lower bound is also attained in, for example, the degenerated case where N1 = {1, . . . , n}. As νp (cid:54) ¯νp, we have the following corollary to Theorem 2.2. Corollary 2.3. Let M be a random subsampling matrix constructed as in (9) using the groups N1, . . . ,NN and the sampling distribution p. For any δ, ξ ∈ (0, 1), with probability at least 1 − ξ, (16) holds for all x ∈ span(Uk) provided that s (cid:62) 3 (25) Proof. As ¯νp (cid:62) νp, (25) implies (17). Theorem 2.2 then proves that (16) holds with probability at (cid:3) least 1 − ξ for all x ∈ span(Uk). (cid:62) k, Condition (25) suggests to always sample more than O(k log(k)) groups, while we know that sampling O(log(k)) can be enough. The interest of this result is thus in the regime where k is small. The sufficient condition (25) can be much more pessimistic than (17). Indeed, as we have ¯ν2 p p log(cid:18) 2k ξ (cid:19) . δ2 ¯ν2 As we have done it with νp, we can also optimise p to minimise ¯νp. The sampling distribution that minimises ¯νp is the distribution q∗ satisfying (26) and for which (27) q∗(cid:96) := (cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 k F , for all l ∈ {1, . . . , N}, ¯ν2 q∗ = k. With this distribution, (25) proves that sampling O(k log(k)) groups is always enough to sample all k-bandlimited signals. This result is particularly interesting at small k's because estimating q∗ is much easier and faster than estimating p∗ (see Section 4). Structured sampling and fast reconstruction of smooth graph signals 9 In some particular cases, this result can be interesting also at large k's. Indeed, (25) is a pessimistic q∗ (cid:54) k so that, according to (17), fewer samples than O(k log(k)) bound. In reality, we may have ν2 are actually sufficient when using q∗. Furthermore, q∗ might actually be close to p∗, in which case we would reach almost optimal results with q∗. The occurrence of these events is however quite difficult to predict and depends on the structure of the groups and of the graph. Note that we need the knowledge of the truncated Fourier matrix Uk to compute the distributions p∗ and q∗. Computing Uk can be intractable for large graphs. We will present a method to overcome this issue in Section 4. We continue first by explaining how to estimate x from its measurements. 3. Fast reconstruction Once the signal has been sampled, we shall also be able to reconstruct it. In [17], the authors propose to estimate the original signal by solving (28) min z∈Rn (cid:107)P(Mz − y)(cid:107)2 2 + γ z (cid:124) g(L)z, where γ > 0 and g is a nonnegative nondecreasing polynomial. We have (29) g(L) := d(cid:88)i=0 αi Li = (cid:124) αi UΛiU , d(cid:88)i=0 where α0, . . . , αd ∈ R are the coefficients of the polynomial and d ∈ N its degree. These polynomial's parameters can be tuned to improve the quality of the reconstruction. The role of g can be viewed as a filter on G and should ideally be a high-pass filter. The matrix P is introduced to account for the RIP. The advantage of this method is that it can be solved efficiently even for large graph for example by conjugate gradient. Indeed, each step of this algorithm can be implemented by using only matrix-vector multiplications with PM and L, which are both sparse matrices. The matrix g(L) does not have to be computed explicitly. For brevity, we do not recall the theorem proving that (28) provides accurate and stable recovery of all k-bandlimited signals. However, this theorem applies as soon as the restricted isometry prop- erty (16) holds, and thus applies here when (17) holds. The reconstruction quality is controlled by g(λk) and the ratio g(λk)/g(λk+1). One should seek to design a filter g such that these quantities are as close as possible to zero to improve the reconstruction quality. We propose now a method to obtain a faster estimation of the original signal when it is nearly piecewise constant. 3.1. Piecewise constant graph signals Before continuing, we want to stress that we consider non-overlapping groups N1, . . . ,NN in this rest of Section 3. If a graph signal is nearly piecewise constant over the groups N1, . . . ,NN then reconstructing the mean values of this signal for each group is enough to obtain a good approximation of the original signal. Instead of estimating n unknowns, we reduce the estimation to N unknowns. When N (cid:28) n, this is a large reduction of dimension yielding a significant speed up and memory reduction. The fact that a signal x is piecewise constant over the groups N1, . . . ,NN is characterized as follows. We construct the averaging row-vectors a((cid:96)) ∈ R1×n that satisfy (30) a((cid:96)) := 1 and the matrix (31) (cid:124) N((cid:96)) N(cid:96)1/2 , A := a(1) ... a(N )  ∈ RN×n. 10 G. PUY AND P. P´EREZ As the groups do not overlap, we have (cid:124) AA = I, (32) hence (cid:107)A(cid:107)2 = 1. Applying A to x provides N values, each one of them corresponding to the sum of the (cid:124) values of x within the group N(cid:96), scaled by N(cid:96)−1/2. Then, applying A to Ax gives an approximation (cid:124) Ax are constant within each group; this is a piecewise constant of x where the values in the vector A (cid:124) Ax appearing within the group N(cid:96) is exactly the average of x vector over the groups. The value of A within N(cid:96). Saying that x is nearly constant within each group corresponds to assume that (cid:124) (cid:107)A (33) where (cid:62) 0 is a small value. The signal model of interest in this section is thus (34) (cid:54) (cid:107)x(cid:107)2 , Ax − x(cid:107)2 (cid:124) A := {x ∈ span(Uk) (cid:107)(A (cid:54) (cid:107)x(cid:107)2} . A − I) x(cid:107)2 3.2. Reducing the dimension To build a fast algorithm exploiting the above property, we use a reconstruction method similar to (28) but involving vectors of smaller dimension. We define the averaged vector x := Ax ∈ RN of dimension N . As x ∈ A, estimating x is enough to get a good approximation of x -- we just need (cid:124) . Furthermore, as x is nearly piecewise constant over the groups N1, . . . ,NN , to multiply it with A by construction of the matrix M, the measurement vector y = Mx is also almost piecewise constant over the sampled groups Nω1, . . . ,Nωs. We thus average y over these groups by multiplying it with the matrix (cid:101)A ∈ Rs×m that satisfies (cid:101)Aji :=(cid:40) (cid:12)(cid:12)Nωj(cid:12)(cid:12)−1/2 We obtain (35) 0 (36) (cid:54) i (cid:54)(cid:80)(cid:96) otherwise. j(cid:48)=1(cid:12)(cid:12)(cid:12)Nωj(cid:48)(cid:12)(cid:12)(cid:12) , j(cid:48)=1(cid:12)(cid:12)(cid:12)Nωj(cid:48)(cid:12)(cid:12)(cid:12) for (cid:80)j−1 y :=(cid:101)Ay =(cid:101)AMx ∈ Rs. (cid:101)Mji :=(cid:26) 1 if i = ωj, otherwise. 0 y = (cid:101)M x. y = (cid:101)M x + n, We now have to link y to x. We create the matrix (cid:101)M ∈ Rs×N that restricts the N mean value of x to the s mean value of the selected groups, i.e., (37) We have therefore (38) The goal is now to estimate x from y. To ensure that the reconstruction method is stable to measurement noise, we do not consider the perfect scenario above but instead the scenario where (39) and n ∈ Rs models noise. We now need a regularisation term to estimate x. We obtain this term by reducing the dimension of the regularisation involving the Laplacian L in (28). We compute (40) (41) (cid:124) (cid:124) = (AU) g(Λ) (AU) ∈ RN×N . (cid:101)L := A g(L) A Note that(cid:101)L is a symmetric positive definite matrix. Like g(L), it can thus be used as a regularisation. We thus propose to estimate x by solving 2 where γ > 0 and (cid:101)P ∈ Rs×s is the diagonal matrix with entries satisfying (42) (cid:124) Let x∗ ∈ RN be a solution of (41). We finally obtain an estimation of x by computing A ωj . x∗. + γ z (cid:124)(cid:101)L z, 2 min z∈RN(cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M z − y)(cid:13)(cid:13)(cid:13) (cid:101)Pjj := p−1/2 Structured sampling and fast reconstruction of smooth graph signals 11 In the particular case where g(·) is the identity, one can notice that (43) (cid:101)L(cid:96)(cid:96)(cid:48) = (cid:88)(i,j)∈N(cid:96)×N(cid:96)(cid:48) Lij N(cid:96)1/2 N(cid:96)(cid:48)1/2 preserves the connections present in the original graph represented by L. The dimension of the unknown vector in (41) is N , which can be much smaller than n. This is non-zero only if there is at least one edge in E joining the groups N(cid:96) and N(cid:96)(cid:48). The Laplacian(cid:101)L(cid:96)(cid:96)(cid:48) leads to a large gain in memory and computation time when either(cid:101)L or matrix-vector multiplications with (cid:101)L can be computed rapidly. If g has a small degree, then (cid:101)L can be computed explicitly in a implementation for matrix-vector multiplications with(cid:101)L without temporary going back to the signal (cid:124) domain of dimension n, i.e., multiplying the vector z with A , filtering the high dimensional signal (cid:124) z, and downsampling the result. Even though solving (41) might still be faster than solving (28) in A this situation, we loose part of the efficiency by working temporarily in the high dimensional domain. We thus have less flexibility in the choice of g with this reconstruction technique. short amount of time. In such a case, we can solve (41) faster than (28) as it involves matrices of (much) smaller dimensions. In general, it is however not always straightforward to find an a efficient Let us also mention that (41) can be used to initialise the algorithm used to solve (28) with a good approximate solution. As in multigrid approaches to solve linear systems of equations, see, e.g., [28, 29]. The following theorem bounds the error between the signal recovered by (41) and the original vector. Theorem 3.1. Let Ω = {ω1, . . . , ωs} be a set of s indices selected independently from {1, . . . , N} using a sampling distribution p ∈ RN , M, P,(cid:101)M,(cid:101)P be the associated matrices constructed respectively (cid:54) Mmax. Let ξ, δ ∈ (0, 1) in (9), (12), (37) and (42), and Mmax > 0 be a constant such that (cid:107)PM(cid:107)2 and suppose that s satisfies (17). With probability at least 1− ξ, the following holds for all x ∈ A, all n ∈ Rs, all γ > 0, and all nonnegative nondecreasing polynomial functions g such that g(λk+1) > 0. Let x∗ be the solution of (41) with y = (cid:101)MAx + n. Define α∗ := UkU x∗ and β∗ := (I − (cid:124) (cid:124) k) A UkU +(cid:32)Mmax(cid:115) g(λk) (cid:107)α∗ − x(cid:107)2 x∗. Then, g(λk+1) (cid:124) (cid:124) k A (cid:54) +(cid:112)γg(λk)(cid:33)(cid:107)x(cid:107)2 1 (cid:112)s(1 − δ) ·(cid:34)(cid:32)2 + Mmax (cid:112)γg(λk+1)(cid:33)(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13)2 + (cid:32)2Mmax + Mmax(cid:115) g(λn) g(λk+1) +(cid:112)γg(λn)(cid:33)(cid:107)x(cid:107)2(cid:35) , (cid:54) (cid:107)β∗(cid:107)2 1 (cid:112)γg(λk+1)(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13)2 + (cid:115) g(λk) g(λk+1) (cid:107)x(cid:107)2 + (cid:115) g(λn) g(λk+1) (cid:107)x(cid:107)2 . (cid:3) (cid:124) The vector α∗ is the orthogonal projection of A x∗ onto span(Uk). The vector β∗ is the projection x∗ onto the orthogonal complement of span(Uk). There are several remarks to make about the Proof. See Appendix B. (cid:124) of A above theorem: (44) and (45) • Theorem 3.1 shows that the result obtained via (41) is similar to the one we would have obtained by solving (28) -- see [17] for the error bounds -- with additional errors controlled by . We recall that characterises how far x is from a piecewise constant signal. As expected, the smaller , the better the reconstruction. • The reconstruction quality improves when g(λk) and the ratio g(λk)/g(λk+1) go to 0, and when g(λn)/g(λk+1) tends to 1. We recall that we have g(λn) (cid:62) g(λk+1) > g(λk) by assumption. • The effect of the noise n decreases when g(λk+1) increases, and, obviously, γ should be adapted to the signal-to-noise ratio. 12 G. PUY AND P. P´EREZ Let us mention that the idea of "coarsening" a graph and the signals that live on it using a partition into different groups of nodes can also be found in [30], where a multiresolution analysis method of graph signals is proposed. The coarsening method is however different than the one used here. 4. Optimal sampling estimation In this section, we come back to the sampling process of Section 2 and leave the reconstruction problem. We explain how to estimate the sampling distributions p∗ and q∗ of Section 2 without computing the truncated Fourier matrix Uk, as this computation is intractable for large graphs. The methods below only involve matrix-vector multiplications with the sparse Laplacian matrix L and are thus computationally tractable even for large n. 4.1. Estimation of p∗ The distribution p∗, defined in (18), that minimises the coherence νp is entirely defined by the values of (cid:107)N((cid:96))Uk(cid:107)2 for (cid:96) = 1, . . . , N , which are thus the quantities we need to evaluate. the signal x filtered by h is We recall that, in graph signal processing, a filter is represented by a function h : R → R, and that (46) (cid:124) xh := U diag(h) U x ∈ Rn, (cid:124) where h = (h(λ1), . . . , h(λn)) Fourier transform of x, we can approximate the function h by a polynomial ∈ Rn. To filter the signal x without actually computing the graph (47) r(t) := d(cid:88)i=0 αi ti ≈ h(t) of degree d, and compute xr instead of xh. The filtered signal xr is computed rapidly using the formula (48) xr = αi U diag(λi 1, . . . , λi (cid:124) n) U x = αi Lix, d(cid:88)i=0 d(cid:88)i=0 that involves only matrix-vector multiplications with the sparse Laplacian matrix L. We let the reader refer to [31] for more information on this fast filtering technique. For any polynomial function r of the form above and any matrix A ∈ Rn×n, we define Let iλk : R → R be the ideal low-pass filter at cutoff frequency λk, i.e., the filter that satisfies αi Ai. d(cid:88)i=0 if t (cid:54) λk, otherwise. (49) r(A) := (cid:124) Note that r(L) = U r(Λ) U . (50) We have UkU (51) (cid:124) k = iλk (L). Then, we notice that (cid:124) kN((cid:96)) 2 0 iλk (t) =(cid:26) 1 =(cid:13)(cid:13)(cid:13)N((cid:96))UkU (cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) 2 (cid:124)(cid:13)(cid:13)(cid:13)2 =(cid:13)(cid:13)(cid:13)N((cid:96)) iλk (L) N((cid:96)) (cid:124)(cid:13)(cid:13)(cid:13)2 . We recall that N((cid:96)) is the matrix that restricts the signal to the nodes belonging to N(cid:96). The matrix appearing on the right hand side of the last equality corresponds to the linear operator that 1) extends a vector on the complete graph by inserting 0 in all groups (cid:96)(cid:48) (cid:54)= (cid:96), 2) low-pass filters the extended signal, 3) restricts the result to the group N(cid:96). This process can be approximated by replacing the ideal low-pass filter iλk with a polynomial approximation iλk of iλk and (52) (cid:13)(cid:13)(cid:13)N((cid:96)) iλk (L) N((cid:96)) (cid:124)(cid:13)(cid:13)(cid:13)2 ≈(cid:13)(cid:13)(cid:13)N((cid:96)) iλk (L) N((cid:96)) (cid:124)(cid:13)(cid:13)(cid:13)2 . (cid:80)N Structured sampling and fast reconstruction of smooth graph signals 13 To estimate p∗, we estimate the spectral norm of the matrix appearing on the right hand side for which matrix-vector multiplication is fast. The quality of the approximation depends obviously on the choice of the polynomial iλk . (cid:124) (cid:124) Estimating (cid:107)N((cid:96)) iλk (L) N((cid:96)) (cid:107)2 amounts to computing the largest eigenvalue of N((cid:96)) iλk (L) N((cid:96)) which can be done, e.g., by using the power method. This method requires matrix-vector multiplica- tion only with N((cid:96)) and iλk (L) and is thus fast. Finally, the approximation ¯p ∈ RN of p∗ satisfies (53) (cid:124) . ¯p(cid:96) := (cid:124) λmax(N((cid:96)) iλk (L) N((cid:96)) ) (cid:96)(cid:48)=1 λmax(N((cid:96)(cid:48)) iλk (L) N((cid:96)(cid:48)) ) Note that an estimation of λk is required beforehand to define the filter iλk . We estimate this value using the dichotomy method presented in [17]. 4.2. Estimation of q∗ Computing ¯p requires the estimation of N eigenvalues. Even though these estimations can be done in parallel, this process might still be too slow for certain applications. As explained before, when k is small, we can use the sampling distribution q∗ in (26) that minimises ¯νp. This distribution is faster to compute than ¯p. We start by noticing that we have 2 (54) (cid:13)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)(cid:13) = (cid:88)i∈N(cid:96) for each group (cid:96) = 1, . . . , N . The vector δi ∈ Rn is the unit vector that is null on all nodes except (cid:124) kδi(cid:107)2 2, i = 1, . . . , n, to estimate q∗. An algorithm at node i. Hence, we only need an estimation of (cid:107)U was already proposed in [17] to estimate these values. We let the reader refer to Algorithm 1 in [17] for the details of the method. We just recall that this estimation is obtained by filtering O(log(n)) random signals with a polynomial approximation of iλk . Finally, our estimation ¯q ∈ RN of q∗ has entries (cid:124) kδi(cid:107)2 (cid:107)U F 2 (cid:124) ¯q(cid:96) := (cid:80)i∈N(cid:96) (cid:107)U kδi(cid:107)2 (cid:124) (cid:80)N (cid:96)=1(cid:80)i∈N(cid:96) (cid:107)U kδi(cid:107)2 F F , (55) 2 are estimated by Algorithm 1 in [17]. (cid:124) kδi(cid:107)2 where each (cid:107)U This estimation is faster than for ¯p because the power method is an iterative method that involves one filtering at each iteration. Furthermore, the power method is run independently for each group N(cid:96). In total, (much) more than N(cid:96) filterings are thus required. On the contrary, for ¯q, we just need to filter O(log(n)) signals to obtain the estimation. In most situations, we already have O(log(n)) (cid:54) N(cid:96) and computing ¯q is thus faster than computing ¯p. 5. Experiments In this last section, we first test our sampling strategies on two different graphs to illustrate the effect of the different sampling distributions on the minimum number of samples required to ensure that the RIP holds. Then, we apply our sampling strategy for user-guided object segmentation. In this application, we also test the different recovery techniques proposed in Section 3. 5.1. Sampling distributions We perform experiments on two different graphs: the Minnesota graph of size n = 2642 and the bunny graph of size n = 2503. Both graphs are presented in Fig. 3 and are available in the GSP toolbox [32]. For each graph, we group the nodes using the spatial coordinates associated to each node. For the Minnesota graph, we divide the space into 100 cells and group the nodes that fall in the same cell. After removing empty cells, we obtain the N = 73 groups represented in Fig. 3. For the bunny graph, we obtain N = 213 groups with a similar procedure (see Fig. 3). 14 G. PUY AND P. P´EREZ Figure 3. First column from the left: Minnesota graph (top); bunny graph (bottom). The groups N1, . . . ,NN are indicated by different colors. Other columns: probability that δk is less that 0.995 as a function of s. The curves in black, red, blue and green are obtained using the sampling distributions u, p∗, ¯q and ¯p, respectively. The top row shows the result for the Minnesota graph. The bottom row shows the result for the bunny graph. The bandlimit k is indicated on top of each curve. For each graph, we compute the combinatorial Laplacian and Uk for different values of k. Then, we compute the lower RIP constant, i.e., the constant δk > 0 that satisfies δk = 1 − 1 s inf x∈span(Uk) (cid:107)x(cid:107)2=1 (cid:107)PM x(cid:107)2 2 . This constant is the smallest value that δ can take such that the left-hand side of the RIP (16) holds. Remark that (cid:124) (cid:124) kM λmin(cid:0)U P2MUk(cid:1) . (56) δk = 1 − 1 s We estimate δk for 500 independent draws of the set Ω, which defines the matrices PM, and different numbers of selected groups s. All samplings are done in the conditions of Theorem 2.2 using the sampling distributions u, p∗, ¯p and ¯q. The vector u denotes the uniform distribution over {1, . . . , N}. When conducting this experiment with the estimated distributions ¯p and ¯q, we re-estimate these distributions at each of the 500 trials. These distributions are estimated using Jackson-Chebychev polynomials of order 50 [33]. For the Minnesota graph, we consider the bandlimits k = 5, 10, 20. For the bunny graph, we consider the bandlimits k = 10, 25, 50. We present the probability that δk is less than 0.995, estimated over the 500 draws of Ω, as a function of s in Fig. 3. For the Minnesota graph, the performance is better when using the optimal distribution p∗ than when using the uniform distribution u for all k, which is in line with the theory. The estimated ¯p and ¯q yield performance equivalent to p∗. This confirms that we can achieve similar sampling performance without having to compute the Fourier matrix Uk, which, we recall, is intractable for large graphs. This also shows that ¯q can lead to nearly optimal results. For the bunny graph, all sampling distributions yield essentially the same results at all bandlimits. We notice a slight improvement at k = 50 when using ¯p, ¯q or p∗ instead of u. For illustration, we present in Fig. 4 examples of computed sampling distributions p∗, ¯p and ¯q. All sampling distributions exhibit similar structures, which explains why they all yield about the same performance in our experiments. 1020304050607000.20.40.60.81k = 5s1 − ε1020304050607000.20.40.60.81k = 10s1 − ε1020304050607000.20.40.60.81k = 20s1 − ε 5010015020000.20.40.60.81k = 10s1 − ε5010015020000.20.40.60.81k = 25s1 − ε5010015020000.20.40.60.81k = 50s1 − εStructured sampling and fast reconstruction of smooth graph signals 15 p∗ ¯p ¯q Figure 4. Example of sampling distributions. Top panels: p∗ (left), ¯p (middle), and ¯q (right) for the Minnesota graph at k = 10. Bottom panels: p∗ (left), ¯p (middle), and ¯q (right) for the bunny graph at k = 25. 5.2. Object segmentation 5.2.1. Protocol We now test our method for interactive object segmentation. We consider the image of size n = 321×481 presented in Fig. 1 for which our goal is to segment the tiger. The ground truth segmentation map x ∈ {0, 1}n is presented in Fig. 5. The value 1 (white) indicates the presence of the tiger and the value 0 (black) stands for the foreground. The original image and the ground truth image are part of the dataset available2 in [34]. Our objective is to recover the original map x from few user-inputs. To facilitate the interactions with the user, we divide the original image into the N = 600 superpixels showed in Fig. 1 and computed with SLIC [23], choose a small number of superpixels at random and ask the user to label these superpixels: 1 if the superpixel belongs to the tiger; 0 otherwise. The graph G used to propagate the user-labels to the complete image is constructed as follows. We build a feature vector for each pixel by extracting a color RGB patch of size 3 × 3 around the pixel, transform this patch in vector form, and augment this vector with the absolute 2D coordinates of the pixels in this extracted patch. This yields n feature vectors gi ∈ R45, i = 1, . . . , n. We then connect each feature vector to its 9 nearest neighbours (in the Euclidean sense), which gives a set of 9n edges 2/σ2], where σ > 0 is the 25th E. The adjacency matrix W ∈ Rn×n satisfies Wij := exp[−(cid:107)gi − gj(cid:107)2 percentile of the set {(cid:107)gi − gj(cid:107)2 : (i, j) ∈ E}. We finally symmetrise the matrix W and compute the combinatorial Laplacian L ∈ Rn×n. We study three strategies to choose the superpixels. The first strategy consists in choosing the superpixels uniformly at random, i.e., using the sampling distribution u. The second and third strategies consist in choosing the superpixels with respectively the optimised distributions ¯q and ¯p, which we evaluate at k0 = 50 using Jackson-Chebychev polynomials of order 75 [33]. For illustration, we present in Fig. 5 the estimated values (cid:107)N((cid:96))Uk0(cid:107)2 2, which define the optimised distributions ¯q and ¯p, respectively. Both distributions indicate that one should label more superpixels around the tiger. The distribution ¯q "focuses" however more on specific regions, like the head of tiger. The distribution ¯p spreads the measurements over the entire tiger more uniformly. F and (cid:107)N((cid:96))Uk0(cid:107)2 2http://www.ntu.edu.sg/home/asjfcai/Benchmark_Website/benchmark_index.html. 0.010.020.030.040.050.060.070.08 0.0050.010.0150.020.0250.030.0350.040.0450.050.055 0.010.020.030.040.050.060.070.080.090.10.11 0.0020.0040.0060.0080.010.0120.0140.0160.0180.02 0.0020.0040.0060.0080.010.0120.0140.0160.018 0.0020.0040.0060.0080.010.0120.0140.0160.0180.020.022Original image G. PUY AND P. P´EREZ (cid:13)(cid:13)(cid:13)N((cid:96))Uk0 (cid:13)(cid:13)(cid:13)2 2 values (cid:13)(cid:13)(cid:13)N((cid:96))Uk0 (cid:13)(cid:13)(cid:13)2 F values u ¯p ¯q 16 p o T ) 1 4 ( g n i v l o s t l u s e R ) 8 2 ( g n i v l o s t l u s e R Figure 5. Top row: ground truth segmentation map (left); local group coherence map (cid:107)N((cid:96))Uk0(cid:107)2 2 evaluated with the spectral norm (middle); local group coherence map (cid:107)N((cid:96))Uk0(cid:107)2 F evaluated with the Frobenius norm (right). Middle row: segmentation map estimated from s = 50 sampled superpixels obtained by solving (41) and using a uniform sampling (left), ¯p (middle), ¯q (right). Bottom row: segmentation map estimated from s = 50 sampled superpixels obtained by solving (28) and using a uniform sampling (left), ¯p (middle), ¯q (right). We emulate user-interactions as follows. For each chosen superpixel, we compute the mean of the ground truth map x within this superpixel. If the mean value is larger than 0.5, we label the superpixel as part of the tiger. Otherwise, we label the superpixel as part of the background. This strategy obviously introduces noise if some superpixels cover part of the background and of the tiger. Once the labelling is done, we have access to the measurement vector y ∈ Rm from which we want to reconstruct x. We repeat this procedure for s ∈ {50, 70, . . . , 250}. For each s, we also repeat the experiments 50 times with independent draws of the superpixels. We draw the superpixels with replacements in all cases. To reconstruct the original map x, we first use the fast reconstruction method (41) and then refine the solution at the pixel level with (28), using the solution of the first minimisation problem as initialisation to solve the second minimisation problem. We choose g(L) = L and solve both problems in the limit where γ → 0. In this limit, the problems (41) and (28) become (57) subject to z min z∈RN (cid:124)(cid:101)L z (cid:101)M z = y and (58) (cid:124) z min z∈Rn g(L)z subject to Mz = y, respectively. Both problems are solved using FISTA [35]. The same stopping criteria are used for all experiments. 0.0150.0300.0450.0600.0750.0900.1050.1200.1350.250.500.751.001.251.501.752.00Structured sampling and fast reconstruction of smooth graph signals 17 Figure 6. Left: the curves represent the mean reconstruction snr as a function of the number of sampled superpixels s. The dash-dotted curves are obtained by solving (28). The continuous curves are obtained by solving (41). Right: mean computation time in seconds as a function of s. The dash-dotted curves are obtained by solving (28). The continuous curves are obtained by solving (41) using the result of (28) as initial point. In both graphs, the errorbars are at one standard deviation. The black curves are obtained with the uniform sampling distribution u. The blue curves are obtained with ¯q. The red curves are obtained with ¯p. 5.2.2. Results We present in the top panel of Fig. 6 the reconstruction snr obtained with the different methods. The reconstruction snr is defined as −20 log10(−(cid:107)x − x∗(cid:107)2 /(cid:107)x(cid:107)2), where x∗ is the reconstructed signal. We notice that the snr attained with the fast decoder (41) is very similar to the snr attained with (28). We also remark that the optimised distributions ¯p and ¯q yield better reconstructions than the uniform distribution u. The mean reconstruction snr is slightly better with ¯p than with ¯q at s (cid:62) 150. We present the computation time of each method in the bottom panel of Fig. 6. We notice that solving (41) is much faster than solving (28), while they yield almost the same quality. This highlight the interest of the fast reconstruction technique. It is also interesting to note that it is faster to solve (28) when the measurements are drawn with ¯p or with ¯q than with u. The reason is probably a better initialisation of (28) or a better "quality" of the measurements with the optimised distributions than with the uniform distribution. Finally, we present in Fig. 5 some examples of reconstructions from s = 150 sampled superpixels for each method. We notice that the optimised sampling distributions improve the reconstruction of x around the head and tail of the tiger, i.e., where the optimised distributions have higher values. With a uniform distribution, the structure of the graph makes it difficult to reconstruct x around the head and tail from the values of other superpixels. The optimised sampling distribution compensate this issue by favouring this area when selecting the measurements. 6. Discussion and Conclusion We presented a structured sampling strategy for k-bandlimited signals where the nodes are selected by groups. We proved that the local group graph cumulative coherence quantifies the importance of sampling each group to ensure a stable embedding of all k-bandlimited signals. Finally, we presented a fast reconstruction technique for k-bandlimited signals which are also nearly piecewise-constant over pre-defined groups of nodes. Among the possible applications of these methods, we believe that they can also be useful to accelerate the compressive spectral clustering method proposed in [36]. After having computed some features vectors for each nodes, this compressive method works by downsampling the set of features of vectors, performing k-means on this reduced set to find k clusters, and interpolating the clustering results on all nodes by solving (28). To accelerate the method, one could 1) pre-group similar nodes to form N groups such that k (cid:54) N (cid:28) n, e.g., by running few iterations of the k-means algorithm; 2) 50100150200250s678910snr50100150200250s0.00.51.01.52.02.53.03.54.0Time (s)18 G. PUY AND P. P´EREZ subsample this set of N groups; 3) cluster this subset to find k clusters; and 4) solve (41) to cluster all nodes. If the overhead of computing the N groups is small, this method has the potential to be faster than the original compressive spectral clustering method. Finally, we would like to discuss two limitations in the proposed methods. First, the optimal sampling distribution depends on the parameter k. In some applications, the final result may change a lot depending on the value of k which was chosen to compute this distribution. Finding a range of values of k which give acceptable and stable results is thus an important step in every application. Second, the estimation of the optimal sampling distribution depends on the quality of the polynomial approximation of the ideal low-pass filter λλk . It is sometimes necessary to use a polynomial of large degree to get a correct estimation, which limits the computational efficiency of the proposed methods. In such cases, it would be especially useful to find more efficient alternatives to estimate the distributions p∗ and q∗. Appendix A - Proof of the Theorem 2.2 As done in [17], the proof is obtained by applying the following lemma obtained by Tropp in [37]. Lemma A.1 (Theorem 1.1, [37]). Consider a finite sequence {Xj} of independent, random, self- adjoint, positive semi-definite matrices of dimension d× d. Assume that each random matrix satisfies λmax(Xj) (cid:54) R almost surely. Define E Xj . E Xj and µmax := λmax(cid:88)j (1 − δ)1−δ(cid:21) µmin (1 + δ)1+δ(cid:21) µmax (cid:54) d(cid:20) (cid:54) d(cid:20) e−δ eδ R R for δ ∈ [0, 1], for δ (cid:62) 0. δ2µmax 3 R (cid:19) . eδ (1 + δ)1+δ(cid:21)µmax/R (cid:54) exp(cid:18)− P(ωj )(cid:17)(cid:16)P(ωj )N(ωl)Uk(cid:17) . (cid:124) and X := Xj = (cid:124) (cid:124) U kM 1 s P2MUk. s(cid:88)j=1 (59) Then (60) (61) (62) µmin := λmin(cid:88)j P λmin(cid:88)j and P λmax(cid:88)j (1 − δ)1−δ(cid:21)µmin/R (cid:54) exp(cid:18)− e−δ (cid:20) (63) We define (64) Xj := (cid:124) U kM 1 s (cid:124) (cid:124) kN(ωj ) 1 s(cid:16)U We will also use the facts that, for all δ ∈ [0, 1], Proof of Theorem 2.2. We start by noticing that Xj (cid:54) (1 − δ)µmin Xj (cid:62) (1 + δ)µmax 3 R (cid:19) and (cid:20) s(cid:88)j=1(cid:16)U P(ωj )(cid:17)(cid:16)P(ωj )N(ωj )Uk(cid:17) (cid:124) kN(ωj ) PPMUk = δ2µmin 1 s (cid:124) The matrix X is thus a sum of s independent, random, self-adjoint, positive semi-definite matrices. We are in the setting of Lemma A.1. We continue by computing E Xj and λmax(Xj). The expected value of each Xj is E Xj = E(cid:20) 1 s(cid:16)U k(cid:32) N(cid:88)(cid:96)=1 1 s = U (cid:124) (cid:124) (cid:124) kN(ωj ) N((cid:96)) P(ωj )(cid:17)(cid:16)P(ωj )N(ωj )Uk(cid:17)(cid:21) = N((cid:96))(cid:33) Uk = (cid:124) U kUk = 1 s 1 s I. (cid:124) (65) 1 s (cid:124) U k(cid:32) N(cid:88)(cid:96)=1 (cid:124) p(cid:96)(cid:16)N((cid:96)) P((cid:96))(cid:17)(cid:16)P((cid:96))N((cid:96))(cid:17)(cid:33) Uk Structured sampling and fast reconstruction of smooth graph signals 19 Therefore, λmin(cid:16)(cid:80)j E Xj(cid:17) = 1 and λmax(cid:16)(cid:80)j E Xj(cid:17) = 1. Furthermore, for all j = 1, . . . , s, we have 2 (66) λmax(Xj) = (cid:107)Xj(cid:107)2 Lemma A.1 yields, for any δ ∈ (0, 1), P((cid:96))N((cid:96))Uk s (cid:54) max 1(cid:54)(cid:96)(cid:54)N(cid:13)(cid:13)(cid:13)(cid:13) 2 2 = p(cid:96) 1 s max (cid:13)(cid:13)(cid:13)(cid:13) (1 − δ)1−δ(cid:21)s/ν2 (1 + δ)1+δ(cid:21)s/ν2 1(cid:54)(cid:96)(cid:54)N(cid:40)(cid:13)(cid:13)N((cid:96))Uk(cid:13)(cid:13)2 p (cid:54) k exp(cid:18)− p (cid:54) k exp(cid:18)− e−δ eδ (cid:41) = ν2 p s δ2s 3 ν2 δ2s 3 ν2 p(cid:19) p(cid:19) . . (cid:3) (67) (68) P{λmin (X) (cid:54) (1 − δ)} (cid:54) k ·(cid:20) and P{λmax (X) (cid:62) (1 + δ)} (cid:54) k ·(cid:20) 1 − δ (cid:54) λmin (X) Therefore, for any δ ∈ (0, 1), we have, with probability at least 1 − ξ, and λmax (X) (cid:54) 1 + δ (69) provided that (17) holds. Noticing that (69) implies that (70) (1 − δ)(cid:107)α(cid:107)2 2 (cid:54) 1 s (cid:107)PMUkα(cid:107)2 2 (cid:54) (1 + δ)(cid:107)α(cid:107)2 2 , for all α ∈ Rk, which is equivalent to (16) for all x ∈ span(Uk), terminates the proof. Appendix B - Proof of Theorem 3.1 In order to prove Theorem 3.1, we need to establish few properties between the different matrices used in this work. The first useful property is (71) Indeed, for any z ∈ Rn, the jth entry of (cid:101)P(cid:101)MAz is (cid:16)(cid:101)P(cid:101)MAz(cid:17)j (72) = (cid:101)P(cid:101)MA =(cid:101)APM. (cid:124) 1 N(ωj )z . (cid:0)pωj (cid:12)(cid:12)Nωj(cid:12)(cid:12)(cid:1)1/2 (cid:124) = I, (73) (cid:101)A(cid:101)A . Therefore, The third property is The second property is PMz, which is p−1/2 ωj N(ωj )z. From the definition of (cid:101)A, the sum is scaled by (cid:12)(cid:12)Nωj(cid:12)(cid:12)−1/2 Then, the jth entry of (cid:101)APMz is the scaled sum of the values in the jth sampled group appearing in ((cid:101)P(cid:101)MAz)j = ((cid:101)APMz)j for all j ∈ {1, . . . , s}, which terminates the proof. which implies (cid:107)(cid:101)A(cid:107)2 = 1. z(cid:107)2 , (cid:124)(cid:101)P(cid:101)M z. Indeed, the entries in (cid:124) for all z ∈ RN . To prove this property, we remark that PMA (cid:124) z corresponding to the first selected group Nω1 are all equal to (pω1 Nω1)−1/2 zω1. There PMA (cid:124)(cid:101)P(cid:101)M z are also equal to are Nω1 such entries. One can also notice that the first Nω1 entries of (cid:101)A (pωj Nω1)−1/2 zω1. Repeating this reasoning for all the sampled groups proves the equality. On the (cid:124)(cid:101)P(cid:101)M z(cid:107)2 = (cid:107)(cid:101)P(cid:101)M z(cid:107)2, where we used (73). On the other side, (cid:124) z(cid:107)2 = (cid:107)(cid:101)A one side, we thus have (cid:107)PMA we have (cid:107)(cid:101)APMA z(cid:107)2 = (cid:107)(cid:101)P(cid:101)M z(cid:107)2, where we used (71) and (32). This terminates the z(cid:107)2 = (cid:107)(cid:101)P(cid:101)MAA (cid:124) z(cid:107)2 = (cid:107)PMA (cid:107)(cid:101)APMA z = (cid:101)A proof. (74) (cid:124) (cid:124) (cid:124) 20 G. PUY AND P. P´EREZ Proof of Theorem 3.1. As x∗ is a minimiser of (41), we have To prove the theorem, we need to lower and upper bound the left and right hand sides of (75), 2 2 2 2 (75) + γ x + γ ( x∗) (cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x∗ − y)(cid:13)(cid:13)(cid:13) (cid:124)(cid:101)L x∗ (cid:54) (cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x − y)(cid:13)(cid:13)(cid:13) respectively. We start with the bound involving(cid:101)L and then with the ones involving (cid:101)M. [Bounding the terms in (75) involving (cid:101)L]. We define the matrices ¯Uk := (uk+1, . . . , un) ∈ Rn×(n−k), Gk := diag (g(λ1), . . . , g(λk)) ∈ Rk×k, ¯Gk := diag (g(λk+1), . . . , g(λn)) ∈ R(n−k)×(n−k). (cid:124) By definition of α∗ and β∗, A (cid:124)(cid:101)L x. (77) (78) (76) x∗ = α∗ + β∗ with α∗ ∈ span(Uk) and β∗ ∈ span(¯Uk). We recall that (79) . Therefore, we obtain (cid:124) (cid:124) kα∗) + (¯U kβ∗) Gk (U (cid:124) In the first step, we used the facts that ¯U kα∗ = 0 and U (cid:124)(cid:101)L x∗ = (U (cid:124) kα∗) (cid:124) kβ∗) (cid:62) g(λk+1)(cid:107)β∗(cid:107)2 2 . (cid:124) (cid:124) ¯Gk (¯U (cid:124) kβ∗ = 0. The second step follows form the (cid:124) ( x∗) (cid:101)L = (AU) g(Λ) (AU) fact that(cid:13)(cid:13)¯U (cid:124)(cid:101)L x = (U (cid:124) A x (cid:124) (80) (cid:124) kβ∗(cid:13)(cid:13)2 = (cid:107)β∗(cid:107)2. We also have x) (cid:54) g(λk)(cid:107)UkA Ax(cid:13)(cid:13)2 (cid:54) g(λk)(cid:107)A (cid:54) g(λk)(cid:107)x(cid:107)2 (cid:124) g(Λ) (U Ax(cid:107)2 2 + 2g(λn)(cid:107)x(cid:107)2 2 . 2 + g(λn)(cid:13)(cid:13)¯U (cid:124) x) (cid:124) (cid:124) (cid:124) kA (cid:124) A 2 (cid:124) x(cid:107)2 (cid:54) g(λk)(cid:107)x(cid:107)2 (cid:124) (cid:124) kA 2 + g(λn)(cid:13)(cid:13)¯U x(cid:13)(cid:13)2 2 + g(λn)(cid:104)(cid:13)(cid:13)¯U 2 +(cid:13)(cid:13)¯U kx(cid:13)(cid:13)2 2 (cid:124) (cid:124) (cid:124) k(A The second inequality follows from the facts that (cid:107)Uk(cid:107)2 = 1 and x = Ax. To obtain the third (cid:124) 2 2 (81) inequality, we used (cid:107)A(cid:107)2 = 1 and the triangle inequality. For the last step, notice that(cid:13)(cid:13)¯U (as x ∈ span(Uk)),(cid:13)(cid:13)¯Uk(cid:13)(cid:13)2 = 1 and use (33). [Bounding the terms in (75) involving (cid:101)M]. By definition of y, it is immediate that (cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x − y)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13) For the other term involving (cid:101)M, the triangle inequality yields (cid:62)(cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)M x∗ −(cid:101)P(cid:101)MAx(cid:13)(cid:13)(cid:13)2 −(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13)2 =(cid:13)(cid:13)(cid:13)(cid:101)APMA x∗ −(cid:101)P(cid:101)MAx(cid:13)(cid:13)(cid:13)2 Ax(cid:13)(cid:13)(cid:13)2 −(cid:13)(cid:13)(cid:13)APMA (cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)M x∗ −(cid:101)P y(cid:13)(cid:13)(cid:13)2 =(cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)MAA (cid:62)(cid:13)(cid:13)(cid:13)APMA (cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)M x∗ −(cid:101)P(cid:101)MAx(cid:13)(cid:13)(cid:13)2 (cid:124) The first equality follows from AA the last step. To summarise, we are at (cid:124) x∗ − APMA Then, we have (82) (83) (cid:124) (cid:124) (cid:124) 2 2 . . = I, the second from (71), and the triangle inequality was used in 2(cid:105) Ax − x)(cid:13)(cid:13)2 kx(cid:13)(cid:13)2 = 0 . (cid:124) x∗ −(cid:101)APMx(cid:13)(cid:13)(cid:13)2 Ax − APMx(cid:13)(cid:13)(cid:13)2 Ax − APMx(cid:13)(cid:13)(cid:13)2 −(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13)2 . (84) (cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)M x∗ −(cid:101)P y(cid:13)(cid:13)(cid:13)2 (cid:124) (cid:124) x∗− APMA We continue by lower bounding (cid:107)APMA separately. (cid:124) x∗ − APMA Ax(cid:107)2]. Equality (74) yields (cid:124) (cid:124) (cid:124) x∗ − APMA x∗ − PMA (cid:124) [Lower bound on (cid:107)APMA (85) Ax(cid:107)2 . (cid:124) Ax(cid:107)2 and upper bounding (cid:107)APMA Ax− APMx(cid:107)2 (cid:62)(cid:13)(cid:13)(cid:13)APMA (cid:13)(cid:13)(cid:13)APMA (cid:124) (cid:124) x∗ − APMA (cid:124) Ax(cid:13)(cid:13)(cid:13)2 −(cid:13)(cid:13)(cid:13)APMA Ax(cid:13)(cid:13)(cid:13)2 (cid:124) = (cid:107)PMA Structured sampling and fast reconstruction of smooth graph signals 21 Using the triangle inequality and the fact that x ∈ A, we obtain (cid:124) (cid:107)PMA (cid:124) x∗ − PMA Ax(cid:107)2 (cid:62) (cid:107)PMA (cid:62) (cid:107)PMA (cid:62) (cid:107)PMA (cid:124) (cid:124) (cid:124) (cid:124) x∗ − PMx(cid:107)2 − (cid:107)PMx − PMA (cid:124) x∗ − PMx(cid:107)2 − (cid:107)PM(cid:107)2 (cid:107)A x∗ − PMx(cid:107)2 − Mmax (cid:107)x(cid:107)2 . Ax(cid:107)2 Ax − x(cid:107)2 The restricted isometry property (16) then yields (cid:124) (cid:107)PMA (87) (cid:124) We used the equality A (88) (cid:124) (cid:107)APMA (cid:124) x∗ − APMA (cid:124) [Upper bounding (cid:107)APMA x∗ − PMx(cid:107)2 = (cid:107)PM (α∗ − x) + PMβ∗(cid:107)2 (cid:62) (cid:107)PM(α∗ − x)(cid:107)2 − (cid:107)PMβ∗(cid:107)2 x∗ = α∗ + β∗. We thus proved that (cid:62)(cid:112)s(1 − δ) (cid:107)α∗ − x(cid:107)2 − Mmax (cid:107)β∗(cid:107)2 . Ax(cid:107)2 (cid:62)(cid:112)s(1 − δ) (cid:107)α∗ − x(cid:107)2 − Mmax (cid:107)β∗(cid:107)2 − Mmax (cid:107)x(cid:107)2 . Ax − APMx(cid:107)2]. We have (89) (cid:54) Mmax (cid:107)x(cid:107)2 . We used the facts that (cid:107)A(cid:107)2 = 1 (see (73)) and that x ∈ A in the last inequality. (cid:54) (cid:107)A(cid:107)2 (cid:107)PM(cid:107)2 (cid:107)A Ax − x(cid:107)2 Using the inequalities (88) and (89) in (84), we arrive at (cid:124) (cid:124) (cid:13)(cid:13)(cid:13)APMA (cid:13)(cid:13)(cid:13)(cid:101)P(cid:101)M x∗ −(cid:101)P y(cid:13)(cid:13)(cid:13)2 [Finishing the proof ]. Remark that (75) implies (86) (90) (91) (92) (93) (94) (95) (96) . 2 2 2 2 2 2 2 2 + γ x + γ x and γ ( x∗) + γ g(λk)(cid:107)x(cid:107)2 (cid:124)(cid:101)L x, (cid:124)(cid:101)L x. Ax − APMx(cid:13)(cid:13)(cid:13)2 (cid:62)(cid:112)s(1 − δ) (cid:107)α∗ − x(cid:107)2 − Mmax (cid:107)β∗(cid:107)2 − 2Mmax (cid:107)x(cid:107)2 −(cid:13)(cid:13)(cid:13)(cid:101)P n(cid:13)(cid:13)(cid:13)2 (cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x∗ − y)(cid:13)(cid:13)(cid:13) (cid:54)(cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x − y)(cid:13)(cid:13)(cid:13) (cid:124)(cid:101)L x∗ (cid:54)(cid:13)(cid:13)(cid:13)(cid:101)P((cid:101)M x − y)(cid:13)(cid:13)(cid:13) (cid:54) (cid:13)(cid:13)(cid:13)(cid:101)Pn(cid:13)(cid:13)(cid:13) (cid:54) 2(cid:13)(cid:13)(cid:13)(cid:101)Pn(cid:13)(cid:13)(cid:13)2 + Mmax (cid:107)β∗(cid:107)2 + ((cid:112)γg(λk) + (cid:112)γg(λn) + 2Mmax)(cid:107)x(cid:107)2 . (cid:54) 2(cid:13)(cid:13)(cid:13)(cid:101)Pn(cid:13)(cid:13)(cid:13)2 (cid:112)γg(λk+1)(cid:13)(cid:13)(cid:13)(cid:101)Pn(cid:13)(cid:13)(cid:13)2 + Mmax(cid:115) g(λn) g(λk+1) (cid:107)x(cid:107)2 +(cid:16)(cid:112)γg(λk) + (cid:112)γg(λn) + 2Mmax(cid:17)(cid:107)x(cid:107)2 =(cid:32)2 + +(cid:112)γg(λk)(cid:33)(cid:107)x(cid:107)2 (cid:112)γg(λk+1)(cid:33)(cid:13)(cid:13)(cid:13)(cid:101)Pn(cid:13)(cid:13)(cid:13)2 + (cid:32)2Mmax + Mmax(cid:115) g(λn) +(cid:32)Mmax(cid:115) g(λk) +(cid:112)γg(λn)(cid:33)(cid:107)x(cid:107)2 . + Mmax(cid:115) g(λk) 2 + 2 γ g(λn)(cid:107)x(cid:107)2 2 , g(λk+1) (cid:107)x(cid:107)2 g(λk+1) g(λk+1) Mmax Mmax + Using (79), (80), (81) in (92), we obtain γ g(λk+1)(cid:107)β∗(cid:107)2 2 which implies (45) in Theorem 3.1. It remains to prove (44) to finish the proof. Using (80), (81) and (90) in (91), we obtain Using (45) to bound (cid:107)β∗(cid:107)2 on the right hand side, we have (cid:112)s(1 − δ)(cid:107)α∗ − x(cid:107)2 (cid:112)s(1 − δ)(cid:107)α∗ − x(cid:107)2 We only rearranged the term in the last step. This proves (44) and terminates the proof. (cid:3) 22 G. PUY AND P. P´EREZ References [1] Y. Boykov and M. Jolly, "Interactive graph cuts for optimal boundary and region segmentation of objects in n-d images," in Proc. Int. Conf. Computer Vision, 2001. [2] L. Grady, "Random walks for image segmentation," IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 28, no. 11, pp. 1768 -- 1783, 2006. [3] X. Chen, D. Zou, Q. Zhao, and P. Tan, "Manifold preserving edit propagation," ACM Transactions on Graphics, vol. 31, no. 6, pp. 132 -- 139, 2012. [4] X. Chen, D. Zou, S. Zhou, Q. Zhao, and P. Tan, "Image matting with local and nonlocal smooth priors," in Proc. Conf. Computer Vision and Pattern Recognition, 2013. [5] L. Xu, Q. Yan, and J. Jia, "A sparse control model for image and video editing," ACM Transactions on Graphics, vol. 32, no. 6, p. 197, 2013. [6] A. Levin, D. Lischinski, and Y. Weiss, "Colorization using optimization," ACM Transactions on Graphics, vol. 23, no. 3, pp. 689 -- 694, 2004. [7] D. Lischinski, Z. Farbman, M. Uyttendaele, and R. Szeliski, "Interactive local adjustment of tonal values," ACM Transactions on Graphics, vol. 25, no. 3, pp. 646 -- 653, 2006. [8] D. Shuman, S. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, "The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains," Signal Processing Magazine, IEEE, vol. 30, no. 3, pp. 83 -- 98, May 2013. [9] I. Pesenson, "Sampling in Paley-Wiener spaces on combinatorial graphs," Transactions of the American Mathe- matical Society, vol. 360, no. 10, pp. 5603 -- 5627, 2008. [10] I. Z. Pesenson and M. Z. Pesenson, "Sampling, Filtering and Sparse Approximations on Combinatorial Graphs," J. Fourier Anal. Appl., vol. 16, no. 6, pp. 921 -- 942, 2010. [11] A. Anis, A. Gadde, and A. Ortega, "Towards a sampling theorem for signals on arbitrary graphs," in Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on, May 2014, pp. 3864 -- 3868. [12] A. Anis, A. E. Gamal, S. Avestimehr, and A. Ortega, "Asymptotic justification of bandlimited interpolation of graph signals for semi-supervised learning," in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), April 2015, pp. 5461 -- 5465. [13] S. Chen, A. Sandryhaila, and J. Kovacevic, "Sampling theory for graph signals," in Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on, April 2015, pp. 3392 -- 3396. [14] S. Chen, R. Varma, A. Sandryhaila, and J. Kovacevic, "Discrete signal processing on graphs: Sampling theory," Signal Processing, IEEE Transactions on, vol. PP, no. 99, pp. 1 -- 1, 2015. [15] A. Anis, A. Gadde, and A. Ortega, "Efficient sampling set selection for bandlimited graph signals using graph spectral proxies," IEEE Transactions on Signal Processing, vol. PP, no. 99, pp. 1 -- 1, 2016. [16] S. Chen, R. Varma, A. Singh, and J. Kovacevi, "Signal recovery on graphs: Random versus experimentally designed sampling," in Sampling Theory and Applications (SampTA), 2015 International Conference on, May 2015, pp. 337 -- 341. [17] G. Puy, N. Tremblay, R. Gribonval, and P. Vandergheynst, "Random sampling of bandlimited signals on graphs," Appl. Comput. Harmon. Anal., in press, 2016. [18] N. Chauffert, P. Ciuciu, and P. Weiss, "Variable density compressed sensing in mri. theoretical vs heuristic sampling strategies," in Proc. IEEE Int. Symp. on Biomedical Imaging, 2013, pp. 298 -- 301. [19] B. Adcock, A. C. Hansen, C. Poon, and B. Roman, "Breaking the coherence barrier: a new theory for compressed sensing," arXiv:1302.0561, 2013. [20] C. Boyer, J. Bigot, and P. Weiss, "Compressed sensing with structured sparsity and structured acquisition," arXiv:1505.01619, 2015. [21] J. Bigot, C. Boyer, and P. Weiss, "An analysis of block sampling strategies in compressed sensing," IEEE Trans. Inf. Theory, vol. 62, no. 4, pp. 2125 -- 2139, 2016. [22] X. Chen, D. Zou, Q. Zhao, and P. Tan, "Manifold preserving edit propagation," in ACM SIGGRAPH Asia, vol. 31, no. 6, 2012. [23] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Ssstrunk, "Slic superpixels compared to state-of-the-art superpixel methods," IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 11, pp. 2274 -- 2282, 2012. [24] C. Korinke, T. C. Stratmann, T. Laue, and S. Boll, "Superselect: An interactive superpixel-based segmentation method for touch displays," in Proc. ACM Int. Conf. Multimedia, 2016. [25] F. Chung, Spectral graph theory. Amer Mathematical Society, 1997, no. 92. [26] G. Puy, P. Vandergheynst, and Y. Wiaux, "On variable density compressive sampling," IEEE Signal Process. Lett., vol. 18, no. 10, pp. 595 -- 598, 2011. [27] F. Krahmer and R. Ward, "Stable and robust sampling strategies for compressive imaging," IEEE Trans. Image Process., vol. 23, no. 2, pp. 612 -- 622, 2013. [28] A. Borzi and V. Schulz, "Multigrid methods for pde optimization," SIAM Rev., vol. 51, no. 2, pp. 361 -- 395, 2009. [29] T. Dreyer, B. Maar, and V. Schulz, "Multigrid optimization in applications," J. Comput. Appl. Mathematics, vol. 128, no. 1-2, pp. 67 -- 84, 2000. [30] N. Tremblay and P. Borgnat, "Subgraph-based filterbanks for graph signals," IEEE Transactions on Signal Pro- cessing, vol. PP, no. 99, pp. 1 -- 1, 2016. Structured sampling and fast reconstruction of smooth graph signals 23 [31] D. K. Hammond, P. Vandergheynst, and R. Gribonval, "Wavelets on graphs via spectral graph theory," Appl. Comput. Harmon. Anal., vol. 30, no. 2, pp. 129 -- 150, 2011. [32] N. Perraudin, J. Paratte, D. Shuman, V. Kalofolias, P. Vandergheynst, and D. K. Hammond, "Gspbox: A toolbox for signal processing on graphs," arXiv:1408.5781, 2014. [33] E. D. Napoli, E. Polizzo, and Y. Saad, "Efficient estimation of eigenvalue counts in an interval," arXiv:1308.4275, 2013. [34] H. Li, J. Cai, T. N. A. Nguyen, and J. Zheng, "A benchmark for semantic image segmentation," in IEEE Int. Conf. on Multimedia and Expo, 2013, pp. 1 -- 6. [35] A. Beck and M. Teboulle, "A fast iterative shrinkage-thresholding algorithm for linear inverse problems," SIAM J. Imaging Sci., vol. 2, no. 1, pp. 183 -- 202, 2009. [36] N. Tremblay, G. Puy, R. Gribonval, and P. Vandergheynst, "Compressive spectral clustering," in Machine Learning, Proceedings of the Thirty-third International Conference (ICML 2016), June 20-22, New York City, USA, 2016. [37] J. A. Tropp, "User-friendly tail bounds for sums of random matrices," Found. Comput. Math., vol. 12, no. 4, pp. 389 -- 434, 2012.
1804.02318	2	1804	2018-04-10T15:23:36	How Constraints Affect Content: The Case of Twitter's Switch from 140 to 280 Characters	[ "cs.SI" ]	It is often said that constraints affect creative production, both in terms of form and quality. Online social media platforms frequently impose constraints on the content that users can produce, limiting the range of possible contributions. Do these restrictions tend to push creators towards producing more or less successful content? How do creators adapt their contributions to fit the limits imposed by social media platforms? To answer these questions, we conduct an observational study of a recent event: on November 7, 2017, Twitter changed the maximum allowable length of a tweet from 140 to 280 characters, thereby significantly altering its signature constraint. In the first study of this switch, we compare tweets with nearly or exactly 140 characters before the change to tweets of the same length posted after the change. This setup enables us to characterize how users alter their tweets to fit the constraint and how this affects their tweets' success. We find that in response to a length constraint, users write more tersely, use more abbreviations and contracted forms, and use fewer definite articles. Also, although in general tweet success increases with length, we find initial evidence that tweets made to fit the 140-character constraint tend to be more successful than similar-length tweets written when the constraint was removed, suggesting that the length constraint improved tweet quality.	cs.SI	cs	How Constraints Affect Content: The Case of Twitter's Switch from 140 to 280 Characters Kristina Gligori´c EPFL [email protected] Ashton Anderson University of Toronto [email protected] Robert West EPFL [email protected] 8 1 0 2 r p A 0 1 ] I S . s c [ 2 v 8 1 3 2 0 . 4 0 8 1 : v i X r a Abstract It is often said that constraints affect creative production, both in terms of form and quality. Online social media plat- forms frequently impose constraints on the content that users can produce, limiting the range of possible contributions. Do these restrictions tend to push creators towards producing more or less successful content? How do creators adapt their contributions to fit the limits imposed by social media plat- forms? To answer these questions, we conduct an observa- tional study of a recent event: on November 7, 2017, Twitter changed the maximum allowable length of a tweet from 140 to 280 characters, thereby significantly altering its signature constraint. In the first study of this switch, we compare tweets with nearly or exactly 140 characters before the change to tweets of the same length posted after the change. This setup enables us to characterize how users alter their tweets to fit the constraint and how this affects their tweets' success. We find that in response to a length constraint, users write more tersely, use more abbreviations and contracted forms, and use fewer definite articles. Also, although in general tweet suc- cess increases with length, we find initial evidence that tweets made to fit the 140-character constraint tend to be more suc- cessful than similar-length tweets written when the constraint was removed, suggesting that the length constraint improved tweet quality. 1 Introduction The enemy of art is the absence of limitations. -Orson Welles It is often thought that constraints affect both the form and quality of creative content. For example, there is anec- dotal evidence across many domains that imposing length constraints can shape and improve the resulting writing: aca- demic authors edit papers to fit a page limit, poets adhere to a prescribed verse form or rhyme scheme, and journal- ists edit articles to fit a word count limit (McPhee 2015). More broadly, research in several fields-spanning product design (Joyce 2009; Moreau and Dahl 2005), process man- agement (Fritscher and Pigneur 2009), and education (Hen- nessey 1989)-suggests that having too much freedom can be paralyzing (epitomized by the feeling of staring at a blank sheet of paper), and that there is a sweet spot with just the right amount of constraints. Copyright © 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Many social media platforms enforce length restrictions for posts, which is why the latter are frequently called mi- croposts. For instance, at the time of writing, Instagram cap- tions were limited to 2,200 characters, LinkedIn updates to 700 characters, Pinterest pins to 500 characters, and Twit- ter posts, tweets, to 280 characters. In the context of social media, a less restrictive character limit could conceivably be either positive or negative. In this work, we investigate the relationship between length constraints on post length and the form and quality of the resulting content. Do length con- straints steer users towards editing their content in a useful way on social media, or does additional space allow for more creative and engaging content? Answering this question in- volves resolving two competing hypotheses: H1 Imposing constraints has a positive effect on creativity, influencing users to create succinct content that is more likely to be appealing to others (Joyce 2009). H2 Relaxing constraints provides users more space for ex- pressing opinion and allows for more potentially in- teresting content. Also, longer posts occupy a larger portion of the update feeds displayed to users (Rosen 2017), further increasing engagement. Choosing the right character limit is not trivial, given these conflicting hypotheses. Resolving them would not only help elucidate the relation between content and constraints; given the multi-billion-dollar economy depending on social me- dia, the question is also of significant financial consequence. The ideal way to answer this question would be through a randomized control study: subject randomly sampled users to various character limits, and observe how the nature and success of their posts depends on the respective limit. Unfor- tunately, however, only social-media providers themselves have full control over their platforms, including the ability to conduct such A/B tests. Furthermore, even this ideal ex- periment might suffer from creating an ecosystem with dif- fering content lengths with clashing norms and conventions. Studying the effect of constraints is therefore difficult. Twitter's November 2017 switch. In order to circumvent these difficulties, we take advantage of a recent event to con- duct an observational study of how length constraints affect microposts: on November 7, 2017, Twitter suddenly and un- expectedly increased its maximum tweet length from 140 to 280 characters. Figure 1: Left: Schema of study setup. Right: Histogram of tweet lengths, before and after the switch. Figure 2: Engagement (retweets and favorites) as a function of length, with 95% confidence intervals. Left: Probability of receiving at least one engagement. Right: Mean number of engagements (for tweets with at least one engagement). The switch was reportedly introduced to allow users to ex- press their thoughts without running out of characters, thus preventing them from finishing a thought (Rosen and Ihara 2017). This change in length constraint-which we hence- forth refer to as the switch-constitutes an exogenous event that was most likely unexpected to most Twitter users, so we may reasonably assume that user behavior did not change in anticipation of this event. After carefully controlling for certain factors, differences between posts tweeted before vs. after the switch can inform us about the effect of different length constraints. Research questions. Given that our aim is to study the ef- fect of length constraints on the style and success of content in online social media platforms, we seek to answer the fol- lowing questions: RQ1 Do length constraints lead to characteristic changes in the writing style of posts? RQ2 Do length constraints push users toward creating con- tent that other users are more likely to engage with? 2 Methodology1 Research design. In this paper, our main methodological contribution is a matched observational study design that al- lows us to use the exogenous shock to the Twitter ecosystem to investigate the relationship between length constraints and content form and quality, even in the absence of full exper- imental control of Twitter. Before Twitter's switch from a 140-character limit to a 280-character limit, tweets that are nearly or exactly 140 characters in length are likely to have been explicitly "squeezed" by the user to comply with the character limit. After the switch, tweets of this length are less likely to have been affected by the character limit, since they are far short of the new 280-character maximum. Our basic design is thus to compare various properties of tweets 1More details in the appendix. of a length between 136 and 140 characters2 just before vs. right after the switch (Fig. 1, left). The major difference be- ing the presence vs. absence of the 140-character limit, com- paring tweets of this length from before vs. after the switch (corresponding to the upper, red arrow in Fig. 1, left) lets us isolate the impact of the character limit. By comparing content written very close in time, we minimize the likeli- hood of external factors changing in the time during which our studied tweets were written. In a control study, we also compare tweets with 91 to 100 characters before vs. after the switch, as these tweets are unaffected by the character limit during both time periods (lower, gray arrow in Fig. 1, left). Data. In order to execute this research design, one would ideally analyze all tweets in our character range of interest that were authored right before and right after the switch. In lieu of access to complete Twitter data, we approximate this ideal. We first collect tweets from the 1% sample Twitter supplies via its Spritzer API, for the time between April and June 2017. From this set of tweets we sample 100K users, where the probability of being sampled is proportional to the number of 140-character-long tweets they posted. This en- sures that we are likely to sample users who disproportion- ately generate tweets for which the 140-character constraint is relevant. We proceed to collect these users' timelines us- ing the Twitter user timeline API, gathering up to the 3,200 most recent tweets per user. We simultaneously keep collect- ing new tweets these users post. All tweets we analyze were posted between November 2017 and January 2018.3 Controlling for users and topics. With the goal of mitigat- ing potential confounds in our comparison of tweets before vs. after the switch, we perform two analyses. First, we con- trol for users by including, for each user, the same number of tweets written before the switch as after the switch (but this number may vary across users). As we are interested in matching tweets actually written by a particular user, we do not consider retweets of other users' tweets. Second, we also conduct an analysis in which we control for topics by including, for each unique set of hashtags, the same number of tweets before as after the switch with that set of hashtags. Controlling for temporal effects. By design, we compare tweets posted before and after the switch, thus introducing a time gap between the observed content. This raises two problems: first, retweet and favorite counts depend on how much time has passed since posting; and second, one might observe differences in the number of retweets and favorites as a side effect of a user's becoming more popular with time. We address the first issue by making sure to collect a tweet at least 48 hours after it was posted, since by that time 99.99% of retweets happen (Sysomos Retweets Study 2010). The second issue is ruled out by making sure that the number of a user's followers did not increase by more than 1% be- tween the user's first and last tweets in our dataset, when controlling for users. 2We use a 5-character range because it is not always possible to strictly optimize for the limit: omitting one word to meet the con- straint may result in a tweet of fewer than exactly 140 characters. 3Data available at https://github.com/epfl-dlab/140 to 280 Before switchLength[136, 140]After switchLength[91, 100]ConstrainedtweetsUnconstrainedtweetsUnconstrainedtweetsUnconstrainedtweets050100150200250Number of characters0.000.010.020.030.040.050.060.07ProbabilityBefore switchAfter switch03570105140175210245280Number of characters0.00.20.40.60.81.0Probability of at least oneFavorites after switchFavorites before switchRetweets after switchRetweets before switch03570105140175210245280Number of characters0204060Mean numberFavorites after switchFavorites before switchRetweets after switchRetweets before switchLinguistic feature Number of hashtags Number of emojis Number of abbreviations Number of definite article Number of indefinite articles Number of and Number of & Number of missing spaces after punctuation Number of auxiliary verbs and negations in long form Number of auxiliary verbs and negations in contracted form Fraction of lexical words (nouns, verbs, adjectives, adverbs) Lexical variation Lexical sophistication Number of words per sentence Matched for user After Before 0.54 0.53 0.25 0.25 0.17 0.13 0.71 0.68 0.55 0.55 0.42 0.38 0.11 0.16 0.64 0.59 0.75 0.73 0.53 0.44 Matched for topic After Before - 0.29 0.07 0.69 0.45 0.32 0.14 0.68 0.66 0.41 - 0.37 0.09 0.66 0.44 0.25 0.18 0.64 0.68 0.52 56.45% 0.9089 0.0014 17.81 56.39% 0.9081 0.0014 17.86 55.44% 0.9136 0.0014 15.86 55.62% 0.9121 0.0014 15.85 Control study Before 0.22 0.31 0.12 0.48 0.38 0.32 0.03 0.48 0.53 0.39 57.34% 0.9405 0.0013 14.82 After 0.22 0.32 0.12 0.48 0.39 0.32 0.03 0.52 0.52 0.29 57.92% 0.9410 0.0013 14.90 Matched for user Before After 0.013 0.011 0.051 0.057 0.128 0.123 0.026 0.033 0.068 0.065 0.021 0.027 0.033 0.034 0.011 0.014 I am I'm have 've will 'll would 'd Table 1: Results of linguistic analysis (RQ1). Left: Linguistic features averaged across tweets, before vs. after switch, when matching for users and topics (Sec. 3.2), and in control study (Sec. 3.3). Right: Mean number of occurrences per tweet of expanded and contracted word forms. Bold values are significant with p < 0.00365 according to Kolmogorov–Smirnov tests. Probability of at least one Quartiles (25%, 50%, 75%) Matched for user Matched for topic Retweets Favorites Retweets Favorites Control study (tweet length [91,100]) Retweets Favorites Before 21.61% (1, 2, 7) After 20.38% (1, 2, 7) Before 45.50% (1, 3, 8) After 45.52% (1, 3, 8) Before 28.78% (1, 2, 5) After 29.94% (1, 2, 4) Before 52.29% (1, 2, 6) After 52.17% (1, 2, 6) Before 19.10% (1, 2, 5) After 18.72% (1, 2, 5) Before 52.04% (1, 2, 6) After 52.94% (1, 2, 6) Table 2: Results of success analysis (RQ2). Top: Probability of receiving at least one engagement (retweet or favorite). Bot- tom: Quartiles of numbers of engagements (for tweets with at least one engagement). Constraint consistency over time. We ensure that all con- tent we analyze was generated under the same character counting policy. Additionally, we are primarily interested in how text is altered under a particular restriction. We thus limit ourselves to tweets not containing URLs and @ replies. Pagination. Twitter users have long been working around the 140-character limit by splitting a long piece of text into a sequence of length-compliant tweets. These tweet threads are usually annotated with the position of the tweet in its se- quence, e.g., 2/3. They are usually not intentionally altered in order to fit the character limit requirement using mecha- nisms we are interested in. Thus, we disregard such tweets. 3 Results 3.1 General impact of length constraints We begin with an analysis of the general impact of tweet length limits. We follow a data-driven approach and ana- lyze 4M tweets written before, and 1.9M written after, the switch. Tweets are filtered according to our overall study setup (Sec. 2). Fig. 1 (right) contains a histogram of tweet lengths before and after the switch. The spike at exactly 140 characters be- fore the switch is indicative of users dealing with the char- acter limit constraint. Additionally, the fact that there is no corresponding spike after the switch reflects the fact that the spike was indeed induced by the constraint-once it is lifted, 140 characters is an unexceptional tweet length. Fig. 2 presents the probability of obtaining at least one engagement (retweet or favorite) and the average number of engagements (given at least one engagement), as a func- tion of tweet length. We observe that, both before and after the switch, longer tweets are more successful on average, as captured by the number of retweets and favorites. This demonstrates the existence of a length effect consistent with hypothesis H2 (Sec. 1). This finding is, however, subject to several potential confounds; e.g., tweet length might corre- late with the importance of what the user has to say, or with the popularity of the user herself, which might in turn be the real cause for the success of the respective tweet. Our matched studies, described next, address these confounds. 3.2 Matched studies We compare tweets of length [136,140] posted before, with tweets of a length in the same range posted after, the switch. Linguistic analysis (RQ1). We are interested in linguis- tic aspects that might conceivably be altered in an attempt to meet length constraints: hashtags, emojis, abbreviations and acronyms used in the Twitter community (Beal, Vangie 2016), articles, conjunctions, spaces after punctuation, and auxiliary verbs in their long, contracted, or incorrectly con- tracted forms. We also evaluate stylistic features of text, in- cluding the frequency of lexical words, lexical variation, and sophistication as defined by Lu (2012). Moreover, we com- pute the number of words per sentence, an indicator of read- ability. Here, we are not interested in content that fails to elicit any engagement at all, as it is unlikely to reflect the op- timizing mechanisms of interest. We thus keep only tweets that have received at least one retweet or favorite, resulting in 12K (4K) tweet pairs when matching for users (topics). Matching for users and for topics, respectively, we com- pare the above linguistic features before vs. after the switch using Kolmogorov–Smirnov significance tests. Given that we test multiple hypotheses (one per feature), the signifi- cance threshold is adjusted via Dunn–Sid´ak correction (Abdi 2007), where the adjusted significance threshold is calcu- lated as 1− (1− α)1/m. We set α = 0.05 and m = 14, result- ing in an adjusted significance threshold of 0.00365. Whether we control for users or for topics, constrained tweets systematically differ in linguistic features (Table 1). They contain fewer hashtags, fewer articles, and more words in their short forms: more abbreviations, more auxiliary verbs in their contracted forms, more &, and fewer and. These discrepancies are indicative of edits taking place given a 140-character limit constraint. We do not observe more missing spaces after punctuation in constrained tweets. This signals that users are unwilling to use punctuation in an incorrect way in order to fit a constraint. Success analysis (RQ2). When quantifying engagement, we evaluate how much other users engage with a tweet in terms retweets and favorites. Controlling for users and for topics, respectively, we evaluate success before vs. after the switch. The results of this study are presented in Table 2. We see that, when controlling for users, constrained tweets are 6.0% (relative) more likely to elicit engagement in the form of at least one retweet, compared to unconstrained tweets. Note that the trend is reversed when controlling for topics (in- stead of users), but the difference is smaller (−3.9% rela- tive), and controlling for users is arguably a stricter crite- rion than controlling for topics, which, taken together, indi- cates that length restrictions slightly improve the success of tweets, in line with hypothesis H1 (Sec. 1). 3.3 Control study for ruling out confounds As our analysis compares two different time periods to each other, one might object that the observed stylistic changes were in fact caused by a community-wide drift of norms between the two time periods. Although unlikely given the condensed time frame, we aim to rule out this alternative ex- planation with a control study in which we repeat our anal- ysis (matching for users; 9K tweet pairs), but this time com- paring tweets of a length much below the 140-character limit (91 to 100 characters) before vs. after the switch (cf. Sec. 2). Such short tweets are unlikely to have been optimized for the length constraint. As a consequence, if any differences are observed for [91,100] tweets, the same mechanisms might be at work for [136,140] tweets, and it might be these mechanisms-rather than the 140-character constraint-that also cause the differences among [136,140] tweets. How- ever, we observe only one statistically significant difference in linguistic features for the [91,100] range (Table 1), hinting at the length constraint as the cause for the other features. As for success, we observe that [91,100] tweets posted before the change are 2.0% (relative) more likely to be retweeted (Table 2), a much smaller difference than for [136,140] tweets (6.0% relative), which we interpret as sup- port for our above finding that length constraints slightly in- crease tweet success. However, more studies are necessary to confirm these findings, ideally with more data. 4 Discussion In this paper, we address the question of how users adapt their contributions to fit the limits imposed by social media platforms and whether these restrictions tend to push users towards producing more or less successful content. We de- velop a matched observational methodology that involves comparing carefully selected sets of tweets posted just be- fore vs. just after the switch. We find that tweets constrained by a 140-character limit contain fewer hashtags and definite articles, and more words in abridged form: more abbreviations, more contracted aux- iliary verbs, more &, and fewer and, implying that, when subject to a length constraint, users write more tersely. We also found initial evidence that length-constrained tweets are slightly more successful in terms of the engage- ment they receive from other users. However, future work needs to develop a better understanding of this phenomenon and determine whether the observed findings are robust. We anticipate that collecting more data from the time after the relaxation of the constraint will be helpful to this end. In our studies, we assume that tweets in the [136,140] range posted after the switch are not constrained. We cannot rule out the existence of users adhering to the old constraint despite the switch (e.g., out of sheer nostalgia, or because they are bots generating tweets via scripts written before the switch), but in the absence of such hypothetical users, the amplitude of the observed effects would be even larger. Future work. This paper presents the first in what we hope will be a series of studies regarding the impact of Twitter's switch from 140 to 280 characters. In particular, our obser- vational studies should be followed by experimental studies, in order to corroborate the initial findings presented here. Also, our results are from right around the switch, when things were potentially still in flux. It would be interesting to revisit our findings in the future once the Twitter ecosystem has settled into a new steady state. Finally, we hope that fu- ture work will generalize our findings to constrained content production beyond Twitter. References Abdi, H. 2007. Bonferroni and Sid´ak corrections for multiple comparisons. In Encyclopedia of Measurement and Statistics, 103–107. Sage Publications. Beal, Vangie. 2016. Twitter dictionary: A guide to understanding Twitter lingo. https://goo.gl/3Wbkkt. Fritscher, B., and Pigneur, Y. 2009. Supporting business model modelling: A compromise between creativity and constraints. In Proc. International Workshop on Task Models and Diagrams. Hennessey, B. A. 1989. The effect of extrinsic constraints on children's creativity while using a computer. Creativity Research Journal 2(3):151–168. Joyce, C. K. 2009. The Blank Page: Effects of Constraint on Creativity. PhD thesis, UC Berkeley. Lu, X. 2012. The relationship of lexical richness to the quality of ESL learners' oral narratives. The Modern Language Journal 96(2):190–208. McPhee, J. 2015. Omission: Choosing what to leave out. The New Yorker. September 14, 2015. https://goo.gl/639Gpr. Moreau, C. P., and Dahl, D. W. 2005. Designing the solution: The impact of constraints on consumers' creativity. Journal of Consumer Research 32(1):13–22. Rosen, A., and Ihara, I. 2017. Giving you more characters to express yourself. https://goo.gl/syjkWs. Rosen, A. 2017. Tweeting made easier. https://goo.gl/DYzBji. Sysomos Retweets Study. 2010. Replies and retweets on Twitter. https://goo.gl/hqDv5J. the limit is approached. As a consequence, the user is not aware of the exact number of characters left until only 20 (out of the total of 280) remain. As this change was made contemporaneously with the switch from 140 to 280 char- acters, we cannot rule out the possibility that our results are affected by this user interface change. However, regardless of whether users see an exact character count or a half-filled circle, it would likely be clear that tweets in the [136,140] range that we study are not near the 280-character limit af- ter the switch. Moreover, by being shown a circle instead of an exact character count, users after the switch cannot real- ize how close exactly they are to the deprecated 140-charac- ter limit and whether they have passed it, which effectively rules out optimization for the old 140-character limit after the switch due to nostalgia or hipsterdom. For the above rea- sons, we argue it is reasonable to assume that the change from displaying an exact count to displaying a circle has a negligible affect on our results. A.3 Final remark about engagement We also make an interesting observation with respect to the relation between tweet length and other users' engagement with a tweet: Fig. 2 shows that the success of tweets is strongly correlated with their length, both before and af- ter the switch, but also that tweets of a length up to 140 characters posted after the switch are vastly less popular than tweets of the same length posted before the switch. In- deed, it seems that the total amount of engagement caused by the entirety of all tweets stays constant before vs. af- ter the switch-it is merely distributed over a wider range of lengths after the switch. It is interesting to see Twitter's switch in this light, given that it was officially motivated by the claim that tweets longer than 140 characters lead to more engagement (Rosen 2017). A Appendix The first four pages of this paper were published in the proceedings of ICWSM 2018. In this appendix, which did not appear in the proceedings, we provide additional de- tails about the implementation of our observational study and make some additional remarks. Implementation details A.1 Controlling for users and topics. In the main studies, we build pairs of tweets posted by the same user or containing the same set of hashtags. Note that a single user might post multiple such pairs of tweets. When matching, we choose the tweets such that the time distance from the switch is min- imized: before the change, we keep the most recent tweet, and after the change, we keep the least recent tweet. As a result, selected pairs of tweets are tightly distributed around the switch. Three quarters of matched tweets posted before the switch are posted after October 9, and the median is Oc- tober 26. In the case of tweets posted after the switch, three quarters are posted before December 9, and the median is November 23. Constraint consistency over time. Twitter has been gradu- ally distancing itself from the original hard limit of 140 char- acters. First, they stopped counting media (including URLs and quoted tweets) and @ replies toward the character count, with these becoming metadata instead. However, after relax- ing the character limit, such content is now counted again. Therefore, in order to study the impact of the exact identical constraint during all time periods, we exclude tweets con- taining URLs, quoted tweets, and @ replies altogether when sampling tweets for our studies. In order to rule out any fur- ther potential variations of the counting policy over time, and since it is impossible to determine the version of the ap- plication that was used to send a given tweet, we moreover consider only tweets that were posted using the Twitter web client. Furthermore, since the character limit is not equal for all languages (Rosen 2017) and in order to maintain a mean- ingful interpretation of lexical features, we limit ourselves to tweets in English. Pagination. In our analysis, we detect and disregard pagi- nated tweets in the following way. From the gathered data, we detect 1,722 instances of pagination by detecting con- secutive tweets posted by the same user containing a label in form of position of a tweet in a sequence/total number of tweets in a sequence, e.g., 1/3, 2/3, 3/3. We infer that pagina- tion is relatively rarely used, as less than 0.1% of all tweets in our dataset are paginated. Furthermore, 90% of subse- quent paginated tweets are posted within four minutes of each other. In the subsequent analysis, we disregard tweets the same user posts within this time frame, but we observe that varying the time frame, and thus the percentage of de- tected paginations, does not qualitatively affect our findings. A.2 Note on a user interface update Alongside the change of the character limit, Twitter updated its user interface such that it no longer counts down the num- ber of characters as the user types. Instead, a circle fills in as
1809.05124	1	1809	2018-09-13T18:15:33	Enhanced Network Embeddings via Exploiting Edge Labels	[ "cs.SI", "physics.soc-ph" ]	Network embedding methods aim at learning low-dimensional latent representation of nodes in a network. While achieving competitive performance on a variety of network inference tasks such as node classification and link prediction, these methods treat the relations between nodes as a binary variable and ignore the rich semantics of edges. In this work, we attempt to learn network embeddings which simultaneously preserve network structure and relations between nodes. Experiments on several real-world networks illustrate that by considering different relations between different node pairs, our method is capable of producing node embeddings of higher quality than a number of state-of-the-art network embedding methods, as evaluated on a challenging multi-label node classification task.	cs.SI	cs	Enhanced Network Embeddings via Exploiting Edge Labels Haochen Chen Stony Brook University [email protected] Xiaofei Sun Stony Brook University [email protected] Bryan Perozzi Google Research [email protected] Muhao Chen University of California, Los Angeles [email protected] Yingtao Tian Stony Brook University [email protected] Steven Skiena Stony Brook University [email protected] 8 1 0 2 p e S 3 1 ] I S . s c [ 1 v 4 2 1 5 0 . 9 0 8 1 : v i X r a ABSTRACT Network embedding methods aim at learning low-dimensional latent representation of nodes in a network. While achieving com- petitive performance on a variety of network inference tasks such as node classification and link prediction, these methods treat the relations between nodes as a binary variable and ignore the rich semantics of edges. In this work, we attempt to learn network em- beddings which simultaneously preserve network structure and re- lations between nodes. Experiments on several real-world networks illustrate that by considering different relations between different node pairs, our method is capable of producing node embeddings of higher quality than a number of state-of-the-art network em- bedding methods, as evaluated on a challenging multi-label node classification task. KEYWORDS network embeddings; network representation learning; social rela- tion ACM Reference Format: Haochen Chen, Xiaofei Sun, Yingtao Tian, Bryan Perozzi, Muhao Chen, and Steven Skiena. 2018. Enhanced Network Embeddings via Exploiting Edge Labels. In The 27th ACM International Conference on Information and Knowledge Management (CIKM '18), October 22 -- 26, 2018, Torino, Italy. ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/3269206.3269270 1 INTRODUCTION Network embedding methods learn low-dimensional latent rep- resentation of nodes in a network. The learned representations encode complex social relations between nodes, which can be used as features for a variety of tasks such as node classification [12], network reconstruction [16] and link prediction [4]. While achieving superior performance on a handful of network inference tasks, network embedding models typically consider the social relation between nodes as a binary variable. Accordingly, the objective of many network embedding algorithms is to predict the existence of edge between two nodes [14]. This objective is sometimes extended to consider the co-occurrences between nodes Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). CIKM '18, October 22 -- 26, 2018, Torino, Italy © 2018 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-6014-2/18/10. https://doi.org/10.1145/3269206.3269270 within k hops away from each other [4, 12], but the relationships between nodes are still considered to be binary. In real-world social networks, the social relations between nodes are often complex, containing much more information than a binary variable. Consider an online social network such as Facebook, where people are connected to their co-workers, family members and colleagues. Ideally, network embedding methods should not only be able to capture the (binary) friendships between people, but also preserve the types of relations. In this paper, we develop a network embedding framework which is capable of preserving both the proximities between nodes and so- cial relations between different node pairs. Concretely, our method jointly minimizes an unsupervised loss of predicting node neigh- borhoods and a supervised loss of predicting edge labels (social relations). Experimental results on two real-world networks demon- strate that our method learns better node embeddings when com- pared to a variety of unsupervised and semi-supervised network embedding methods, as evaluated on a challenging multi-label node classification task. 2 RELATED WORK Unsupervised Network Embedding Methods. Unsupervised network embedding methods only utilize network structure in- formation to construct node embeddings. The learned node em- beddings are treated as generic feature vectors, which can then be used for downstream tasks in networks. A representative method is DeepWalk [12], which is a two-phase method for learning node embeddings. In the first step, DeepWalk constructs node neigh- borhoods by performing fixed-length random walks in the input network. Then, DeepWalk employs the skip-gram [11] model to preserve the co-occurrences between nodes and their neighbors. Many later studies follow this two-step framework for learning network embeddings, by proposing different strategies for con- structing node neighborhoods [4, 14] or modeling co-occurrences between nodes [1, 16]. There are also a number of methods that utilize edge labels for learning better node representations. SNE [18] presents a log- bilinear model that incorporates two vectors to capture the positive or negative relationships between nodes respectively. TransR [9] learns a projection matrix for each relation type in knowledge graph, which is used for projecting entity embeddings from entity space to relation space. These work generally assume that each edge is associated with a single type of label: it could be the sign of edges in signed networks, or relation types in knowledge graphs. Different Algorithm 1 Model Optimization Input: graph G walks per node r, walk length l, window size w total number of batches T , weight factor λ Sample a batch of N1 random walks RW = {rwi} Ls = − 1 N1 Take a gradient step for Ls u∈C(v) log Pr(uv) v ∈rwi i =1 for i = 0 to (1 − λ) · T do N1 N2 1: while not converged do 2: 3: 4: 5: 6: 7: 8: 9: 10: end for 11: 12: end while end for for i = 0 to λ · T do i =1 L Sample a batch of N2 edges E = {ei} Lr = − 1 N2 Take a gradient step for Lr j=1 H(eij , eij) from these work, our method is capable of dealing with edges with multiple labels, which is a common scenario in real-world networks. Semi-supervised Network Embedding Methods. Another direction of work utilize label information to enhance network embedding methods. These works typically assume that there is a feature vector associated with each node, and that nodes in the input network are partially labeled. Planetoid [17] simultaneously minimizes the unsupervised loss over network structure and the supervised loss of predicting node labels. GCN [8] and GraphSAGE [5] adopt a neural message passing strategy to propagate node features in the network, with supervision from the labeled nodes. Different from these work, our method does not rely on extra node features to work. Additionally, we aim at preserving social re- lations associated with edges, while these methods are only capable of learning from node labels. DeepGL [13] is an exception in that it is capable of utilizing edge labels for node / edge representation learning, but it assumes that label information is available for all edges, which is not often the case. 3 PROBLEM DEFINITION AND NOTATION In this section, we formalize the problem of learning enhanced network embeddings by exploiting edge labels. Let G = (V , E) be an undirected graph, where V is the set of nodes and E is the set of edges. Let L = (l1, l2, · · · , lL) be the set of relation types (labels). A partially labeled network is then defined as G = (V , EL, EU , YL), where EL is the set of labeled edges, EU is the set of unlabeled edges with EL ∪ EU = E. YL represents the relation types associated with the labeled edges in EL. We assume that an edge can have multiple labels; in other words, for all YL(i) ∈ YL, we have YL(i) ⊆ L. The objective of network embeddings is to learn a mapping function Φ : V → RV ×d, where d ≪ V . 4 METHOD The basic idea of our framework is to simultaneously minimize structural loss and relational loss. Structural loss Ls refers to the loss of predicting node neighborhoods, while relational loss Lr is the loss of predicting the labels of edges. Formally, the overall loss function of our method is: L = (1 − λ)Ls + λLr (1) where λ is a hyperparameter that controls the weights of the two parts. 4.1 Structural Loss The first part of our model minimizes the unsupervised structural loss in the network. Given a center node v ∈ V , we maximize the probability of observing its neighbors C(v) given v. Note that the neighborhood of a node is not necessarily the set of nodes it connects to; we will present our definition of a node's neighborhood later. Formally, we minimize the following objective function: log Pr(uv) (2) Ls = − u∈C(v) We adopt the skip-gram objective to compute Pr(uv). We note that in the skip-gram model, each node v has two different embedding vectors Φ(v) and Φ′(v) when serving as center node and context node. Accordingly, Pr(uv) is computed as follows: Pr(uv) = exp(Φ(u) · Φ′(v)) u′∈V exp(Φ(u′) · Φ′(v)) (3) However, calculating the denominator in Eq. 3 requires a proba- bility summation over all vertices, which is too computationally expensive. To alleviate this problem, we use the negative sampling strategy [11] to enable faster model optimization. Once the method for computing Pr(uv) is decided, the only problem left is the construction of node neighborhood C(v) for each v ∈ V . The most straightforward approach is to use the nodes that are adjacent to v in G: C(v) = {u ∈ E (u, v) ∈ E}. However, real-world networks are often sparse, containing a large portion of nodes which have very few neighbors. To alleviate the data sparsity problem, we adopt a random walk strategy similar to that of DeepWalk [12] to enlarge the neighborhood of nodes. We first start r random walks of length l from each node. Within each random walk sequence S = {v1, v2, · · · , vl}, we iterate over each vi ∈ S and use nodes within a window of w as its neighborhood: C(v) = {vi−w , · · · , vi−1}{vi +1, · · · , vi +w}. 4.2 Relational Loss One challenge with modeling social relations between nodes is that the labels we seek to predict are associated with edges, instead of nodes. Although some semi-supervised network embedding meth- ods are capable of utilizing node labels, they fail to deal with edge labels [5, 8]. To bridge this gap, we propose to compose edge repre- sentations from the representations of their endpoints. Given an edge e = (u, v) ∈ E, its representation is given as: Φ(e) = д(Φ(u), Φ(v)) (4) where д is a mapping function from node embeddings to edge em- beddings: д : Rd × Rd → Rd′. In practice, we find that a simple concatenation function works well for constructing edge embed- dings: Φ(e) = Φ(u) ⊕ Φ(v) (5) Once the embedding vector of e is obtained, we feed it into a feed-forward neural network for edge label prediction. Formally, the k-th hidden layer hk of the network is defined as: (k) = f (W (k) (k−1) + b (k)) h h (6) where W (k) and b(k) are the weight and bias of the k-th hidden layer respectively, f is a nonlinear activation function, and h(0) = Φ(e). Specifically, we employ ReLU as the activation function for intermediate layers, and sigmoid as the activation for the last layer since the training objective is to predict edge labels. Assume that the ground-truth label vector of e is y ∈ RL and the output of the last layer is y, we minimize the following binary cross entropy loss for all labels: Lr = H(yi , yi) = yi · log yi + (1 − yi) · log(1 − yi) (7) L i =1 L i =1 4.3 Model Optimization The overall training process of our model is summarized in Algo- rithm 1. For each training step, we first sample λ · T batches of node-context pairs from the random walk sequences and optimize the structural loss on them. Then, we sample (1 − λ) · T batches of edge-label pairs from the training data and optimize the relational loss on them. We perform mini-batch Adam [7] to optimize the objective functions defined above. 5 EXPERIMENT In this section, we describe the datasets being used and compare our method against a number of baselines. 5.1 Dataset We use two datasets from different domains where edge labels are available, which we detail below: • ArnetMiner [15]. We use a processed ArnetMiner dataset provided by TransNet [15]. ArnetMiner is a large-scale col- laboration network between researchers, with over a mil- lion authors and four million collaboration relations. The social relations (edge labels) between researchers are the research topics that they collaborate in. For each co-author relationship, representative keywords are extracted from the abstracts of the co-authored publications as edge labels. The node labels are the research interests of researchers extracted from their homepages. This network has 13,956 nodes, 23,854 edges and 100 different node labels. • AmazonReviews [6, 10]: This dataset contains product re- views and metadata for Amazon products in the category of "Musical Instruments". We take the bipartite graph between users and items as the input network, and use the LDA [2] model to extract topics from review texts as edge labels. The node labels are the detailed categories each product belongs to on Amazon. This network has 25,674 nodes, 60,077 edges and 100 different node labels. 5.2 Baseline Methods The baseline methods we compare against are as follows: DeepWalk [12]: This is an unsupervised network embedding method that captures node co-occurrences via performing short Algorithm GCN DeepWalk LINE node2vec Our Method Training Ratio 20% 17.91 23.98 22.72 24.80 43.64 10% 15.79 21.34 19.98 21.77 40.97 5% 13.56 19.00 16.90 19.15 36.99 Table 1: Node classification results on ArnetMiner. random walks in the input graph. DeepWalk employs the Skip-gram [11] model to learn node representations. LINE [14]: This is a network embedding method that preserves first-order and second-order proximities in networks. LINE uses Skip-gram with negative sampling to learn node representations. node2vec [4]: This is a network embedding method that im- proves DeepWalk by introducing a biased random walk algorithm. GCN [8]: This is a semi-supervised neural network method for network representation learning. We note that GCN requires a feature vector for each node as input. Thus, as suggested by the original GCN paper, we use a one-hot encoding of the node's identity for each node. 5.3 Experimental Setup for Node Classification For DeepWalk, LINE and node2vec, we follow the experimental setup in DeepWalk. Firstly, we obtain the latent representations of nodes. Then, a portion of nodes and their labels are randomly sampled from the graph as training data, and the goal is to predict the labels of the rest of the nodes. We train a one-vs-rest logistic regression model with L2 regularization implemented in LibLinear [3] using the node representations for prediction. To ensure the reliability of our experiments, the above process is repeated for 10 times, and the average Macro F1 score is reported. 5.4 Hyperparameter Settings For all baseline methods, we use the default hyperparameters as described in their papers. For our method, we set λ to 0.8, batch size N1 and N2 both to 400, and embedding dimensionality to 128. For generating random walks, we use r = 80, l = 10, w = 10 in both our method and all baseline methods. For relational loss, we use a feedforward neural network with one hidden layer. We set the learning rate of Adam to 0.01 and perform early stop with a window size of 5; that is, we stop training if the loss on the validation set does not decrease for 5 consecutive iterations. 5.5 Results and Analysis We summarize the node classification results in Tables 1 and 2. We report the Macro F1 scores of each method using 5%, 10%, and 20% of labeled nodes. In real-world networks, edge labels are usu- ally scarce. Thus, for our method, we only use the labels of 10% of the edges. We can see that even with such a small amount of labeled edges, our method still achieves significant improvement over the baseline methods. Also, the extent of improvement on the ArnetMiner dataset is much larger than that on the AmazonRe- views dataset. Part of the reason is that in a collaboration network (a) Effect of the ratio of labeled edges on node classification Macro F1 score. (b) Effect of λ on node classification Macro F1 score. (c) Effect of embedding dimensionality on node classification Macro F1 score. Figure 1: Parameter Sensitivity Study. relationships between nodes, our method is capable of learning better node embeddings than the previous methods, as evaluated on a challenging multi-label node classification task. 7 ACKNOWLEDGEMENTS This work is partially supported by NSF grants IIS-1546113 and DBI-1355990. REFERENCES [1] Sami Abu-El-Haija, Bryan Perozzi, and Rami Al-Rfou. 2017. Learning edge representations via low-rank asymmetric projections. In Proceedings of CIKM. ACM, 1787 -- 1796. [2] David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Algorithm GCN DeepWalk LINE node2vec Our Method Training Ratio 20% 19.07 27.66 25.63 28.26 28.41 10% 12.96 19.45 18.97 19.95 21.15 5% 8.79 14.08 13.67 14.48 16.12 Table 2: Node classification results on AmazonReviews. like ArnetMiner, the labels of an edge often indicate the shared characteristics of its endpoints. In this scenario, a high correlation between edge labels and node labels can be expected. 5.6 Parameter Sensitivity Analysis In this part, we conduct a brief parameter sensitivity analysis to see how does the performance of our method change with hyper- parameters. We use the default parameter settings as in Section 5.4 unless otherwise stated. We report the Macro F1 score our method achieves with 10% labeled edges and 5% labeled nodes. In Figure 1a, we vary the amount of labeled edges to see its effect on the quality of node embeddings. We can see that the Macro F1 score significantly increases when a larger number of labeled edges are available. This indicates that our method is capable of utilizing the rich semantic information of edges. In Figure 1b, we change λ to see how does the ratio between structural loss and relational loss affect the learned node embeddings. When λ is set to 0, our method is equivalent to DeepWalk. As λ increases, we can see a steady increase in Macro F1 which shows the importance of minimizing relational loss. Also, our method is not sensitive to the choice of λ: similar performance are achieved for λ = 0.4, 0.6, 0.8. In Figure 1c, we analyze the impact of the dimensionality of embedding vectors. We can see that the best performance is achieved when embedding dimensionality is 128 or greater. 6 CONCLUSION We propose a semi-supervised framework for enhancing network embeddings with edge labels. Different from traditional network embedding methods which only model network structural informa- tion, our method simultaneously preserves the local neighborhoods of nodes and the social relations between node pairs. Experimental results on real-world networks show that by capturing the different JMLR 3, Jan (2003), 993 -- 1022. [3] Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. 2008. LIBLINEAR: A library for large linear classification. JMLR 9 (2008), 1871 -- 1874. [4] Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for Networks. In Proceedings of KDD. [5] Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In NIPS. 1025 -- 1035. [6] Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In Proceedings of WWW. 507 -- 517. [7] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014). [8] Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR. [9] Yankai Lin, Zhiyuan Liu, Maosong Sun, Yang Liu, and Xuan Zhu. 2015. Learning entity and relation embeddings for knowledge graph completion. In AAAI, Vol. 15. 2181 -- 2187. [10] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. 2015. Image-based recommendations on styles and substitutes. In Proceedings of SIGIR. ACM, 43 -- 52. [11] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In NIPS. 3111 -- 3119. [12] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. Deepwalk: Online learning of social representations. In Proceedings of KDD. ACM, 701 -- 710. [13] Ryan A Rossi, Rong Zhou, and Nesreen K Ahmed. 2018. Deep Inductive Network Representation Learning. In Companion of the Web Conference 2018. 953 -- 960. [14] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. 2015. Line: Large-scale information network embedding. In Proceedings of WWW. International World Wide Web Conferences Steering Committee, 1067 -- 1077. [15] Cunchao Tu, Zhengyan Zhang, Zhiyuan Liu, and Maosong Sun. 2017. TransNet: translation-based network representation learning for social relation extraction. In Proceedings of IJCAI. [16] Daixin Wang, Peng Cui, and Wenwu Zhu. 2016. Structural Deep Network Em- bedding. In Proceedings of KDD. [17] Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. 2016. Revisiting semi- supervised learning with graph embeddings. arXiv preprint arXiv:1603.08861 (2016). [18] Shuhan Yuan, Xintao Wu, and Yang Xiang. 2017. SNE: signed network embedding. In PAKDD. Springer, 183 -- 195. 8000016000024000032000040000048000056000064000072000080000016000002400000320000040000004800000560000064000007200000# of Training Examples0.050.100.150.200.250.300.35Macro F1Ratio of Labeled Edges15108000016000024000032000040000048000056000064000072000080000016000002400000320000040000004800000560000064000007200000# of Training Examples0.050.100.150.200.250.300.35Macro F1λ0.00.20.40.60.88000016000024000032000040000048000056000064000072000080000016000002400000320000040000004800000560000064000007200000# of Training Examples0.050.100.150.200.250.300.350.40Macro F1Embedding Dimensionality163264128256512
1611.07454	1	1611	2016-11-22T18:44:08	An Agent-Based Model of Message Propagation in the Facebook Electronic Social Network	[ "cs.SI", "cs.MA" ]	A large scale agent-based model of common Facebook users was designed to develop an understanding of the underlying mechanism of information diffusion within online social networks at a micro-level analysis. The agent-based model network structure is based on a sample from Facebook. Using an erased configuration model and the idea of common neighbours, a new correction procedure was investigated to overcome the problem of missing graph edges to construct a representative sample of the Facebook network graph. The model parameters are based on assumptions and general activity patterns (such as posting rate, time spent on Facebook etc.) taken from general data on Facebook. Using the agent-based model, the impact of post length, post score and publisher's friend count on the spread of wall posts in several scenarios was analyzed. Findings indicated that post content has the highest impact on the success of post propagation. However, amusing and absorbing but lengthy posts (e.g. a funny video) do not spread as well as short but unremarkable ones (e.g. an interesting photo). In contrast to product adoption and disease spread propagation models, the absence of a similar "epidemic" threshold in Facebook post diffusion is observed.	cs.SI	cs	An Agent-Based Model of Message Propagation in the Facebook Electronic Social Network Hamid Reza Nasrinpour, Marcia R. Friesen, and Robert D. McLeod, Member, IEEE 1  Abstract-A large scale agent-based model of common Facebook users was designed to develop an understanding of the underlying mechanism of information diffusion within online social networks at a micro-level analysis. The agent-based model network structure is based on a sample from Facebook. Using an erased configuration model and the idea of common neighbours, a new correction procedure was investigated to overcome the problem of missing graph edges to construct a representative sample of the Facebook network graph. The model parameters are based on assumptions and general activity patterns (such as posting rate, time spent on Facebook etc.) taken from general data on Facebook. Using the agent-based model, the impact of post length, post score and publisher's friend count on the spread of wall posts in several scenarios was analyzed. Findings indicated that post content has the highest impact on the success of post propagation. However, amusing and absorbing but lengthy posts (e.g. a funny video) do not spread as well as short but unremarkable ones (e.g. an interesting photo). In contrast to product adoption and disease spread propagation models, the absence of a similar "epidemic" threshold in Facebook post diffusion is observed. Index Terms- Agent-Based Modeling, Facebook, Information Diffusion, Online Social Network I. INTRODUCTION M arketers have widely accepted the importance of Word- Of-Mouth (WOM) for a product success [1]-[4]. For example, Philips in 2006, Hewlett-Packard (HP) in 2008, Microsoft in 2009 and Ford in 2009 all developed different types of word-of-mouth seed marketing campaigns to promote their sales [5]. The objective is to seed a marketing campaign with the intention of fostering message propagation or spread. This kind of viral marketing is not necessarily only for promoting a product; it can also help in acquiring new members or broadcasting a message or trends in general. In comparison with traditional marketing, WOM has a longer lasting impression to a new member [6]. Moreover, word-of- mouth can work well in cyberspace such as online communities, emails, product ratings, blogs, forums and The authors gratefully acknowledge financial support from the NSERC Discovery Grants program and the University of Manitoba for GETS funding. H. R. Nasrinpour and R. D. McLeod are with the Department of Electrical and Computer Engineering, University of Manitoba, Winnipeg, MB R3T 5V6, Canada (e-mail: [email protected]; [email protected]). M. R. Friesen is with Design Engineering, University of Manitoba, University of Manitoba, Winnipeg, MB R3T 5V6, Canada (e-mail: [email protected]). electronic social networks. In particular, online WOM could be more attractive to companies because of associated lower costs, wider accessibility, immediate distribution, ease of use [7], and better tracking analytics. The importance of online social networks for venders to advertise their products can also be verified by considering the numerous blogs and publications regarding online social networks in marketing science [8]-[13]. Among online social networks, Facebook is currently the most well-known social network on the internet. However, the answer to the mysterious question of how to make a post go viral on Facebook still remains somewhat of an enigma. There are many recommendations, anecdotes and hints on different blogs to help one reach maximum influence. Apart from the complex news feed algorithm of Facebook, including a very large network of different people with complex psychologies and so many soft factors, a better understanding of message propagation mechanisms through this type of online social network is needed. Some detailed studies of the Digg social news website [14] and Twitter social network [15] exist that pay specific attention to the rules and structure of those social networks. However, many of the other works in the context of social networks and diffusion are either too generic or at a macro-level aggregate, without utilizing a detailed model of an online social network (see section IV). In this paper, the Facebook message propagation process is examined at a micro-level. We focus on two properties of a Facebook post (post length and post interest) and one attribute of the sender (friend count within Facebook). As an instance, an attempt is made to determine which of the two properties of post length and post content (interest) plays a more impactful role in the success of post diffusion. By inspecting Facebook at the level of individuals, a better understanding of the underlying dynamics of message propagation process in Facebook is hopefully obtained. Moreover, ultimately the goal is to understand the similarities and differences between the propagation of relatively intangible entities such as messages and memes and the propagation of tangible entities such as infectious diseases within a population (see section IV). An Agent-Based Model (ABM) of aspects of the Facebook social network was created to delineate and understand the patterns of message propagation. Agent-based modeling [16], [17] is a natural way of simulating systems where individual agents (e.g. people) play significant roles. In this bottom-up approach, the system contains a set of autonomous individuals, i.e. agents, interacting based on a set of rules within an environment. From the micro-level inter-actions of 2 reading a friend's post and sharing the post among friends, the macro-level patterns of diffusion of the post evolves. This makes ABMs a suitable method of analyzing Facebook where both micro-level and macro-level analyses are of interest. In addition, we have a heterogeneous population of Facebook users, which is inherently suitable to an agent-based model where each user can have their own profile, in this case, preferences of when to sign in or share a post. ABMs are generally well suited to model social networks when either the agents or the topology of the interactions is heterogeneous or complex [18]. The idea of deploying an ABM is not only for analysing the impact of post length versus post score, but rather for making a tool potentially capable of including a variety of features to set up different experiments that empirical work cannot address. Cellular Automata as a limited form of ABM has previously been applied for modeling a seeding program [5]. In our study, similar to cellular automata, agents have an internal state machine but are not modeled as (limited) cells. The agents are people modeled as nodes within a graph. Each has its own internal parameters such as the number of friends within the social network. They can asynchronously and independently act (see section II). As an example, agents can decide when to log in/out, read and share a note among friends. The detailed behaviour of agents is explained in section II. ABMs are also inherently extendable to almost infinite levels of details, and in this case the ABM can be extended to study other possible actions by Facebook users (see section V). Depending on the scope of the research, one can add or remove rules and states to or from agents. An advantage of ABM over Differential Equation (DE) or Statistical models is its lack of complex math, which means we do not have to understand relatively complex model formulations [18]. In other words, we just need to be able to describe the system and agent behaviour in detail with a set of "what-if" rules, in a problem-specific and natural lexicon. These characteristics make ABMs well suited to problems that are computationally irreducible [19]. This has however left ABMs open to critique as being more difficult to validate. Obviously, there are also other limitations within ABMs. ABMs can easily be slow and computationally intensive. In fact, the speed performance was one of the main obstacles encountered here in obtaining the results which are presented in section III. These limitations are acknowledged in section V. The remainder of this paper addresses the following:   initial set of simulations, A realistic agent-based model of Facebook posts diffusion was created. This framework can easily be extended to include more features of Facebook and its users. In an relative importance of each input factor such as post score, post length and publisher's friend count is compared.  A second set of simulations explores the impact of the details of post score versus post length particularly for shorter posts like URLs or photos more specifically. the  A third set of simulations sheds light on two seeding strategies relative to a mass of users versus a few hub users.  Generally, post content has the highest impact for information propagation within the electronic social network; however, among the posts which spread fairly well through the network, post length is of more importance than the post content (interest) and the initial seeder's friend count.  Surprisingly, it is shown that there is no tipping point [20] for post diffusion analogous to the transition to epidemic spread observed in infectious diseases. On average, the moment a post is submitted is when it reaches its peak of the probability of being shared or read by friends. It is also shown, unlike product adoption or disease spread, that it is unlikely for a Facebook post to go viral and reach a fair percentage of the entire network. In this case (like other celebrity phenomen- a), the fact that some posts obviously do go viral may skew a typical Facebook user's perception of the probability of their own post doing so.  II. ABM ARCHITECTURE The ABM is implemented in the Java-Based educational version of the Anylogic software toolkit, which supports Agent-Based, Discrete Event and System Dynamics Modeling. In this section, the agent-based model, agents' properties, the structure of their environment, and the governing rules are explained in detail. A. The Big Picture People using Facebook can either visit the webpage on their browser or use the Facebook application on their mobile device. In either case, once you open your Facebook profile, you may receive a list of notifications of what has previously happened since your last login. The difference is that in the second case, you can stay signed into your Facebook profile with your Facebook phone or tablet application, which results in receiving notifications when they occur. Once a person truly decides to check their Facebook profile, they usually go through the notifications and then generally switch to their news feed (home) page to see the activities and posts from friends or other Facebook pages/groups to which they are affiliated. At any time during the visit on Facebook, a user might decide to post a text note or upload a photo/video on their Facebook (wall/timeline) page. They might also copy a post previously shared by a friend and paste it on their own page in order to share it with their own friends. Through this feature, a post would spread over the network. The other common way to interact with a post is to "like" which is invoked by clicking a like button below a post. Currently, there are also many other features available on Facebook such as private messaging and applications that all act like incoming stimuli to a Facebook user to draw their attention. The current scope of this study is general posts by users on their own wall page. We recognize that this is a simplification 3 of actual social networking via Facebook. The simplification was necessary in the first attempt at creating a model. Detailed simplifying assumptions are presented in section II.D where the rules of behaviors by agents are explained. Herein, akin to reality, time passes continuously in minutes and seconds. Agents (Facebook users) are connected to one another within a virtual social network. Each agent, independently from all the other events, decides when to log in and when to log off. During the interval they are logged into the system, they go and check their friends' Facebook wall pages, each post one by one, until they decide to switch to another friend's page. Each post has some interest score and requires its own unique time to be read. Once an agent finds out something interesting on a friend's page, they might decide to share it again on their own page. Also at any time when an agent is online, they can publish a new post of their own. This agent is denoted the initial seeder/publisher of the post. Agents - when they are online - are able to receive notification of recent activities from their immediate friends. In the current agent-based model, this activity only includes the case where a friend shares a post on their own wall page. In this agent-based model, similar to the real Facebook where users check their notifications, upon receiving a notification by an online agent, they go through the notification and read the post shared by their friend. B. Agents and Parameters The agent-based model consists of only one type of agent which is a Facebook user or individual. Agents can create and publish different posts with two important properties of Post Length and Post Score. Each agent has a set of internal parameters including Activity, Average Login Time, Average Post Rate, Friend Time and Friend Count. Wherever possible, published reports on Facebook user statistics were used to set agent parameters, and where a given parameter was reported in several references, reasonable inferences and consolidations were made [21]-[28], etc. is the main parameter The Activity parameter that characterizes the heterogeneity in individual preferences for posting/sharing notes. In other words, each user posts new notes at a given rate, which depends on the user's Activity parameter. Also, the chance of sharing a post already shared by a friend is related to this parameter. This parameter is assigned a uniform random value from 0% to 100%. Obviously, the higher the value, the more posts the user generates. Generally, the more active users (users who post more often) are also online more often. The exact association between the frequency of activity and online time is described below. The Average Login Time parameter indicates how many minutes a user is online in a day, on average. This parameter has a normal distribution and is used to calculate the Login Time parameter by the following formula: Login Time = 2 × Activity × Average Login Time. This means that a typical user, whose activity parameter is 50%, has a Login Time parameter value around the value of Average Login Time parameter. Also the formula states that active users spend more time online on average. Furthermore, the login time follows a long tail distribution that also accounts for 'lurkers' (high online time but low Activity). The exact mathematical distribution of a product of a uniform random variable and a normal random variable can be found in the Appendix. The Average Login Time parameter has a normal distribution. There are different choices for its normal distribution properties. In 2011, Facebook Press Room reported the average Facebook user spends more than 11 hours per month on Facebook [22]. Assuming 30-day months, this means over 22 minutes per day. In addition, [23] claims the "average user spends an average 15 hours and 33 minutes on Facebook per month," which equates to roughly 31 minutes per day. There are also some reports on mobile usage, such as an average of 441 minutes per visitor in each month (i.e., 14.7 minutes per day) reported in [24]. Consolidating these sources, the average online time was set at 23 minutes per day, which may change in the future work. Considering the fact that most obsessed Facebook users spend daily average of 8 hours on the site [25], the average online time is bounded by a maximum of 10 hours per day. Assuming a day is 1440 minutes, a login time of 23 minutes results in having a logout time of 1440 – 23 = 1417 minutes. This means if the login time for a user is calculated to be 23 minutes, they will be online for 23 minutes and will be offline for 1417 minutes in a day. However, since most people tend to check their Facebook profile more than once a day (e.g., It is reported that "on average, [a person] visits the Facebook app/site 13.8 times during the day, for two minutes and 22 seconds each time" [26]) these 23 and 1417 minutes have to be divided into different intervals. These intervals are drawn from an exponential distribution with the mean value of 1440/1417 and 1440/23 minutes per a day (1440 minutes), for the login and logout times respectively. The choice of exponential distributions for login/logout rates where the probability to login is the highest immediately after logging out may sound irrelevant. However, similar to working out at gym, people are likely to check their social profiles within a certain period of day (e.g., it is reported that "peak Facebook time is during the evening, just before bed" [26]). Although for some smartphone users with Facebook application installed on their phone, this period might be from 6 am to 11 pm. In either case, it would be safe to assume that people would not check their Facebook profiles when they are asleep! In the agent-based model, the login intervals should not be too far away from one another, and there has to be a limit to control when people log into their profile. The choice of exponential distributions attempts to keep the online intervals close to each other. In addition, a login rate of an average of 23 minutes/day compared to a logout rate of an average of 1417 minutes/day is low enough to span the daytime. The choice of the exponential distribution is also related to performance issues. Since it is the default distribution for rate triggers in Anylogic, deploying another distribution would significantly have decreased the speed of simulation. Part of the reason may be that the triggers scheduler in Anylogic can be set ahead in outer code loops of the program. The Average Post Rate parameter defines the average number of new notes published per month by each user. It has a truncated normal distribution based on the parameters shown in Table I. The actual distribution of the number of posts is calculated as: Post Rate = 2 × Activity × Average Post Rate. This distribution is plotted in Fig. 1. The mathematical formula of this probability density function is derived in the Appendix. Assuming a month is 30 days, the exact time when an agent publishes a new note is drawn from an exponential distribution every time the user logs in with a rate of 30/Post Rate. There are several different reports on the rate of different posts in Facebook. Additionally, these numbers are intuitively known to keep changing. However, we had to adopt one of these reports, which is that in every 20 minutes, over 1 million links are shared, 1.8 million statuses are updated and 2.7 million photos are uploaded [27]. This results in having 11.9 billion posts per month. Add to that, the knowledge that Facebook had 845 million monthly active users as of December 31, 2011 [28], implies that there are approximately 14 posts per month for each person. This justifies the choice of an average of 13.8 posts per month for each user. Each user is connected to some other users known as friends. The number of friends a user has is controlled by the Friend Count parameter. The number of seconds dedicated for each friend to check their posts is set by the Friend Time parameter. Every time a user begins checking their friend's posts, this timeout value is generated based on the associated normal distribution. Once the time is up, the user leaves the current friend and looks at another friend's recent posts. Complementary to this parameter, the Post Length parameter dictates how many seconds are required to completely read a certain post. When a user publishes a new note, this value and the Post Score parameter are calculated and assigned to the post. The Post Score parameter represents how interesting and appealing a post is. Table I summarizes the model parameters and their distribution properties. As shown in the table, the Activity and Post Score parameter are drawn from standard uniform distributions between 0 and 1. For the other parameters, sampling from truncated normal distributions in Anylogic is employed. This kind of distribution is essentially the standard normal distribution which is stretched by the Mean coefficient, then shifted to the right by Std. Dev., after that it is truncated to fit in [Min, Max] interval. Truncation is performed by discarding every sample outside this interval and taking a subsequent try. C. Environment Akin to online social networks, the environment is a graph where each node represents a user/agent whose social friends are neighbour nodes in the graph. In this subsection, we describe this graph and its properties in detail. The Facebook network graph can be viewed as a small- world network [29], [30] as most nodes can be reached from every other by a small number of hops. Generally, however, scale-free networks are a better choice to model a social network graph, as they have a more realistic degree for the power law distribution. In fact, it is shown that scale-free networks themselves are ultra-small worlds, where the shortest 4 Fig. 1. Post Rate density function TABLE I SIMULATION PARAMETERS AND THEIR DISTRIBUTIONS Parameter Distribution Values Activity Average Login Time* (minutes/day) Average Post Rate* (times/month) Friend Time (seconds) Post Length (seconds) Post Score Mean Std. Dev. Min Max Uniform Normal 0.5 23 1/√12 120 Normal 13.8 13.8 Normal Normal Uniform 30 30 0.5 30 30 1/√12 0 0 0 10 10 0 1 1080 300 600 600 1 * Multiplied by 2 times Activity for calculating the associated rates paths become even smaller [31]. Yet the strict power-law distribution is not accurate enough to represent Facebook's degree distribution [32]. More precisely, if the power-law distribution of 𝑃(𝑘) = 𝑘−𝛼 is accepted to be the degree distribution of nodes, two sections for 1 ≤ k <300 and 300 ≤ k ≤ 5000 can be approximated by a power law with exponents αk<300 = 1.32 and αk≥300 = 3.38, respectively [21]. Therefore it was decided to synthesize the graph by directly sampling from the Facebook graph. Ideally one should generate multiple sample graphs and run the experiments on all the sampled graphs to ensure the results are not specific to just one graph. But due to processing time-limitations, only a few graphs were sampled; as all the sampled graphs shared comparable properties, one representative graph was selected for the basis of the agent-based model. A dataset of the Facebook social graph released by the Networking Group of the University of California, Irvine [21] was used. Two datasets of uniform sampling and Metropolis- Hastings Random Walks (MHRW) were available. The Metropolis-Hastings algorithm is a Markov Chain Monte Carlo (MCMC) method to simulate a complex distribution from which direct sampling is difficult. The MHRW option was chosen, as the Facebook IDs within this dataset are consecutive numbers, which makes it easier to construct the graph. Gjoka et al. obtained this dataset by 28 Facebook-wide independent MHRWs in April of 2009 [21]. The dataset contains the number of friends and their Facebook IDs for approximately 957K unique users. It was not possible to 5 directly build our network structure on the dataset itself, as it was missing a large number of edges. Strictly speaking, the dataset has a uniform random sample of users in Facebook. Here, however, a dataset was required which would densely cover only some small region of the Facebook graph. Such a dataset could be obtained by a Breadth First Search (BFS) crawling method. Hence the following approach was taken: 1) The first 500 sampled Facebook user IDs were assigned to the 500 primary agents in the model. 2) Each primary agent was connected to a number of new secondary agents according to the number of friends of their corresponding Facebook user in the dataset. This resulted in a network of total 89,977 agents, but the number of edges was not sufficient. 3) Extra links between the secondary agents were inserted based on a custom distribution of all the sampled users in the dataset. This resulted in a network with a total of 7,528,164 edges. The network was created and used (saved) for further experiments. After the second step, each node on average had only one connection, which is not at all the case in the real- world Facebook graph. The reason is that the dataset tells us the node x is connected to x1,x2,…,xn; but hardly ever is any information available for each of xi's. As a result, the third step is necessary. The first and second steps are quite straightforward. The challenge is in the third, where creating an undirected graph of n nodes according to a certain degree distribution (here denoted distribution F) is desired. In this case n = 89,477 and the probability distribution F is given by the dataset. Furthermore, it was desirable to create the graph in such a way that nodes with more mutual friends have higher probability of making friends than that of total random nodes. Therefore, the approach is based on the idea of common neighbours as well as the Erased Configuration Model [33]. It is defined as follows. For each node ni, draw a degree di according to the probability distribution F. Make di half-links (or stubs) and connect them to the node ni. When all the stubs are created, start the following loop procedure through the stub pool: 1) Choose three distinct random stubs d1, d2 and d3 and remove them from the stub pool. Notice that each stub is already connected to a node. 2) Among these three stubs, find the two stubs di and dj with more mutual neighbours between their corresponding connected nodes in such a way that they neither make multiple edges between the nodes ni and nj nor make a loop on the node ni or nj. If the step 2) was successful, join the two stubs di and dj to make a connection between the two different nodes of ni and nj; otherwise return one of the stubs (e.g. d1) to the stub pool. 3) The above procedure removes two stubs from the stub pool at each repetition. Therefore, after exactly ⌊∑ 2⁄ ⌋ iterations it terminates; it could be that one stub is left behind in the stub pool. This is quite natural in the Erased Configuration Model, and the proof of its convergence to the desired degree distribution can be found in [33]. 𝑑𝑖 𝑖=𝑛 𝑖=1 To ensure that the graph represents the 'six degrees of separation' phenomenon, the diameter of the synthesized graph (i.e., the longest shortest path) was estimated. To measure such a statistic, 6K nodes (out of roughly 90K nodes) were randomly selected and the Breadth-First Search (BFS) algorithm was run for each node to count the number of reachable nodes at each hop. It was found that all the nodes after three to five hops reach 100% connectivity to all the rest. The percentage of reachable pairs within a certain distance is shown in Fig. 2, which is similar to the graph of degrees of separation of Facebook reported in [32]. The average distance was calculated to be 4.0074, which is comparable to 4.7 and 4.3 of the global and U.S. population of Facebook users in May 2011, respectively [32], [34]. The other statistics regarding the constructed graph, the dataset [21] and the Facebook graph [32] are illustrated in Table II, demonstrating that the dataset with a smaller number of nodes was sampled successfully. As the constructed graph has a much smaller number of nodes (but retains the same number for maximum possible number of friends), it is much less sparse than the real Facebook graph. This is part of the reason the mean distance has been reduced, compared to Facebook. Most conservatively, the synthesized graph can be considered as an acceptable example of Facebook connectivity structures of a sub-region of Facebook network graph. Nevertheless, 90K is almost the maximum possible number of users to handle during the simulations. The reason for this is that all the nodes have their own autonomous behaviour and processes, and their interaction with each other is a function of the number of edges in the graph. The time complexity of only constructing the network graph itself is O(E), where E is on the order of a million. Finally, the complementary cumulative degree distribution function (CCDF) of the dataset and our graph is illustrated in Fig. 3, displayed on a log-log scale. As can be seen in this line graph, the distributions do not strictly follow power-law distributions, which are straight lines on a log-log plot. D. Rules In delineating the scope of the model, the following assumptions have to be made: 1) As we know, there are many different pieces of content, such as photos, videos, links, status updates, event invites, notes and etc. [35] that can be published on a Facebook wall/timeline page. However, throughout the rest of this paper the word "post" or "note" is used to indicate generic content that is posted on a wall page. This is because all content shared on a Facebook page act as incoming stimuli to a Facebook user. The objective in this 6 study does not aim to cover the detailed properties of a successful or uneventful post. Regardless of the specific type of note, each post requires some time to be noticed by user agents. This is why the post length is of interest to us. In addition, each post is trying to convey a message to the viewer/reader. The message could be a warning, funny picture, personal news, an inspirational quote, amazing fact, etc. Each might be appealing to specific types of people. However, to keep the model simple, we considered a single scalar value to represent the general interest in a post as the post score. Each user, depending on their activity level, may share an interesting post. 2) Users cannot send private messages to one another. They can neither make comments on any post nor share any note on somebody else's wall except for their own wall; in other words, they are only able to share something on their own wall. Adding the private messaging property could have only increased the complexity of the model as the model is limited to public means of sharing a post over the network. Posting notes directly on someone else's page is not at all as common as sharing posts on personal wall pages. Incorporating these features would have demanded obtaining more statistical data about the activity behaviour of people on Facebook. Here an effort was made to not increase the number of model parameters, as tuning these parameters is one of the most challenging parts of designing an agent-based model. Even at the current model, values of some parameters rational assumptions rather than actual data. Commenting on posts is a very important feature of Facebook social network. Through comments or lack thereof, a post can stay alive or die. Comments and Likes in Facebook have a direct relationship with the news feed algorithm of Facebook. In simple words, users are more likely to receive more (recent) posts on their home page from those friends to whom they had most interaction in the past. One way of making interaction with friends is via liking or commenting on their posts. The exact Facebook algorithm to rank the news feed page is unknown. In our model, we let each user agent choose which friends to have interaction with. The commenting feature is removed from the model to decrease its complexity. Adding these features back into the model could be a very nice extension of current model. intuitive are based on 3) The only possible relationship between users is bidirectional friendship. This means subscriptions to Pages or Groups held in common between two or more agents, and any other similar features are ignored in the model. Page and groups can be considered as normal user agents in the graph with higher number of friends and activity level. In social science, they are referred to as hubs. So there is no need to distinguish pages from people in our model. Subscription to a Facebook user is similar to a Fig. 2. Degrees of Separation: Percentage of user pairs within x hops of each other Fig. 3. Degree Distribution: The fraction of users who have degree x or greater TABLE II NETWORK GRAPH STATISTICAL PROPERTIES Mean Degree Median Degree Min Degree Max Degree No. of Nodes Mean Distance Facebook [32] Dataset [21] Our Graph 190 99 0 5000 721 M 4.74 168 110 0 4979 957 K N/A 169 111 1 3734 90 K 4.0072 unidirectional friendship. This feature was added later to Facebook. One can think of the current agent- based model as a model of Facebook in its first years without this feature. 4) Similar to the notification feature in Facebook, when a user is online, if a friend of theirs shares a note, the user will receive a notification message within the model. As a result, considering assumption 2) above, there is no way to receive or potentially save a notification for an offline user. In this model, when users go online they begin to check recent posts by friends; as such, adding a notification feature to the model would not have made a significant change to 7 Fig. 4. User state machine the results; but it would make a difference in time complexity of the model. 5) The network is static. This means users cannot make or remove any friendship connections so as not to change the reality-based network graph. The scope of the current agent-based model does not include effects of dynamics of the network graph. There is a viral marketing study where the evolution of network graph and changes in preferences of users for different subjects have been analyzed [36]. This heterogeneity of preferences in different topics is controlled by the activity parameter in our agent- based model. Changing a user's friends does not significantly change chances of repost. Because the activity parameter is uniformly distributed and there is no similarity between friends' activities. One might say more active people are more likely to be friends with one another. But there are many more important demographics characteristics between friends such as nationality and age which trends to be similar. Thus, the activity parameter cannot be considered as a significant factor in this list. The agent-based model keeps an inner state for each user, controlling their behaviour. This stochastic hierarchical state machine is shown in Fig. 4. All the users are in one of the two general states of Online or Offline. Every time the model restarts, the login and logout rate for switching between these two general states are assigned to each user. The intervals when a user is in the Online/Offline state are controlled by login and logout rate. When the login/logout transition is triggered, they log into/log out of the system. Login transition times are drawn by an exponential distribution with the mean value of 𝐿𝑜𝑔𝑖𝑛 𝑇𝑖𝑚𝑒 1440 minutes per a day (1440 minutes). This was explained in more detail in section AII.A. As seen in Fig. 4, all the interactions and events for a user happen when they are in the Online state. Users in the Online state can either produce a note and share it or repost a note which was shared before by another user (which then appears on their own wall, in keeping with assumption 2) above). As mentioned earlier, each user posts new notes with a given rate based on the user's activity. During these intervals between publishing a new note, users check their friends' wall pages. It was possible to employ a stack to sort recently posted notes for each user. However, in order to increase the level of autonomy of users, the freedom to choose which friend to see their posts is given to users. In the preliminary builds of the model, users sorted their friends based on nationality and differences in their ages. As we did not aim to analyze clustering effects of nationality and ages, in the final version, they randomly select friends to review their most-recent wall posts to speed up the simulation. This is also accordant with the Facebook policy to make a ranked list of friends' posts for each user in their homepage (News Feed). This feed is an algorithmically ranked list of friends' posts based on a number of optimization criteria [37]. In its simplest case, the feed contains recent posts of those friends with higher probability of having interaction with the user. For each friend, a normally distributed random value is given as the maximum time threshold to check their wall page. When the time is passed, the user selects next friend to check their wall posts. This loop continues until one of high-level timeouts, such as logging off or publishing a new post, happens. As shown in Fig. 4, when a user is checking friends' wall posts, if one of their friends publishes or reposts a note, they will be notified. Consequently, the user stops their current task of checking friends' page and begins reading the new shared note, which may be considered a result of human curiosity. 8 Users are able to recognize a duplicate post and skip it in one second. Otherwise they would spend time in the amount of the post's length to read it. When a user completely reads a note, they may decide to copy the note from the friend's wall page and repost it. The likelihood of this decision increases with the product of the post score and the user's activity. As an instance, a super active user with the activity value equal to 100% would definitely share an interesting post with the score of one (with probability 1 × 100% =1). One could have picked other formulas to define the probability of sharing. Multiplication, however, was found to be the simplest (and fastest) to proceed with the simulation. By this mechanism, posts propagate over the network, which is the subject of this study. It was also possible to let users share the post before they themselves read it completely. This scenario usually happens when people see a public warning that they think it might be of use to others. Again, adding a new parameter and possibility into the system would have decreased the speed of simulation. It was decided to keep the model as simple as possible in the current version. One might ask if it is not more likely that the marginal score of a post also increases as the post is shared over and over by users. This is the case for product adoption models, mostly inspired by the classical Bass diffusion model [38] or disease epidemic modeling approaches such as the Susceptible- Infected-Removed (SIR) models [39]. In the product adoption models, it makes sense that as the number of people/friends using a certain product grows, others would become more interested in purchasing the product as well. Also, in epidemic models, as the number of infected people with a certain virus increases, the chance of transmission of the virus to others increases. However, propagation of a meme through a social network might be different. Similar to product adoption, some individuals might find sharing a popular post very cool or helpful, such as posts asking to unify people on some belief or giving alerts. On the other hand, some may consider repeated exposure to a specific peace of content as being boring or démodé. The second case is verified in Digg social news website where people have a tendency of not sharing repeated news [40]. Thus, as herein all various types of content are modeled as posts with different scores, we neither increase nor decrease post score as it disseminates through the users; each user, depending on their activity level, might pick a post and share it. A post score, which is fixed, can be determined by external factors which are not part of this study. For example, assuming a post is a product advert, a combination of the market conditions and psychological and sociological characteristics of consumers would determine this factor for each person. There are different studies concerning this aspect such as the decoy effect in [41]. Having implemented an agent-based model, individual consumer's judgment on purchasing a product changes from brand A toward brand B after the introduction of a decoy brand, in a competitive market. showed how an they In the current simulations, for a given post, the post length time must be passed until we can conclude that the user has read the post. For instance, in the situation where a user decides to log out while reading a post, the post will not be checked as a read post. E. Verification and Validation As per Rand and Rust [42], verification of a model ensures the simulated model matches the conceptual model. This procedure is mainly through documentation, program testing and test cases [42]. In the present case, the exact written assumptions and rules of behaviour defined above were coded. The programming code is also extensively commented. Each function and module of the agent-based model was solely tested on a small network of users to receive a known output for a given input. After all the debugging was done, corner cases with extreme values (such as full connectivity, no connectivity, zero activity-level users, one activity-level user, zero post score, one post score, etc.) were run. As per Rand and Rust [42], the four major steps of ABM validation were followed. These are micro-face validation, macro-face validation, empirical input validation and empirical output validation. At the micro-face level, actions of users are a limited form of a real Facebook user's possible actions. Also, the mechanism through which a post propagates is a type of cascade model corresponding to the real world. This paradigm has been validated by various studies over the past. At the macro-face level, it is observed that most posts will not be shared by a friend, which is the same scenario for a typical Facebook user. Also, our aggregate pattern of the post share distribution on-face is similar to Facebook and other social network statistics. At the empirical input level, the ranges of all the parameters are drawn from either Facebook network reasonable assumptions. Further explanation of the input parameters is discussed in section V. statistics or Relative to empirical output validation, it is not possible to exactly validate the agent-based model against reality at this stage, nor is this the claim in this work. The correct way is to run our experiments on Facebook; firstly however, it is not clear how to set a post's probability of being shared (i.e. its score), therefore it cannot be validated in this way in practice, apart from the fact we should be able to monitor exactly how many users have indeed read a post completely. Again this cannot be measured in practice. We may be able to only distinguish if someone has seen the post for at least a few seconds. Unless Facebook or similar social media service providers were using, for example, built-in counters to calculate how many seconds each user spends on a specific post, which is very unlikely for non-video posts; and if there exists such a counter, it would only work when you actually open a post (e.g. a photo) but could not be calculated when you are at your Facebook home page displaying more than one post, as it is not known which post you are exactly looking at. Having said that, it is still possible to demonstrate that real world data are possible outputs of our agent-based model, meaning that our average results match average results in reality. A set of experiments was performed in order to validate the results here against the Facebook statistics reported in [37]. According to the statistics for users' most recent post, "the median post reached 24% of a user's friends (mean = 24%, SD = 10%)," provided that the most recent post was at least 48 hours old. Their population size is 589 different users with the median friend count of 335 (mean = 457, SD = 465). 9 To do such an experiment, 20 unique users were selected such that their friend counts were drawn from a normal distribution with the same mean and standard deviation. As a result, the population of our selected users had a median friend count of 335 (mean = 464, SD = 337) which is quite close to the sample they used in [37]. Then their Activity parameter was set to be 1%, to prevent them from publishing any other note, since we wanted to perform the test for their most recent post; after that a post with score of zero and a short length of 10 seconds was published by the user. The score is set to zero to make sure the post does not spread over the network as we are interested in the number of immediate friends who read the post. The length was set to 10 seconds as this is our minimum post length at the current model. As mentioned above, the statistics reported by Facebook is in fact the number of people who have seen the post, although not necessarily read a long post completely. Then we ran the model. The model ran for 24 simulated hours as an initial warm up stage. Then a chosen user published a specific post. The simulation ran for another 48 simulated hours, after that the result was saved. The whole experiment for all 20 users was repeated two times, for a total of 40 runs. The results were very close in both runs for each user. Real time computation for each run was 9-12 hours on workstations equipped with an Intel Xenon CPU W3679 @3.2GHz with 16 GB RAM or higher configuration. We observed that the median perceived audience size was 19% of a user's friends (mean = 25%, SD= 18%), which is near although slightly less than the expected 24% median of Facebook [37]. One difference in our experiment and Facebook statistics is that our results are the statistics after exactly 48 hours, whereas the Facebook ones are the statistics after at least 48 hours. So it is reasonable to reach a lower percent of immediate friends. Strictly speaking, in our agent- based model, the post life time is defined to be the last moment when the post is read. The average and median post life time in this set of experiments was 47.08 and 46.81 hours, respectively. III. SIMULATION STUDIES In this section, the agent-based model was run with different input settings to explore the impacts of each factor in the post diffusion process. In the current model, the targeted input parameters include: (1) Post length; (2) Post score; (3) Friend count which is the number of friends of the first publisher of the note. As each minute in the simulated world takes 7-10 seconds to run computationally in the real world, we have limited our initial exploration to these three parameters. Future simulations will explore the impact of other parameters, including the day/time to disseminate the note over the graph. The following statistics as the model output were recorded: (1) The number of users who have read the note; (2) The number of users who have reposted the note; (3) The times when the note was read / reposted; (4) The last time when the note was read as its life time. The Analysis Of Variance (ANOVA) procedure between different scenarios is employed to find the importance of each parameter relative to the spread of a message within the online social network. In some cases, multivariate regression was run to test the magnitude of each factor. However as the outputs of an agent-based simulation should not be interpreted quantitatively [17], [18], the numerical values of a linear regression coefficients are interpreted qualitatively. This means even if we obtained numbers as the magnitude of importance of each parameter, they have to be discussed at the qualitative level. A. Study 1: General Insight on Input Parameters Within the first study, a total of 19 unique scenarios (simulations) were set up and each simulation was repeated 20 times, for a total of 380 runs for eight simulated days each. The real-world computing time required for each run was one to two days. The total computing time required for these simulations was over 380 days. In each scenario, our chosen user publishes a certain post after one simulated day of warmup phase, then one week after the spread, the outputs are saved and the simulation ends. The results of the first set of simulations are illustrated in Table III. According to the first (six) rows of the table, it is immediately clear that as long as the number of repost is near zero, the post score does not have much impact on the number read. Because the post is not shared by anyone except for the first publisher, the number read is directly related to the number of immediate friends of the publisher, which is trivial. For the simulation IDs 1-6 plus IDs 9-11, with the fixed post length of 30 seconds, where the post score is relatively small, Analysis Of Variance (ANOVA) yielded a significant effect of friend count (F7.53 = 25.07, p < 0.01), no significant effect of post score (F7.53 = 0.22, p = 0.8) and no significant interaction effect on friend count × post score (F5.12 = 0.23, p = 0.9). Among these simulations with almost no reposts, higher variances in the number of reads are observed for the experiment IDs of 9-11. This is in line with the Facebook observation that "a post produced by a user with many friends has more variability in the audience size than one produced by a user with few friends" [37]. Here, part of the reason is due to the differences in the first publisher's activity parameter. For example, in simulation #9 with zero reposts, it was observed that one repetition of the experiment had a very low activity parameter of 5%; for this reason, the post remained as the most recent post of the publisher and did not slide down the wall page. Consequently, more friends had chance to read this post. In this simulation, we had the highest number of reads which in turn increased the variance of the number of reads. Also, the time of publishing a note relative to other events at the time is another important factor to receive a high number of reads. In fact, the activity parameter and timing both represent the complexity and heterogeneity of users and their interaction within the system. We did not explicitly analyze the impact of the activity parameter in the number of reads. However we did observe that for a certain post length in cases where the post score is low, the friend count parameter has the greatest impact and the publisher's activity is the second dominant parameter in determining the audience size. The reason is that, if the post is not interesting enough to be shared by others, it would only be read by the immediate RESULTS IN FIRST SET OF SIMULATIONS AFTER ONE WEEK TABLE III Post Score Post Length Friend Count No. of Reads No. of Reposts Mean Median Std. Dev. Mean Median Std. Dev. 10 0.001 0.01 0.1 0.001 0.01 0.1 0.3 0.5 0.001 0.01 0.1 0.3 0.5 0.1 0.3 0.5 0.1 0.3 0.5 30 30 30 30 30 30 30 30 30 30 30 30 30 60 60 60 90 90 90 9 9 9 139 139 139 139 139 530 530 530 530 530 530 530 530 530 530 530 2.3 1.7 2.0 13.0 10.5 17.1 33.0 101.1 45.8 57.0 55.9 73.3 2075.5 30.4 33.4 41.6 12.3 12.8 14.1 2 2 1 10 8 13 25 13 28 41 42 50 1588 20 24 23 10 10 11 1.4 0.7 1.2 10.3 7.7 16.5 28.8 341.7 46.6 45.9 50.6 68.0 2109.5 24.1 25.9 52.7 8.2 11.7 12.5 0 0 0.2 0 0 1.30 6.2 32.1 0 0.3 4.2 14.0 702.1 2.3 6.1 13.8 0.6 2.6 3.7 0 0 0 0 0 1 5 4 0 0 2 8 532 2 6 7 0 2 2 0 0 0.4 0 0 1.9 6.5 110.2 0 0.5 4.8 13.6 724.1 2.2 4.6 18.5 1.1 3.1 3.8 0 ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 friends of the publisher. So in order to increase the number of reads in this case, a higher friend count would help. Secondly, a lower activity level by the publisher keeps the post recent and top on their wall page. This in turn increases chances of being seen by others. The initial implication is that if one cannot make an attractive post with high interest, at minimum, one needs to have it posted by a user with large number of friends in order to reach its maximum audience. Of further interest is knowing more about the importance of a post score versus a publisher's friend count. Assuming the note is interesting, the question is whether one should focus on finding a hub user with many friends to post it or should one improve the quality of note as much as possible. Consider the simulation IDs 6-8 and 11-13, all of which have the same post length but relatively high scores published by users with different friend counts. Among these experiments, ANOVA yielded a major effect of post score (F4.8 = 19.04, p < 0.01), a bit weaker but still strong effect of friend count (F6.86 = 18.43, p < 0.01) and major interaction effect on post score × friend count (F4.8 = 16.36, p < 0.01). We can conclude that in cases where the post interest and the friend count are large enough, the former parameter is more influential than the latter. Keep in mind that friend count is still important, and one needs to consider the combined effect of both together. However, if one is able to find a user with an acceptable number of friends, it is recommended to focus more on post content rather that necessarily finding a hub user with many friends. This phenomenon implies that having a good seeder may help reach/saturate a local cluster of the network faster, but ultimately a higher post score is needed to reach further regions of the network. The next question is the trade-off between post quality and post length when we have a well-connected user with relatively high number of friends to publish the desired post. To compare the impact of post score versus post length, for simulation IDs 11-19, the number of reads is shown as a heat- Fig. 5. Impact of Post Score versus Post Length on number of reads (medium to long notes) map in Fig. 5. According to the figure, post length dominates post score within all the range of post scores and lengths simulated. This means in order to reach a maximum audience, keeping the message brief is more significant rather than making an impressive but lengthy one. For example, by comparing the simulations #11 and #16, both of which have similar number of reads, it is observed that for a long note, more reposts (and consequently more time) is required to reach a similar audience size of a short note with a lesser score. Statistically speaking, for the simulation IDs 11-19, ANOVA yielded a strong effect of post score (F4.73 = 18.33, p < 0.01), a bit stronger effect of post length (F4.73 = 20.35, p < 0.01) and a significant interaction effect on post score × post length (F3.43 = 18.03, p < 0.01). Therefore both post length and post score are important properties, and their combined effect has to be considered when making a post; yet, depending on the situation, the post length can be considered of more importance as if the length extends over a certain threshold it 11 RESULTS IN SECOND SET OF EXPEREMINTS AFTER ONE DAY TABLE IV Post Score Post Length Friend Count No. of Reads No. of Reposts Mean Median Std. Dev. Mean Median Std. Dev. 0.001 0.1 0.25 0.5 0.75 1 0.001 0.1 0.25 0.5 0.75 1 30 30 30 30 30 30 10 10 10 10 10 10 319 319 319 319 319 319 319 319 319 319 319 319 33.8 35.2 48.4 1857.2 7651.8 23559 46.7 92.2 120.7 12449.9 32473.5 41165.8 28 19 41 58 6670 23468 46 101 91 16964 32297 41036 25.4 32.9 26.4 2465.4 8085.1 507.6 12.3 55.2 95.7 8,608 1,007.3 1071.6 0 2.6 7.6 635.8 3815.3 15325.6 0 5.1 18.1 4077.9 15375 25424.1 0 2 7 17 3324 15213 0 4 10 5565 15256 25330 0 2.5 3.9 852.7 4035.3 340.8 0 4.7 18.9 2,825.3 457 587.5 ID 20 21 22 23 24 25 26 27 29 30 31 32 Fig. 6. Impact of Post Score versus Post Length on number of reads (short notes) severely affects the post reachability no matter what the post score is. The reason is that users generally do not spend much time to judge a post. For example, a post might be very amusing, but as it is lengthy a typical user never spend sufficient time to recognize true score of the post. B. Study 2: Importance of Score versus Length in Short Posts After inferring that a post should not be too long if it is to propagate, in the second study, 12 more scenarios were simulated to test a broad scale of different post scores versus relatively short post lengths. In this set of simulations, the user who publishes the post for first time is a new user with a different position in the network and a subsequently changed local network structure. This ensures our results are not specific only to some local region of the network structure. This time, a total of 12 unique simulations were set up and each simulation was repeated 10 times, for a total of 120 runs for two simulated days each. In each simulation, the post spread began after a 24-hour initial warm-up phase, and then one day later the outputs were saved and simulation ended. In these experiments, smaller post lengths of 10 and 30 seconds with higher chances of sharing were tested. The post score changed from 0.001 to 1 to study a broader scale of scores. The results of our second set of experiments are shown in Table IV. According to Table IV, as a post score increases from 0.25 to 0.5, a significant change in the number of reads is observed. Obviously, post score should likely have the greatest impact in general. For a more detailed comparison, the simulation is split into two subsections of low and high post scores. One heat-map for each part is shown in Fig. 6. From the heat-map on the left corresponding to low score posts, it can be seen that post length dominates the post score. On average, all the scenarios of shorter (i.e., 10 seconds) posts reach comparable size or larger audiences than longer (i.e., 30 seconds) posts with any scores, as long as post score is not very high (less than 25%). This verifies the previous result with a different seeder and time to collect the result. However, for the heat- map on right, this inference does not hold true any longer. For higher score posts, shorter posts do not necessarily reach more users, and it depends on both post score and post length together. For example, simulation #25 with a post length of 30 seconds and score of 100% has reached a larger audience than simulation #30 with a shorter post length of 10 seconds but lower score of 50%. Yet simulation #31 with a short post length of 10 seconds and score of 75% has larger audience than simulation #30. From an ANOVA perspective, a significant effect of post length (F6.88 = 190.99, p < 0.01), a more significant effect of post score (F3.19 = 293.77, p < 0.01) and a significant interaction effect on post score × post length (F3.37 = 46, p < 0.01) is observed. Therefore, generally both post score and post length are very important in the success of a post propagating. However, the relationship with these two parameters and audience size is not linear. Assuming it would be desirable to explain impacts of each factor in the number of reads with a linear model, Table V 12 COEFFICIENTS OF LINEAR MODEL OF NUMBER OF READS FITTED TO SECOND STUDY TABLE V Model input Coefficients Std. Error Intercept (constant) 3552.302 1981.203 Post Score* 33077.74 1788.742 Post Length* Activity * Significant at an alpha level of 0.01 -441.586 1923.83 63.51394 2441.076 TABLE VI Standardized Coefficients 0.820 -0.308 0.035 t Stat P-value 1.793 0.076 18.492 2.2E-36 -6.953 2.25E-10 0.432 0.788 COEFFICIENTS OF TRANSFORMED INPUTS OF LINEAR MODEL OF NUMBER OF READS FITTED TO SECOND STUDY Model input Coefficients Std. Error Standardized Coefficients t Stat P-value Intercept (constant) -9029.287 1353.794 -6.670 ≤0.001 (Post Score)2 3.396 0.158 0.849 21.443 ≤0.001 1 / (Post Length) 132908.250 17021.300 0.309 7.808 ≤0.001 TABLE VII COEFFICIENTS OF GENERAL MODEL OF NUMBER OF READS FOR THOSE POSTS FOR WHICH NUMBER OF REPOSTS ≥ 1 Model input Coefficients Std. Error Intercept (constant) 27952.530 69.463 Standardized Coefficients t Stat P-value 402.410 √ (Friend Count) 149.849 1.636 .094 91.583 ≤0.001 Log (Post Length) -21236.924 42.333 -.510 -501.659 ≤0.001 (Post Score)2 1.905 .004 .553 542.076 ≤0.001 shows the result of fitting a linear model to our second study. The linear regression model has an adjusted R2 of 76%. However, expectedly, the plot of residuals does not suggest a linear model as being a suitable model for these scenarios. The coefficients statistics confirm our previous inference using ANOVA and heat-maps. According to Table V, considering a α significance level of 1%, the activity parameter does not have any significant impact on the number of reads with a p- value of 0.4. Both post score and post length have significant impacts. Post score has higher influence within all the simulations studied in the second set of simulations together. In order to obtain a better model of number of reads, each of the following (five) functions was applied to our model inputs: Inverse, Logarithm, Square root, Square and Cube. So instead of three inputs of Post Score, Post Length and Activity, there are now 3×6 = 18 inputs to choose from. The best linear model based on all these inputs was found to be a linear model of post score squared and inverse of post length. The new model has a better adjusted R2 of 81% with the coefficients shown in Table VI. Properties of over one million posts were saved after two days from the starting point of simulations in different runs for another linear regression. The basic inputs were post score, post length and friend count. However, once again the five functions of Inverse, Logarithm, Square root, Square and Cube were applied to each of the basic inputs to find the best combination of inputs for the regression model. The best model was found to have an adjusted R2 of only 32%. Subsequently, posts with zero repost were excluded from our input data to have more coherent input data, which resulted in almost half a million posts remaining. With the objective to have only one function of each of the basic inputs as an input to our linear model, the best fitted model was found to be a linear model of post score squared, logarithm of post length and square root of friend count. This model has an adjusted R2 of 54%. The coefficients of the model are shown in Table VII. The scatterplot of standardized residuals versus standardized predicted value is shown in Fig. 7. Ideally it would be desirable to have a uniform scattering of points around the zero reference line; but as the agent-based model is highly non-linear, a better linear model could not be fit to its outcome. All the inputs have coefficients significantly different from zero. Additionally, by looking at the standardized coefficients, it is clear that generally both 'logarithm of post length' and 'post score squared' have significant impacts on number read. Yet post score squared is a stronger predictor. According to the second study, one needs to avoid lengthy posts, and the post still needs to have some minimum score. There are certain thresholds for post lengths and post scores that post properties should lie within. Initially, it is better to focus on post score to gain some interest and chances of sharing, then one should try to shorten the post length considering the limited time of users. Lastly, when the message is brief enough, fine-tuning the post score can achieve better results than slightly reducing of post length. 13 There is an uncommon difference between the median and average number of reads in simulation #23. The low median number of reads can suggest that a typical user may not read a long post. However if, with the help of good timing, users read a long (but interesting) post, the post would propagate very well through the network increasing the total number of reads. Therefore in simulation #23 when the timing is matched, we observe a high number of reads increasing the average number of reads, and when timing does not cooperate well, users show little interest in the post. C. Study 3: Comparison of two Seeding Strategies In the third and last study, the objective was to gain insight by comparing a classical seeding strategy of a small number of hub users with many friends versus seeding a large of number users with few friends. First, a special post was published by four randomly chosen users in the network, each of which has exactly 50 friends. Then, the simulation was repeated for two randomly chosen users with 100 friends each. The details are described below. All four users had a fixed activity level of 50%. At a certain time after the initial warm-up phase, they all shared a unique short post with 50% score and 10 seconds length. Then 24 simulation hours after the spread, the numbers of reads were collected and saved. The simulation was repeated 10 times. Then the whole simulation was repeated for two more sets of four random users with exactly 50 friends, for a total of 30 runs for the 4×50-friend case. The same simulation was carried out for three different sets of two users with exactly 100 users each, for a total of 30 runs for the 2×100-friend case. The average number read for the 4×50-friend case and the 2×100-friend case is 13300.7 and 9540.1, respectively. The statistics are shown in Table VIII. The average results suggest that a mass of small seeders may broadcast a certain post better than a few hub users. However, assuming the more general case of not-equal variances for these two cases, one cannot reject the null hypothesis of having equal means in these two cases. Technically, the null cannot be rejected by t-test with a t stat of 1.679 and the degrees of freedom of 57.426. This is equivalent to a non-significant p-value of 0.099. In other words, the difference between the two cases is not statistically significantly different. To see the trend of message propagation, Fig. 8 plots the number of reads and number of reposts versus time for a run of simulation #13 on a log-log scale. All other simulations where the post is shared by some users have a similar trend. The only time required for the outbreak to propagation is the seconds required to read the post. Unlike other diffusion patterns such as product adoption or the spreads of infectious disease, this trend is not s-shaped for Facebook posts. In other words, there is no classical tipping point or epidemic threshold for post propagation after which we could expect an outbreak in the number read. The reasons are discussed in section IV. This trend may sound surprising at first, yet it is consistent with real observations of Facebook. According to a select Fig. 7. Scatterplot of standardized residuals of linear model versus standardized predicted value of number of reads TABLE VIII STATISTICS OF NUMBER OF READS IN THIRD STUDY Case Mean Std. Deviation Std. Error Mean 4×50-friend 13300.73 8228.25 2×100-friend 9540.10 9096.65 1502.27 1660.81 group of brand posts data of Facebook in November 2012, each post, on average, reached half of their target audience within 30 minutes after publication [43]. Also a quite similar trend is reported for the number of retweets in Twitter [44]. Another surprising result is that the median number of reposts is zero. This means it is quite common that posts by a typical user on Facebook do not attain even a single repost. We confirm that there exist posts with high number of reposts in both Facebook and our agent-based model. These highly shared contents are mostly (high scored) posts published by popular pages, which can be thought of as hub users in our agent-based model. Yet many of the posts, especially the ones submitted by typical users receive few or very limited reposts. Most of the published results about social networks are generally focused on successful posts and their properties, and as such, statistics regarding the failed ones submitted by random usual users could not be found. The total distribution of number of reposts is shown in Fig. 9. This long-tailed distribution indicates only a few posts gain a huge number of reads. The distribution is also consistent on the surface with that of retweets (popularity) for twitter reported by [44]. Section IV discusses this consistency. Initially, we expected that having a score of 10% would be enough for a post to be broadcast and seen by everybody. Firstly, in this case, the average probability of reposting is 1/20. Secondly, when the post is published for the first time, only the immediate friends will see the post. The immediate friends should be online in a certain time period to catch the post. Therefore, if the friend count is not sufficient, the post will never have any chances to spread further. In other words, if the network were fully connected so that users observe the whole population, the reposting probability of 1/20 might have been adequate, however this is not the case in this agent-based model. IV. RELATED WORK AND DISCUSSION Diffusion is a core process in many areas such as energy flows in physics, disease spread in biology, behavioral contagion in sociology and product adoption in economics. Several researchers have studied aspects of this phenomenon in human systems. In human systems, a general way of social contagion is through the word-of-mouth (WOM) mechanism which is quite similar to spread of an infectious disease within a population. Below, the current simulation results are discussed in relation to other common diffusion processes of product adoption (marketing), disease spread (epidemiology), and electronic networks. Within product adoption research, Moldovan et al. studied the importance of WOM to new product success [1]. In their research, a product score is divided into two dimensions of originility and usefulness of a product, and the number of online reviews is used as a proxy for the amount of WOM. Reference [5], using an ABM, studies the actual value of a seeding program for WOM in terms of market expansion and purchase acceleration for a certain product. Reference [36] attempted to predict users' adoption of a given product on the Digg social news website. They studied the effect of network- level dynamics and changes in individual preferences for different topics, and proposed a viral marketing strategy which was tested with an agent-based model. Reference [45] studies social commerce where individual sellers are connected to each other through an online social network. Using time series analysis at the marketplace level and Bayesian statistical analysis at the shop level, they explore whether connecting the sellers to each other increase the sale, and how the position of a seller within the networks influences their value. Reference [46] studies impact of different connection patterns among individuals on the diffusion process in a European online social network. Using a hazard-rate model, they investigate characteristics and local structure of potential adopters and their neighbours. Tirunillai and Tellis also confirm the importance of user- generated content as a type of WOM and in particular study the impact of online product reviews and ratings on stock market performance using multivariate time-series models [7]. Similarly, [3] finds that consumer generated online product ratings has a direct relation with the product sales; although the previously submitted ratings affect the future ratings. Reference [47] also investigates the evolution of online reviews of books over time and sequence. Weblogs as a part of the larger set of online social media can also influence purchase/adoption of a product. An analytical model for a blogger is made and studied in [48]. In particular, they explain why blogs may link to rivals, and what the relative benefits and costs of such linking are. Goldenberg et al. study the role of hubs in the diffusion process of products over the Cyworld social networking site in Fig. 8. Trend of message propagation 14 Fig. 9. Reposts distribution Korea [49]. They define two types of innovator hubs who need little exposure to adopt a product and follower (social) hubs who are well-connected. Using an ABM, they analyzed the impact of each type on the market eventual size and the speed of adoption. Cellular automata modeling and aggregate level modeling together are used in [50] to study growth rate of a new product. By separating network externalities effects such as mass media and advertising from internal interactions (i.e., word of mouth through the network), they found a chilling effect on growth rate of a product. This is a "wait-and-see" state for a product when potential consumers wait for others to adopt a certain product and then decide whether to purchase or not. This partially explains why product adoption has an s- shaped growth rate. Thus growth of products can follow a two-stage process which includes a slow start due to the chilling effect and then a fast growth because of the bandwagon effect. However, a similar concept for post- sharing does not exist. One does not need to see if others have shared a post to decide whether one should or not; basically because one is not purchasing a post with real currency. In the context of contagious disease, [51] studies SIR models of epidemic disease spread in networks. Rahmandad and Sterman compare AB and DE models of epidemic spread on different networks [52]. There are some similarities and differences between the spread of an electronic post and an infectious disease. The propagation is similar in the way in 15 which it cascades through the network. Unlike posts and similar to products, disease has an s-shaped spread rate (for basic review, see [53]). In a disease epidemic, the rate of people entering from a susceptible into the infected state depends on the number of people in other states (such as the infective state). As people become more and more infected, chance of transmission of disease is increasing until the epidemic threshold is passed and the disease rapidly spreads throughout the population. In our agent-based model, it is assumed that a post score, corresponding to chances of reposting, is constant. In epidemiology, a quicker recovery rate makes a smaller epidemic; it takes a longer time to happen as people recover faster and the number of infected in the beginning of the process is not enough to form an outbreak. In our agent-based model, users need to read a post completely (i.e., become infected) then may spread it. With a short post, a shorter time is expected for the outbreak to occur, and it has a higher peak, as the time needed to read the post is smaller and more people are likely to read it. During the infection period of the disease epidemic case, the chance of transmission of the disease is constantly present; so the epidemics with a slower recovery (removal) rate (i.e., a long infected period) that slowly kill people are more dangerous to the populations having more deaths at the end (for discussion see [53] pages 21-22). In social networks, users generally share a post only once and their friends usually read the post once and decide whether to share or not. As the assumption is people do not share a post while reading it, a longer post (unlike a longer infected period of a disease) does not provide a better opportunity to spread the post. Lastly, the ultimate objective in disease modelling is mitigating against the spread of the disease such as targeted vaccinations as a means of achieving herd immunity; but in social networks, more spread and penetration are desired. Diffusion of applications in Facebook is studied in [54] using a customized commercial application about the movie industry to track user behaviour. Using hazard modeling, they test effectiveness of passive-broadcast messaging versus active-personalized messaging. In another study on Facebook, susceptibility of various type of users (e.g. young, women, married etc.) to influence their adoption decisions is measured [55]. Application adoption and social influence in Facebook is also studied in [56] using fluctuation scaling (FS) method. They track popularity of a set of applications among all users in their dataset collected in 2007. Their observation is limited to the adoption of an application, and not necessarily the usage. Similarly, sharing a post needs an increased level of engagement, rather than simply reading it. Reference [57] analyzed the prevalence data reported by a computer virus for a time window of 50 months. They found the absence of an epidemic threshold for virus spread on scale- free networks due infinite connectivity effect phenomenon in large scale-free networks. This effect happens because of a heterogeneous rising and falling in the number of links of nodes in the scale-free graph. This factor is also applicable to the current agent-based model as the synthesized graph shares this feature of scale-free networks. to an are and There various sometimes conflicting recommendations regarding choosing the optimal set of users to publish a message (see [58] for summary of previous research). Reference [58] analyzes four different seeding strategies for both messages and products using two small- scale field experiments and one real-life viral marketing campaign. Their four seeding strategies are: 1) seeding well- connected hub users; 2) seeding low-degree members, called the fringes; 3) seeding high-betweenness users, called the bridges that connect different sub-networks; and 4) random seeding. They find high-degree and high-centrality strategies are preferable in general. In contrast, [59] finds large cascades are driven by critical mass of individuals, but not necessarily influentials (hubs). In the current third simulation, some evidence in favor of this strategy of mass of individuals also emerged. However, the test was limited and the results were not significantly different from seeding a few hub users. In summary, it is safe to conclude each strategy may work well under certain conditions. in Furthermore, the current agent-based model, by investigating properties of the most successful posts, it was found that as long as a post is not qualified enough to be reposted by typical users, it does not have any chance to diffuse across the whole network. In other words, relying only on reposts by active users (with high tendency to share) does not guarantee message propagation. Because wall pages of active users who repost many notes tend to be crowded with different notes, the note would only have a minimal chance to be read by someone else. Nevertheless, if the note is reposted by a selective hub user with many connections but low activity to post other notes, a significant growth in the number of reads and reposts will be achieved. As the post remains as the most recent post of the hub, it gains maximum exposure to the hub's friends, although one still needs the post to be highly scored to be ultimately reposted by a critical mass of users. Therefore, is recommended to produce the desired post and begin its publishing by different hubs who have many connections but (who) are quite discerning about their willingness to share other notes. Reference the degree distribution of a network on adoption process, using an agent- based model. They discovered that while most researchers simply assume the adoption process propagates over the entire visible (overt) network, the actual active subset of the network over which the propagation occurs may have quite different properties from the entire network. They provided evidence that the degree distribution of the active network is generally different from that of the entire network. The degree distribution has a significant effect on contagion properties of nodes within the network. to achieve a maximum audience, [60] analyzes impact of the it In an inspiring paper by Lerman and Ghosh, the active network of users on Digg and Twitter is extracted, and then it is studied to see how the structure of the (active) network affects the dynamics of information flow on each network [44]. The general mechanism of the spread of information in both sites are similar to Facebook, where users watch their 16 friends' activity, and they may share/tweet/vote for a post to make the post visible to their own friends/fans/followers. The precise underlying details vary from case to case though. As mentioned in section III, the evolution of number of tweets received by each post in Twitter reported in [44] is utterly similar to that of reposts in the current work as displayed in Fig. 8; (successful) posts display a burst of growth at the beginning, and the growth saturates after a while. The point at which growth saturates is the cascade size indicating how far and wide the post has penetrated though the network. Reference [44] reports that Digg's (active) social network of their dataset has a larger clustering coefficient than Twitter's, meaning that its network is denser. As such, initially posts spread easier in the more highly interconnected Digg's network, but they eventually spread more distance in Twitter's less densely connected network. The distribution reported for the number of retweets (or reposts) in [44] is somewhat analogous to the distribution of number of reposts in the current agent-based model displayed in Fig. 9. Retweets distribution in Twitter has a small number of posts (or tweets) with almost zero retweets, followed by an exponential peak and then a gradual decrease of retweets to zero again, creating a long tail. This means a majority of posts gain a few (but non- zero) retweets in Twitter and very few posts exist with many retweets. Fig. 9 states the same fact finding, except that it claims that the majority of posts in Facebook have almost zero reposts. The reason is that the Twitter dataset used in [44] does not include any non-popular tweets (with low score). They collected only frequently-retweeted posts on Twitter using the Tweetmeme.com site at the time. In addition, tweets can only have a limited number of characters. As such, they all can be considered as short-length posts as in the current agent- based model, with higher chances of being reposted. However the distribution reported in Fig. 9 corresponds to all the posts with various length and score simulated within the agent- based model. Reference [61] employs an agent-based model to simulate reposting behaviour of users in the Twitter social network. They explicitly model the competition for humans' limited attention among different posts (memes), and how it affects memes' popularity. They developed a memory mechanism to reflect users' past behaviour to what they would share in the future, as users are likely to share posts similar to what they shared before. Similar to the result we presented in this paper, long-tailed distributions for memes' popularity and lifetime are reported. They found extremely heterogeneous behaviour; a few memes are extremely successful while most of them die out quickly. Reference [62] proposes a coupled hidden Markov model to capture neighbour users' influence on users' posting activity. Their model is also tested on Twitter. An interesting study on information cascades on Digg by Steeg et al. revealed that a high level of clustering structure of the Digg network limits the overall growth of cascade [40]. In highly clustered networks, people are usually exposed to a certain post multiple times through multiple friends, which in turn lowers the epidemic threshold, speeding the spread up initially. However, the surprising effect reported by [40] is that repeated exposure to the same post does not encourage people to repost it. This is a fundamental difference between the spread of information and classical spread of products or diseases. Reference [40] shows this effect drastically limits the overall cascade/epidemic size. As an example, while many posts with a fast starting spread exist in their dataset, only one post about Michael Jackson's death reached a significant fraction of 5% of active Digg users. They observed that the effective number of people who have not been exposed to a post is gradually decreasing. In addition, other effects such as decay of visibility and novelty could be other reasons why the epidemic stops [40]. In one more study on Twitter, rapid decay of visibility combined with the limited attention of users are determined to be the primary reasons for preventing the growth of propagation of online information [15]. The way the current model network described in this paper is synthesized does not produce a high (and desired) level of clustering coefficient. This means that the number of reads for successful posts may have been overestimated. That being said, most successful posts in the current simulations end up reaching around 30% of the entire network. The percentage of viral posts is less than 2.5% of all the posts generated within the model. Moreover, it was discovered that all these posts had extremely high scores of mostly over 90% and short lengths of mostly 10-15 seconds published by users with various friend counts. Precisely speaking, these viral posts have scores ranged from 70% to 100% with a mean and median of 89% and 91%, respectively. The post lengths are varied from 10 to 25 seconds with a mean and median of 15 and 14.5 seconds, respectively. It implies that viral messages can be published by a user with low number of friends, but certainly various users including hubs would have to repost it during the spread. More importantly, the post must be brief while extremely highly qualified. We used a singular numerical value to represent a post score, but a post can have different aspects to be engaging to different kinds of people. There are various categories of posts: promotional offers/deals, advice and tips, warnings, amusing video clips, amazing pictures, personal news, motivating speeches, campaigns recalls etc. Thus, a score of 90% in our simulations represents a high quality post in a variety of features. But in reality, it is nearly impossible to make such a universally fascinating post. Certainly, viral posts exhibit a variety of features. For example, death news of a famous celebrity has a wide range of viral features making it likely to go viral. Such a post has information, novelty, may contain stimulating quotes or represent a group of people's mourning or respect. V. CONCLUSION AND LIMITATIONS There has been a great deal of work on dissemination of information in online social networks especially on Twitter and Digg. Studies on application adoption in Facebook also exist. This paper presents a large scale stochastic agent-based approach for modeling wall post propagation within the Facebook network. Network and other input parameters have been drawn from and tuned to published sources of Facebook 17 statistics. Other studies on the realm of social media diffusion have confirmed the importance of various factors including underlying network structure, local network structure of following spreaders, influence degree and activity level of each spreader, type and novelty of a post and people's response to repeated exposure. At the current stage, analyzing all these factors in a single study would be too complex or too generic. The agent-based platform created here is potentially capable of testing all factors by some code modifications. In some cases, one would have to insert a few more parameters into the system, such as a social influence parameter for each person for example. In this paper, various scenarios have been explored to investigate the impact of post length, post score, and the post publisher's friend count on post diffusion. It is observed that posts with small scores hardly ever spread farther than the immediate friends of the publisher, meaning that higher friend count for the publisher has a stronger impact on audience size than higher post score does. However, beyond a certain level of message quality, the post score has a larger influence on message propagation through the network than publisher's friend count. This intuitively means the content of a message is more important than who has delivered it. In cases with relatively medium post interests and medium friend counts where there definitely exists some reposting, the post length is the most influential parameter followed by the post score and the friend count as the second and third influential factors, respectively. This implies that creating a long post makes it boring and significantly affects its chances of getting shared by others. Whereas for relatively short posts, increasing the quality of post contributes more to the audience size than cutting the post length any shorter, which does not necessarily boost the growth of number of reads. The intuition behind this result is that people spend a minimum amount of time on each post, and once a post length is below that minimum length, there would be no need to make the post any shorter. Keeping these hints in mind could help marketers to find a balance between length and content of post making their ideal post advert for example. Adjustments are needed to study other online social networks using this agent-based model. For example, reconstructing the network, limiting post lengths and the way users look through the posts from friends have to be changed in order to study the Twitter social network. However, the results reported here may not be limited only to the Facebook online social network. Less intuitive findings about the dynamic of post spread mechanism such as lack of epidemic threshold in the propagation of information have been previously confirmed in other online social networks. Along with other work on seeding strategies, it is observed that both hub-seeding and having a large number of individual seeders could result in having a viral post reaching an epidemic portion of population of at least a sub-region of the network graph. The simulations performed to compare these two strategies were limited to conclude a general statement though. Achieving an 'epidemic reach' of the entire Facebook network is nearly impossible. People are online at different times and it is vital for a post to catch their attention when they are online. Assuming the timing can be handled through the interface design of online social networking websites, a viral post still needs to be highly scored in a variety of features. Each person is likely to become engaged in a certain category of posts. Modeling a post score based on a single scalar value may be an over simplistic assumption. A much more realistic way to define a post score is using a vector where each component describes post content in a different perspective. Then users' activity parameters also need to be vectorized to capture the heterogeneity of population in different directions. Add to this the fact that certain categories of contents (e.g. politics, fashions, and sports) are being shared more on certain social networking sites (e.g. Twitter, Pinterest, and Facebook) [62]. The limitation in the current paper is that an active user is likely to share any type/category of posts, and also a high-scored post has all the features of all sorts of appealing posts which cannot be true in practice. One more realistic extension of current research would be the insertion of dynamic scores for post. There are different ways and reasons to change a post's score during the spread. A rational reason is that after users spend a few seconds on a post, they get a better idea of how interesting the post is. So they may decide to continue reading/watching the post or disregard it. This is especially true for video posts. Once the dynamic scoring feature is added into the model, one can let users share a post even before they read it completely. There are other limitations with the current study which bring opportunities for further extensions of this research. Sensitivity analyses for input parameters such as average post length or login rate were limited only to the primary stages of making the agent-based model. The parameters and distributions were tuned on the basis of a smaller network and Facebook's known statistics. The parameters were fixed once it was observed that the main agent-based model of the larger network had rational functionality in line with the conceptual model. Testing sensitivity of all the outputs to all input parameters on such a large set of data would require a lot of time and effort. Such reports could offer insights on what to expect if people begin spending twice as much time on Facebook for example, or investigate robustness of results across a range of parameters and distributions. In addition, the exact correlations, threshold values and interactive effects of simulation inputs (post length, post score and publisher's friend count) still need to be determined through more simulations. for future work One crucial direction is a more comprehensive way of modeling the Facebook news feed ranking algorithm, which used to be called the EdgeRank algorithm. Our current model of EdgeRank emphasizes Recency of posts too highly, meaning that if a post is recently published, it has the highest chance to be seen by users. Facebook tries to identify those who have most interaction with a user. Currently the algorithm is considering numerous factors (including recency) to decide which posts to show for each person in their home news page. For example, if one likes or comments on posts by a person/group, chances of 18 receiving more posts from those persons/groups will be increased. So modeling these two Facebook features could be an essential feature to complement the current agent-based model. APPENDIX Let 𝑈 = 𝐴𝐵, where 𝐴 = 𝑈[0,2] and 𝐵 = 𝑁(𝜇, 𝜎2) are independent random variables (rvs). Using the following formula for the Probability Density Function (PDF) of product of two independent rvs, [11] G. Peng and J. Mu, "Technology adoption in online social networks," Journal of Product Innovation Management, vol. 28, no. s1, p. 133– 145, 2011. [12] C. R. T. Shintaro Okazaki, "Social media and international advertising: theoretical challenges and future directions," International Marketing Review, vol. 30, no. 1, pp. 56-71, 2013. [13] H. Park and H. Cho, "Social network online communities: information sources for apparel shopping," Journal of Consumer Marketing, vol. 29, no. 6, pp. 400-411, 2012. [14] T. Hogg and K. Lerman, "Stochastic models of user-contributory web sites," Proc. of the 3rd Intl Conf on Weblogs and Social Media (ICWSM2009), pp. 50-57, 2009. 𝑓𝑈=𝐴𝐵(𝑢) = ∫ ∞ −∞ 1 𝑥 𝑓𝐴(𝑥)𝑓𝐵 ( 𝑢 𝑥 ) 𝑑𝑥 we can find that in our case 𝑓𝑈(𝑢) 1 √2𝜋 ∞ ∫ 𝑢−𝜇 𝜎 1 𝑥∗𝜎+𝜇 𝑥2 2 𝑑𝑥 𝑒− ; 𝑢 > 0 { −1 √2𝜋 𝑢−𝜇 ∫ 𝜎 −∞ 1 𝑥∗𝜎+𝜇 𝑥2 2 𝑑𝑥 𝑒− ; 𝑢 < 0 (1) [15] N. O. Hodas and K. Lerman, "How visibility and divided attention constrain social contagion," in Proceedings of the 2012 ASE/IEEE International Conference on Social Computing (SocialCom-2012), Washington, DC, 2012. [16] R. Axelrod, The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration, Princeton, NJ: Princeton University Press, 1997. [17] E. Bonabeau, "Agent-based modeling: methods and techniques for simulating human systems," Proceedings of the National Academy of Sciences of the United States of America, vol. 99, no. Suppl 3, pp. 7280-7287, 2002. [18] R. Garcia, "Uses of Agent-based modeling in innovation/new product development research," Journal of Product Innovation Management, vol. 22, no. 5, p. 380–398, 2005. [19] T. Rowland, "Computational Irreducibility," [Online]. Available: http://mathworld.wolfram.com/ComputationalIrreducibility.html. [Accessed 23 Dec 2014]. [20] M. Gladwell, The Tipping Point: How Little Things Can Make a Big Difference, Little, Brown and Company, 2000. . (2) This density is given in terms of the integral that cannot be evaluated explicitly but can be approximated numerically. ACKNOWLEDGMENT The authors would like to thank Professor Miroslaw Pawlak [21] M. Gjoka, M. Kurant, C. T. Butts and A. Markopoulou, "Walking in for his input and guidance for the mathematical analysis. REFERENCES [1] [2] S. Moldovan, J. Goldenberg and A. Chattopadhyay, "The different roles of product originality and usefulness in generating word-of- mouth," Intern. J. of Research in Marketing, vol. 28, pp. 109-119, 2011. J. A. Chevalier and D. Mayzlin, "The effect of word of mouth on sales: Online book," Journal of Marketing Research, vol. 43, no. 3, pp. 345 - 354, 2006. [3] W. W. Moe and M. Trusov, "The value of social dynamics in online product ratings forums," Journal of Marketing Research, vol. 48, no. 3, pp. 444-456, 2011. [4] D. Godes and D. Mayzlin, "Using online conversations to study word- of-mouth," Marketing Science, vol. 23, no. 4, pp. 545-560, 2004. [5] B. Libai, E. Muller and R. Peres, "Decomposing the value of Word-of- Mouth seeding programs: acceleration versus expansion," Journal of Marketing Research, vol. 50, no. 2, pp. 161-176, 2013. [6] M. Trusov, R. E. Bucklin and K. Pauwels, "Effects of word-of-mouth versus traditional marketing: Findings from an internet social networking site," Journal of Marketing, vol. 73, no. 5, pp. 90-102, 2009. [7] S. Tirunillai and G. J. Tellis, "Does chatter really matter? Dynamics of user-generated content and stock performance," Marketing Science, vol. 31, no. 2, pp. 198-215, 2012. Facebook: a case study of unbiased sampling of OSNs," in INFOCOM'10 Proceedings of the 29th conference on Information communications, 2010. [22] M. Kagan, "12 Essential Facebook Stats [Data]," 31 May 2011. [Online]. Available: http://blog.hubspot.com/blog/tabid/6307/bid/14715/12-Essential- Facebook-Stats-Data.aspx. [Accessed 30 March 2013]. [23] K. Burbary, "Facebook demographics revisited – 2011 statistics," 7 March 2011. [Online]. Available: http://socialmediatoday.com/kenburbary/276356/facebook- demographics-revisited-2011-statistics. [Accessed 31 March 2013]. [24] S. Radwanick, "comScore Introduces Mobile Metrix 2.0, Revealing that Social Media Brands Experience Heavy Engagement on Smartphones," 7 May 2012. [Online]. Available: http://www.comscore.com/Insights/Press_Releases/2012/5/Introducing _Mobile_Metrix_2_Insight_into_Mobile_Behavior. [Accessed 31 March 2013]. [25] B. London, "Virtual vanity: Nation's most obsessed Facebook users spend a staggering EIGHT hours a day on the site," 10 September 2012. [Online]. Available: http://www.dailymail.co.uk/femail/article- 2200962/Nations-obsessed-Facebook-users-spend-staggering-hours- day-site.html. [Accessed 31 March 2013]. [26] F. Keating, "Got a smartphone? You probably check Facebook FOURTEEN times a day," 2013. [Online]. Available: http://www.dailymail.co.uk/sciencetech/article-2300466/Smartphone- users-check-Facebook-14-times-day-admit-looking-movies.html. [Accessed 18 9 2014]. [8] A. Patino, D. A. Pitta and R. Quinones, "Social media's emerging importance in market research," Journal of Consumer Marketing, vol. 29, no. 3, pp. 233-237, 2012. [27] C. Pring, "100 social media statistics for 2012," 11 January 2012. [Online]. Available: http://thesocialskinny.com/100-social-media- statistics-for-2012/. [Accessed 31 March 2013]. [9] S. Burton and A. Soboleva, "Interactive or reactive? Marketing with Twitter," Journal of Consumer Marketing, vol. 28, no. 7, pp. 491-499, 2011. [28] D. A. Ebersman, "Amendment No. 3 to REGISTRATION STATEMENT on Form S-1," Facebook Inc., Menlo Park, 2012. [29] D. J. Watts and S. H. Strogatz, "Collective dynamics of 'small-world' [10] C. R. Hollenbeck and A. M. Kaikati, "Consumers' use of brands to networks," Nature, vol. 393, no. 6684, pp. 440-442, 1998. reflect their actual and ideal selves on Facebook," International Journal of Research in Marketing, vol. 29, no. 4, pp. 395-405, 2012. [30] S. Schnettler, "A structured overview of 50 years of small-world research," Social Networks, vol. 31, no. 3, pp. 165-178, 2009. 19 [31] R. Cohen and S. Havlin, "Scale-free networks are ultrasmall," Physics Review Letter, vol. 90, no. 5, 2003. [32] J. Ugander, B. Karrer, L. Backstrom and C. Marlow, "The anatomy of [50] J. Goldenberg, B. Libai and E. Muller, "The chilling effect of network externalities," International Journal of Research in Marketing, vol. 27, no. 1, pp. 4-15, 2010. the Facebook social graph," Arxiv preprint arXiv:11114503, 2011. [51] M. E. J. Newman, "The spread of epidemic disease on networks," [33] T. Britton, M. Deijfen and A. Martin-Lof, "Generating simple random graphs with prescribed degree distribution," Journal of Statistical Physics, vol. 124, no. 6, September 2006. [34] L. Backstrom, P. Boldi, M. Rosa, J. Ugander and S. Vigna, "Four degrees of separation," arXiv:1111.4570, 2012. [35] S. Gunelius, "25 Basic Facebook Terms You Need to Know," 29 Dec. 2011. [Online]. Available: http://sproutsocial.com/insights/2011/12/facebook-terms-definitions/. [Accessed 26 03 2013]. [36] H. Sharara, W. Rand and L. Getoor, "Differential adaptive diffusion: Phys. Rev. E, vol. 66, pp. 016128-1 – 016128-11, 2002. [52] H. Rahmandad and J. Sterman, "Heterogeneity and network structure in the dynamics of diffusion: comparing Agent-based and Differential Equation models," Management Science, vol. 54, no. 5, pp. 998-1014, 2008. [53] T. Tassier, "The economics of epidemiology," in SpringerBriefs in Public Health, 1, Ed., Springer, 2013. [54] S. Aral and D. Walker, "Creating social contagion through viral product design: a randomized trial of peer influence in networks," Management Science, vol. 57, no. 9, pp. 1623 - 1639, 2011. understanding diversity and learning whom to trust in viral marketing," in 5th International AAAI Conference on Weblogs and Social Media, 2011. [55] S. Aral and D. Walker, "Identifying influential and susceptible members of social networks," Science, vol. 337, no. 6092, pp. 337-341, 2012. [37] M. S. Bernstein, E. Bakshy, M. Burke and B. Karrer, "Quantifying the invisible audience in social networks," 2013. [38] F. M. Bass, "A new product growth for model consumer durables," [56] J.-P. Onnela and F. Reed-Tsochas, "Spontaneous emergence of social influence in online systems," Proceedings of the National Academy of Sciences, vol. 107, no. 43, pp. 18375 - 18380, 2010. Management Science, vol. 15, no. 5, pp. 215-227, January 1969. [57] R. Pastor-satorras and A. Vespignani, "Epidemic Spreading in Scale- [39] W. O. Kermack and A. G. McKendrick, "A Contribution to the mathematical theory of epidemics," Proceedings of the Royal Society of London. Series A, vol. 115, no. 772, pp. 700-721, 1927. [40] G. V. Steeg, R. Ghosh and K. Lerman, "What stops social epidemics?," in 5th International AAAI Conference on Weblogs and Social Media, Barcelona, 2011. [41] T. Zhang and D. Zhang, "Agent-based simulation of consumer purchase decision-making and the decoy effect," Journal of Business Research, vol. 60, p. 912–922, 2007. [42] W. M. Rand and R. T. Rust, "Agent-based modeling in marketing: guidelines for rigor," International Journal of Research in Marketing, 2011. [43] MarketingCharts, "Facebook Posts Get Half Their Reach Within 30 Free Networks," Phys. Rev. Lett., vol. 86, pp. 3200-3203, 2001. [58] O. Hinz, B. Skiera, C. Barrot and J. U. Becker, "Seeding strategies for viral marketing: an empirical comparison," Journal of Marketing, vol. 75, no. 6, pp. 55-71, 2011. [59] D. J. Watts and P. S. Dodds, "Influentials, networks, and public opinion formation," Journal of Consumer Research, vol. 34, no. 4, pp. 441-458, 2007. [60] Y. Dover, J. Goldenberg and D. Shapira, "Network traces on penetration: uncovering degree distribution from adoption data," Marketing Science, vol. 31, no. 4, p. 689–712, 2012. [61] L. Weng, A. Flammini, A. Vespignani and F. Menczer, "Competition among memes in a world with limited attention," Scientific Reports, vol. 2, no. 335, 2012. Minutes of Being Published," 2012. [Online]. Available: http://www.marketingcharts.com/online/facebook-posts-get-half-their- reach-within-30-minutes-of-being-published-24453/. [Accessed 16 Oct 2014]. [62] V. Raghavan, G. Steeg, A. Galstyan and A. Tartakovsky, "Modeling temporal activity patterns in dynamic social networks," Computational Social Systems, IEEE Transactions on, vol. 1, no. 1, pp. 89-107, March 2014. [44] K. Lerman and R. Ghosh, "Information Contagion: an empirical study [63] MarketingCharts, "Which Content Categories Are Being Shared on Which Social Networks?," 9 July 2014. [Online]. Available: http://www.marketingcharts.com/online/which-content-categories-are- being-shared-on-which-social-networks-43953/. [Accessed 10 12 2014]. of the spread of news on Digg and Twitter social networks," in Proceedings of 4th International Conference on Weblogs and Social Media (ICWSM-10), 2010. [45] A. T. Stephen and O. Toubia, "Deriving value from social commerce networks," Journal of Marketing Research, vol. 47, no. 2, pp. 215-228, 2010. [46] Z. Katona, P. P. Zubcsek and M. Sarvary, "Network effects and personal influences: the diffusion of an online social network," Journal of Marketing Research, vol. 48, no. 3, pp. 425-443, 2011. [47] D. Godes and J. C. Silva, "Sequential and temporal dynamics of online opinion," Marketing Science, vol. 31, no. 3, pp. 448 - 473, 2011. [48] D. Mayzlin and H. Yoganarasimhan, "Link to success: How blogs build an audience by promoting rivals," Management Science, vol. 58, no. 9, pp. 1651 - 1668, 2012. [49] J. Goldenberg, S. Han, D. R. Lehmann and J. W. Hong, "The role of hubs in the adoption process," Journal of Marketing, vol. 73, no. 2, pp. 1-13, 2009.
1304.4837	1	1304	2013-04-17T14:26:59	Friends, Strangers, and the Value of Ego Networks for Recommendation	[ "cs.SI", "physics.soc-ph" ]	Two main approaches to using social network information in recommendation have emerged: augmenting collaborative filtering with social data and algorithms that use only ego-centric data. We compare the two approaches using movie and music data from Facebook, and hashtag data from Twitter. We find that recommendation algorithms based only on friends perform no worse than those based on the full network, even though they require much less data and computational resources. Further, our evidence suggests that locality of preference, or the non-random distribution of item preferences in a social network, is a driving force behind the value of incorporating social network information into recommender algorithms. When locality is high, as in Twitter data, simple k-nn recommenders do better based only on friends than they do if they draw from the entire network. These results help us understand when, and why, social network information is likely to support recommendation systems, and show that systems that see ego-centric slices of a complete network (such as websites that use Facebook logins) or have computational limitations (such as mobile devices) may profitably use ego-centric recommendation algorithms.	cs.SI	cs	Friends, Strangers, and the Value of Ego Networks for Recommendation Amit Sharma Dept. of Computer Science Cornell University Ithaca, NY 14853 USA [email protected] Mevlana Gemici Dept. of Computer Science Cornell University Ithaca, NY 14853 USA [email protected] Dan Cosley Information Science Cornell University Ithaca, NY 14853 USA [email protected] 3 1 0 2 r p A 7 1 ] I S . s c [ 1 v 7 3 8 4 . 4 0 3 1 : v i X r a Abstract Two main approaches to using social network infor- mation in recommendation have emerged: augmenting collaborative filtering with social data and algorithms that use only ego-centric data. We compare the two ap- proaches using movie and music data from Facebook, and hashtag data from Twitter. We find that recom- mendation algorithms based only on friends perform no worse than those based on the full network, even though they require much less data and computational resources. Further, our evidence suggests that locality of preference, or the non-random distribution of item pref- erences in a social network, is a driving force behind the value of incorporating social network information into recommender algorithms. When locality is high, as in Twitter data, simple k-nn recommenders do better based only on friends than they do if they draw from the entire network. These results help us understand when, and why, social network information is likely to support rec- ommendation systems, and show that systems that see ego-centric slices of a complete network (such as web- sites that use Facebook logins) or have computational limitations (such as mobile devices) may profitably use ego-centric recommendation algorithms. An increasingly common approach in recommender sys- tems research has been to use network ties in recommen- dation. One main strategy for this is to augment collabo- rative filtering algorithms with social data (Ma et al. 2011; Konstas, Stathopoulos, and Jose 2009). However, in many cases, computation on the full dataset may be undesirable or impossible. For instance, most online social networks reveal only a user's first degree connections through their APIs, meaning that many third-party recommendation providers only have access to the immediate friends of a user. Thus, another main strategy for using social informa- tion has been to focus on making recommendations using only ego-centric slices of the network (Chen et al. 2010; Guy et al. 2009). Ego-centric algorithms have clear advan- tages in terms of the data and computational resources re- quired; however, not much is known about the relative per- formance of the two approaches. In this paper, we empirically compare these two ap- proaches to social recommendation using datasets from Copyright c(cid:13) 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Facebook and Twitter. We find that algorithms using only friends' data are comparable in performance to those using the full network -- and for Twitter data, they perform better. We hypothesize that this surprising result may be because of the prevalence of locality in these networks. By this we mean that the preferences of people close to each other in the network are more similar than one would expect if they were independently distributed, an intuition called out by Ma et al. in early social recommendation work (Ma et al. 2008). To explore this, we develop a set of metrics that char- acterize the extent of locality in a network. For our datasets, we find that Twitter data exhibits the highest locality with all metrics; that, combined with the fact that ego-centric rec- ommendations are most effective on Twitter, suggests that preference locality is an important factor in the success of such recommendations. Overall, our results suggest ego-centric approaches may be good enough for recommendation in social contexts, to the extent that locality exists. Further, the extent of local- ity varies between datasets, which suggests that the pay- off of using social data may also vary with domain. For practitioners, our proposed locality metrics can be helpful to characterize the expected benefits of social recommen- dation (at least in terms of recommender performance). For researchers, our results point to the need to better understand the distribution of preferences over individuals and networks in developing effective recommendation algorithms. Data sources. We use three datasets: two from Facebook (one for movies, one for music) and one from Twitter (hash- tag use). The Facebook datasets were collected as a part of a user study involving university students (Sharma and Cosley 2013), while the Twitter data was collected using their pub- lic API (McAuley and Leskovec 2012). The datasets have an ego-centric structure: each core user is guaranteed to have all his first-degree connections. The preference data is a set of user-item pairs, where items are movies or musical artists Liked in Facebook or hashtags used in Twitter. We use the term friends to refer to first- degree connections or followees of a user, ego networks to refer to a core user and her friends, non-friends to refer to ev- eryone not in a given ego network, items to refer to movies, artists, or hashtags, and likes to refer to Likes on Facebook or usage of a hashtag. Total users Total core users Friends/user (µ; σ) Total items Total likes Likes/user (µ; σ) Likes/item (µ; σ) Artists 63230 153 (499; 341) 139986 1289340 (20; 34) (9; 117) Movies Hashtags 51365 69414 935 149 (126; 61) (352; 156) 214941 78244 1230169 873261 (17; 33) (17; 17) (6; 29) (11; 96) Table 1: Overview of datasets. Both the Facebook artists and movies datasets have a higher friend average and average number of likes per item compared to Twitter. Figure 1: Distribution of item likes for the three domains, showing the cumulative percentage of total likes for items in the kth percentile of popularity. Compared to hashtags, likes for artists and movies are skewed toward popular items. Table 1 shows the statistics for the three datasets. The two Facebook networks have both a higher average number of friends than Twitter and a higher average likes per item. The distribution of likes for items (Figure 1) shows that likes for artists and movies are concentrated toward the most popu- lar items: the 10% most popular artists and movies receive around 5/6 of the likes, versus about half for hashtags. Artists and movies have fairly similar profiles with a long tail and an uneven distribution of popularity. We suspect that this happens because of exogenous effects such as media ex- posure for the most popular artists and movies that generates large amounts of attention for a relatively small number of items. Hashtag use in Twitter, on the other hand, is much more evenly distributed, perhaps because hashtags receive less exposure outside of Twitter itself and so must spread largely through the Twitter network. Goodness of ego network recommendations We now turn to our main question, around how ego network- based recommendations compare to recommendations that use the full dataset. We use a simple k-nn recommender as a reasonable baseline algorithm that also allows us to directly discuss similarity between users1. algorithms k-nn similarity. Recommender typically choose the most similar neighbors, so we first look at how a user's k most similar friends compare to the most similar k non-friends in the network. We use the Jaccard similarity coefficient, a common measure for unary ratings in recommender systems (Candillier, Meyer, and Fessant 2008). For two users u1 and u2, it is given by: Likes(u1) ∩ Likes(u2) Likes(u1) ∪ Likes(u2) JS = Figure 2 shows the results. As expected, average similar- ity decreases as k increases. For artists and movies, the top-k friends are less similar than the top-k non-friends. Hashtags offer a different scenario: a core user's top 10 friends are more similar than the top 10 others. Recommendation quality. We next evaluate how useful friends are compared to the whole network for recommen- dation. We divide each core user's preference data into 70:30 train-test splits (using 30% test items means even users with few likes have at least one in the test set) and make top- 10 recommendation lists using items liked by the k nearest neighbors, weighted by Jaccard similarity. For evaluation, we use normalized discounted cumulative gain (NDCG), a common metric used to compare ranked results: Rel1 +(cid:80) 1 +(cid:80) NDCG = 2..N Reli/ log2 i 2..N 1/ log2 i where N = {min(10,T estSet)} and Reli = 1 if the ith ranked item is in the test set and 0 otherwise. To focus on the effect of friends, we first compare ego- centric recommendations based on only friends to recom- mendations that use only non-friends from the full network. Figure 3 shows the results averaged across 10 random 70:30 splits. Compared to non-friends, friends are better for artists, worse for movies, and much better for hashtags. For both artists and movies, performance differences between friends and non-friends are much lower than might be expected based on the large differences in k-nn similarities from Fig- ure 2. In practice, algorithms would include friends as part of the full network as well. Using the full network (includ- ing friends) improves performance (Table 2), but ego-centric recommendations are still comparable, and for hashtags, the ego network-based recommender performs slightly better than the one using the full network. Locality of preferences We now look at properties of the datasets that might explain these results, focusing on comparing ego networks with the full dataset in terms of similarity, sparsity, and item cover- age. In all three cases, we find evidence that preferences are in fact unevenly distributed, as proposed by Ma et al. (2008). 1We tried a number of algorithms; the patterns were similar to k-nn, so, like Fermat, we omit the details due to lack of space. Figure 2: Average Jaccard similarity between a user, her top-k friends, and her top-k non-friends. For the Facebook datasets, non-friends are more similar at all values of k; friends appear to be more informative for hashtags. Friends Non-Friends Full Network 0.17 0.07 0.24 0.17 0.11 0.22 0.15 0.11 0.13 Artists Movies Hashtags Table 2: NDCG values for 50-nn recommenders using the ego network, non-friends, and the full network. The full net- work recommender shows improved performance, but is still comparable to the algorithm based only on ego network data. Similarity. We first look at how similar a person is to his friends versus randomly selected people in the full dataset. We measure this directly by comparing the average Jaccard similarity between a core user and his friends versus that user and an equal number of randomly chosen non-friends. Table 3 shows the results averaged over 10 random sets of non-friends. In all three datasets, friends are in fact more similar. The effect is strongest with hashtags, consistent with our expectation of relatively strong endogenous effects in the adoption of hashtags compared to artists or movies. These results suggest that for any set of people chosen at random from a user's friends or the full network, friends are expected to be more similar. Similarity Sparsity Artists Movies Hashtags Friends Non-Friends 0.040 0.020 0.043 0.023 0.013 0.002 Network Ego 0.73% 0.01% 0.87% 0.02% 3.60% 0.01% Table 3: Matrix-based locality metrics. "Similarity" is the average Jaccard similarity between a user and her k friends or k randomly selected non-friends. "Sparsity" refers to rat- ings density in the user-item matrix in ego networks versus the network as a whole. Figure 3: NDCG for k-nn recommender using only friends versus only non-friends. Friends are better for artists and hashtags, worse for movies. The black marker within each bar represents a recommender choosing people randomly. Error bars represent standard deviation of the mean. Sparsity. Ratings sparsity is common in preference datasets, including ours, which have an average user-item matrix density of 0.02% or less. However, preference local- ity suggests that this sparsity should be unevenly distributed. Table 3 shows that this is in fact the case: ratings for the items in a given ego network are two orders of magnitude denser than in the full dataset. Again, the effect is stronger for hashtags than for movies or artists, reflecting their higher locality. Ego Coverage Metrics. Another way to think about pref- erence locality is to look at coverage, or what percentage of items can be recommended to a user. Intuitively, as an item's preferences are more localized, it will be liked in fewer ego networks, reducing coverage. There are a number of ways we might formalize this notion. The simplest approach is to look at the percentage of ego networks in which at least one person likes a given item. Av- eraging this over all items gives us a measure of ego network coverage: what percent of possible ego network-item pairs exist in the dataset? We then subtract from 100 so that higher numbers correspond to increased locality, and call this the Uncovered Ego of the network. Uncovered Ego, however, does not account for item popu- larity. For instance, items with one like will look maximally local -- true but uninformative -- while an item with many likes should appear in many networks, and thus appear less local even if its preferences are more concentrated than we expect. Thus, we might want to account for the expected number of ego networks an item should appear in. To do this, we create a network with identical friend connections and number of likes per node, but with the items randomly distributed (subject to the constraint that each node can only like a given item once). Dividing the number of ego net- works that contain a given item in the randomized network by the number of ego networks that contain it the real net- Uncov. Ego Item/Ego Friend/Ego Artists Movies Hashtags 97.4% 96.3% 99.7% 1.19 1.20 1.78 1.16 1.28 1.79 Table 4: Ego network coverage by dataset. Compared to ran- dom, hashtags exhibit the highest locality for all three met- rics. Between artists and movies, the metrics are divided. work gives a measure of how much preferences deviate from what we would expect if they were distributed without ref- erence to the friend network. We call this metric Random Item/Ego, with higher numbers indicating greater locality. In some ways, that approach is too random because it doesn't account for patterns of individual preferences -- that, as Amazon reminds us, people who bought X also tend to buy Y. One way to account for this is to randomize at the network level rather than the item level: keeping the same number of friends for each core user and the same itemsets, but randomly reassigning the friend links. We call this met- ric Random Friend/Ego. Table 4 shows these metrics averaged across all items in each dataset. All datasets exhibit more locality than random. As with the sparsity and similarity measures, hashtags are more local than movies or artists. The effect of randomizing by friend versus by item varies by network, indicating that the amount of item-item correlation in the networks is also different. These results suggest that the dynamics of prefer- ences differ across networks in ways that merit further study. Discussion We show that on balance, recommendation algorithms that use only ego network information do fairly well compared to algorithms that use an entire dataset, although coverage for ego network-only recommendations may be a concern. This result adds evidence to earlier work that argued that recom- mendations based on only ego networks may be effective in some contexts (Sharma and Cosley 2011). For example, when data or computation is local to a mobile device (such as peer to peer recommenders, or mobile recommenders that keep data local to support privacy, as in PocketLens (Miller, Konstan, and Riedl 2004)), it may be infeasible or undesir- able to compute on a large dataset. Likewise, many websites such as Flixster, TripAdvisor, and CNET use existing social networks such as Facebook to support their user accounts. These sites essentially see individual ego networks drawn from the underlying full Facebook network; our results show that those views may be valuable for recommendation. We also empirically demonstrate the existence of prefer- ence locality using a variety of metrics. These results align nicely with the stream of work around understanding diffu- sion in networks, although we focus on preferences rather than memes (Leskovec, Backstrom, and Kleinberg 2009). Further, we show that locality differs across preference types and networks. This suggests the need to examine disparate datasets to support generalizable claims about diffusion pro- cesses and their effects on preferences. Putting our recommendation and locality results together implicates preference locality as an important reason why using social network information can improve recommenda- tions (Chen et al. 2010; Ma et al. 2011), as ego-centric rec- ommender performance roughly correlates with preference locality. Locality isn't the only reason why social informa- tion matters -- some metrics show higher locality for artists versus movies, and vice versa, despite the higher perfor- mance of local k-nn for artists over movies -- but it does mat- ter. It seems likely that social network information becomes more valuable for recommendation as locality increases. Exploring locality and its drivers is a promising direc- tion for future research. Our proposed metrics are imper- fect reflections of locality, capturing some elements that vary across datasets but saying little about why. Further, in- dividual metrics provide different relative orderings of the datasets, suggesting that each captures only part of the story. Properties of the network such as the frequency of item preferences, distribution of friendship links, and visibility of preference information (e.g., through explicit sharing in Twitter or seeing Likes in a Facebook newsfeed) all likely af- fect the diffusion and locality of preferences. Better models of these properties, along with the endogenous and exoge- nous processes by which people adopt preferences, would provide both theoretical understanding and practical value for recommender systems. Acknowledgements This work was supported by NSF grant 0910664. References [Candillier, Meyer, and Fessant 2008] Candillier, L.; Meyer, F.; and Fessant, F. 2008. Designing specific weighted similarity measures to improve collaborative filtering systems. In Proc. ICDM. [Chen et al. 2010] Chen, J.; Nairn, R.; Nelson, L.; Bernstein, M.; and Chi, E. 2010. Short and tweet: Experiments on recommending content from information streams. In Proc. CHI. [Guy et al. 2009] Guy, I.; Zwerdling, N.; Carmel, D.; Ronen, I.; Uziel, E.; Yogev, S.; and Ofek-Koifman, S. 2009. Personalized recommendation of social software items based on social relations. In Proc. RecSys. [Konstas, Stathopoulos, and Jose 2009] Konstas, I.; Stathopoulos, V.; and Jose, J. M. 2009. On social networks and collaborative recommendation. In Proc. SIGIR. [Leskovec, Backstrom, and Kleinberg 2009] Leskovec, J.; Back- strom, L.; and Kleinberg, J. 2009. Meme-tracking and the dy- namics of the news cycle. In Proc. KDD. [Ma et al. 2008] Ma, H.; Yang, H.; Lyu, M. R.; and King, I. 2008. SoRec: Social recommendation using probabilistic matrix factor- ization. In Proc. CIKM. [Ma et al. 2011] Ma, H.; Zhou, D.; Liu, C.; Lyu, M. R.; and King, I. 2011. Recommender systems with social regularization. In Proc. WSDM. [McAuley and Leskovec 2012] McAuley, J. J., and Leskovec, J. 2012. Learning to discover social circles in ego networks. In NIPS. [Miller, Konstan, and Riedl 2004] Miller, B. N.; Konstan, J. A.; and Riedl, J. 2004. Pocketlens: Toward a personal recommender sys- tem. ACM TOIS 22(3). [Sharma and Cosley 2011] Sharma, A., and Cosley, D. 2011. Network-centric recommendation: Personalization with and in so- cial networks. In Proc. IEEE SocialCom. [Sharma and Cosley 2013] Sharma, A., and Cosley, D. 2013. Do social explanations work? Studying and modeling the effects of so- cial explanations in recommender systems. In Proc. WWW.
1905.03038	2	1905	2019-05-09T19:43:25	Maximin share allocations on cycles	[ "cs.SI", "cs.DM", "math.CO" ]	The problem of fair division of indivisible goods is a fundamental problem of social choice. Recently, the problem was extended to the case when goods form a graph and the goal is to allocate goods to agents so that each agent's bundle forms a connected subgraph. For the maximin share fairness criterion researchers proved that if goods form a tree, allocations offering each agent a bundle of at least her maximin share value always exist. Moreover, they can be found in polynomial time. We consider here the problem of maximin share allocations of goods on a cycle. Despite the simplicity of the graph, the problem turns out to be significantly harder than its tree version. We present cases when maximin share allocations of goods on cycles exist and provide results on allocations guaranteeing each agent a certain portion of her maximin share. We also study algorithms for computing maximin share allocations of goods on cycles.	cs.SI	cs	MAXIMIN SHARE ALLOCATIONS ON CYCLES ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI ABSTRACT. The problem of fair division of indivisible goods is a fundamental problem of social choice. Recently, the problem was extended to the case when goods form a graph and the goal is to allocate goods to agents so that each agent's bundle forms a connected subgraph. For the maximin share fairness criterion researchers proved that if goods form a tree, allocations offering each agent a bundle of at least her maximin share value always exist. Moreover, they can be found in polynomial time. We consider here the problem of maximin share allocations of goods on a cycle. Despite the simplicity of the graph, the problem turns out to be significantly harder than its tree version. We present cases when maximin share allocations of goods on cycles exist and provide results on allocations guaranteeing each agent a certain portion of her maximin share. We also study algorithms for computing maximin share allocations of goods on cycles. 1. INTRODUCTION Fair allocation of indivisible goods is a fundamental problem of social choice [5, 3]. It assumes a set of elements, referred to as goods, and a collection of agents each with her own utility function on the sets, or bundles, of goods. The utility functions are commonly assumed to be additive and so it is enough to specify their values on individual goods only. The objective is to assign to agents disjoint subsets of goods in a way that meets some fairness criteria. Among the most commonly studied ones are proportionality and envy-freeness, adapted to the case of indivisible goods from the problem of fair allocation of divisible goods or cake-cutting [5, 11], as well as recently proposed maximin and minimax share [6, 4]. For each of the criteria, it is of interest to identify classes of instances when fair allocations exist, to establish the complexity of deciding the existence of fair allocations, and to design algorithms for computing them. In this paper, we focus on the maximin share criterion [6]. It is a relaxation of envy-freeness and proportionality, and has a natural interpretation in terms of some allocation protocols. In a few years since it was first proposed it has already received substantial attention. The criterion turns out to be highly non-trivial. First, while it is easy to see that envy-free and proportional allocations do not always exist, it is not at all clear whether the same is true for the less restrictive maximin criterion. We now know it is. But it took a few years before the first examples showing that maximin share allocations are not guaranteed to exist were found [12]. Moreover, they turned to be quite intricate. Further, the complexity of deciding the existence of maximin share allocations has not yet been determined [4]. To get around the difficulty of constructing maximin allocations, researchers proposed to relax the maximin share criterion by requiring that the value of each agent's share in an allocation be at least equal to some positive fraction of the maximin share. Procaccia and Wang [12] proved that an allocation guaranteeing agents at least 2/3 of their maximin share always exists, and that it can be found in polynomial time if the number of agents is fixed (not part of input). Both the result and the algorithm are based on deep combinatorial insights. Building on that work, Amanatidis, Markakis, Nikzad and Saberi [1] proved that for every constant ǫ, 0 < ǫ < 2/3, there is a polynomial-time algorithm finding an allocation guaranteeing to each agent at least (2/3−ǫ) of A PRELIMINARY VERSION OF THE PAPER APPEARED IN THE PROCEEDINGS OF IJCAI 2018. 1 2 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI their maximin share (with the number of agents a part of the input). Ghodsi, Hajiaghayi, Seddighin, Seddighin and Yami [8] proved that in the results by Procaccia and Wang and by Amanatidis et al., the constant 2/3 can be replaced with 3/4. We study maximin share allocations of indivisible goods in the setting proposed by Bouveret, Cechl´arov´a, Elkind, Igarashi and Peters [2]. In the original problem, there are no restrictions on sets of goods that can be allocated to agents. This ignores important practical constraints that may make some sets highly undesirable. For instance, goods may be rooms and labs in a building to be allocated to research groups [2], or plots of land to be consolidated [9]. In such cases, legal sets of goods that could be allocated to an agent might be required to form connected subgraphs in some graph describing the neighborhood relation among goods (offices spanning segments of a hall, plots forming contiguous areas of land). Bouveret et al. [2] studied envy-free, proportional and maximin share allocations for that setting obtaining several interesting complexity and algorithmic results. Our paper extends their study for the maximin share criterion. In a striking positive result, Bouveret et al. [2] proved that maximin share allocations of goods on trees always exist and can be found in polynomial time. In our work we look beyond trees and show that as soon as the underlying graph has a single cycle, the picture becomes much more complicated. Our main contributions are as follows. 1. We show that for goods on a cycle the maximin share value for an agent can be computed in polynomial time. In two cases, when m ≤ 2n and when agents can be grouped into a fixed number of types (agents are of the same type if they have the same utility function), this allows us to design polynomial time algorithms for computing maximin share allocations of m goods (on cycles) to n agents or determining that such allocations do not exist. 2. We show that deciding the existence of maximin share allocations of goods on an arbitrary graph is in the class ∆P 2 . For complete graphs (the setting equivalent to the original one) this result improves the bound given by Bouveret and Lemaıtre [4]. We further improve on this upper bound for cycles and more generally, unicyclic graphs by showing that for such graphs the existence of a maximin share allocation is in NP. 3. We obtain approximation results on the existence of allocations of goods on a cycle that guar- antee all agents a specified fraction of their maximin share value. While it is easy to show that for any number of agents there are allocations guaranteeing each agent at least 1/2 their maximin share, improving on the guarantee coefficient of 1/2 is a non-trivial problem. We show that it can be improved to (√5 − 1)/2 (> 0.618). Further improvements are possible if we limit the number of agents or the number of types of agents. In particular, we show that for the three-agent case, there are allocations guaranteeing each agent 5/6 of their maximin share; for an arbitrary number of agents of up to three types there are allocations guaranteeing each agent 3/4 of their maximin share; and for any number of agents of t ≥ 4 types there are allocations guaranteeing each agent t/(2t− 2) of their maximin share. In each case, these allocations can be found in polynomial time. Moreover, the constants 5/6 and 3/4 for the cases of three agents and any number of agents of up to three types, respectively, are best possible, that is, they cannot be replaced with any larger ones. 2. DEFINITIONS AND BASIC OBSERVATIONS A utility function on a set V of goods (items) is a function assigning non-negative reals (utilities) to goods in V . We extend utility functions to subsets of V by assuming additivity. Following Bouveret et al. [2], we consider the case when goods form nodes of a certain con- nected graph G = (V, E). We adopt the assumption of the connectedness of the graph for the entire paper and do not mention it explicitly again. In particular, whenever we use the term graph MAXIMIN SHARE ALLOCATIONS ON CYCLES 3 we mean a connected graph. We write V (G) and E(G) for the sets of nodes and edges of G and refer to G as the graph of goods. Given a graph G of goods, we will often refer to utility functions on (the set) V (G) as utility functions on (the graph) G. Let V = V (G) be a set of goods. A G-bundle is a subset of V that induces in G a connected subgraph. A (G, n)-split is a sequence P1, . . . , Pn of pairwise disjoint G-bundles such that n Pi = V. [i=1 We use the term split rather than partition as we allow bundles to be empty. When G or n are clear from the context we drop them from the notation and speak about bundles and splits (occasionally, G-splits and n-splits). Let G be a graph of goods, u a utility function on G, q a non-negative real, and n a positive integer. We call a split q-strong if every bundle in the split has value at least q under u. The maximin share for G, u and n, written mms (n)(G, u), is defined by setting mms (n)(G, u) = max{q : there is a q-strong n-split of G}. An equivalent definition is given by mms (n)(G, u) = max P1,...,Pn min i=1,...,n u(Pi), where the maximum is taken over all n-splits P1, . . . , Pn of G. A split for which the maximum is attained is a maximin share split or an mms-split for n and u or for n and an agent with a utility function u. We often leave n and G implicit when they are clearly determined by the context. By the definition, for every bundle P in an mms-split for n and u, u(P ) ≥ mms (n)(G, u). Sim- ilarly as above, when G and n are clear from the context, we leave them out and write mms(u) for mms (n)(G, u). When considering an agent i with a utility function ui we write mms(i) for mms(ui). Let G be a graph of goods. A (G, n)-allocation is an assignment of pairwise disjoint G-bundles of goods to n agents 1, 2, . . . , n so that all goods are assigned. Clearly, (G, n)-allocations can be represented by (G, n)-splits, where we understand that the ith bundle in the split is the bundle assigned to agent i. Let ui be the utility function of agent i, 1 ≤ i ≤ n. A (G, n)-allocation P1, . . . , Pn is a maximin share allocation, or an mms-allocation, if for every i = 1, . . . , n we have ui(Pi) ≥ mms (n)(G, ui). To illustrate the concepts introduced above, let us consider the graph in Figure 1. We will call this graph G. The graph G defines the set of goods {v1, . . . , v8} and their adjacency structure. The table in the figure shows three utility functions u1, u2 and u3 on the set of goods. The values of the utility function u1 are also shown by the corresponding goods in the graph (it will facilitate our discussion below). In our example, the set {v1, v2, v6, v7} of goods is a G-bundle (or, simply a bundle, as G is clear). Indeed, the subgraph of G induced by {v1, v2, v6, v7} is connected. On the other hand, the set {v1, v2, v5, v6} of goods is not a bundle as the corresponding induced graph is not connected. Fur- ther, the sequence {v1, v7, v8},{v2, v3},{v4, v5, v6} is a (G, 3)-split (or simply a 3-split). Indeed, all sets in the split are bundles. It is also easy to see that the sequence {v1, v7, v8},{v2, v4},{v3, v5, v6} is not a split as the set {v2, v4} is not a bundle. Let us now consider the utility function u1. The total utility of all goods under u1 is 18. Further, it is easy to see that 3-splits into bundles of value 6 (under u1) do not exist. To this end, we note that v1 and v8 cannot be in the same bundle valued at 6 as their total value is 7. Thus, the bundle containing 4 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI v1 v8 3 4 v2 v7 1 1 v3 v6 1 1 v4 v5 4 3 v1 v2 v3 v4 v5 v6 v7 v8 4 3 3 2 1 3 1 3 3 4 3 3 3 1 0 1 1 2 1 2 3 1 0 2 u1 u2 u3 FIGURE 1. A graph of goods and three utility functions. v8 must be either {v6, v7, v8} or {v2, v7, v8}. In each case the value of the bundle containing v1 is not 6. Thus, mms (3)(G, u1) < 6. On the other hand, the sequence {v1, v2, v3},{v4, v5},{v6, v7, v8} is a 3-split with its bundles having values 5, 7 and 6, respectively. Thus, mms (3)(G, u1) = 5. Reasoning in a similar way one can show that maximin share values for the two other utility functions are also 5, that is, mms (3)(G, u2) = mms (3)(G, u3) = 5. Let us now observe that the sequence {v1, v7, v8},{v4, v5, v6},{v2, v3} is a 3-split (indeed, its components are bundles). Moreover, u1({v1, v7, v8}) = 8, u2({v4, v5, v6}) = 7 and u3({v2, v3}) = 5. Thus, this 3-split defines an mms-allocation of goods on G to three agents with the utility functions u1, u2 and u3. Later, we will also see examples when mms-allocations of goods on graphs (specifically, cycles) do not exist. We now are ready to present a few basic results on the problem of the existence of mms- allocations. Two agents are of the same type if they have the same utility functions. The maximin share allocation problem is easy when all but one agent are of the same type. Proposition 2.1. Let G = (V, E) be a graph of goods. If we have n agents and at most one of them is of a different type than the others, then an mms-allocation exists. Proof. The result is obvious for n = 1. Thus, we will assume n ≥ 2. Let u be the utility function of agents 1, . . . , n − 1 and u′ a utility function of n. Let Π be an mms-split for u, and P a bundle of Π most valuable under u′. Then, u′(P ) ≥ Pv∈V u′(v) ≥ mms(u′). Assign P to agent n. Next, allocate the remaining parts of Π to agents 1, . . . , n − 1 in an arbitrary way. Since for each bundle Q in Π, u(Q) ≥ mms(u), the resulting allocation is an mms-allocation. (cid:3) n In particular, this result applies to the case of two agents. Corollary 2.2. Mms-allocations of goods on a graph to two agents always exist. ✷ Let G = (V, E) be a graph of goods. An agent with a utility function u on G (a utility function u on G) is n-proportional if mms(n)(G, u) = Pv∈V u(v) n . We often omit n from the notation if it is clear from the context. A set {1, . . . , n} of agents (a set {u1, . . . , un} of utility functions) is proportional if every agent i (utility function ui) is n- proportional. MAXIMIN SHARE ALLOCATIONS ON CYCLES 5 An agent (a utility function) is n-regular if it is n-proportional and its maximin share (with respect to n) is 1. A collection of n agents (utility functions) is regular if every agent in the collection is n-regular. It is clear that for every agent in an n-element regular collection of agents, the total value of all goods for that agent is n. Let c be a positive real. A bundle P ⊆ V is c-sufficient for an agent i if ui(P ) ≥ c · mms(i). An allocation Π = P1, . . . , Pn is c-sufficient if for every i = 1, . . . , n, Pi is c-sufficient for i. Clearly, a 1-sufficient allocation is an mms-allocation. The next result shows that when studying the existence of c-sufficient allocations (and so, in particular, mms-allocations) one can restrict considerations to the case when all agents are regular. We use it to prove Theorem 2.4 later on in this section but its full power becomes clear in Section 4. Proposition 2.3. Let G be a graph of goods and c a positive real such that for every regular collection of n agents (for every regular collection of n agents of at most t-types) a c-sufficient allocation exists. Then a c-sufficient allocation exists for every collection of arbitrary n agents (for every collection of arbitrary n agents of at most t types). Proof. Let N = {1, . . . , n} be the set of agents and {u1, . . . , un} the set of their utilities. If for every i ∈ N , mms (n)(ui) = 0, then the assertion holds. Any split of G into n bundles is a c- sufficient allocation. Thus, let us assume that for at least one agent, say k, mms (n)(uk) > 0. For each agent i ∈ N such that mms (n)(ui) = 0, we change her utility function to uk. We denote the new set of utilities by v′′ 1 , . . . , v′′ n. It is clear that if P1, . . . , Pn is a c-sufficient allocation with respect to the utility functions v′′ i , then it is also a c-sufficient allocation with respect to the original utilities ui. It is then enough to prove that a c-sufficient allocation with respect to the functions v′′ i exists. i (x), mms (n)(v′ i. Since for every item x, v′ i . For each bundle of Πi valued more than mms (n)(v′′ For every agent i ∈ N , we denote by Πi any mms-split into n bundles with respect to the utility i ) we decrease the value of some i ). In this way, we produce a new i ). On the i ). Thus, each utility function v′′ elements in that bundle to bring its value down to mms (n)(v′′ utility function, say v′ i(x) ≤ v′′ other hand, by the construction, for every bundle P ∈ Πi, v′ i) = mms (n)(v′′ function v′ Next, for each agent i ∈ N we set ki = n/(Py∈V (G) v′ Clearly, each utility function vi, i ∈ N , is n-regular. Moreover, it follows directly from the con- struction that if the original utility functions are of at most t types, then the resulting utility functions are of at most t types, too. i(P ) = mms (n)(v′′ i ) > 0. i(y)) and, for every x ∈ V (G), we define i is n-proportional and mms (n)(v′ i) ≤ mms (n)(v′′ vi(x) = kiv′ i(x), Let Π = P1, . . . , Pn be a c-sufficient allocation for v1, . . . , vn (its existence is guaranteed by the assumption). Therefore, for every agent i ∈ N , vi(Pi) ≥ c · mms (n)(vi). Since mms (n)(vi) = kimms (n)(v′ i) and vi(Pi) = kiv′ i). Next, we recall that for every x ∈ V (G), v′′ i (P ) ≥ v′ i (x) ≥ v′ i(P ). Since for every i ∈ N , mms (n)(v′ i) = mms (n)(v′′ v′′ i (Pi) ≥ v′ i(Pi), it follows that v′ i(Pi) ≥ c · mms (n)(v′ i(x). Consequently, for every bundle P , v′′ i ), it follows that i(Pi) ≥ c · mms (n)(v′ i) = c · mms (n)(v′′ i ). Thus, Π is a c-sufficient allocation of goods to agents in N with respect to the utility functions v′′ i and the result follows. (cid:3) Corollary 2.2 states that for two agents and for any connected graph of goods an mms-allocation exists. On the other hand, Bouveret et al. [2] gave an example of nonexistence of an mms-allocation 6 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI for a cycle and four agents. In fact, as we show in Figure 2, even for three agents it may be that mms-allocations of goods on a cycle do not exist. In that figure, v1, . . . , v9 denote consecutive nodes of a cycle and the numbers in each row represent the utility functions. Observe that the maximin shares for the agents 1 and 2 are 5. For the agent 3, it is 6. Moreover, no two consecutive nodes have the total value satisfying any of the agents. Therefore, if an mms-allocation existed, it would have to be a split into three bundles of three consecutive nodes each. It is simple to check that none of the three possible partitions of this type is an mms-allocation. agent 1 agent 2 agent 3 v1 0 2 1 v2 3 2 3 v3 1 0 2 v4 3 3 3 v5 1 1 0 v6 3 3 3 v7 0 1 2 v8 2 3 3 v9 2 0 1 FIGURE 2. An example of nonexistence of an mms-allocation for a cycle with 9 nodes and 3 agents. On the other hand, for three agents and at most 8 goods on a cycle mms-allocations always exist. Theorem 2.4. For three agents and at most 8 goods on a cycle, mms-allocations always exist. Proof. Let C be a cycle with at most 8 nodes and let N = {1, 2, 3} be the set of agents. By Proposition 2.3, we assume without loss of generality that the agents are 3-regular. In particular, they are proportional and mms(i) = 1, for every i ∈ N . Let us consider mms-splits for agents 1, 2 and 3, each into three bundles. By regularity, all bundles in those splits are non-empty. Thus, each split determines three edges that connect adjacent bundles of the split. Since C has 8 edges and there are three splits, there is an edge in C that connects adjacent bundles in two different splits, say for agents 1 and 2. It follows that for some bundles A and B of the mms-splits for agents 1 and 2, respectively, A ⊆ B or B ⊆ A. Let us assume without loss of generality that A ⊆ B. Clearly, the elements of the set B − A can be added to one of the two other parts of the mms-split for the agent 2 to form a 2-split of the path C − A, say B′, B′′, with both parts of value at least 1 for the agent 2. We now construct an mms-allocation as follows. If the set A has value at least 1 for the agent 3, then she receives it. Clearly, the value of C − A for the agent 1 is equal to 2. Thus, one of the parts B′, B′′ has value at least 1 to agent 1. We allocate this part to agent 1. Finally, the agent 2 receives the part that remains. If A does not satisfy the agent 3, then the agent 1 gets this part. The value of C − A for the agent 3, is larger than 2. Thus, we allocate the parts B′, B′′ to the agents 3 and 2 in the same way as to the agents 1 and 2 in the previous case. (cid:3) We conclude this section with a comment on the notation. Formally, whenever we talk about allocations of goods on a subgraph of a graph we consider, we should use restrictions of the utility functions ui to the set of nodes of the subgraph. Continuing to refer to ui's is simpler and does not introduce ambiguity. Therefore, we adopt this convention in the paper. 3. COMPLEXITY AND ALGORITHMS We assume familiarity with basic concepts of computational complexity and, in particular, with the classes NP, ∆P 2 . We refer to the book by Papadimitriou [10] for details. We start with general comments on how instances of fair division of indivisible goods on graphs are given. First, we assume any standard representation of graphs. As our primary objective is to understand when 2 and ΣP MAXIMIN SHARE ALLOCATIONS ON CYCLES 7 problems we consider have polynomial solutions and not most efficient algorithms to solve them, the details of how graphs are represented are not critical. Moreover, many of our results concern particular graphs such as cycles, when explicit representations of nodes and edges are not even needed. Finally, we assume that all utilities and other parameters that may be part of the input are rational, the least restrictive assumption for studies of algorithms and complexity. We now formally define several problems related to maximin share allocations of indivisible goods on graphs and derive bounds on their complexity. MMS-VALUES-G: Given a graph G on m goods, a utility function u, that is, a sequence u = (u1, . . . , um) of non-negative rational numbers, an integer n > 1 and a rational number k ≥ 0, decide whether mms (n)(G, u) ≥ k. The problem was studied in the original case (with G being a complete graph) by Bouveret and Lemaıtre [4]. Their complexity result and its proof extend to the general case. The hardness argument uses a reduction from the well-known NP-complete problem PARTITION [7]: PARTITION: Given a set U of m non-negative integers, decide whether there is a set X ⊆ U such thatPu∈X u =Pu∈U \X u. Proposition 3.1. The problem MMS-VALUES-G is NP-complete. Proof. The problem is clearly in NP. Indeed, to solve the problem, we guess an n-split and verify that each of its parts induces in G a connected subgraph that has value at least k (under u). The problem is NP-hard even in the case when G is a complete graph, n = 2 and the utility function is integer valued. We can show this by a reduction from PARTITION. Indeed, an instance of PARTITION consisting of a set of non-negative integers {u1, . . . , um} yields an instance to MMS- VALUES-G with the complete graph of goods, n = 2, and k = ⌈(Pm It is easy to observe that an instance of PARTITION is a YES instance if and only if the corresponding instance of MMS-VALUES-G is a YES instance. Thus, NP-hardness follows. (cid:3) i=1 ui)/2⌉. Another related problem concerns the existence of an allocation meeting given bounds on the values of its individual bundles. ALLOC-G: Given a graph G on m goods, n utility functions u1, . . . , un (each ui is a sequence of length m consisting of non-negative rational numbers) and n rational numbers q1, . . . , qn, decide whether an allocation A1, . . . , An for G exists such that for every i = 1, . . . , n, ui(Ai) ≥ qi. Proposition 3.2. The problem ALLOC-G is NP-complete. Proof. The membership in NP is evident. The problem is NP-hard even if G is a complete graph, n = 2, q1 = q2, and u1 = u2. Indeed, it becomes then the MMS-VALUES-G problem with G being a complete graph, n = 2, k = q1 (= q2), and u = u1 (= u2). Thus, the proof of hardness for Proposition 3.1 can be applied here, too. (cid:3) We use these observations to show an upper bound on the complexity of the problem of the existence of an mms-allocation in the graph setting. We formally define the problem as follows. MMS-ALLOC-G: Given a graph G of goods and n utility functions u1, . . . , un (each ui is a se- quence of length m of non-negative rational numbers), decide whether an mms-allocation for G and u1, . . . , un exists. Theorem 3.3. The problem MMS-ALLOC-G is in the class ∆P 2 . Proof. The key step in the proof is the observation that we can rescale each utility function so that all its values are integers. To this end, given a utility function ui, 1 ≤ i ≤ n, we compute the 8 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI product, say ki, of all denominators of the rational numbers used by ui as its values. Next, we multiply all values of ui by ki to produce a new utility function u′ i. By the construction, all values of u′ i are indeed integers. Let us denote by I the original input instance to the problem and by I ′ the instance obtained by rescaling the utility functions in the way described above. We note that the products ki and the rescaled utility functions u′ i can be computed in time O(nm2M 2), where M is the number of digits in the binary representation of the largest integer appearing as the numerator or the denominator in rational numbers specifying the original utility functions ui. Since the size, say S, of the input instance satisfies S = Ω(m + n + M), I ′ can be computed in polynomial time in S. Finally, we note that I is a YES instance of the problem MMS-ALLOC-G if and only if I ′ is a YES instance of the problem. In fact, an allocation Π is an mms-allocation for I if and only if Π is an mms-allocation for I ′. Let us now consider an auxiliary problem where, given a graph G on m nodes and a sequence s = s1, . . . , sm of m non-negative integers (that is, a utility function), the goal is to compute mms (n)(G, s). We will design for this problem an algorithm with an NP-oracle. We will then show that our algorithm runs in polynomial time in the size of the representation of s, where we consider each call to an oracle to be a single step taking a constant amount of time. To this end, we observe that mms (n)(G, s) ≤ (Pm j=1 sj)/n. Thus, we can compute mms (n)(G, s) by binary search on the range [0..⌊(Pm j=1 sj)/n⌋] of possible values for mms(n)(G, s). Each step starts with a range, say [p..r], that contains mms (n)(G, s). We narrow down this range by making a call to an oracle for the problem MMS-VALUES-G, which we proved to be NP-complete above. The call to the oracle is made on the instance consisting of G, s and k = ⌈(p + r)/2⌉. Depending on the oracle output, it results in a new smaller range -- either [p..k − 1] or [k..r]. The process stops when the range is reduced to just one element and that element is returned as mms (n)(G, s). The number of range-narrowing steps is given by log2(⌊(Pm sj + 1) ≤ j=1 sj)/n⌋ + 1). Clearly, Xj=1 sj)/n⌋ + 1) ≤ log2( log2(sj + 1). log2(⌊( Xj=1 m Xj=1 m m Since the number of bits needed to represent a non-negative integer x is given by max(1,⌈log2(x + 1)⌉, it follows that the size, say S′, of the representation of a problem instance satisfies S′ ≥ Pm j=1 max(1,⌈log2(sj + 1)⌉. Thus, the number of range-narrowing steps is bounded by S′. Con- sequently, the oracle algorithm we described runs indeed in polynomial time in S′. It follows that the problem MMS-ALLOC-G can be solved for an instance I by a procedure consisting of these three steps: (1) Compute the instance I ′. (2) Compute the values qi = mms(n)(G, u′ n are the rescaled utility functions computed in step (1). This is done using the oracle algorithm described above. It can be applied as all utilities in I ′ are integer-valued. i), where u′ 1, . . . , u′ (3) Decide whether there is an allocation A1, . . . , Am for G such that u′ i = 1, . . . , m, by invoking once an NP-oracle for the problem ALLOC-G. i(Ai) ≥ qi, for every We recall that I ′ can be computed in polynomial time in S (the size of the original instance I). Based on that and on our discussion of the auxiliary problem, step (2) of our algorithm also runs in polynomial time in S, counting each oracle call as taking a unit amount of time. Finally, the last step takes a single call to an oracle. Thus, the entire algorithm runs in polynomial time in S, where MAXIMIN SHARE ALLOCATIONS ON CYCLES 9 we count each oracle call as taking the unit amount of time. Since the oracles used by the algorithm are for NP-complete problems, the problem MMS-ALLOC-G is in ∆P 2 . (cid:3) The upper bound ∆P 2 established by Theorem 3.3 applies also in the case when G is assumed implicitly, for instance, when it is a path, a cycle or a complete graph represented by its set of nodes (but not edges). It is because the number of oracle calls is bounded by the size of the representation of the utility functions only. In the case of complete graphs, Theorem 3.3 yields an improvement on the bound ΣP 2 obtained by Bouveret and Lemaıtre [4]. We do not know if the upper bound of ∆P 2 can be improved in general and, in particular, whether it can be improved for complete graphs. On the other hand, we do know that it can be improved for trees. Bouveret et al. [2] proved the following two results. Theorem 3.4 (Bouveret et al. [2]). There is a polynomial time algorithm that computes mms (n)(T, u) and a corresponding mms-split given a tree of goods T , a non-negative rational utility function u on the nodes (goods) of T , and an ingeter n ≥ 1. Theorem 3.5 (Bouveret et al. [2]). For every tree T of goods and every set of n agents with non- negative rational utility functions on the goods in T , an mms-allocation exists. Moreover, there is a polynomial-time algorithm to find it. The results we presented suggest a question of the relationship between the complexity of the MMS-ALLOC-G problem and the properties of the underlying graph, as it becomes more complex than trees. The first step towards understanding how the complexity grows is to analyze the case of cycles and unicyclic graphs. First, we show that as in the case of trees (cf. [2]), the maximin share values and mms-splits for the case when goods form a cycle (or a unicyclic graph) can be computed in polynomial time. Theorem 3.6. There is a polynomial time algorithm for computing mms (n)(U, u) and a corre- sponding mms-split, where U is a unicyclic graph and u is a rational-valued utility function. Proof. Using the rescaling technique discussed in the proof of Theorem 3.3, we can reduce in polynomial time the general problem to the problem when all utilities are integers. To this end, we compute the product, say k, of all denominators of the rational numbers that are values of goods under u (if all values are integers, we assume the denominator is 1; this applies, in particular, to the case when all utilities are 0). We then define a new utility function by multiplying the values of u by k. Once the maximin share value for the rescaled utility function is computed, it is then used as the numerator of the maximin share for u and k is used as the denominator. Thus, to complete the proof it suffices to describe a method of computing mms (n)(U, u) and an mms-split under the assumption that all values of u are integers. Let C be the unique cycle of U . Every U -split has a bundle that contains C or is a split of the graph U − e for some edge e ∈ C. Thus, mms (n)(U, u) = max(max e∈C mms (n)(U − e, u), mmsC), where mms C stands for max P1,...,Pn min i=1,...,n u(Pi), with the maximum taken over all splits P1, . . . , Pn that have a bundle containing C. To compute mms C, we proceed as follows. We construct a tree UC by contracting C in U to a single "supern- ode," say c. Thus, the nodes of UC are all nodes of U − C and the supernode c. We define a utility 10 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI function u′ on the nodes of UC by setting u′(x) =(u(x) Py∈C u(y) if x is a node of U − C (not the supernode) if x = c (is a super node). Clearly, mms C = mms (n)(UC, u′). We note that UC and u′ can be computed in polynomial time. Moreover, UC and all graphs U − e, e ∈ C are trees. We now apply the algorithm by Bouveret et al. [2] (cf. Theorem 3.4) to compute mms C and mms (n)(U − e, u), for all edges e ∈ C, as well as the corresponding mms-splits. This takes polynomial time. We then select the largest value among them and return it together with its mms-split (in the case, the largest value is mms C, in the bundle containing the superode, we replace it with the nodes of C). (cid:3) It is now a matter of routine to show that the MMS-ALLOC-G problem for cycles and, more generally, for unicyclic graphs is in NP. Corollary 3.7. The problem MMS-ALLOC-G for unicyclic graphs is in NP. Proof. Let U be a unicyclic graph and u1, . . . , un rational-valued utility functions defined on the nodes of U . The following non-deterministic polynomial-time algorithm decides the problem MMS-ALLOC-G: (1) guess an allocation P1, . . . , Pn, (2) compute the values mms (n)(U, ui), i = 1, . . . , n, and (3) verify that ui(Pi) ≥ mms (n)(U, ui), i = 1, . . . , n. By Theorem 3.6, the step (2) can be accomplished in polynomial time. Thus, the entire algorithm runs in polynomial time. (cid:3) We do not have an argument for NP-hardness of the problem MMS-ALLOC-G. Thus, we do not know whether the bound obtained in Corollary 3.7 is tight in general. We do know, however, that in some special cases it can be improved. We will now discuss these results. Our first group of results is concerned with the case when the number m of goods is small with respect to the number n of agents, specifically, when m ≤ 2n. Our results rely on the following simple observation. Proposition 3.8. Let C be a cycle, u a utility function on C, and n ≥ 2 an integer. For every node x of C, mms(n−1)(C − x, u) ≥ mms(n)(C, u). Proof. Let Π be an mms-split of C for u and n. Let P be the bundle in Π containing x, and let P ′ and P ′′ be the two bundles in Π inducing in C segments neighboring the one induced by P , respectively preceding it and succeeding it when traversing C clockwise (P ′ = P ′′ if n = 2). We move all goods in P that precede x to P ′ and those that succeed x to P ′′. In this way each bundle still spans a connected segment in C. Next, we remove the "new" P (at this point, P consists of x only). The result is a split Π′ of C − x into n − 1 bundles, in which every bundle has value at least mms (n)(C, u). Thus, mms (n−1)(C − x, u) ≥ mms (n)(C, u). (cid:3) We now are ready to show that, when there are m goods on a cycle and n agents, and m < 2n, then mms-allocations are guaranteed to exist. Moreover, we also show that this result is sharp -- having exactly m = 2n goods allows for situations when mms-allocations do not exist. Theorem 3.9. If m < 2n, then an mms-allocation of m goods on a cycle C to n agents exists and it can be computed in polynomial time. Proof. Let us consider agents 1, . . . , n with utility functions u1, . . . , un. Let Π be an mms-split for the agent n. Clearly, for every bundle P ∈ Π, un(P ) ≥ mms (n)(C, un). Since m < 2n, at least one bundle in Π consists of only one element, say x. It follows that un(x) ≥ mms (n)(C, un). Since C − x is a path (and so, a tree), Theorem 3.5 implies that there is an allocation giving each agent j, MAXIMIN SHARE ALLOCATIONS ON CYCLES 11 1 ≤ j ≤ n− 1, a bundle valued at least mms (n−1)(C − x, uj). By Proposition 3.8, each such bundle is valued at least mms (n)(C, uj). Thus, this allocation extended by the bundle {x}, allocated to the agent n, forms an mms-allocation of the goods on C to n agents. For an algorithm, we (1) compute the maximin share mms (n)(C, un); (2) select an item x such that un(x) ≥ mms (n)(C, un) (as argued above, such an x exists); (3) construct an mms-allocation of goods on the path C − x to agents 1, . . . , n − 1 (possible as C − x is a tree); and (4) extend the allocation constructed in (3) by giving {x} to n. By our argument above, the allocation constructed in step (4) is an mms-allocation. Steps (1) and (3) can be accomplished in polynomial time by Theorems 3.6 and 3.5, respectively. The same obviously holds for steps (2) and (4). Thus, the algorithm we described runs in polynomial time. (cid:3) This result is sharp. If m = 2n and n > 3 then an mms-allocation of m goods on a cycle to n agents may not exist as shown by the example in Figure 3. v1 v2 v3 v5 . . . v2n−1 v2n v4 2 v6 3 agents 1, . . . , n − 2 n agents n − 1, n FIGURE 3. An example showing that the result in Theorem 3.9 is sharp. n − 2 . . . n − 1 1 n − 1 n − 1 n − 2 2 n n . . . 1 n 1 1 Indeed, all maximin share values are n + 1. Thus, every bundle in any mms-allocation would have to contain at least two elements. Since m = 2n, every bundle would have to consist of exactly two elements. There are only two splits of a cycle with m = 2n nodes into n bundles of two consecutive goods. None of them is an mms-allocation. The assumption that n > 3 is essential. We have seen earlier that if n = 2 then mms-allocations exist for every number m of goods. Thus, they exist for m = 2n = 4. If n = 3, mms-allocations of m goods on a cycle exist for every m ≤ 8 (cf. Theorem 2.4). In particular, they exist if m = 2n = 6. As our example shows, if m = 2n, there are cases when mms-allocations do not exist. However, whether an mms-allocation exists in the case when m = 2n can be decided in polynomial time and, if so, an mms-allocation can be computed efficiently. Corollary 3.10. There is a polynomial time algorithm deciding the existence of an mms-allocation of 2n goods on a cycle to n agents, and computing one, if one exists. Proof. Let C be the cycle and u1, . . . , un the rational-valued utility functions of the agents 1, . . . , n. We first compute the values mms (n)(C, ui), i = 1, . . . , n (cf. Theorem 3.6). If for some item x and agent i, ui(x) ≥ mms (n)(C, ui), then reasoning as in the proof of Theorem 3.9 one can show that an mms-allocation for C exists, and that it can be found in polynomial time. Otherwise, if there is an mms-allocation of goods on C to n agents, every bundle in this allocation consists of two consecutive goods. There are only two candidates for such allocations and one can check whether any of them is an mms-allocation in polynomial time (as the values mms (n)(C, ui) are known). (cid:3) The next result of this section concerns the case of n agents of t types, where t is fixed and is not a part of the input. Before we present our result and prove it, we discuss how inputs and outputs are represented. As we already noted earlier, instances to an mms-allocation problem on a cycle are specified by a non-negative integer m representing the number of goods, and n m-element sequences of non- negative rational numbers, each sequence representing a utility function. We note that m implicitly 12 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI defines the goods as v1, v2, . . . , vm, as well as the cycle as having these goods appear on it in this order. Further, we adopt a natural convention that for every i = 1, . . . , m, the ith element in every utility sequence provides the utility value for the good i. This representation consists of nm + 1 rational numbers (one of them, m, an integer).1 In the case when agents can be grouped into t types, where t is fixed and not part of input, the case we are now considering, input instances can be represented more concisely by a non-negative integer m, t sequences of m non-negative rational numbers (t utility functions) and t non-negative integers s1, . . . , st, where each sr represents the number of agents of type r (that is, having ur as their utility function). Thus, an instance consists of tm = Θ(m) rational numbers represented by pairs of integers, and t + 1 = Θ(1) integers. We will use M to denote the length of the binary representation of the largest of the integers appearing in this representation, and S to denote the total size of the binary representations of the integers in the instance. In particular, we have that S = Ω(m + M). We observe that if n > m, then the maximin share for each agent is 0 and, consequently, every allocation of m goods to n agents is an mms-allocation. Since t is fixed, computing n = s1+. . .+st takes time O(M) and, consequently, also O(S). Thus, this case can be recognized in time O(S) and then handled directly. Namely, we return a sequence of t 0's as the maximin shares for agents of types 1, 2 . . . , t and, if an mms-allocation is required, we allocate all goods to agent 1 and empty bundles of goods to all other agents. We do not generate any explicit representation for this allocation. It is implicitly identified by the all-0's output of the maximin shares. This is to avoid having to output explicitly an n-element sequence of bundles, as n, the length of this sequence, may be exponential in the size of input. It follows that the key case is then the case n ≤ m. Theorem 3.11. For every integer t ≥ 1, there is a polynomial time algorithm deciding existence of an mms-allocation of m goods on a cycle to n agents of t types (and computing one, if one exists). Proof. Let us consider an instance to the problem given by an integer m (the number of goods), t sequences u1, . . . , ut, each consisting of m non-negative rational numbers (t utility functions), and t non-negative integers s1, . . . , st (the numbers of agents of type 1, . . . , t, respectively). We start by computing n = s1 + s2 + . . . + st, which can be accomplished in O(S) time, as t is fixed (we recall that we write S for the size of the input instance). The case n > m has been discussed above. Thus, in what follows we assume that n ≤ m. An important consequence of this assumption is that for every r = 1, . . . , t, sr ≤ m. To decide whether there is an mms-allocation and, if so, to compute it, we compute the values qr = mms (n)(C, ur), r = 1, . . . , t. Since t is fixed, Theorem 3.6 implies that all these values can be computed in time bounded by a polynomial in u, where u represents the largest size of the representation of an input utility function and, consequently, also by a polynomial in S. We now observe that an allocation Π of goods on C to agents of t types given by u1, . . . , ut is an mms-allocation on C if and only if for some edge e of C, Π is an allocation of goods on the path C − e to these agents such that for every agent of type r, the bundle P ∈ Π allocated to that agent satisfies ur(P ) ≥ qr. Thus, to prove the assertion, it is enough to show that the following problem has a polynomial- time solution: Given a path F of m goods v1, v2, . . . , vm, appearing in this order on F , t rational- valued utility functions u1, . . . , ut, t non-negative integers s1, . . . , st such that n = s1 + . . . + st ≤ m, and t non-negative integers q1, . . . , qt, find an allocation Π of goods on F to s1 + . . . + st agents such that for every agent of type r, the bundle P assigned to that agent satisfies ur(P ) ≥ qr. 1We note that we do not include the number of agents, n, in the input; if we need it, we can compute it by counting in the input the sequences representing the utility functions. MAXIMIN SHARE ALLOCATIONS ON CYCLES 13 To describe our method, we define H to be the set of all sequences (h1, . . . , ht) such that for every r = 1, . . . , t, 0 ≤ hr ≤ sr. Clearly, the number of sequences in H is bounded by (s1 + 1) · . . .· (st + 1) ≤ (m + 1)t = O(mt) = O(St). If H denotes a sequence (h1, . . . , ht) ∈ H and hr > 0, where 1 ≤ r ≤ t, then we write H − r for the sequence (h1, . . . , hr−1, hr − 1, hr+1, . . . , ht) which, we note, also belongs to H. For a sequence (h1, . . . , ht) ∈ H, we define T (h1, . . . , ht) to be the smallest j such that there is an allocation of goods v1, . . . , vj to h1 + . . . + ht agents, exactly hr of them of type r, 1 ≤ r ≤ t, so that each agent of type r obtains a bundle worth at least qr. If such an allocation does not exist, T (h1, . . . , ht) is set to ∞. Next, for every j = 1, . . . , m and every r = 1, . . . , t, we define k(j, r) to be the minimum k ≥ j such that ur({vj+1, . . . , vk}) ≥ qr. We set k(j, r) = ∞ if such a k does not exist. Clearly, all values k(j, r) can be computed in time O(tmM) = O(mM) = O(S2). Let (h1, . . . , ht) ∈ H. We will show that T (h1, . . . , ht) can be efficiently computed. Clearly, T (0, . . . , 0) = 0 (there are no agents to get any bundles and so, even the empty path suffices). Thus, let us consider H = (h1, . . . , ht) ∈ H and let us assume that h1 + . . . + ht > 0. Let us define I = {r : hr > 0, 1 ≤ r ≤ t}. It is easy to see that (1) T (H) = min{k(T (H − r ), r) : r ∈ I}. Assuming that all values k(T (H − r ), r), for r ∈ I, are known, T (H) can be computed in time O(t) = O(1). It follows that with the initial value of T (0, . . . , 0) = 0, considering sequences H ∈ H in any order consistent with non-decreasing sums of their elements and using the formula (1) to compute T (H), one can compute T (s1, . . . , sr) in polynomial time, in fact, in time O(Smax(2,t)) (we recall that all values k(j, r) can be computed in time O(S2); also, the number of entries in T is O(St) and each entry can be computed in O(1) time). If T (s1, . . . , st) = ∞, then there is no split Π of F such that for every agent of type r, its bundle P ∈ Π satisfies ur(P ) ≥ qr. Otherwise, T (s1, . . . , st) 6= ∞ and a split solving the problem exists. Moreover, it can be computed in the standard way for dynamic programming algorithms. To this end, each time the formula (1) is applied we have to record an index r ∈ I that minimizes the expression k(T (H − r ), r)). This information allows us to construct the split in polynomial time in S. As all algorithmic tasks we presented can be accomplished in time bounded by a polynomial in (cid:3) S, the assertion follows. We close this section with yet another corollary to Theorem 3.6. It concerns a possibility of regularizing utility functions on unicyclic graphs in polynomial time by converting a method used in the proof of Proposition 2.3 into an algorithm. Let us call a collection {u1, . . . , un} trivial if for every i, mms (n)(ui) = 0. We will call all other collections non-trivial. Corollary 3.12. There is a polynomial time algorithm that, given a unicyclic graph U of goods, and a non-trivial collection of rational-valued utility functions ui, 1 ≤ i ≤ n, on U , produces a rational-valued regular collection of utility functions u′ i on U such that for every c, if a split Π is a c-sufficient allocation with respect to u′ i's then it is a c-sufficient allocation with respect to ui's. If the original utility functions are of at most t types, then the resulting utility functions are of at most t types. Proof. The algorithm follows the method we used in the proof of Proposition 2.3. It consists of the following key steps. 14 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI (1) Rescale the utilities ui as described in the proof of Theorem 3.3 to produce an equivalent collection of integer-valued utilities wi (by equivalent we mean determining the same splits as mms-allocations). (2) Compute the maximin shares qi = mms (n)(U, wi), 1 ≤ i ≤ n. (3) Select one utility function wk such that qk > 0. For each agent i such that qi = 0, replace (4) For a good x and a utility function vi, use the binary search method, similar to that used in ) = qi; replace vi ; repeat for all goods as long as it is possible to decrease a value of some utility wi with wk; denote the new utility functions vi, 1 ≤ i ≤ n. the proof of Theorem 3.3, to find the smallest a such that mms (n)(U, vx→a with vx→a function on some good.2 i i (5) Rescale the computed utility functions so that the values of each of them sum up to n. Call the resulting functions u′ i. The correctness of the method, both for the main statement and under the restriction to utility functions of at most t types, follows from the argument used to prove Proposition 2.3. For the running time, we note that step (1) runs in polynomial time (we argued this in the proof of Theorem 3.3). Step (2) runs in polynomial time (Theorem 3.6). Step (3) runs in time O(nm). Next, we note that the argument we used in the proof of Theorem 3.3 to estimate the running time of the "range narrowing" binary search and the fact that the maximin shares can be computed in polynomial time (Theorem 3.6) together imply that step (4) runs in polynomial time. Step (5) consists of O(nm) additions and multiplications involving utility values produced in step (4) and the integer n. Thus, it also runs in polynomial time in the size of the original instance. (cid:3) 4. APPROXIMATE MAXIMIN SHARE ALLOCATIONS ON A CYCLE We start with a simple observation that 1 2 -sufficient allocations always exist and, moreover, can be found in polynomial time. Proposition 4.1. There is a polynomial time algorithm which, for any number of agents and any number of goods on a cycle, constructs a 1 2 -sufficient allocation. Proof. We remove any edge e from the cycle C and get a path, say P . The maximin share for each agent on P is at least half of the original maximin share for C. Indeed, let us consider an arbitrary agent i and her mms-split of C. By removing e, no more than one part of this split may break into two pieces. The value for the agent i of one of these pieces is at least 1 2mms(i). We adjoin the other piece, if it is present, to its neighboring part in the mms-split of C. In this way, we obtain a split of P . Clearly, in this split of P every piece is worth to i at least 1 2 mms(i), so the claim follows. 2 -sufficient allocation can be found by (1) removing any edge from the cycle, (2) apply- ing to the resulting path the algorithm by Bouveret et al. [2] that constructs in polynomial time an mms-allocation for trees (cf. Theorem 3.5). (cid:3) Thus, a 1 of agent types. To find a better guarantee than 1 2 turns out to be non-trivial. We will show that it can be improved to (√5 − 1)/2 ≈ 0.61803..., and that further improvements are possible if we restrict the number Let c be a positive real and let N = {1, . . . , n}, where n ≥ 2, be a set of agents with arbitrary utility functions u1, . . . , un on a set of goods on a path, say P . Figure 4 shows an algorithm allocate that assigns to some (possibly all) agents in N bundles they value at c or more. 2Here, v x→a i stands for the utility function obtained from vi by setting its value on x to a and keeping all other values as they are in vi. MAXIMIN SHARE ALLOCATIONS ON CYCLES 15 allocate(N, P, Q, c) % P is a path; we fix its direction so that prefixes of P are well defined % N = {1, . . . , n}, n ≥ 2, is a set of agents, each with a utility function on P % c is a positive real % Q is a prefix of P valued at least c by at least one agent in N S := {i ∈ N : there is an (n − 1)-split of P − Q that is c-strong for i}; R := N − S; j := 1; while j ≤ n and P has value at least c for some i ∈ S ∪ R do Qj := the shortest prefix of P worth at least c for some agent i ∈ R ∪ S if Qj is worth at least c to an agent i ∈ R then else assign Qj to i; R := R − {i} assign Qj to an agent i ∈ S such that Qj is worth at least c for i; S := S − {i}; 1 2 3 4 5 6 7 8 9 10 11 12 13 P := P − Qj; j := j + 1 FIGURE 4. Algorithm allocate We will use this algorithm in our theoretical considerations on the existence of c-sufficient al- locations. Under the restriction to rational-valued utilities and a rational c, the algorithm runs in polynomial time. We will use it to argue the existence of polynomial-time algorithms for construct- ing c-sufficient allocations. First, let us discuss the algorithm allocate informally. The algorithm starts by defining a set S to consist of all agents i in N for whom an (n − 1)-split of P − Q that is c-strong for i can be found. The algorithm then sets R = N − S. We note that N = S ∪ R, S ∩ R = ∅ and that it may be that one of the sets S and R is empty. Next, the algorithm sets j to 1 and proceeds to the loop (4 -- 13). Each time the body of the loop is executed, it assigns a bundle to an "unassigned" agent. Throughout the execution of the loop, P denotes the path consisting of unallocated goods, and R ∪ S contains all agents that are as yet unassigned. In each iteration j, the algorithm attempts to allocate a bundle to an "unassigned" agent. At the start of that iteration j − 1 agents have received bundles selected as prefixes of the paths being considered in earlier iterations. If j > n, then the loop terminates and all agents are assigned bundles. If j ≤ n, some agents are unassigned, R∪S contains all unassigned agents and P is a path on unallocated goods. If the value of P for each unassigned agent is less than c, no further assignments are possible and the loop terminates. Otherwise P has value at least c for at least one unassigned agent and an assignment can be made. The bundle for the assignment, denoted by Qj is chosen as a shortest prefix of P that has value at least c for some unassigned agent. Selecting Qj as a prefix ensures that the remaining goods form a path. Selecting for Qj a shortest prefix that has value c or more for some unassigned agent is essential for a key property of the algorithm, which we will discuss later. Once the bundle Qj is constructed, it is assigned. By construction, there are unassigned agents that value Qj at c or more. We select one such agent, say i. We first check if such an agent i can be found in R and if so, assign Qj to i. Only if no agent in R values Qj at c or 16 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI more, we select i from S (this selection is possible as at least one unassigned agent values Qj at c or more), and assign Qj to that i. The following proposition gives the key property of the algorithm allocate. Proposition 4.2. Let P be a path of goods and N = {1, . . . , n}, where n ≥ 2, a set of agents, each with a utility function on P , and c a positive real. Further, let Q be a prefix of P valued at least c by at least one agent in N , and S the set computed by the algorithm allocate(N, P, Q, c) in line (1). When the algorithm allocate(N, P, Q, c) terminates, S = ∅, that is, all agents included in S in line (1) are assigned bundles they value at c or more. Proof. We will denote by P0 the original path P and by S0 the set S as computed in line (1). Let us assume that when the algorithm allocate(N, P0, Q, c) terminates, S 6= ∅. Let i ∈ S. Since after line (1) the algorithm never includes elements in S, it follows that i ∈ S0. Let us consider the value, say jt of the variable j when the loop (4 -- 13) terminates. Clearly, the body of the loop was executed jt−1 times and jt−1 agents are assigned bundles Q1, . . . , Qjt−1. By the conditions on the input parameters, the body of the loop executes for j = 1 and defines a bundle Q1. This bundle is a prefix of Q (because of how we select prefixes Qj). Since i ∈ S0, P0 − Q has an (n − 1)-split that is c-strong for i. Since Q1 is a prefix of Q, P0 − Q1 has an (n − 1)-split that is c-strong for i. Let us denote this split by D2, . . . , Dn. Since i has not been assigned a bundle in iteration 1 (in any iteration, in fact) and D2 has value at least c for i, at the beginning of iteration 2 we have that i ∈ S ∪ R and the value of the path P for i is at least c. Thus, the body of the loop executes for the second time. It follows that 3 ≤ jt. Further, since i is not assigned a bundle, jt − 1 ≤ n − 1. Thus, 3 ≤ jt ≤ n. Let us assume that (Q1 ∪ . . .∪ Qjt−1)∩ Djt = ∅. It follows that in the iteration of the loop when j = jt, Djt ⊆ P . Thus, P has value at least c for i. Consequently, the body of the loop would execute for j = jt, a contradiction. Therefore, we have (Q1 ∪ . . . ∪ Qjt−1) ∩ Djt 6= ∅. Let k be the smallest integer such that 3 ≤ k ≤ jt and (Q1 ∪ . . . ∪ Qk−1) ∩ Dk 6= ∅. From our observation above it follows that k is well defined. Moreover, (Q1 ∪ . . . ∪ Qk−2) ∩ Dk−1 = ∅. Let P be the path of unallocated goods when j = k − 1, that is, P = P0 − (Q1 ∪ . . .∪ Qk−2). It follows that Dk−1 is a subpath of P . Let D be the shortest prefix of P containing Dk−1. It follows that D is a strictly shorter prefix of P than Qk−1 and D has value at least c to i. This is a contradiction with the algorithm selecting Qk−1 when j = k − 1. (cid:3) Extending the notation we used for splits, we call an allocation q-strong if it assigns to each agent a bundle worth at least q to that agent. We note that for regular sets of agents, q-strong and q-sufficient allocations coincide. Theorem 4.3. Let P be a path of goods, N = {1, . . . , n}, n ≥ 2, a regular set of agents, and c a real such that c ≤ 1. Further, let Q be a prefix of P valued at least c by at least one agent in N . Let R be the set computed by the algorithm allocate(N, P, Q, c) in line (2). If no single good in P has value at least c for any agent in N and R ≤ 1−c c n + 1, then the algorithm allocate(N, P, Q, c) finds a c-strong allocation of goods in P to agents in N . Proof. Let us assume that when the algorithm terminates, some agents are left without a bundle. Let us denote s = S, r = R, where S and R are computed in the lines (1) and (2) of the algorithm allocate(N, P, Q, c) and let k be the number of agents that have no bundle when the algorithm terminates. By Proposition 4.2, these agents are members of R. Moreover, by our assumption, k ≥ 1. Let ℓ be an unassigned agent. Clearly, in each iteration that assigns a bundle to an agent from S, the value of that bundle is smaller than c for the agent ℓ (it is so because agents in R have a MAXIMIN SHARE ALLOCATIONS ON CYCLES 17 preference over agents in S when bundles are assigned). Moreover, in each iteration j when an agent from R is assigned a bundle, we recall this bundle is referred to as Qj, the value of Qj for the agent ℓ is smaller than 2c. Indeed, otherwise the prefix of Qj formed by removing the last node of Qj would have value at least c for ℓ (since, by assumption, that node is worth less than c to ℓ). This contradicts the property that Qj is a shortest prefix of the path P in the iteration j. It follows that the total value for ℓ of all bundles constructed and allocated by the algorithm is less than cs + 2c(r − k). Consequently, the value for ℓ of all unallocated goods when the algorithm terminates, say vℓ, satisfies vℓ > n − cs − 2c(r − k) (since N is a regular collection of agents, for every agent in N , the total value of goods on P to that agent is n). On the other hand, by the stopping condition, vℓ < c. Thus, c > n − cs − 2c(r − k). Since s = n − r, it follows that r > 1−c Thus, when the algorithm terminates, all agents are assigned bundles and each bundle the algo- rithm allocates to an agent has value at least c for that agent. In other words, the allocation defined by the algorithm is c-strong. (cid:3) c n + 2k − 1 ≥ 1−c c n + 1, a contradiction. Theorem 4.4. Let C be a cycle of goods and N = {1, . . . , n}, where n ≥ 2, a set of agents, each with a utility function on C. Let Then, there is a c-sufficient allocation of goods on C to agents in N . c = max d=n,n+1,... min n d , n d (cid:7) + n − 2! (cid:6) n2 Proof. By Proposition 2.3, it suffices to show the result under the assumption that N is a regular collection of agents. In particular, c-sufficient allocations and c-strong allocations coincide. If some node x of C has value at least c to some agent j then a c-sufficient allocation exists. Indeed, we assign x to the agent j. By Proposition 3.8, the values of the maximin shares for the remaining n − 1 agents and the path C − x do not drop. Applying the algorithm of Bouveret at al. [2], we construct an mms-allocation of goods on the path C − x to those n − 1 agents. This allocation together with an assignment of x to the agent j is a c-sufficient allocation of goods on C to agents in N (j receives a bundle worth at least at c and all other agents receive bundles worth at least their maximin share). So, we assume that no single-element bundle is c-sufficient for any agent in N . For every agent i ∈ N we select any of her mms-splits of C. By regularity, all bundles in these splits are non-empty. Thus, each selected split can be obtained by removing n different edges, say ei 1, . . . , ei j, i, j = 1, . . . , n, into a sequence E = e0, e1, . . . , en2−1. To this end, we start in any place in C and inspect the edges of C moving clockwise. Each time we find an edge e used to obtain mms-splits for, say, k agents, we place k occurrences of e in the sequence. n, from C. We arrange all edges ei For an integer d = n, n + 1, . . ., we define f (d) = min n d , n d (cid:7) + n − 2! . (cid:6) n2 We then define p to be that integer d ≥ n, for which f (d) achieves its maximum (in case of ties we pick for p the smallest of those values d). Since f (n) > 1/2 and, for d ≥ n2, f (d) ≤ 1/n, it follows that n ≤ p < n2. It is also clear that f (p) = c (as defined in the statement of the theorem). Moreover, c ≤ ⌈n2/p⌉+n−2 ≤ 1. n 18 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI p m, h = j n2 p k and r = n2 − p · h. Clearly, we have h ≥ 1 and 0 ≤ r < p. Let us define h = l n2 We define a split of the cycle C into p parts (some of them possibly empty) by removing p edges ei, where i = 0, h, 2h, . . . , rh, rh + h, rh + 2h, . . . , rh + (p − r − 1)h. It is easy to check that between two consecutive removed edges there are h − 1 or h − 1 edges of the sequence E. Let Q be a part of this split with the largest value to agent n (any other agent could be chosen for n, too). This value is at least n p ≥ f (p) = c. Let P be the path obtained from C by removing an edge so that Q is a prefix of P . We will show that the call allocate(N, P, Q, c) produces a c-strong allocation for N . First, we note that Q contains at most h − 1 edges of the sequence E. This means that there are at most h − 1 agents such that Q intersects more than one part of their mms-split. For each of the remaining agents their mms-split gives rise to an (n − 1)-split of the path P − Q that is 1-strong for them. Since c ≤ 1, these splits are c-strong for the corresponding agents and, consequently, all these agents are in S -- the set defined in line (1) of the algorithm allocate(N, P, Q, c). It follows that S ≥ n − (h− 1). Let R be the set defined in line (2) of allocate(N, P, Q, c). Clearly, R = n − S ≤ h − 1. Hence, by the definition of p, R ≤ h − 1 =(cid:24)n2 p (cid:25) − 1 ≤ n f (p) − n + 1 = n c − n + 1 = 1 − c c n + 1. By Theorem 4.3, there is a c-strong allocation for P . This allocation is also a c-strong allocation for C. (cid:3) This result yields a corollary that displays a specific value c for which the existence of c-sufficient allocations of goods on cycles is guaranteed. Let √5 + 1 2 ϕ = and ψ = √5 − 1 2 1 ϕ = ≈ 0.61803.. . Corollary 4.5. For any number n of agents there is a ψ-sufficient allocation for a cycle. Proof. Clearly, the corollary holds for n = 1, so we assume that n ≥ 2. Since ⌊ϕn⌋ ≥ n, by Theorem 4.4, it suffices to show that where d = ⌊ϕn⌋. Clearly, n ⌊ϕn⌋ ≥ n ϕn = ψ. Thus, it remains to prove that min n d , n d (cid:7) + n − 2! ≥ ψ, (cid:6) n2 ψ ≤ n . ⌊nϕ⌋m + n − 2 l n2 To this end, we observe that n2 ≤ n2 + n − 1 = (ϕn − 1)(ψn + 1), so ⌊ϕn⌋(cid:25) + n − 2(cid:19) ≤ ψ ·(cid:18)(cid:24)(ϕn − 1)(ψn + 1) ≤ ψ · (ψn + 2 + n − 2) = ψ(ψ + 1)n = n. ψ ·(cid:18)(cid:24) n2 ϕn − 1 (cid:25) + n − 2(cid:19) (cid:3) MAXIMIN SHARE ALLOCATIONS ON CYCLES 19 We now turn our attention to the problem of c-sufficient allocations when some agents have the same utility function, that is, are of the same type. Specifically, we will consider allocations of goods on cycles to n agents of at most t types. Lemma 4.6. Let C be a cycle of goods, N a regular set of agents, t ≥ 2 an integer, and Ni, 1 ≤ i ≤ t, pairwise disjoint subsets of N such thatSt i=1 Ni = N , N1 ≥ N2 ≥ . . . ≥ Nt and, for every i = 1, . . . , t, all agents in Ni are of the same type. If N1, N2 6= ∅ and there are splits t Π1 and Π2 such that Πi is 2t−2 -strong for agents in Ni, for i = 1, 2, and there is a bundle in Π1 t 2t−2 to some agent in N , then there is a and a bundle in Π2 whose intersection has value at least t 2t−2 -strong allocation of goods on C to agents in N . Proof. Let us define n = N, ct = t 2t−2 , and ni = Ni, 1 ≤ i ≤ t. We will call agents in a set Ni, 1 ≤ i ≤ t, to be of type i. If there is a single good, say x, in C of value at least ct to some agent k ∈ N , then agent k receives x. It follows from Theorem 3.5 that there is an mms-allocation for the path C − x and the remaining n − 1 agents. We distribute the goods of C − x to these n − 1 agents according to this allocation. By Proposition 3.8, each agent i ∈ N \ {k} receives a bundle worth to her at least mms (n)(i). Since each agent is regular, for every i ∈ N we have mms (n)(i) = 1 ≥ ct. Thus, each agent in N is allocated a bundle worth at least ct to her. In other words, the allocation we constructed is ct-strong. Thus, let us assume that no single good in C has value at least ct for any agent in N . Let Q be the intersection of two bundles A and B, where A is a bundle from Π1 and B is a bundle from Π2, such that Q has value at least ct for some agent in N . It is clear that Q is a proper subpath of C. Let P be the path obtained from C by removing an edge so that Q is a prefix of P . We will consider the call allocate(N, P, Q, ct) and follow the notation introduced in the description of the algorithm. Since Q ⊆ A, B, all agents of types 1 and 2 are in S. It follows that R ≤ n − (n1 + n2). Since n1 ≥ n2 ≥ . . . ≥ nt, n1 + n2 ≥ 2nj, for j = 3, . . . , t. Thus, (t − 2)(n1 + n2) ≥ 2(n3 + . . . + nt) = 2(n − (n1 + n2)). Consequently, we have n1 + n2 ≥ 2 t n. We use this inequality to estimate R getting R ≤ n − 2 t n = t − 2 t n = 1 − ct ct n. By Theorem 4.3, there is a ct-strong allocation of goods on C to agents in N . (cid:3) We will use this lemma to obtain a general result about the existence of c-sufficient allocations for any number of agents of t ≥ 4 types. Afterwards, we will obtain results for the two specific cases of t = 2 and t = 3. Theorem 4.7. Let C be a cycle of goods, N a set of agents of at most t types, where t ≥ 4. Then, a t 2t−2 -sufficient allocation of goods on C to agents in N exists. Proof. By Proposition 2.3 we may assume that all agents in N are regular. Thus, we have that the maximin share for all agents is 1. Let n1, . . . , nt be the numbers of agents of types 1, . . . , t, respectively, and let n = n1+. . .+nt be the number of all agents in N . As before, we define ct = t 2t−2 and assume that n1 ≥ n2 ≥ . . . ≥ nt. If all agents are of the same type, Proposition 2.1 ensures the existence of an mms-allocation. Since 2t−2 -sufficient allocation. Thus, we assume that n2 > 0 (and ct < 1, this mms-allocation is also a so, obviously, n1 > 0, as well). t 20 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI Let A1, . . . , An and B1, . . . , Bn be mms-splits of the cycle for agents of type 1 and 2, respectively. In particular, since ct < 1, the two splits are ct-strong for agents of type 1 and 2, respectively. If for some i and j, where 1 ≤ i, j ≤ n, the path Ai ∩ Bj has value at least ct to some agent in N , then we are done by Lemma 4.6. Indeed, by the regularity of the set of agents, ct-strong and ct-sufficient allocations coincide. So, let us assume that no path Ai ∩ Bj has value ct or more to any agent. Then, in particular, Ai 6⊆ Bj and Bj 6⊆ Ai for all i and j, 1 ≤ i, j ≤ n. Thus, each of the sets A1, . . . , An intersects exactly two consecutive sets of the mms-split B1, . . . , Bn. We can assume without loss of generality that for every i, 1 ≤ i ≤ n, the set Ai intersects the sets Bi and Bi+1 (the arithmetic on indices is modulo n, adjusted to the range [1..n]). We claim that the mms-split A1, . . . , An for agents of type 1 is a ct-sufficient split for agents of type 2. To prove it denote by u2 the utility function for agents of type 2. Since for every i, 1 ≤ i ≤ t, Ai ⊆ Bi ∪ Bi+1, it follows that 2 = u2(Bi ∪ Bi+1) = u2(Ai−1 ∩ Bi) + u2(Ai) + u2(Ai+1 ∩ Bi+1). By our assumption the sets Ai−1 ∩ Bi and Ai+1 ∩ Bi+1 have value less than ct to any agent. Since for t ≥ 4 we have ct ≤ 2 3 , it follows that u2(Ai) > 2 − 2ct ≥ ct. We proved that A1, . . . , An is a ct-strong split for any agent of type 2. As we noted, it is also a ct-strong split for agents of type 1. Therefore, we can take A1, . . . , An for Π1 and Π2 in Lemma 4.6. Moreover, A1 is clearly the intersection of a bundle in Π1 with a bundle in Π2 simply because A1 is a bundle in each of these splits. Moreover, as A1, . . . , An is an mms-split for agents of type 1, A1 has value at least 1 and, consequently, at least ct for agents of type 1. Thus, the assumptions of Lemma 4.6 are satisfied and a ct-strong allocation exists. This completes the proof as ct-strong and ct-sufficient allocations coincide. (cid:3) Next, we consider the case of agents of two types and show that 3 4 -sufficient allocations of goods on a cycle to n agents of two types are always possible. Later on we will present a result showing that 3 4 -sufficient allocations of goods on a cycle also exist when agents are of three types. The proof of this more general result is highly technical, complex and long (enough so that we provide it in an appendix). In contrast, the proof in the case of agents of two types is simple yet still of substantial interest in its own right to be presented. Theorem 4.8. Let C be a cycle of goods and N a set of agents of no more than two types. Then, a 3 4 -sufficient allocation of goods on C to agents in N exists. Proof. By Proposition 2.3 we may assume that the set N of agents is regular. In particular, 3 4 - strong and 3 4 -sufficient allocations coincide. Let n be the number of agents and let n1 and n2 be the numbers of agents of type 1 and type 2, respectively. Clearly, n = n1 + n2. We will write u1 and u2 for the utility functions for the agents of type 1 and type 2, respectively. Without loss of generality, we will assume that n1 ≥ n2. If n2 = 0, then all agents are of the same type and an mms-allocation (or, equivalently, a 1-sufficient allocation) exists by Proposition 2.1. This implies the assertion. Therefore, we will assume that n2 > 0. Consequently, n1 > 0, too. Let A1, . . . , An and B1, . . . , Bn be mms-splits for agents of type 1 and type 2, respectively. If some set Ai ∩ Bj, where 1 ≤ i, j ≤ n, is of value at least 3 4 to some agent, then we allocate this set Ai ∩ Bj to this agent. Let P = C − (Ai ∩ Bj). Clearly, P is a path. Moreover, since every bundle in the split A1, . . . , An other than Ai is included in P , mms (n−1)(P, u1) ≥ mms (n)(C, u1) = 1. Similarly, we have mms (n−1)(P, u2) ≥ 1. Thus, we can assign the goods in P to the remaining n − 1 agents so that each agent receives a bundle that she values at least at 1. To this end, we may MAXIMIN SHARE ALLOCATIONS ON CYCLES 21 4 -sufficient. use the algorithm by Bouveret et al.[2] (cf. Theorem 3.5). The resulting allocation of goods on C is 3 4 -strong and so, 3 Therefore, we assume from now on that no set Ai ∩ Bj, 1 ≤ i, j ≤ n, is of value 3 4 or more for any of the agents. In particular, Ai 6⊆ Bj and Bj 6⊆ Ai for all i and j, 1 ≤ i, j ≤ n. Thus, without loss of generality, we can assume that each set Ai has a nonempty intersection with the sets Bi and Bi+1 (addition modulo n adjusted for [1..n]) and with no other set Bj. 4 to this agent. Suppose that fewer than n We will show that the sets A1, . . . , An can be allocated so that each agent receives a set of value at least 3 4 to agents of type 2. Then, there are two sets Ai, Ai+1 such that each of them is valued at less than 3 4 by agents of type 2. Thus, we have 2 of the sets A1, . . . , An have value at least 3 3 = u2(Bi ∪ Bi+1 ∪ Bi+2) = u2(Ai−1 ∩ Bi) + u2(Ai ∪ Ai+1) + u2(Ai+2 ∩ Bi+2) < + u2(Ai−1 ∩ Bi) + u2(Ai+2 ∩ Bi+2). 3 2 It follows that at least n This inequality implies that at least one of the sets Ai−1 ∩ Bi and Ai+2 ∩ Bi+2 has value larger that 3 4 for agents of type 2, a contradiction. 4 to agents of type 2. Since 2 agents of this type (we recall that n1 ≥ n2), we have sufficiently many such sets 4 -strong (cid:3) there are at most n for them. The remaining sets are allocated to agents of type 1. The resulting allocation is 3 and so, also 3 2 of the sets A1, . . . , An are of value at least 3 4 -sufficient. Theorem 4.8 is tight. Let us consider a cycle v1, v2, . . . , v12, and six agents of two types with the 4 -sufficient allocation exists utility functions shown in Figure 5. It is easy to check that in this case 3 but c-sufficient allocation for c > 3 4 does not. agents 1, 2, 3 agents 4, 5, 6 v1 3 3 v2 3 1 v3 1 2 v4 2 2 v5 2 1 v6 1 3 v7 3 3 v8 3 1 v9 1 2 v10 2 2 v11 2 1 v12 1 3 FIGURE 5. An example showing that the result in Theorem 4.8 is sharp. As we pointed out above, the 3 4 fraction of maximin shares can be guaranteed in a more general setting when agents are of three types. Specifically, the following theorem holds. We provide its proof in the appendix. Theorem 4.9. Let C be a cycle of goods and N a set of agents of at most three types. Then, a 3 4 -sufficient allocation of goods on C to agents in N exists. Theorem 4.9 is tight too. In Figure 6 we present an example where a 3 4 -sufficient allocation of goods on a cycle exists but there is no c-sufficient allocation for any c > 3 4 . agents 1, 2 agents 3, 4 agents 5, 6 v1 2 2 2 v2 0 1 1 v3 2 2 2 v4 1 1 0 v5 2 2 2 v6 1 0 1 v7 2 2 2 v8 0 1 1 v9 2 2 2 v10 1 1 0 v11 2 2 2 v12 1 0 1 v13 2 2 2 v14 0 1 1 v15 2 2 2 v16 1 1 0 v17 2 2 2 v18 1 0 1 FIGURE 6. An example showing that the result in Theorem 4.9 is sharp. 22 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI An interesting property of the example in Figure 6 is that the set {0, 1, 2} of values of the utility functions is very small. The problem of existence and construction of an mms-allocation with this set of values was studied in the original version of the mms-allocation problem (i.e. for complete graphs in our terminology). Amanatidis et al. [1] proved that unlike in the case of a cycle (see Figure 6) for a complete graph an mms-allocation with {0, 1, 2} as the set of values of the utility functions always exists. It is also easy to show that if {0, 1} is the set of values of the utility functions, then an mms- allocation for a cycle always exists and can be constructed in polynomial time. An analogous statement for complete graphs was observed earlier by Bouveret and Lemaıtre [4]. As we have seen earlier, even for three agents mms-allocations of goods on a cycle may not exist. Theorem 4.9 implies that when allocating goods on a cycle to three agents, we can guarantee that each agent receives a bundle worth at least 3 4 of her maximin share. We will now show that this guarantee can be strengthened to 5 6 . Re-examining the example in Figure 2 shows that this is the best guarantee we can get. To see this, we recall that in this case the maximin share value for agents 1 and 2 is 5 and for agent 3 is 6. In particular, any split in which an agent 1 or 2 obtains a bundle consisting of two or fewer items is at best a 4 5 -sufficient allocation (indeed, any two consecutive items have the total value of no more than 4). Thus, in any c-sufficient allocation with c ≥ 5 6 , agents 1 and 2 receive bundles of at least three items. If agent 3 receives at least three items, then all agents obtain bundles of exactly three items. There are only three such allocations. It is easy to see that one of them is 4 6 -sufficient. Since any two consecutive items have the total value at most 5 for agent 3, any allocation in which agent 3 receives no more than 2 items is at best 5 6 -sufficient; for instance, the allocation {v4, v5, v6}, {v7, v8, v9, v1},{v2, v3}. 6 -sufficient allocation of goods on a cycle always exists for 3 agents we start by introducing some additional terminology and two lemmas. Let C be a cycle, N = {1, 2, 3} a set of agents and ui the utility function of an agent i = 1, 2, 3. To simplify notation, we call an mms-split for (C, 3, ui) an mms(i)-split and we recall that we write mms(i) as a shorthand for mms (3)(C, ui). Lemma 4.10. Let C be a cycle and N = {1, 2, 3} a set of agents. Let c, 0 < c ≤ 1, be a real number. If for some agent i ∈ N two different bundles of an mms(i)-split of C are of value at least c· mms(j) to an agent j, where j 6= i, then there is a c-sufficient allocation of goods on C to agents in N . 6 -sufficient (some of them are actually 5 5 -sufficient and the other two are 5 To show that a 5 Proof. We assume without loss of generality that i = 1 and j = 2. Let us observe that for any agent and any 3-split of the cycle there is a bundle of value at least one third of the total value of all goods for this agent. The value of this bundle is larger than or equal to the maximin share for this agent. In particular, the value of one of the bundles of any mms(1)-split is greater than or equal to mms(3). The following protocol finds a c-sufficient allocation. Agent 3 picks a bundle of the mms(1)-split that has value at least mms(3) to her. At least one of the remaining two bundles has value at least c · mms(2) to agent 2. That bundle is allocated to agent 2. The remaining bundle of the split is assigned to agent 1. Lemma 4.11. Let C be a cycle and N = {1, 2, 3} a set of agents. Let c, 0 < c ≤ 1, be a real number. If for some two different agents i, j ∈ N , the intersection of a bundle of an mms(i)-split and a bundle of an mms(j)-split of C has value at least c · mms(i) to the agent i, then there is a c-sufficient allocation of goods in C to agents in N . (cid:3) Proof. Without loss of generality, we assume that i = 1 and j = 2. Let A and B be bundles of mms-splits for agents 1 and 2, respectively, such that Q = A ∩ B has value at least c · mms(1) to MAXIMIN SHARE ALLOCATIONS ON CYCLES 23 the agent 1. The set B − Q consists of one or two intervals, the latter in the case when A ⊂ B. Clearly, the elements in B − Q can be added to the remaining two bundles (different from B) of the mms(2)-split so that to form a 2-split of C − Q, say B′, B′′, with both parts of value at least mms(2) for the agent 2. We construct a c-sufficient allocation as follows. If u3(Q) ≥ c · mms(3), then we allocate Q to agent 3. Since Q ⊆ A, C − Q contains the remaining two bundles (different from A) of the mms(1)-split. Hence, u1(C − Q) ≥ 2 · mms(1). Since C − Q = B′ ∪ B′′, at least one of the bundles B′ and B′′ is worth mms(1) or more to the agent 1. We allocate that bundle to that agent and we allocate the remaining one to the agent 2. If u3(Q) < c · mms(3), then the agent 1 gets Q. Since u3(C) ≥ 3 · mms(3) and c ≤ 1, u3(C − Q) > 2 · mms(3). Thus, at least one bundle of B′ and B′′ is worth mms(3) or more to agent 3. That bundle goes to the agent 3 and the remaining one to agent 2. (cid:3) We are now ready to prove that if goods on a cycle are allocated to three agents, then each agent can be guaranteed a bundle worth at least 5 Theorem 4.12. Let C be a cycle and N = {1, 2, 3} a set of agents. Then, a 5 of goods in C to agents in N exists. 6 of her maximin share. 6 -sufficient allocation Proof. Because of Proposition 2.3, we restrict attention to the case when the set N of agents is regular. It follows that each agent values all goods on C at 3 and that all agents are proportional. In particular, mms(i) = 1, for every agent i ∈ N . Let Π1 = A1, A2, A3, Π2 = B1, B2, B3 and Π3 = C1, C2, C3 be mms-splits for agents 1, 2 and 3, respectively. If a bundle of one of these splits is contained in a bundle of another split, then an mms-allocation (and consequently a 5 6 -sufficient allocation) exists, by Lemma 4.11 (applied to the two bundles and c = 1). From now on we assume that there are no such containments. Then, we can relabel the parts of the splits Π1, Π2 and Π3 so that the sets A⌈(k+2)/3⌉ ∩ B⌈(k+1)/3⌉ ∩ C⌈k/3⌉, k = 1, 2, . . . , 9 (indices are computed modulo 3 with respect to the set {1, 2, 3}) form a split of the cycle. Let us call these 9 sets chunks. Since each agent i ∈ N values C at 3, any split of C into three bundles contains at least one bundle of value at least 1 to i, that is, of value at least mms(i) to i. In particular, such a bundle can be found in the mms-splits of the two agents other than i. Moreover, if for some two agents i, j ∈ N , i 6= j, more than one part in Πi has value at least 1 (= mms(j)) for j, then we are done by Lemma 4.10. Let i, j, k ∈ N be three different agents. We observe that if a bundle of Πk of value at least 1 (= mms(i)) for the agent i is different from a part of Πk of value at least 1 (= mms(j)) for the agent j, then we can easily distribute the parts of Πk among the agents to get an mms-allocation. So, we assume that for every agent i there is a unique bundle in Πi whose value for each of the remaining two agents is at least 1. Let us call this bundle i-significant. One can readily observe that, for i, j ∈ N , i 6= j, the intersection of the i-significant bundle and the j-significant bundle consists of either two chunks, one chunk or is empty. Moreover, it is not possible for each of the three pairs of different significant bundles to intersect on one chunk. Thus, there is a pair i, j ∈ N of agents such that the i-significant bundle and the j-significant bundle are disjoint or intersect on two chunks. We will consider these two cases separately. Clearly, we can assume without loss of generality that i = 1, j = 2 and that As and Bt, s, t ∈ {1, 2, 3}, are the 1-significant and 2-significant bundles, respectively, Case 1 As and Bt are disjoint. 24 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI We will not lose generality, either, if we assume that As = A1 and Bt = B2. Let us consider the mms-split Π3. By the definition of significant bundles, the value for the agent 3 of each of the sets A1 and B2 is at least 1. If the value of A2 or A3 for the agent 3 is at least 5 6 , then a 5 6 -sufficient allocation exists by 6 , then a 5 6 - Lemma 4.10. Moreover, if the value for the agent 3 of the set A1 ∩ C3 is at least 5 sufficient allocation exists by Lemma 4.11. than 5 6 . Then the value of the chunk A1 ∩ B1 ∩ C1 is larger than 3 − 3 · 5 So, let us assume that the values for the agent 3 of the sets A2, A3 and A1 ∩ C3 are all smaller In a very similar way we show that if any of the sets B1, B3, B2 ∩ C2 has value at least 5 6 for 6 -sufficient allocation exists. Otherwise, the value for the agent 3 of the chunk the agent 3, then a 5 A2 ∩ B2 ∩ C1 is larger than 1 2 . and A2 ∩ B2 ∩ C1 is larger than 1, a contradiction because the agents are proportional. Case 2 As and Bt intersect on two chunks. Then, however, the value for the agent 3 of the part C1, which contains the chunks A1 ∩ B1 ∩ C1 6 = 1 2 . We assume without loss of generality that As = A1 and Bt = B1. By the definition of significant bundles, the value for the agent 3 of each of the bundles A1 and B1 is at least 1. Let us then assume the latter. We observe that as in Case 1, if any of the bundles B2, B3, B1 ∩ C1 6 -sufficient allocation exists (by Lemma 4.10 or 4.11). Reasoning in exactly the same way as in the first case, we conclude that either a 5 6 -sufficient allocation exists or the value for the agent 3 of the chunk A1 ∩ B1 ∩ C1 is larger than 1 2 . has value at least 5 Otherwise, the value for the agent 3 of the chunk A1 ∩ B1 ∩ C3 is larger than 1 2 . Then, however, the value for the agent 3 of the set A1 ∩ B1 = (A1 ∩ B1 ∩ C1) ∪ (A1 ∩ B1 ∩ C3) is larger than 1. In this case an mms-allocation exists because we can assign A2 to the agent 1, B3 ∪ (A3 ∩ B2 ∩ C2) to the agent 2 and A1 ∩ B1 to the agent 3. (cid:3) 6 for the agent 3, then a 5 For more than three agents a 5 6 -sufficient allocation of goods on a cycle may not exist. The exam- ple in Figure 5 shows that this may be the case even for agents of two types only. For this example, we have already observed that there is a 3 4 -sufficient allocation but no c-sufficient allocation for c > 3 4 . We close this section by commenting on the problem of computing c-sufficient allocations. First, we note that all proofs we presented in this section, as well as the proof of Theorem 4.9, which we give in the appendix, yield algorithms for constructing c-sufficient allocations. These constructions apply to the case of regular sets of agents and require that both maximin shares and mms-splits be computed for every agent. Given those, the actual construction of a c-sufficient allocation provided in each proof can clearly be implemented to run in polynomial time. We recall that in the case of goods on a cycle, maximin shares and mms-splits can be computed in polynomial time. It follows then by Corollary 3.12 that in all cases we discussed in this section when c-sufficient allocations exist, they can be computed by algorithms running in polynomial times. 5. CONCLUSIONS We investigated maximin share allocations in the graph setting for the fair division problem of indivisible goods proposed by Bouveret et al. [2]. That paper settled the case of trees by showing that maximin share allocations of goods on trees always exist and can be computed in polynomial time. It also gave an example of goods on a cycle to be distributed to four agents where no maximin share allocation exists. MAXIMIN SHARE ALLOCATIONS ON CYCLES 25 Our work focused on cycles. We found several cases when maximin share allocations on cycles exist and can be found in polynomial time. For some other cases, when maximin share allocations are not guaranteed to exist, we found polynomial-time algorithms deciding the existence of max- imin share allocations and, if they do exist, computing them, too. Interestingly, we do not know the complexity of deciding the existence of maximin share allocations on cycles. We proved that the problem is in NP but whether it is NP-hard is open. In general, understanding the complexity of deciding the existence of maximin share allocations of goods on graphs is a major challenge. We improved an earlier upper bound from ΣP 2 down to ∆P 2 but, as in the case of cycles, we do not have any hardness results. Establishing such results and characterizing classes of graphs for which deciding the existence of maximin share allocations is in P, is NP-complete or goes beyond NP (under the assumption, of course, that the polynomial hierarchy does not collapse) are important open problems. Perhaps our most interesting results concern the existence of allocations guaranteeing each agent a given fraction of her maximin share. For instance, we show that for three agents one can always find an allocation giving each agent at least 5/6 of her maximin share. For an arbitrary number of agents of two or three types, we show allocations giving each agent 3/4 of the maximin share, and we also obtain some results for the case when the number of types is larger than three. Moreover, in each case, these allocations can be found by polynomial-time algorithms. We conjecture that in the general case of any number of agents there exist allocations guaranteeing all agents at least 3/4 of their maximin share. Theorems 4.8 and 4.9 show the conjecture holds if agents are of three or fewer types. However, the methods we developed so far seem too weak to prove it in full generality. The work of the second author was partially supported by the NSF grant IIS-1618783. ACKNOWLEDGMENTS REFERENCES [1] G. Amanatidis, E. Markakis, A. Nikzad, A. Saberi, Approximation algorithms for computing maximin share allocations, in: M. M. Halld´orsson, K. Iwama, N. Kobayashi, B. Speckmann (eds.), Proceedings of the 42nd International Colloquium on Automata, Languages, and Programming, LNCS 9134, Springer, 2015, pp. 39 -- 51. [2] S. Bouveret, K. Cechl´arov´a, E. Elkind, A. Igarashi, D. Peters, Fair division of a graph, in: C. Sierra (ed.), Proceedings of the 26th International Joint Conference on Artificial Intelligence, IJCAI 2017, ijcai.org, 2017, pp. 135 -- 141. [3] S. Bouveret, Y. Chevaleyre, N. Maudet, Fair allocation of indivisible goods, in: F. Brandt, V. Conitzer, U. Endriss, J. Lang, A. D. Procaccia (eds.), Handbook of Computational Social Choice, Cambridge University Press, 2016, pp. 284 -- 310. [4] S. Bouveret, M. Lemaıtre, Characterizing conflicts in fair division of indivisible goods using a scale of criteria, Autonomous Agents and Multi-Agent Systems 30 (2) (2016) 259 -- 290. [5] S. J. Brams, A. D. Taylor, Fair Division: From Cake-Cutting to Dispute Resolution, Cambridge University Press, 1996. [6] E. Budish, The combinatorial assignment problem: Approximate competitive equilibrium from equal incomes, Journal of Political Economy 119(6) (2011) 1061 -- 1103. [7] M. R. Garey, D. S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness, W. H. Freeman and Co., 1979. [8] M. Ghodsi, M. Hajiaghayi, M. Seddighin, S. Seddighin, H. Yami, Fair allocation of indivisible goods: Improve- ments and generalizations, in: Proceedings of the 2018 ACM Conference on Economics and Computation, EC- 2018, ACM, New York, NY, USA, 2018, pp. 539 -- 556. [9] R. King, S. Burton, Land fragmentation: Notes on a fundamental rural spatial problem, Progress in Human Geography 6(4) (1982) 475 -- 94. [10] C. H. Papadimitriou, Computational Complexity, Pearson, 1993. 26 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI [11] A. Procaccia, Cake-cutting algorithms, in: F. Brandt, V. Conitzer, U. Endriss, J. Lang, A. D. Procaccia (eds.), Handbook of Computational Social Choice, Cambridge University Press, 2016, pp. 311 -- 329. [12] A. D. Procaccia, J. Wang, Fair enough: guaranteeing approximate maximin shares, in: M. Babaioff, V. Conitzer, D. Easley (eds.), Proceedings of the ACM Conference on Economics and Computation, EC-2014, ACM, 2014, pp. 675 -- 692. APPENDIX - PROOF OF THEOREM 4.9 In this section we prove Theorem 4.9. We adhere to the notation we used throughout the paper. In particular, we consider a set N = {1, 2 . . . , n} of agents of three types and a cycle C of m goods. In the proof, as in several other places in the paper, we work under the assumption that agents are regular. Thus, their mms-splits consist of non-empty bundles. Therefore, in our auxiliary results, concerning properties of splits, we restrict attention to splits into non-empty bundles, that is, to splits that are partitions. We select and fix an element a in C and call it an anchor. Whenever we say that X = X1, . . . , Xn is a split we we mean that it is a split of C and, assume that a ∈ X1 and that parts Xi are enumerated according to their location on the cycle as we traverse it clockwise. Further, for every i = 1, . . . , n, we write Xi for the set of all elements in the segment of C extending from the anchor a clockwise to the last element of Xi (inclusive). We also assume that arithmetic expressions appearing in indices labeling bundles in splits are evaluated modulo n; in particular, for every i = 1, 2, . . . , n, i + n = i. Finally, we call any set of a split X an X-set. Assuming these conventions, let X = X1, . . . , Xn and Y = Y1, . . . , Yn be two splits. A bundle Xi is a jump to the split Y if Xi ⊆ Yi. In such case, we will also say that Xi is a jump to Yi+1. We usually omit the reference to the target split of a jump, if it is clear from the context. The following property follows directly from the definition. Lemma 5.1. Let X = X1, . . . , Xn and Y = Y1, . . . , Yn be two splits. Then for every i, 1 ≤ i ≤ n, at least one of the sets Xi, Yi is a jump (to the other split). Lemma 5.2. Let X = X1, . . . , Xn and Y = Y1, . . . , Yn be two splits. If for some i, 1 ≤ i < n, Xi is a jump and Xi+1 is not a jump, then Yi+1 ⊆ Xi+1 and Yi+1 is a jump. Proof. Since Xi is a jump, Xi ⊆ Yi. Further, since Xi+1 is not a jump, Yi+1 is a jump (Lemma 5.1). Thus, Yi+1 ⊆ Xi+1. Since Xi+1 = Xi+1 \ Xi and Yi+1 = Yi+1 \ Yi, Yi+1 ⊆ Xi+1 follows. (cid:3) of up to n consecutive elements of X (with X1 being a successor of Xn). of C is XY-useful if for every i = 1, 2, . . . , n, at least one of the following conditions holds: Let X = X1, . . . , Xn be a sequence of n subsets of C. An X -interval is any sequence Xi, Xi+1 . . . , Xj Let X = X1, . . . , Xn and Y = Y1, . . . , Yn be two splits. A sequence Z = Z1, . . . , Zn of subsets (1) Zi = Xi and Zi+1 = Xi+1 (2) Zi = Yi and Zi+1 = Yi+1 (3) Zi = Xi and Xi is a jump to Yi+1 (4) Zi = Yi and Yi is a jump to Xi+1. Lemma 5.3. Let X = X1, . . . , Xn and Y = Y1, . . . , Yn be splits. If a sequence Z = Z1, . . . , Zn is XY-useful then the sets in Z are pairwise disjoint. Proof. Let us consider sets Zi and Zj, where 1 ≤ i < j ≤ n. From the definition of a XY-useful sequence, it follows that for every k, 1 ≤ k ≤ n, Zk = Xk or Zk = Yk. Thus, Zi = Xi and Zj = Xj, Zi = Yi and Zj = Yj, Zi = Xi and Zj = Yj, or Zi = Yi and Zj = Xj. In the first two cases, the sets Zi and Zj are disjoint, because they are two different members of a split. The MAXIMIN SHARE ALLOCATIONS ON CYCLES 27 other two cases are symmetric. Without loss of generality, we will restrict attention to the last one: Zi = Yi and Zj = Xj.Clearly, we may assume that Zi 6= Xi and Zj 6= Yj (otherwise, Zi and Zj would be members of the same split, the case we already considered). Let k be the largest integer such that i ≤ k < j, Zk = Yk, and Zk+1 6= Yk+1. Since Zi = Yi, k is well defined. It follows that Yk is a jump. Thus, Yk ⊆ Xk. Since k + 1 ≤ j ≤ n, Xj ∩ Xk = ∅. Consequently, Xj ∩ Yk = ∅. Since i ≤ k, if i ≥ 2 then Zi ⊆ Yk (indeed, it follows from the equality Zi = Yi). Thus, Zi and Zj are disjoint. Let us assume then that i = 1. Since the anchor a belongs to Y1, Y1 is the union of two segments, A = C \ Yn and B = Y1. Since Xj ∩ Yk = ∅, Xj ∩ B = ∅. Let p be the smallest integer such that j ≤ p ≤ n, Zp = Xp and Zp+1 6= Xp+1. Since Zj = Xj and Z1 6= X1 (we proved Zi 6= Xi above and here i = 1), p is well defined. It follows that Xp is a jump. Therefore, Xp ⊆ Yp. Since Yp ⊆ Yn, A ∩ Yp = ∅. Since Xj ⊆ Xp ⊆ Yp, Xj ∩ A = ∅. Thus, Zj(= Xj) and Z1(= Y1) are disjoint. (cid:3) Splits X and Y are proper relative to each other if for every X ∈ X and Y ∈ Y, X 6⊆ Y and Y 6⊆ X. The following lemma provides some basic properties of splits that are proper relative to each other (they are easy to see and we omit the proof). Lemma 5.4. Let splits X = X1, . . . , Xn and Y = Y1, . . . , Yn, with both sequences enumerated according to our convention clockwise starting with sets containing the anchor a, be proper relative to each other. Then for every i = 1, . . . , n, (1) Xi ⊆ Yi ∪ Yi+1 and Yi ⊆ Xi−1 ∪ Xi, and (2) Xi−1 ∩ Yi 6= ∅, Xi ∩ Yi 6= ∅, and Xi ∩ Yi+1 6= ∅. or for every i = 1, . . . , n, (1) Xi ⊆ Yi−1 ∪ Yi and Yi ⊆ Xi ∪ Xi+1, and (2) Xi ∩ Yi−1 6= ∅, Xi ∩ Yi 6= ∅, and Xi+1 ∩ Yi 6= ∅. From now on we consider splits that are mms-splits for a set N of regular agents. In particular, all agents are proportional and for every agent the total utility of all goods in C is n. We also use the following terminology. A subset X of C that induces in C a connected subgraph is acceptable to an agent x if the total value of the goods in X to x is at least 3 4 . Lemma 5.5. Let x and y be two regular agents from N . Let X and Y be mms-splits of a cycle for agents x and y, respectively, and let X and Y be proper relative to each other. If no set X ∩ Y , where X ∈ X and Y ∈ Y, is acceptable to y, then at least one of each pair of consecutive sets of X is acceptable to y. Proof. To prove the lemma let us consider any two consecutive sets in X , say X and X ′, with X ′ directly following X in X . Since X and Y are proper relative to each other, Lemma 5.4 implies that there are three consecutive sets Y, Y ′, Y ′′ ∈ Y such that X ∪ X ′ ⊆ Y ∪ Y ′ ∪ Y ′′. Let X−1 and +1 be the predecessor of X in X and the successor of X ′ in X , respectively. By our assumption, X ′ uy(X−1 ∩ Y ) < 3 4 , where uy is the utility function for agent y. It follows 4 and uy(X ′ that +1 ∩ Y ′′) < 3 3 = uy(Y ∪ Y ′ ∪ Y ′′) = uy(X−1 ∩ Y ) + uy(X) + uy(X ′) + uy(X ′ < 2 · + uy(X) + uy(X ′). +1 ∩ Y ′′) 3 4 4 or uy(X ′) ≥ 3 Thus, uy(X) ≥ 3 4 , that is, at least one of X and X ′ is acceptable to y. (cid:3) 28 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI In the next lemma and in the proof of Theorem 4.9, we use the following notation. We write n1, n2, n3 for the numbers of agents of types i1, i2, i3, respectively. In particular, n = n1 + n2 + n3. Further, we assume without loss of generality that n1 ≥ n2 ≥ n3 ≥ 1. Lemma 5.6. Let A = A1, . . . , An (resp. B = B1, . . . , Bn and C = C1, . . . , Cn) be mms-splits of the cycle C for agents of type i1 (resp. i2 and i3). If n1 = n2, n3 ≥ 1, no set Ai ∩ Bj is acceptable to any agent, and some C-set is contained in some A-set or B-set, then there is a 3 4 -sufficient allocation for a cycle and any number of agents of three types. Proof. Since no set Ai ∩ Bj is acceptable to any agent, we have that for every i and j, where 1 ≤ i, j ≤ n, Ai 6⊆ Bj and Bi 6⊆ Aj. Thus, the splits A and B are proper relative to each other. Our assumptions do not distinguish between the splits A and B. Therefore, without loss of generality, we may assume that a C-set, say Ck, is contained in an A-set, say Ai. Let Bj and Bj+1 be two B-sets that intersect Ai. It follows that Ck ⊆ Bj∪Bj+1. If Ck ⊆ Bj+1, then Ck ⊆ Ai∩Bj+1. But then Ai ∩ Bj+1 would be acceptable to agents of type i3 (and there is at least one such agent). This is a contradiction. Thus, Ck ∩ Bj 6= ∅. Consequently, Ck ∩ Bj ∩ Ai 6= ∅. Without loss of generality, we may assume that the anchor a belongs to Ck ∩ Bj ∩ Ai. Assuming the sets in the splits A, B and C are enumerated with respect to that anchor, we have k = j = i = 1, C1 ⊆ A1 and B1 ⊆ An ∪ A1. The latter implies (we recall that A and B are proper relative to each other and so, Lemma 5.4 applies) that for every i = 1, . . . , n, In what follows we will be applying Lemmas 5.1 - 5.5. We can do so, as we follow here the same Ai ⊆ Bi ∪ Bi+1 and Bi ⊆ Ai−1 ∪ Ai. convention for enumerating sets in splits with respect to the same fixed anchor. Case 1 Some two consecutive A-sets are jumps to C. If all A-sets are jumps to C, then the sequence Z = C1, A2, . . . , A2n1+1, C2n1+2, . . . , Cn is AC-useful. Indeed, C1 is a jump as it is a subset of A1 (thus, C1 ⊆ A1, in the notation we introduced to define jumps), and A2n1+1 is a jump by assumption. By Lemma 5.3, all sets in Z are pairwise disjoint. By assumption, no set Ai ∩ Bj is acceptable to any agent of type i2. Thus, by Lemma 5.5, some n1(= n2) of different sets in the interval A2, . . . , A2n1+1 in Z are acceptable to agents of type i2. We assign them to those agents. We assign the remaining n1 sets in A2, . . . , A2n1+1 to agents of type i1. Finally, we assign the sets Cj, j = 1, 2n1 + 2, . . . , n, to agents of type i3. In this way all agents are allocated sets that are acceptable to them. If there are any unallocated goods (Z does not have to cover all goods in C), they can be distributed to agents so that all agents receive bundles inducing in C a path. This, yields an assignment that is a 3 4 -sufficient allocation. So, assume that some of A-sets are not jumps. Since some two consecutive A-sets are jumps, there is an integer i such that Ai−1 and Ai are jumps and Ai+1 is not. By Lemma 5.2, Ci+1 is a jump and Ci+1 ⊆ Ai+1. The latter property implies that we may relabel sets Ai and Ci so that An−1 and An are jumps and A1 is not. Indeed, under this relabeling we have C1 ⊆ A1, which we assumed to hold when we started the proof (of course, we also need to relabel sets Bi correspondingly). We now consider the case when Cn3 or Cn3−1 is a jump. In the first case, we set and, in the second case, Z = C1, . . . , Cn3, An3+1, . . . , An Z = Cn, C1, . . . , Cn3−1, An3, . . . , An−1. MAXIMIN SHARE ALLOCATIONS ON CYCLES 29 Since An and An−1 are jumps, in either case Z is AC-useful. Thus, by Lemma 5.3, its elements are pairwise disjoint. Reasoning as above, we distribute the sets An3+1, . . . , An (respectively, An3, . . . , An−1) among 2n1 agents of types i1 and i2 (since n = 2n1 + n3, in each case there are 2n1 sets in that sequence), and then allocate the remaining n3 sets, all of them C-sets, to agents of type i3. Clearly, the resulting allocation (after possibly attaching goods not "covered" by Z to appropriate sets in Z) is 3 Next, let us assume that neither Cn3 nor Cn3−1 is a jump. Then let k be the largest integer such that 1 ≤ k < n3 − 1 and Ck is a jump (such k exists because C1 is a jump). Moreover, let ℓ be the smallest integer j such that n3 < j ≤ n and Cj is a jump, if such j exists, or ℓ = n, otherwise. Observe that the sequence 4 -sufficient. Z = C1, . . . , Ck, Ak+1, . . . , Aℓ−(n3−k), Cℓ−(n3−k)+1, . . . , Cℓ, Aℓ+1 . . . , An is AC-useful. Indeed, for k < j < ℓ the sets Cj are not jumps. By Lemma 5.1, the sets Aj, where k < j < ℓ are jumps. Since k < ℓ − (n3 − k) < ℓ, Aℓ−(n3−k) is a jump. Moreover, Ck is a jump based by our choice of k, Cℓ is a jump by our choice of ℓ, if ℓ is not assigned to n by default, and An is a jump by assumption. By Lemma 5.3, all sets in Z are pairwise disjoint. Clearly, if ℓ and n(= 2n1 + n3) have the same parity, then the numbers of sets in the intervals Ak+1, . . . , Aℓ−(n3−k) and Aℓ+1, . . . , An are even (the latter is present and needs to be considered only if ℓ < n). Moreover, the total number of sets in these two intervals is n − n3 = 2n1 = 2n2. By Lemma 5.5, we can select among these A-sets in Z n1(= n2) sets that are acceptable to agents of type i2. We allocate these sets to those agents. We then allocate the remaining n1 of these A-sets to the agents of type i1. Finally, agents of type i3 receive the n3 C-sets of Z. Extending Z to an allocation of C, yields an allocation that is 3 So, assume now that ℓ and n have different parity. In this case, we define ℓ′ to be the smallest integer j such that n3 < j ≤ n − 1 and Cj is a jump, if such j exists, or ℓ′ = n − 1 otherwise. Obviously, ℓ′ = ℓ or ℓ′ = n − 1, so in both cases ℓ′ and n have different parity. Since An−1 is a 4 -sufficient. jump, the same reasoning as above yields that the sequence Z = Cn, C1, . . . , Ck, Ak+1, . . . , Aℓ′−(n3−k)+1, Cℓ′−(n3−k)+2, . . . , Cℓ′, Aℓ′+1 . . . , An−1 is AC-useful. Thus, its elements are pairwise disjoint. Moreover, the A-intervals Ak+1, . . . , Aℓ′−(n3−k)+1 and Aℓ′+1, . . . , An−1 (if ℓ′ < n − 1) consist of even numbers of sets and the total number of sets in these two intervals is 2n1 = 2n2. Therefore, we can define 3 4 -sufficient allocation in a similar way as when ℓ and n have the same parity. Case 2 Some two consecutive C-sets are jumps to A. This case is symmetric to the previous one if we enumerate sets counterclockwise rather than clockwise. Indeed, if a C-set X is a jump to an A-set Y under the clockwise enumeration, then Y is a jump to X under the counterclockwise enumeration. Case 3 No two consecutive A-sets are jumps to C, and no two consecutive C-sets are jumps to A. Since C1 ⊆ A1, C1 is a jump. It follows that Cn ∩ A1 6= ∅ as otherwise, Cn would be a jump. In particular, Cn 6⊆ An. Applying Lemma 5.2 repeatedly, yields that for all odd i, 1 ≤ i ≤ n, Ci ⊆ Ai and for every even i, 1 ≤ i ≤ n, Ai ⊆ Ci. Further, since Cn 6⊆ An, it follows that n is even. We will now show that the splits B and C are proper relative to each other. To this end, we need to show that for every i and j, 1 ≤ i, j ≤ n, Ci 6⊆ Bj and Bj 6⊆ Ci. First, let us assume that Ci ⊆ Bj, for some i and j, 1 ≤ i, j ≤ n. If Ai ⊆ Ci, then Ai ⊆ Bj, a contradiction as the splits A and B are proper relative to each other. If Ci ⊆ Ai, then Ci ⊆ Ai ∩ Bj. Therefore, Ai ∩ Bj is 30 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI acceptable to agents of type i3, a contradiction with the assumptions of the lemma. Thus, for every i, j such that 1 ≤ i, j ≤ n, Ci 6⊆ Bj. Next, let us assume that Bj ⊆ Ci, for some i and j, 1 ≤ i, j ≤ n. If i is odd, then Ci ⊆ Ai. It follows that Bj ⊆ Ai, a contradiction with the fact that A and B are proper relative to each other. Thus, i is even, so Bj ⊆ Ci ⊆ Ci−1 ∪ Ci ∪ Ci+1 ⊆ Ai−1 ∪ Ai ∪ Ai+1. We recall that we have Bj ⊆ Aj−1 ∪ Aj. Hence, j = i or j = i + 1. Let us assume that j = i. Since i is even, Ci−1 ⊆ Ai−1 ⊆ Bi−1 ∪ Bi. Moreover, Bi ⊆ Ci, so Bi ∩ Ci−1 = ∅. Thus, Ci−1 ⊆ Bi−1, a contradiction with what we already proved above. It follows that we must have j = i + 1. In this case we similarly get, Ci+1 ⊆ Ai+1 ⊆ Bi+1 ∪ Bi+2 and Bi+1 ⊆ Ci (so, Bi+1 ∩ Ci+1 = ∅). Thus, Ci+1 ⊆ Bi+2, a contradiction as in the case j = i. This completes the proof that for every i, j such that 1 ≤ i, j ≤ n, Bj 6⊆ Ci, and show that the splits B and C are proper relative to each other. Let us observe that B1 ⊆ An ∪ A1, An ⊆ Cn, B1 ∩ C1 6= ∅ and C1 6⊆ B1. It follows that B1 ⊆ Cn ∪ C1. Since the splits B and C are proper relative to each other, Lemma 5.4 implies that for every i, 1 ≤ i ≤ n, Bi ⊆ Ci−1 ∪ Ci. Let us suppose that some set Bi ∩ Cj, where 1 ≤ i, j ≤ n, is acceptable for agents of type i2. Then, there is i, 1 ≤ i ≤ n, such that Bi ∩ Ci−1 or Bi ∩ Ci has this property. Let us assume the former. We define a 3 4 -sufficient allocation as follows. We consider the se- quence Ci, Ci+1, . . . , Ci+n3 of n3 + 1 C-sets. One of the sets Ci, Ci+1 contains an A-set and that set is assigned to an agent of type i1. The remaining n3 sets of that sequence go to agents of type i3. The sequence Bi−1, Bi−2, . . . , Bi−(2n1−2) contains at least n1 − 1 sets acceptable for agents of type i1 (Lemma 5.5). We select some n1−1 of such sets and assign them to agents of type i1. We allocate the remaining n1−1 sets from this sequence and the set Bi∩Ci−1 to agents of type i2. In this way, all agents are allocated sets that are acceptable to them. Moreover, these sets are pairwise disjoint. In- deed, the sets Ci+n3 and Bi−(2n1−2) do not overlap because i−(2n1−2) = i+n−2n1+2 = i+n3+2 (we recall that the arithmetic of indices is modulo n), and Bi+n3+2 ⊆ Ci+n3+1 ∪ Ci+n3+2. Thus, the assignment is a 3 The latter case, when Bi ∩ Ci is acceptable to an agent of type i2 can be handled similarly by Assume now that no set Bi ∩ Cj is acceptable for agents of type i2. By Lemma 5.5 one in every two consecutive sets of the mms-split C1, . . . , Cn is acceptable for agents of type i2. Clearly, An is a jump to the split C because C1 ⊆ A1. Moreover, Cn3 ⊆ An3 or Cn3+1 ⊆ An3+1. Therefore, Cn3+j is a jump for some j ∈ {0, 1}. It follows that the sequence considering sequences Ci−1, Ci−2, . . . , Ci−(n3+1) and Bi+1, Bi+2, . . . , Bi+2n1−2. 4 -sufficient allocation. Z = C1, . . . , Cn3+j, An3+j+1, . . . , An is AC-useful. By the comment above, at least j sets in C1, . . . , Cn3+j are acceptable to agents of type i2 (since n3 ≥ 1). We assign j of these sets to agents of type i2. The remaining n3 sets of C1, . . . , Cn3+j are assigned to agents of type i3. Next, the sets An3+j+1, . . . , An are distributed among agents of types i1 and i2. Agents of type i2 take n2 − j sets of this interval (there are this many sets acceptable to agents of type i2 because (cid:4) n−n3−j 2 (cid:5) ≥ n2 − j). The agents of type i1 receive the remaining n1 sets of the A-interval. By construction, this allocation is 3 4 - sufficient. (cid:3) (cid:5) = (cid:4) 2n2−j 2 Theorem 4.9. Let C be a cycle of goods and N a set of agents of at most three types. Then, a 3 4 -sufficient allocation of goods on C to agents in N exists. MAXIMIN SHARE ALLOCATIONS ON CYCLES 31 Proof. We adhere to the notation from Lemma 5.6. Further, as there, we assume that n3 ≥ 1 (otherwise, we are in the case of agents of two types settled by Theorem 4.8). We now consider several cases. In each case we assume that none of the cases considered earlier applies. Case 1. Some set Ai ∩ Bj is acceptable to some agent. In this case, the existence of a 3 4 -sufficient allocation follows directly from Lemma 4.6. From now on, we assume Case 1 does not apply, that is, no set Ai ∩ Bj is acceptable to any agent. It follows that for every 1 ≤ i, j ≤ n, Ai 6⊆ Bj and Bj 6⊆ Ai. Thus, the splits A and B are proper relative to each other. This means, in particular, that from now on we may apply Lemma 5.5 to splits A and B, as well as to B and A. Case 2. Some C-set is contained in some A-set. Given our comments above, Lemma 5.6 applies and we can assume that n1 > n2 ≥ n3. We can also assume without loss of generality that C1 ⊆ A1. Then, the sets An and C1 are jumps. If the set Cn3 is a jump too, then the sequence Z = C1, . . . , Cn3, An3+1, . . . , An is AC-useful. By Lemma 5.5 (which also applies now), at least ⌊ n1+n2 ⌋ ≥ n2 of the sets that form the sequence An3+1, . . . , An are acceptable for agents of type i2. We assign n2 of them to these agents. We assign the remaining n1 sets of this interval to agents of type i1. Agents of type i3 receive the sets C1, . . . , Cn3. In this way, we construct a 3 4 -sufficient allocation. If Cn3 is not a jump, then let k be the largest integer such that 1 ≤ k < n3 and Ck is a jump (such k exists because C1 is a jump). Moreover, let ℓ be the smallest integer j such that n3 < j ≤ n and Cj is a jump, if such j exists, or ℓ = n otherwise. We define 2 Z = C1, . . . , Ck, Ak+1, . . . , Aℓ−(n3−k), Cℓ−(n3−k)+1, . . . , Cℓ, Aℓ+1 . . . , An 2 (cid:5) +(cid:4) n−ℓ 2 (cid:5) ≥ ℓ−n3−1 2 + n−ℓ−1 By Lemma 5.5, there are at least (cid:4) ℓ−n3 and observe that it is AC-useful. Indeed, Ck is a jump by the construction. Further, k < ℓ − (n3 − k) < ℓ and for k < j < ℓ the sets Cj are not jumps. Thus, Aℓ−(n3−k) is a jump by Lemma 5.1. Finally, if ℓ < n, Cℓ is a jump by the choice of ℓ and An is a jump by assumption. 2 = n1+n2 2 − 1 > n2 − 1 sets in the intervals Ak+1, . . . , Aℓ−(n3−k) and Aℓ+1 . . . , An (the latter present only if ℓ < n) that are acceptable to agents of type i2. We assign n2 of them to these agents. The remaining sets of these intervals (there are n1 of them) are assigned to agents of type i1. Finally, agents of type i3 receive the sets in the intervals C1, . . . , Ck and Cℓ−(n3−k)+1, . . . , Cℓ. This yields a 3 Case 3. Some A-set is contained in some C-set. Without loss of generality we may assume that A1 ⊆ C1. Since no C-set is contained in an A-set (that possibility is excluded by Case 2), a simple inductive argument shows that for every i = 2, . . . , n, Ai ends strictly before Ci does. In particular, An ends strictly before Cn does. This implies that A1 ∩ Cn 6= ∅, a contradiction with A1 ⊆ C1 (and the fact that different sets in a split are disjoint). Given Cases 2 and 3, from now on we may assume that the splits A and C are proper relative to each other. Case 4. Some set Ai ∩ Cj is acceptable to agents of type i3. As we noted, we may assume that the splits A and C proper are relative to each other. Thus, we can relabel the sets in the split C if necessary so that for all i, Ai ⊆ Ci ∪ Ci+1 and, consequently, also Ci ⊆ Ai−1 ∪ Ai. 4 -sufficient allocation. 32 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI We can assume without loss of generality that A1 ∩ C1 or An ∩ C1 is acceptable to agents of type i3. Let us assume the former. We allocate A1 ∩ C1 and the sets C2, C3, . . . , Cn3 to agents of type i3. Clearly, all these sets are included in A1 ∪ . . .∪ An3 . The sequence An3+1, . . . , An of the remaining A-sets contains n1 + n2 sets. By Lemma 5.5, at least (cid:4) n1+n2 (cid:5) ≥ n2 of them are acceptable for agents of type i2. We allocate any n2 of those sets to agents of type i2 and the remaining n1 of them to the agents of type i1. The allocation we defined in this way extends to a 3 4 -sufficient allocation for C. 2 2⌋ = ⌈ n−1 2 ⌉, we have ⌊ n−1 2 ⌋ ≥ n2+n3−1. Thus, ⌊ n−1−2n2 2 considering sequences Cn, Cn−1, . . . , Cn−n3+2 and A1, A2, . . . , An−n3. The other case, when An ∩ C1 is acceptable to agents of type i3, can be handled similarly by Excluding Cases 1-4 means in particular that A and C are proper relative to each other and that no set Ai ∩ Cj is acceptable to agents of type i3. Therefore, from now on we may apply Lemma 5.5 also to the splits A and C. Case 5. n1 ≥(cid:4) n 2(cid:5). Let Ai be any A-set acceptable to an agent of type i3. Such sets exist by Lemma 5.5. Without loss of generality we may assume that An is acceptable to an agent of type i3. By Lemma 5.5, the sequence A1, A2, . . . , A2n2 contains n2 sets that are acceptable to agents of type i2. We allocate these sets to them. Next, we consider the sequence A2n2+1, A2n2+2, . . . , An−1 of n−1−2n2 sets that follow A2n2. Since n1 ≥ ⌊ n ⌋ ≥ n3−1. By Lemma 5.5, it follows that some n3 − 1 sets in the sequence A2n2+1, A2n2+2, . . . , An−1 are acceptable to agents of type i3. We allocate these sets and the set An to agents of type i3. We allocate the remaining n1 sets in A1, . . . , An to agents of type i1. This defines a 3 Case 6. Some set Ai ∩ Cj is acceptable to agents of type i1 and n1 > n2. By the same argument as in Case 4, we may assume that Ci ⊆ Ai−1 ∪ Ai for all i and that A1 ∩ C1 or An ∩ C1 is acceptable to agents of type i1. If A1 ∩ C1 is acceptable to agents of type i1, we proceed as follows. We allocate the sets C2, C3, . . . , Cn3+1 to agents of type i3. Next, we note that every other set in the sequence An3+2, An3+3, . . . , An is acceptable to agents of type i2. Since the sequence contains n − n3 − 1 = n1 + n2 − 1 ≥ 2n2 sets (recall that we assume here that n1 > n2), some n2 of the sets in the sequence are acceptable to agents of type i2. We allocate these n2 sets to agents of type i2. We allocate the remaining n1 − 1 sets in that sequence and the set A1 ∩ C1 to agents of type i1. In this way, all agents get sets that are acceptable to them. Further, we note that C2 ∪ . . . ∪ Cn3+1 ⊆ A1 ∪ . . . ∪ An3+1. It follows that all sets used in the allocation are pairwise disjoint and so give rise to a 3 The case when A1 ∩ Cn is acceptable to an agent of type i1 follows from the one we just consid- 4 -sufficient allocation for C. 4 -sufficient allocation for C. ered by the same argument as that used in Case 4. Case 7. Some C-set is contained in some B-set. By Lemma 5.6, we can assume that n1 > n2 (because Case 1 is excluded, A and B are proper relative to each other and no set Ai ∩ Bj is acceptable to any agent). We now recall that splits C and A are proper relative to each other. Moreover, no set Ai ∩ Cj is acceptable for agents of type i1 (Case 6 is excluded). Thus, Lemma 5.5 applies to splits C and A and implies that every pair of consecutive sets of the mms-split C1, . . . , Cn contains at least one set acceptable to agents of type i1. We can assume without loss of generality that C1 ⊆ B1. It follows that the sets Bn and C1 are jumps. If the set C2n3 is a jump, then the sequence Z1 = C1, . . . , C2n3, B2n3+1, . . . , Bn MAXIMIN SHARE ALLOCATIONS ON CYCLES 33 is BC-useful. Similarly, if the set C2n3−1 is a jump, then the sequence Z2 = C1, . . . , C2n3−1, B2n3, . . . , Bn is BC-useful. Assume now that the sets C2n3−1 and C2n3 are not jumps. Let k, 1 ≤ k < 2n3 − 1, be the largest index such that Ck is a jump (since C1 is a jump, k is well defined), and let ℓ be the smallest integer j such that 2n3 < j ≤ n and Cj is a jump, if such j exists, or ℓ = n otherwise. We observe that the sequence Z3 = C1, . . . , Ck, Bk+1, . . . , Bℓ−(2n3−k), Cℓ−(2n3−k)+1, . . . , Cℓ, Bℓ+1 . . . , Bn is BC-useful. In particular the set Bℓ−(2n3−k) is a jump. Indeed, k < ℓ − (2n3 − k) < ℓ and for k < j < ℓ the sets Cj are not jumps. By Lemma 5.1, the sets Bj are jumps. It can be shown similarly that the sequence Z4 = C1, . . . , Ck, Bk+1, . . . , Bℓ−(2n3−k)+1, Cℓ−(2n3−k)+2, . . . , Cℓ, Bℓ+1 . . . , Bn is BC-useful. We now observe that the sequence Z1 is a special case of the sequence Z3 for k = 2n3, and the sequence Z2 is a special case of the sequence Z4 for k = 2n3 − 1. Thus, to complete this case, it is enough to describe allocations based on sequences Z3 and Z4. For k even, we construct an allocation based on the sequence Z3 which, we recall is BC-useful. Since no set Ai∩Cj is acceptable for agents of type i1, by Lemma 5.5 half of the sets C1, . . . , Ck and half of the sets Cℓ−(2n3−k)+1, . . . , Cℓ (the number of sets in each of these intervals is even), n3 sets in total, are acceptable for agents of type i1. We allocate these sets to agents of type i1. Agents of type i3 receive the remaining n3 sets from these intervals. Since at least every other set of the intervals Bk+1, . . . , Bℓ−(2n3−k) and Bℓ+1 . . . , Bn in Z5 is acceptable for agents of type i1 (by Lemma 5.5), these two intervals contain at least(cid:4) ℓ−2n3 2 (cid:5) +(cid:4) n−ℓ 2(cid:5)− 1− n3 ≥ n1− n3 sets (because Case 5 is excluded,(cid:4) n 2(cid:5) ≥ n1 +1) that are acceptable to agents of type i1. We allocate n1 − n3 of these sets to the agents of type i1 that have not been allocated any set in the previous stage. The remaining sets, there are n2 of them, are allocated to agents of type i2. Clearly, this allocation gives rise to a 3 2 (cid:5)− n3 ≥(cid:4) n 2(cid:5) +(cid:4) n−ℓ 2 (cid:5) =(cid:4) ℓ 4 -sufficient allocation for C. 2⌋ of the sets C1, . . . , Ck and at least ⌊ 2n3−k−1 For k odd, we construct an allocation based on the BC-useful sequence Z4. By the same argument as above, at least ⌊ k ⌋ of the sets Cℓ−(2n3−k)+2, . . . , Cℓ are acceptable for agents of type i1. Since k is odd, the total number of sets in the two sequences that are acceptable to agents of type i1 is at least k−1 = n3−1. We select some n3−1 of these sets and allocate them to n3−1 agents of type i1. The number of sets remaining in the two sequences is n3. We allocate them to agents of type i3. Moreover, the intervals Bk+1, . . . , Bℓ−(2n3−k)+1 and Bℓ+1 . . . , Bn contain at least(cid:4) ℓ−2n3+1 2 (cid:5)− n3 ≥(cid:4) n 2(cid:5)− n3 ≥ n1 − n3 + 1 sets that are acceptable to agents of type i1 (as before, we use the fact that (cid:4) n 2(cid:5) ≥ n1 + 1). We allocate some n1 − n3 + 1 of these sets to agents of type i1 that have not been allocated a set before. Finally, we allocate the remaining sets in these two sequences, there are n2 of them, to agents of type i2. This allocation gives rise to a 3 2 (cid:5) =(cid:4) ℓ+1 2 (cid:5) +(cid:4) n−ℓ (cid:5) +(cid:4) n−ℓ 2 + 2n3−k−1 4 -sufficient allocation for C. From now on, we will assume that no C-set is included in any B-set. Reasoning as in Case 3, we can prove that also no B-set is included in any C-set. Thus, in the remaining part of the proof we may assume that the splits B and C are proper relative to each other. Case 8. Some set Bi ∩ Cj is acceptable to agents of type i3. Since the splits B and C are proper relative to each other, we may relabel the sets if necessary so that for all i, Bi ⊆ Ci ∪ Ci+1, and that B1 ∩ C1 or Bn ∩ C1 is acceptable for agents of type i3. We 2 2 2 34 ZBIGNIEW LONC AND MIROSLAW TRUSZCZYNSKI 2 (cid:5) ≥ (cid:4) n will consider the first case, that is, that B1 ∩ C1 is acceptable to agents of type i3. As in several places before, the other case can be handled in a similar way. Suppose first that n1 > n2. Since the conditions formulated in Cases 1 and 6 are not satisfied, by Lemma 5.5 every other set in mms-splits B and C is acceptable to agents of type i1 (recall that B and A are proper relative to each other and so are C and A). Thus, there are at least n3 − 1 sets in the interval C2, C3, . . . , C2n3−1 that are acceptable to agents of type i1. We allocate these sets to those agents. We allocate the remaining n3 − 1 the sets in this sequence and the set B1 ∩ C1 to agents of type i3. Similarly, we note that the sequence B2n3, B2n3+1, . . . , Bn contains at least (cid:4) n−2n3+1 2(cid:5) − n3 ≥ n1 − n3 + 1 sets that are acceptable to agents of type i1. We allocate some n1 − n3 + 1 of such sets to those agents of type i1 that have not yet been allocated a set. The remaining sets in the sequence, there are n2 of them, are allocated to agents of type i2. All sets in the assignment are pairwise disjoint as B2n3 ⊆ C2n3 ∪ C2n3+1. Thus, the assignment we defined gives rise to a 3 In this case agents of type i3 receive B1 ∩ C1 and the sets C2, C3, . . . , Cn3. Agents of types i1 and i2 distribute among themselves the sets Bn3+1, Bn3+2, . . . , Bn. This is possible as the number of those sets is even, in fact, equal to 2n1 (= 2n2), and every other set of this sequence is acceptable for agents of type i1. Since Bn3+1 ⊆ Cn3+1 ∪ Cn3+2, these sets are disjoint with the sets assigned to agents of type i3 and, clearly, they are pairwise disjoint themselves. It follows that this allocation gives rise to a 3 Case 9. None of the conditions formulated in Cases 1-8 holds. We may assume that any pair of splits A, B, C are proper relative to each other. Moreover, relabeling the sets if necessary, we may also assume that Ai ⊆ Ci ∪ Ci+1 and Bi ⊆ Ci ∪ Ci+1. Thus, Ai and Bi overlap and, consequently, Ai ⊆ Bi ∪ Bi+1 or Ai ⊆ Bi−1 ∪ Bi. Both cases can be dealt with similarly, so we assume that Ai ⊆ Bi ∪ Bi+1. 4 -sufficient allocation for C. 4 -sufficient allocation for C. Thus, assume that n1 = n2. We note that for every i Ci ∪ Ci+1 ∪ Ci+2 = (Ai−1 ∩ Ci) ∪ Ai ∪ (Ai+1 ∩ Bi+1) ∪ (Bi+2 ∩ Ci+2). The value of Ci ∪ Ci+1 ∪ Ci+2 for agents of type i3 is 3. The conditions formulated in Cases 1, 4 and 8 do not hold, so the sets Ai−1 ∩ Ci, Ai+1 ∩ Bi+1 and Bi+2 ∩ Ci+2 are not acceptable for agents of type 3. Thus, the value of Ai for agents of type i3 is larger than 3 − 3 · 3 4 . As i was arbitrary, it follows that every A-set is acceptable for agents of type i3. We have n2 ≤ n1 < (cid:4) n 2(cid:5), where the latter inequality holds because of Case 5 being excluded. Since at least one in every two consecutive sets Ai, Ai+1 is acceptable to agents of type i2, we assign n2 of such sets to agents of type i2. All remaining A-sets are acceptable for agents of types i1 and i3. Thus, they can be allocated among them arbitrarily yielding a 3 (cid:3) 4 -satisfying allocation. 4 = 3 WARSAW UNIVERSITY OF TECHNOLOGY, POLAND E-mail address: [email protected] UNIVERSITY OF KENTUCKY, USA E-mail address: [email protected]
1610.09769	1	1610	2016-10-31T03:15:02	Meta-Path Guided Embedding for Similarity Search in Large-Scale Heterogeneous Information Networks	[ "cs.SI", "cs.LG" ]	Most real-world data can be modeled as heterogeneous information networks (HINs) consisting of vertices of multiple types and their relationships. Search for similar vertices of the same type in large HINs, such as bibliographic networks and business-review networks, is a fundamental problem with broad applications. Although similarity search in HINs has been studied previously, most existing approaches neither explore rich semantic information embedded in the network structures nor take user's preference as a guidance. In this paper, we re-examine similarity search in HINs and propose a novel embedding-based framework. It models vertices as low-dimensional vectors to explore network structure-embedded similarity. To accommodate user preferences at defining similarity semantics, our proposed framework, ESim, accepts user-defined meta-paths as guidance to learn vertex vectors in a user-preferred embedding space. Moreover, an efficient and parallel sampling-based optimization algorithm has been developed to learn embeddings in large-scale HINs. Extensive experiments on real-world large-scale HINs demonstrate a significant improvement on the effectiveness of ESim over several state-of-the-art algorithms as well as its scalability.	cs.SI	cs	Meta-Path Guided Embedding for Similarity Search in Large-Scale Heterogeneous Information Networks Jingbo Shang1, Meng Qu1, Jialu Liu2, Lance M. Kaplan3, Jiawei Han1, Jian Peng1 1Department of Computer Science, University of Illinois Urbana-Champaign, IL, USA 3Sensors & Electron Devices Directorate, Army Research Laboratory, Adelphi, MD, USA 1{shang7, mengqu2, hanj, jianpeng}@illinois.edu [email protected] [email protected] 2Google Research, New York City, NY, USA 6 1 0 2 t c O 1 3 ] I S . s c [ 1 v 9 6 7 9 0 . 0 1 6 1 : v i X r a ABSTRACT Most real-world data can be modeled as heterogeneous in- formation networks (HINs) consisting of vertices of multiple types and their relationships. Search for similar vertices of the same type in large HINs, such as bibliographic networks and business-review networks, is a fundamental problem with broad applications. Although similarity search in HINs has been studied previously, most existing approaches neither explore rich semantic information embedded in the network structures nor take user's preference as a guidance. In this paper, we re-examine similarity search in HINs and propose a novel embedding-based framework. It models ver- tices as low-dimensional vectors to explore network structure- embedded similarity. To accommodate user preferences at defining similarity semantics, our proposed framework, ESim, accepts user-defined meta-paths as guidance to learn ver- tex vectors in a user-preferred embedding space. Moreover, an efficient and parallel sampling-based optimization algo- rithm has been developed to learn embeddings in large-scale HINs. Extensive experiments on real-world large-scale HINs demonstrate a significant improvement on the effectiveness of ESim over several state-of-the-art algorithms as well as its scalability. 1. INTRODUCTION A heterogeneous information network (HIN) is a network that consists of multi-typed vertices connected via multi- typed edges. Modeling data in the real world as heteroge- neous information networks (HINs) can capture rich data semantics and facilitate various applications [26, 11, 8, 29, 23, 25]. Similarity search in HINs is a fundamental prob- lem for mining large HINs and much digital ink has been spilled over it in the community (e.g., [26, 25, 29, 28]). In this paper, we are particularly interested in utilizing HIN to conduct similarity search among the objects of the same type. For example, given a social media network in Yelp with connections between reviews, users and businesses, we can find similar restaurants (i.e., similarity search among ACM ISBN 978-1-4503-2138-9. DOI: 10.1145/1235 businesses), and recommend potential friends with similar preferences (i.e., similarity search among users). Naturally, people tend to employ different semantics in their network-based search even towards the same appli- cation. Take a bibliographic network as an example that consists of papers, authors, venues, and terms. To find simi- lar authors to a given author, some users may weigh more on shared technical terms in papers, while others more on shared publication venues. PathSim [26] proposes to use meta-paths to define and guide similarity search in HIN, with good success. A meta-path is represented by a path (i.e., a connected sequence of vertices) at the schema level (e.g., hauthor−paper−authori). However, PathSim has not explored similarities embedded deeply in the HIN structure. When hunting for similar authors, if the given meta-path is hauthor−paper−venue−paper−authori, PathSim can only build bridges for authors publishing papers in the same venues (e.g., "WSDM"), but cannot efficiently explore the semantic similarity between venues (e.g., "WSDM" and "WWW") to improve the search. However, such kind of semantic similarity can be easily implied by embedded se- mantic structure of the HIN. For an extreme example, if one author only publish in "WSDM" while the other only has publications in "WWW", their PathSim similarity will be 0. Although a bridge can be built between the two similar venues by traversing some long paths, it becomes much more costly and does not seem to be an elegant way compared with the embedding-based approach studied in this paper. Along another line of study, network-embedding techniques have been recently explored for homogeneous information networks, which treat all vertices and edges as of the same type, represented by LINE [29]. An alternative and better way is to first project the HIN to several bipartite (assum- ing user-given meta-paths are symmetric) or homogeneous networks and then apply the edge-wise HIN embedding tech- nique, such as PTE [28]. However, the network projection itself is count-based which does not preserve the underlying semantics, and the cost for computing such projected net- works is high. For example, taking a user-selected meta-path hauthor−paper−venue−paper−authori (i.e., shared-venues) to find similar authors, the projection will generate a ho- mogeneous network consisting of only authors, but it loses important network structure information (e.g., venues and papers) [25] and leads to a rather densely connected network (since many authors may publish in many venues). Clearly, direct modeling of the original heterogeneous information network will capture richer semantics than exploiting the projected networks. Acknowledging the deficiency of the above two types of approaches, we propose ESim, a novel embedding-based sim- ilarity search framework, with the following contributions. • A general embedding-based similarity search framework is proposed for HINs, where an HIN may contain undi- rected, directed, weighted, and un-weighted edges as well as various types of vertices; • Our framework incorporates a set of meta-paths as an input from a user to better capture the semantic meaning of user-preferred similarity; and • It handles large-scale HINs efficiently due to a novel sam- pling method and a parallel optimization framework. To the best of our knowledge, this is the first work that proposes a general meta-path guided embedding framework for similarity search in heterogeneous information networks. 2. RELATED WORK 2.1 Meta-Path Guided Similarity Search The concept of meta-path, which represents a connected sequence of vertices at the schema level, plays a crucial role in typed and structured search and mining in HINs. PathSim [26] defines the similarity between two vertices of the same type by the normalized count of path instances following a user-specified meta-path between any pair of ver- tices. [26] shows that the PathSim measure captures better peer similarity semantics than random walk-based similarity measures, such as P-PageRank [12] and SimRank [11]. More- over, [27] shows that user guidance can be transformed to a weighted combination of meta-paths. However, PathSim does not explore the similarity embedded in the structure of a HIN. Moreover, PathSim doesn't have the embedding vectors of vertices, which can make the further analysis more efficient, such as clustering. 2.2 Embedding-based Similarity Search Recently, embedding technique, which aims at learning low-dimensional vector representations for entities while pre- serving proximities, has received an increasing attention due to its great performance in many different types of tasks. As a special and concrete scenario, embedding of homoge- neous networks containing vertices of the same type has been studied recently. LINE [29] and DeepWalk [20] utilize the net- work link information to construct latent vectors for vertex classification and link prediction. DCA [4] starts from the personalized PageRank but does further decomposition to get better protein-protein interaction predictions in biology networks. However, these homogeneous models cannot cap- ture the information about entity types nor about relations across different typed entities in HINs. There are also embedding algorithms developed for HINs. For example, Chang et al. propose to incorporate deep neural networks to train embedding vectors for both text and images at the same time [3]. Under a supervised setting, PTE [28] utilizes labels of words and constructs bipartite HINs to learn predictive embedding vectors for words. Embedding techniques have been also applied to knowledge graphs to resolve question-answering tasks [6] and retain knowledge re- lations between entities [32]. However, these are all specially designed for specific types of networks and tasks and thus difficult to be extended to incorporate user guidance. The vector spaces constructed by different methods have different semantic meanings due to the statistics they emphasize. In many real-world scenarios, it is often difficult to find an Figure 1: A real small bibliography network for il- lustrations of the definitions. appropriate model, and the models have to be revised to fit into the desired usage. 2.3 Similarity Search in Vector Spaces Various similarity metrics have been designed for vector spaces in different tasks, such as cosine similarity [24], Jac- card coefficient [14], and the p-norm distance [5]. In our work, we directly utilize cosine similarity based on embed- ding vectors of vertices. Many efforts have been paid on optimizing the efficiency of top-k nearest neighbor search [33, 13, 9, 18, 2]. We adopt these existing efficient similarity search techniques to support online queries. 3. PRELIMINARIES In this section, a series of definitions and notations are presented. Definition 1. Heterogeneous Information Network is an information network where both vertices and edges have been associated with different types. In a heterogeneous information network G = (V, E, R), V is the set of typed vertices (i.e., each vertex has its own type), R is the set of edge types in the network, and E ⊂ V × V × R is the edge set. An edge in a heterogeneous information network is an ordered triplet e = hu, v, ri, where u and v are two typed vertices associated with this edge and r is the edge type. In a general heterogeneous information network, there might be multiple typed edges between the same two vertices and the edge could be either directed or undirected. The defi- nition above naturally supports all these cases. To better explain it, we use a bibliographic network as an example. Example 1. In the bibliographic network as shown in Fig. 1, the vertex set V consists of three types, {"author", "paper", "venue"}, the edge type set R contains "the author of the paper" and "the paper was published in the venue". The edge set E consists of concrete edges like hu = "Jon Kleinberg", v = "HITS", r = "the author of" i, where Jon Kleinberg is one of the authors of the paper HITS. For ease of representation, edge type r is denoted as A−P once there is only one edge type between author and paper. Another edge is hu = "HITS", v = "JACM", r = "be published in" i, which means the paper was published in the venue "JACM". In this case, r is denoted as P−V . Note that both edge types are undirected here. That is, A−P and P−A are actually referring to the same edge type. So as P−V and V −P . We define the concepts of meta-path and sub-meta-path by concatenating edge types in a sequence, as follows. Definition 2. In a HIN G = (V, E, R), a meta-path is a sequence of compatible edge types M = hr1, r2, . . . , rLi with length L, ∀ri ∈ R, and the outgoing vertex type of ri should match the incoming vertex type of ri+1. For any 1 ≤ s ≤ t ≤ L, we can induce a sub-meta-path Ms,t = hrs, rs+1, . . . , rti. Jon KleinbergGraphsover timeHITSSIGKDDJACMJure LeskovecThe author ofBe published inAuthor (A)Paper (P)Venue (V)Vertex TypesEdge TypesPossible Meta-pathsco-authorshipA-P-AA path instance following APAshared venueA-P-V-P-AParticularly, an edge type r can be viewed as a length-1 meta-path M = hri. A sequence of edges following M is called a path instance. Because there might be multiple edges between the same pair of vertices, instead of the vertex sequence, the edge sequence is used to describe a path instance. Formally speaking, Definition 3. Given a HIN G = (V, E, R) and a meta- path M = hr1, r2, . . . , rLi, any path Pe1(cid:32)eL = he1, e2, . . . , eLi connecting vertices u1 and uL+1 (i.e., vL), is a path in- stance following M, if and only if ∀1 ≤ i ≤ L, the i-th edge is type-ri, and ∀1 ≤ i ≤ L, vi = ui+1. We continue the example of the bibliographic network and list some meta-paths and path instances. Example 2. In the bibliographic network as shown in Fig. 1, a length-2 meta-path hA−P, P−Ai (abbrev. as A−P−A) expresses the co-authorship relation. The collaboration be- tween "Jure Leskovec" and "Jon Kleinberg" on the paper "Graphs over time" is a path instance of A−P−A. Similarly, a length-4 meta-path A−P−V −P−A captures the shared venues and any two authors published a certain paper in a same venue could be its path instance. Besides, P−V −P is a sub-meta-path of A−P−V −P−A. Definition 4. Meta-path Guided Similarity Search is a similarity search task on HINs, where the semantic meanings of the similarity are determined by n meta-paths {M1,M2, . . . ,Mn} specified by the user. Example 3. In the bibliographic network, to find similar authors, a user may choose two meta-paths A−P−V −P−A and A−P−A as guidance. In the subsequent discussion, we focus on the case of a single meta-path M. This is because (1) the principal ideas for exploring multiple weight-assigned meta-paths are essentially the same; (2) in our experiments, the performance gain of the optimal weighted combination of meta-paths is not significant. Moreover, we leave the study on derivation of a weighted combination of multiple meta-paths based on user's high-level guidance (i.e., user providing examples instead of path weights explicitly) to future work. Such a user-guided approach without adopting the embedding framework has been studied in [34, 27, 15]. 4. METHODOLOGY In this section, to incorporate meta-paths, we formulate a probabilistic embedding model inspired from many previous studies. We propose an efficient and scalable optimization algorithm relying on the sampling of path instances follow- ing the given meta-path. Our proposed efficient sampling methods are the most crucial steps and thus are discussed separately. In addition, we also provide thorough time com- plexity analysis. 4.1 Probabilistic Embedding Model Incorpo- rating Meta-Paths Model Formulation. The basic idea of our approach is to preserve the HIN structure information into the learned embeddings, such that vertices which co-occur in many path instances turn to have similar embeddings. To preserve the structure of a HIN G = (V, E, R), we first define the conditional probability of vertex v connected to vertex u by some path instances following the meta-path M as: P exp(f(u, v,M)) v0∈V exp(f(u, v0,M)) (1) P r(vu,M) = where function f is a scoring function modeling the rele- vance between u and v conditioned on the meta-path M. In particular, we encode the meta-path through the following formulation inspired from [19, 23]: f(u, v,M) = µM + pMT xu + qMT xv + xuT xv Here, µM ∈ R is the global bias of the meta-path M, pM and qM ∈ Rd are local bias vectors which are d dimensional. xu and xv ∈ Rd are d dimensional embedding vectors for vertices u and v respectively. Under such definition, if the embeddings of two vertices have a larger dot product, the two vertices are likely having a larger relevance score, and thus co-occuring in many path instances. Note that if users want a symmetric score function, ∀u, v, f(u, v,M) = f(v, u,M), we can restrict pM = qM. For a better understanding of the scoring function f, we can rewrite it as follows f(u, v,M) = (µM − pMT qM) + (xu + qM)T (xv + pM) where we can see that pM and qM shift xu and xv according to the semantic of the meta-path M while µM adjusts the score to an appropriate range. For a path instance Pe1(cid:32)eL = he1 = hu1, v1, r1i, e2 = hu2, v2, r2i, . . . , eL = huL, vL, rLii following the meta-path M = hr1, r2, . . . , rLi, we adopt the following approximation, by approximating the probability of the first vertex. P r(Pe1(cid:32)eLM) = P r(u1M) × P r(Pe1(cid:32)eLu1,M) ∝ C(u1, 1M)γ × P r(Pe1(cid:32)eLu1,M) (2) where C(u, iM) represents the number of path instances following M with the ith vertex being u. γ is a widely used parameter to control the effect of overly-popular vertices, which is usually 3/4 inspired from [16]. In Sec. 4.2.1, we will show an efficient dynamic programming algorithm to compute C(u, iM). The conditional probability, P r(Pe1(cid:32)eLu1,M), is now the last undefined term. The simplest definition is P r(vLu1,M), which assumes that the probability only depends on the two ends of the path instance and directly applies Eq. (1). How- ever, it omits the intermediate information and is equivalent to projection-based models. In this paper, we propose two possible solutions as follows, and later show "pairwise" is more effective than "sequential", since it exploits the meta-path guidance in a more thorough way. • Sequential (seq): In this setting, we assume that a ver- tex is highly relevant to its left/right neighbors in the k=1 P r(vkuk,Mk,k). • Pairwise (pair): In this setting, we assume all vertices in a path instance are highly relevant to each other, and thus the probability of the path instance is defined as t=s P r(vtus,Ms,t). As a s=1 result, vertices co-occur in many path instances turn to have large relevance scores. sequence: P r(Pe1(cid:32)eLu1,M) =QL QL P r(Pe1(cid:32)eLu1,M) =QL Noise-Contrastive Estimation (NCE). Given the condi- tional distribution defined in Eqs. (1) and (2), the maximum likelihood training is tractable but expensive because com- puting the gradient of log-likelihood takes time linear in the number of vertices. Since learning rich vertex embeddings does not require the accurate probability of each path instance, we adopt the NCE for optimization, which has been proved to significantly accelerate the training without cutting down the embedding qualities [7, 17]. The basic idea is to sample some observed path instances associated with some noisy ones, and it tries to maximize the probability of each observed instance while minimize the probability of each noisy one. Specifically, we reduce the problem of density estimation to a binary classification, discriminating between samples from path instances following the user selected meta-path and samples from a known noise distribution. In particular, we assume these samples come from the mixture. 1 K + 1 P r +(Pe1(cid:32)eLM) + K K + 1 P r −(Pe1(cid:32)eLM) where P r+(Pe1(cid:32)eLM) denotes the distribution of path in- stances in the HIN following the meta-path M. P r−(Pe1(cid:32)eLM) is a noise distribution, and for simplicity we set −(Pe1(cid:32)eLM) ∝ P r C(ui, iM)γ i=1 We further assume the noise samples are K times more frequent than positive path instance samples. The posterior probability that a given sample came from the path instance samples following the given meta-path is L+1Y P r+(Pe1(cid:32)eLM) P r+(Pe1(cid:32)eLM) + K · P r−(Pe1(cid:32)eLM) i h Since we would like to fit P r(Pe1(cid:32)eLM) to P r+(Pe1(cid:32)eLM), we simply maximize the following expectation. LM =E P r+ log + K E P r− h P r(Pe1(cid:32)eLM) P r(Pe1(cid:32)eLM) + KP r−(Pe1(cid:32)eLM) log KP r−(Pe1(cid:32)eLM) P r(Pe1(cid:32)eLM) + KP r−(Pe1(cid:32)eLM) i P r+ P r− i + K E LM,seq =E log σ(∆e1(cid:32)eLM) The loss function derived from NCE becomes Suppose we are using the sequential definition as Eq. (3). h h i=1 f(ui, vi, M)−log(cid:0)K·P r−(Pe1(cid:32)eLM)(cid:1) where ∆e1(cid:32)eLM =PL log(cid:0)1 − σ(∆e1(cid:32)eLM))(cid:1)i and σ(·) is the sigmoid function. Note that when deriving the above equation, we used exp(f(u, v,M)) in place of P r(vu,M), ignoring the normalization term in Eq. (1). We can do this because the NCE objective encourages the model to be approximately normalized and recovers a perfectly normalized model if the model class contains the data distri- bution [7, 17]. The above expectation is also studied in [16], i=1 f(ui, vi,M) for ease of computation and names the method negative sampling. We follow this idea and have the approximation as follows. which replaces ∆e1(cid:32)eLM withPL X LX log(cid:0)1 − σ(PL log σ( i=1 h LM,seq ≈ PK Pe1(cid:32)eL following M EPk e1(cid:32)eL ∼P r−u1,M k=1 f(ui, vi, Mi,i)) + i , Mi,i))(cid:1)i (3) i=1 f(uk i , vk Algorithm 1: ESim Training Framework Require: HIN G = (V, E, R), a user-specified meta-path M, sampling times t, and negative sampling ratio K Return: Vertex Embedding Vectors xu,∀u initialize parameters µ, p·, q·, x· while not converge do for i = 1 to t do p+ ← a sampled positive path instance following the meta-path M Optimize for a path instance p+ with label 1. s ← the first vertex on p+ for k = 1 to K do p− ← a sampled negative path instance following the meta-path M starting from s Optimize for a path instance p− with label 0. return x· The following loss function under the pairwise setting can be derived from NCE utilizing the same approximation. X LX LX log(cid:0)1 − σ(PL PL log σ( i=1 j=i h f(ui, vj , Mi,j)) + j , Mi,j))(cid:1)i i=1 j=i f(uk i , vk LM,pair ≈ PK Pe1(cid:32)eL following M EPk e1(cid:32)eL k=1 ∼P r−u1,M (4) Online Similarity Search. For any interested pairs of vertices u and v, their similarity is defined by the cosine similarity between xu and xv, i.e., sim(u, v) = xu T xv kxuk·kxvk. We choose the cosine similarity metric instead of the function f in Eq. (1) because the norm of vectors xu and xv do not help the similarity search task [22]. Moreover, cosine similarity is equivalent to Euclidean distance when xu = xv = 1, which makes the top-k similar vertices of the given vertex u able to be efficiently solved using approximate nearest neighbors search [18] after normalizations. 4.2 Optimization Algorithm The number of vertex pairs hu, vi that are connected by some path instances following at least one of user-specified meta-paths can be O(V 2) in the worst case, which is too large for storage or processing when V is at the level of millions and even billions, and thus makes it impossible to directly handle the projected networks over the meta-paths. Therefore, sampling a subset of path instances according to their distribution becomes the best and most feasible choice when optimizing, instead of going through every path in- stance per iteration. Thus even if the network itself contains a large number of edges, our method is still very efficient. The details of our training framework is shown in Algorithm 1. Once a path instance following the meta-path M has been sampled, the gradient decent method is used to up- date the parameters xu, xv, pM, qM, and µM one by one. As a result, our sampling-based training framework (Algo- rithm 1) becomes a stochastic gradient decent framework. The derivations of these gradients are easy and thus are omitted. Moreover, many prior studies have shown that the stochastic gradient descent can be parallelized without any locks. For example, Hogwild [21] provides a general and lock-free strategy for fully parallelizing any stochastic gradi- ent descent algorithms in a shared memory. We utilize this technique to speed up our optimization via multi-threads. Algorithm 2: Pre-computation of C(u, iM) Require: HIN G = (V, E, R) and meta-path M = hr1, r2, . . . , rLi Return: C(u, iM) /* initialization for each vertex u ∈ V do C(u, L + 1M) ← 1 C(u, L + 1M) ← 0 else if u is possibly as the second vertex in rL then /* dynamic programming for i ← L to 1 do for each vertex u ∈ V do for each type ri edge hu, v, rii do C(u, iM) ← 0 C(u, iM) ← C(u, iM) + C(v, i + 1R) Return: C(u, iM) Algorithm 3: Sample a positive path instance p+ Require: HIN G = (V, E, R), meta-path M = hr1, r2, . . . , rLi, C(u, iM), and weighting factor γ Return: a positive path instance following M u1 ← a random vertex ∝ C(u1, 1M)γ from V for i = 1 to L do Vi ← {vhui, vi ∈ Eri} /* γ is only applied at the first vertex when vi ← a random vertex ∝ C(ui, iM) from Vi if i < L then ui+1 ← vi sampling positive path instances. return Pe1(cid:32)eL = he1 = hu1, v1, r1i, e2 = hu2, v2, r2i, . . . , eL = huL, vL, rLii / / */ 4.2.1 Efficient Sampling Given a length-L meta-path M = hr1, r2, . . . , rLi, there might be O(V L) different path instances in total. It be- comes an obstacle for storing all the path instances while directly sampling over them takes a constant time. We pro- pose to run a dynamic programming algorithm computing auxiliary numbers so that the online sampling part can be done in a constant time. Pre-computation. As mentioned in Sec. 3, the probability of sampling a path instance following the meta-path M is only related to C(u, iM), which represents the count of path instances following the meta-path M with the ith vertex being u. First, we need to figure out the boundary case. When i = L + 1, for any vertex u, if it is possible to be the next vertex in an edge of rL (i.e., it could be vL), we have C(u, L + 1M) = 1. Otherwise, it should be 0. 1 ≤ i ≤ L for any vertex u. Then, we derive the following recursion formula when C(u, iM) = X C(v, i + 1M) (5) vhu,v,rii∈E An efficient way to do the summation in this formula is to traverse all type-ri edges starting from vertex u as shown in Algorithm 2. Its detailed time complex analysis will be presented later. Online Sampling. Based on the pre-computed C(u, iM), one can easily figure out an efficient online sampling method for path instances following the user-given meta-path M. The key idea is to sample the vertices on the path instance one by one. That is, the i-th vertex is conditioned on the previous i−1 vertices. As shown in Algorithm 3, the sampling pool for the i-th vertex is restricted to the adjacent vertices (via type-ri edges) of the previous (i − 1)-th vertex. However, things are a little different when dealing with the negative path instances. First, the negative path instances are associated with a positive path instance and thus the first vertex is fixed. Second, the remaining vertices on the negative path instances are independent. Therefore, they are all sampled from V based on ∝ C(u, iM)γ, ∀i > 1. 4.2.2 Weighted Combination Sometimes, due to the subtle semantic meanings of the (∀i, λi > 0 and Pn Pn i=1 λiLMi,seq and Lpair = Pn similarity, instead of a single meta-path, the weighted com- bination of n meta-paths could enhance the performance of similarity search. Suppose {λ1, λ2, . . . , λn} are the weights i=1 λi = 1), the unified loss function becomes the weighted sum over the loss functions of indi- vidual meta-paths based on the weights. That is, Lseq = i=1 λiLMi,pair. The Al- gorithm 1 can be modified accordingly by first sampling a meta-path M from ∀j, P r(Mj) = λj in the beginning of the "while" loop. The weighted combination of meta-paths can be either explicitly specified by users or learned from a set of simi- lar/dissimilar examples provided by users. Such user-guided meta-path generation has been studied in [27] without con- sidering embedding. Because weight learning is beyond the scope of this paper, we leave such an extension to embedding- based similarity search to a future work, and adopt grid searches to obtain the best weighted combination of meta- paths in Sec. 5.3 assuming we have the groundtruth. 4.2.3 Complexity Analysis Pre-computation. For a given length-L meta-path M, we have u ∈ V and 1 ≤ i ≤ L + 1, which means the memory complexity is O(V L). For each given i, we only have to consider all type-ri edges starting from different vertices, which implies the time complexity is O((V + E)L). Online sampling. We have adopted the alias method [31] to make the online sampling from any discrete distribution O(1). Therefore, we have to precompute all discrete distri- butions and restore them in the memory. Given a length-L mete-path M, for the negative sampling, we have O(L) dif- ferent distributions and each of them is over V discrete values; for the positive sampling there are O(V L) different distributions but the number of variables depends on the number of edges of the certain type. The total discrete val- ues they have is O(EL). In summary, both the time and memory complexities of the preparation of the alias method are O((V + E)L), while the time complexity of sampling any path instance becomes O(L). Optimization for a path instance. For a specific path instance Pe1(cid:32)eL, the time complexity is O(dL), and O(dL2) for different loss functions Lseq and Lpair respectively. Overall. In our framework, for Algorithm 1, there are O(tK) times of path optimization per iteration. Suppose there are T iterations before convergence, considering the choices of different loss functions, Lseq and Lpair, the over- Table 1: Dataset Statistics. Whole networks in DBLP and Yelp are utilized in training, while group- ing labels are only used for evaluation. DBLP Yelp Dataset Node Types Edge Types # of Vertices # of Edges paper(P), author(A), term(T), venue(V) P−A, P−T , P−V 2,762,595 103,059,616 review(R), name(N), business(B), word(W) B−N, R−B, R−W 1,616,341 76,708,201 B−N−B B−R−W−R−B, Interesting Meta-paths A−P−A, A−P−V −P−A, A−P−T−P−A Two Grouping Labels Research Area/Group Business/Restaurant Type all time complexity is O(T tKLd) and O(T tKL2d) respec- tively. In addition, considering the complicated case of n user-specified meta-paths, O(n(V +E)L) has to be paid for pre-computations before sampling, where L is the maximum length of given meta-paths. Parallelization. It has been proved that stochastic gradient descent can be fully parallelized without locks [21]. If we use k threads, although the pre-computations remain the same, the time complexity of training can be k times faster. 5. EXPERIMENTS In this section, we evaluate our proposed model ESim, comparing with several state-of-the-art methods on two real- world large-scale HINs both quantitatively and qualitatively. Also, we evaluate the sensitivity of parameters and the effi- ciency of ESim. 5.1 Datasets Table 1 shows the statistics of two real-world large-scale HINs: DBLP and Yelp. The first dataset, DBLP, is a bibliographic network in computer science, which includes papers (P), authors (A), venues (V), and terms (T) as four types of vertices and takes P as the center in a star net- work schema. There are 3 types of undirected edges: P−A, P−V , and P−T . The interesting meta-paths that may be specified by users are the co-authorship meta-path A−P−A, the shared venue meta-path A−P−V −P−A, and the shared term meta-path A−P−T−P−A. We have the following two groupings labeled by human experts. • Research Area. We use the 4-area grouping in DBLP labeled by human experts, which was used when eval- uating PathSim [26]. There are 3,750 authors from 4 diverse research domains of computer science: "data min- ing", "database", "machine learning" and "information retrieval". • Research Group. This grouping in DBLP is also la- beled by human experts and is more fine-grained compar- ing to the Research Area grouping. There are 103 authors from 4 research groups: "Christos Faloutsos", "Jiawei Han", "Michael I. Jordan", and "Dan Roth". The second dataset, Yelp, is a social media network of Yelp, released in Yelp Dataset Challenge1. This network includes businesses (B), words in business names (N), reviews of businesses (R), and words in reviews (W) as vertices. There are 3 different types of undirected edges: B−N, B−R, and R−W . The interesting meta-paths that may be specified by users are the shared review word meta-path B−R−W−R−B and the shared name word meta-path B−N−B. We have the following two groupings extracted from the meta-data provided in Yelp dataset. 1https://www.yelp.com/academic_dataset • Business Type. There are various business types in the Yelp dataset, such as "restaurants", "hotels", "shopping", and "health and medical". Businesses with multiple cate- gories have been discarded to avoid ambiguity. To keep the results from being biased by some dominating types, we randomly sample 881 businesses from each of these four types as labeled data, because the 4-th popular type contains that many businesses. • Restaurant Type. Since the majority of businesses in the Yelp dataset are restaurants, we look deep into them by dividing them into different types of cuisines. More specifically, we have sampled 270 restaurants from 5 cuisines respectively: "Chinese", "Japanese", "Indian", "Mexican", and "Greek". As a result, there are in total 1350 labeled restaurants in our labeled dataset. 5.2 Experimental Setting Meta-path. We select different meta-paths for different datasets to see how the meta-paths will reflect the user- preferred similarity and affect the performance. In addition, we run grid searches against different groupings to obtain the best weights of different meta-paths for ESim models. Compared Algorithms and Notations. We select the previous state-of-the-art algorithm in the meta-path guided similarity search problem, PathSim, which has been reported to beat many other simiarity search methods, for example, SimRank [11], P-PageRank [10], random walk, and pairwise random walk. In addition, we also consider (heterogeneous) network embedding methods, such as LINE and PTE, which beat other embedding methods like graph factorization [1] and DeepWalk [20]. More details about these methods are as follows. • PathSim [26] is a meta-path guided similarity search al- gorithm which utilizes the normalized count of path in- stances following the user selected meta-path between any pair of vertices. When the meta-path involves text (e.g., A−P−T−P−A), PathSim becomes a text-based similarity - the cosine similarity using bag-of-words. • LINE [29] is an embedding algorithm specifically designed for homogeneous networks, which considers both first and second order information in a network (i.e., the neighbors and the neighbors of the neighbors). By treating all vertices and edges in the HIN as homogeneous ones, we can directly apply LINE and denote the model with first order only as LINE-1st and the model using the second order information as LINE-2nd respectively. One can also project the HIN to a weighted homogeneous network based on the user selected meta-path and apply LINE. However, based on our experiments, the results are always worse than PTE and thus omitted. • PTE [28] decomposes a HIN to a set of edgewise bipartite networks and then learn embedding vectors. To adapt this method to our settings, the way with the best perfor- mance we discovered is to project the HIN to a weighted bipartite HIN based on the user selected meta-path. For example, if the selected meta-path is the shared venue meta-path A−P−V −P−A in the bibliographic network, we construct a bipartite HIN consisting of A and V , where the weight of edges between any pair of a type-A vertex and a type-V vertex equals to the numbers of path in- stances following A−P−V between them. ESim refers to our proposed meta-path guided embedding model. Considering the choice of loss functions LM,seq and Table 2: AUC Evaluation on DBLP dataset. Meta-path Research Group Research Area Table 3: AUC Evaluation on Yelp dataset. Meta-path Business Type Restaurant Type Model LINE-1st LINE-2nd PathSim PTE ESim-seq ESim-pair PathSim PTE ESim-seq ESim-pair PathSim PTE ESim-seq ESim-pair ESim-pair N/A A−P−A A−P−V −P−A A−P−T−P−A Best Weighted Combination (grid search) 52.32% 51.82% 52.07% 50.90% 52.97% 52.87% 80.51% 74.40% 77.06% 83.58% 55.22% 68.38% 61.17% 69.18% 83.81% 52.89% 51.53% 76.75% 77.07% 73.97% 81.46% 72.60% 65.87% 74.83% 73.07% 61.16% 73.28% 65.73% 74.96% 82.27% N/A B−R−W−R−B B−N−B Model LINE-1st LINE-2nd PathSim PTE ESim-seq ESim-pair PathSim PTE ESim-seq ESim-pair 78.89% 80.20% 83.49% 85.67% 84.39% 89.22% 53.77% 61.69% 62.53% 59.61% 82.39% 70.20% 74.66% 83.77% 78.62% 88.73% 55.26% 63.18% 61.27% 59.39% collaborations are more important in Research Groups grouping, and thus A−P−A works the best. In the Yelp dataset, although business names carry some semantics, words in review are more telling. In both groupings, B−R−W−R−B always better fits the user-preferred simi- larity. These phenomena imply that a better fit meta-path will lead models such as ESim to better performance. • ESim-pair performs better than ESim-seq, because the pairwise loss function exploits the meta-path guidance in a more thorough way. We will focus on ESim-pair in later experiments. • User-guided models perform better. As long as the selected meta-path is reasonable (e.g., A−P−V −P−A in the Re- search Area grouping , A−P−A in the Research Group grouping, and B−R−W−R−B in the Yelp dataset), ESim, PathSim, and PTE always perform better than LINE, which proves the importance of following the guidance from user-selected meta-paths. • ESim-pair performs the best with significant advantages over PathSim and PTE, which demonstrates the power of embedding techniques and proper usage of network structures respectively. Meta-path Combination. We choose the best perform- ing meta-paths A−P−A and A−P−V −P−A in the DBLP dataset and run grid searches for best weights to achieve high- est AUC scores in Research Area and Research Group group- ings respectively. Because the best performing meta-path in Yelp dataset is always B−R−W−R−B, any combination is useless in this case. Note that this grid search against grouping labels shows an upper bound of the highest possible AUC score, which can be rarely achieved without knowing la- bels. As shown in Table 2, the improvement of best weighted meta-paths is marginal. Therefore, weighted combination might be necessary to achieve the best performance but the choice of meta-paths is more important. 5.4 Visualizations With embedding vector for each vertex, we can show mean- ingful visualizations, which layout the vertices of the same type in a two-dimensional space, and check whether the boundaries between different groups are relatively clear. Taking the DBLP dataset as an example, we visualize the vertices with different colors regarding to their group labels in the Research Area grouping and the Research Group group- ing. Their embedding vectors are projected to a 2-D space using the t-SNE package [30], which is a nonlinear dimen- sionality reduction technique and well suited for projecting high-dimensional data into a low dimensional space. Laying out these vertex embedding vectors is challenging, especially for the vertices in the four closely related research areas: "data mining", "database", "machine learning" and "information retrieval". For different embedding-based meth- 0.1 A−P−A + 0.9 A−P−A + 0.9 A−P−V −P−A 0.1 A−P−V −P−A LM,pair, the corresponding model are denoted as ESim-seq and ESim-pair respectively. Default Parameters. The parameter γ controlling the effect of overly-popular vertices is set to 3/4 inspired from [16]. The dimension of the vertex embedding vectors, d, is set to 50. The negative sampling ratio K is set to 5, whereas the sampling times t is set to 1 million by default. The number of working cores is set to 16. Talking about the initialization of global bias, local bias vectors, and embedding vectors, we assign all parameters as a random real value uniformly in [−1, 1]. The learning rate in stochastic gradient descent is initialized as 0.25 and later linearly decreased. Machine. The following experiments on execution time were all conducted on a machine equipped two Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz with 20 physical cores in total. Our framework is fully implemented in C++2. 5.3 AUC Evaluations Although it is hard to obtain the labels of the detailed rankings among all pairs of vertices, it is relatively easy to give an estimation based on the labels of the vertex groupings l(·). Considering the ranking problem for each individual vertex u, if we rank the other vertices based on the similarity scores, it is very natural that we expect the vertices from the same group (similar ones) are at the top of the ranking list whereas the dissimilar ones are at the bottom of the list. More specifically, we define the AUC score as follows. For a better similarity metric, the AUC score should be larger. AU C = 1 V v,v0∈V ∧l(u)=l(v)∧l(u)6=l(v0) 1 sim(u,v)>sim(u,v0) v,v0∈V ∧l(u)=l(v)∧l(u)6=l(v0) 1 Note that the models that generate similarity measures are learned from the whole dataset, whereas the AUC metric is calculated only in the subset of vertices where we have group labels. The subset is usually small because computing AUC needs pairwise similarities among the subset. We have the following observations from Tables 2 and 3. First, about single meta-path we have: • User-guidance is crucial. The choice of the user-selected meta-path is really important and affects the perfor- mance of user-guided models significantly. For exam- ple, A−P−V −P−A works the best in the Research Area grouping, where the shared venue is more telling. However, 2The source code will be published in the author's GitHub after acceptance. P X u∈V P (a) Varying d. Figure 2: Parameter Sensitivity of ESim-pair. (b) Varying K. (c) Varying total samples. (a) Varying network sizes. (b) Varying cores. Figure 3: Efficiency of ESim-pair. (a) ESim-pair (b) PTE (c) LINE-1st Figure 4: Visualization of embedding vectors of 10% random sampled authors in DBLP Research Area grouping when M=A−P−V −P−A. Colors cor- respond to research areas: Database, Data Mining, Information Retrieval, and Machine Learning. (a) ESim-pair (b) PTE (c) LINE-1st Figure 5: Visualization of embedding vectors of all authors in DBLP Research Group grouping when M=A−P−A. Colors correspond to research groups: Christos Faloutsos, Jiawei Han, Michael I. Jordan, and Dan Roth. ods (i.e., ESim, PTE, and LINE), we choose to visualize their variants holding the best AUC performance, i.e., ESim- pair using A−P−V −P−A, PTE using A−P−V −P−A, and LINE-1st. As shown in Fig. 4, we visualize 10% random samples of 3, 750 authors from 4 research domains of com- puter science. Better embedding vectors should lead to a clearer figure where the boundaries between different col- ored points should be clean and almost not interfering with each other. Based on the visualizations, one can easily ob- serve that ESim-pair using A−P−V −P−A provides the best embedding vectors judged from this criterion. Similarly, the visualizations of the Research Groups group- ing based on the models with best AUC performance are shown in Fig. 5. Our proposed model ESim-pair using A−P−A clearly beats PTE using A−P−A and LINE-1st. The significant improvements over PTE and LINE observed via visualization are consistent with the previous evaluations. 5.5 Parameter Sensitivity We select the best performing models ESim-pair to study the parameter sensitivity, such as using A−P−V −P−A in the Research Area grouping, using A−P−A in the Research Group grouping, and using B−R−W−R−B in both Business Type and Restaurant Type groupings. We vary the parameter values and see how the AUC performance changes. Based on the curves in Fig. 2(a), we can observe that setting the dimension (d) of vertex embedding vectors as 50 is reasonable, because too small d cannot sufficiently capture the semantics, while too large d may lead to some overfitting problem. Fig. 2(b) indicates that the AUC scores are not sensitive to the negative sample ratio K and K = 5 is a good choice. As shown in Fig. 2(c), as more samples are optimized during training, the AUC scores keep an increasing trend and finally converge. 5.6 Scalability and Parallelization We investigate the efficiency of ESim by considering both the scalability and the parallelization as shown in Fig. 3(a). We try different portions of network sizes (i.e., V + E) in the two networks and run our best performing models, i.e., ESim-pair using A−P−V −P−A and using A−P−A on the DBLP dataset, as well as ESim-pair using B−R−W−R−B on the Yelp dataset. Based on these curves, the running time is linear to the size of networks while the longer meta- path costs a little more time, which are consistent with our previous theoretical time complexity analysis. We vary the number of working cores and run our models on the DBLP and Yelp datasets. The results are plotted in Fig. 3(b). The speedup is quite close to linear, which shows that ESim is quite scalable to the number of working cores. 6. CONCLUSIONS In this paper, we propose a general embedding-based simi- larity search framework for heterogeneous information net- works (HINs). Our proposed model, ESim, incorporates given meta-paths and network structures to learn vertex embedding vectors. The similarity defined by the cosine similarity be- tween vertex embeddings of the same type has demonstrated its effectiveness, outperforming the previous state-of-the-art algorithms on two real-world large-scale HINs. The efficiency of ESim has also been evaluated and proved to be scalable. There are several directions to further extend this work. First, instead of similarities between vertices of the same type, one can also explore the relevances between vertices of different types. Second, a mechanism could be developed to automatically learn and extract a set of interesting meta- paths or their weighted combinations from user-provided rankings or preferences. Third, similarity is the fundamental operation for mining and exploring HINs. This study on similarity measure, defined in HINs based on meta-path guided embedding, and its efficient computations will impact other searching and mining problems in HINs. For example, it is necessary to re-examine clustering, classification and prediction functions in HINs by reconsidering the similarity measures defined based on meta-path guided embedding. Also, mining outliers in networks can be formulated as finding a small subset of vertices with extremely low similarities to other vertices or clusters. lllll60657075808590DimensionAUC122550100200lResearch Area A−P−V−P−AResearch Group A−P−ABusiness Type B−R−W−R−BRestaurant Type B−R−W−R−Blllll7075808590Negative RatioAUC13579lResearch Area A−P−V−P−AResearch Group A−P−ABusiness Type B−R−W−R−BRestaurant Type B−R−W−R−Bllllllllll2468107075808590Samples (million)AUClResearch Area A−P−V−P−AResearch Group A−P−ABusiness Type B−R−W−R−BRestaurant Type B−R−W−R−Bllllllllll0.20.40.60.81.00100030005000PortionRunning Time (s)lDBLP A−P−V−P−ADBLP A−P−AYelp B−R−W−R−Blllll51015051015# of coresSpeedup RatiolDBLP A−P−V−P−ADBLP A−P−AYelp B−R−W−R−B7. REFERENCES [1] A. Ahmed, N. Shervashidze, S. Narayanamurthy, V. Josifovski, and A. J. Smola. Distributed large-scale natural graph factorization. In WWW, pages 37–48, Republic and Canton of Geneva, Switzerland, 2013. International World Wide Web Conferences Steering Committee. [2] A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions. volume 51, pages 117–122, New York, NY, USA, Jan. 2008. ACM. [3] S. Chang, W. Han, J. Tang, G.-J. Qi, C. C. Aggarwal, and T. S. Huang. Heterogeneous network embedding via deep architectures. In SIGKDD, pages 119–128, New York, NY, USA, 2015. ACM. [4] H. Cho, B. Berger, and J. Peng. Diffusion component analysis: Unraveling functional topology in biological networks. In Research in Computational Molecular Biology, pages 62–64. Springer, 2015. [5] P. L. Duren. Theory of Hp spaces, volume 38. IMA, 1970. [6] K. Gu, J. Miller, and P. Liang. Traversing knowledge graphs in vector space. In EMNLP 2015, 2015. [7] M. U. Gutmann and A. Hyvärinen. Noise-contrastive estimation of unnormalized statistical models, with applications to natural image statistics. JMLR, 13(1):307–361, 2012. [8] D. Hallac, J. Leskovec, and S. Boyd. Network lasso: Clustering and optimization in large graphs. In SIGKDD, pages 387–396, New York, NY, USA, 2015. ACM. [9] P. Indyk and R. Motwani. Approximate nearest neighbors: Towards removing the curse of dimensionality. In STOC, pages 604–613, New York, NY, USA, 1998. ACM. [10] G. Iván and V. Grolmusz. When the web meets the cell: using personalized pagerank for analyzing protein interaction networks. Bioinformatics, 27(3):405–407, 2011. [11] G. Jeh and J. Widom. Simrank: A measure of structural-context similarity. In SIGKDD, pages 538–543, New York, NY, USA, 2002. ACM. [12] G. Jeh and J. Widom. Scaling personalized web search. In WWW, pages 271–279, New York, NY, USA, 2003. ACM. [13] N. Katayama and S. Satoh. The sr-tree: An index structure for high-dimensional nearest neighbor queries. In SIGMOD, pages 369–380, New York, NY, USA, 1997. ACM. [14] M. Levandowsky and D. Winter. Distance between sets. Nature, 234(5323):34–35, 1971. [15] C. Meng, R. Cheng, S. Maniu, P. Senellart, and W. Zhang. Discovering meta-paths in large heterogeneous information networks. In WWW, pages 754–764. International World Wide Web Conferences Steering Committee, 2015. [16] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean. Distributed representations of words and phrases and their compositionality. In NIPS, pages 3111–3119, 2013. [17] A. Mnih and Y. W. Teh. A fast and simple algorithm for training neural probabilistic language models. In ICML, 2012. [18] M. Muja and D. G. Lowe. Fast approximate nearest neighbors with automatic algorithm configuration. VISAPP (1), 2, 2009. [19] J. Pennington, R. Socher, and C. D. Manning. Glove: Global vectors for word representation. In EMNLP, volume 14, pages 1532–1543, 2014. [20] B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk: Online learning of social representations. In SIGKDD, pages 701–710, New York, NY, USA, 2014. ACM. [21] B. Recht, C. Re, S. Wright, and F. Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In NIPS, pages 693–701, 2011. [22] A. M. Schakel and B. J. Wilson. Measuring word significance using distributed representations of words. arXiv preprint arXiv:1508.02297, 2015. [23] J. Shang, T. Chen, H. Li, Z. Lu, and Y. Yu. A parallel and efficient algorithm for learning to match. In ICDM, pages 971–976. IEEE, 2014. [24] A. Singhal. Modern information retrieval: A brief overview. IEEE Data Eng. Bull., 24(4):35–43, 2001. [25] Y. Sun and J. Han. Mining heterogeneous information networks: A structural analysis approach. SIGKDD Explor. Newsl., 14(2):20–28, Apr. 2013. [26] Y. Sun, J. Han, X. Yan, P. S. Yu, and T. Wu. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. VLDB 2011, 2011. [27] Y. Sun, B. Norick, J. Han, X. Yan, P. S. Yu, and X. Yu. Integrating meta-path selection with user-guided object clustering in heterogeneous information networks. In SIGKDD, pages 1348–1356, New York, NY, USA, 2012. ACM. [28] J. Tang, M. Qu, and Q. Mei. Pte: Predictive text embedding through large-scale heterogeneous text networks. In SIGKDD, pages 1165–1174. ACM, 2015. [29] J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei. Line: Large-scale information network embedding. In WWW, pages 1067–1077, Republic and Canton of Geneva, Switzerland, 2015. International World Wide Web Conferences Steering Committee. [30] L. Van der Maaten and G. Hinton. Visualizing data using t-sne. JMLR, 9(2579-2605):85, 2008. [31] A. J. Walker. An efficient method for generating discrete random variables with general distributions. ACM Trans. Math. Softw., 3(3):253–256, Sept. 1977. [32] R. Xie, Z. Liu, J. Jia, H. Luan, and M. Sun. Representation learning of knowledge graphs with entity descriptions. In AAAI, 2016. [33] P. N. Yianilos. Data structures and algorithms for nearest neighbor search in general metric spaces. In SODA, pages 311–321, Philadelphia, PA, USA, 1993. Society for Industrial and Applied Mathematics. [34] X. Yu, Y. Sun, B. Norick, T. Mao, and J. Han. User guided entity similarity search using meta-path selection in heterogeneous information networks. In CIKM, pages 2025–2029, New York, NY, USA, 2012. ACM.
1806.02201	2	1806	2018-06-25T20:50:07	Scholarly Twitter metrics	[ "cs.SI" ]	Twitter has arguably been the most popular among the data sources that form the basis of so-called altmetrics. Tweets to scholarly documents have been heralded as both early indicators of citations as well as measures of societal impact. This chapter provides an overview of Twitter activity as the basis for scholarly metrics from a critical point of view and equally describes the potential and limitations of scholarly Twitter metrics. By reviewing the literature on Twitter in scholarly communication and analyzing 24 million tweets linking to scholarly documents, it aims to provide a basic understanding of what tweets can and cannot measure in the context of research evaluation. Going beyond the limited explanatory power of low correlations between tweets and citations, this chapter considers what types of scholarly documents are popular on Twitter, and how, when and by whom they are diffused in order to understand what tweets to scholarly documents measure. Although this chapter is not able to solve the problems associated with the creation of meaningful metrics from social media, it highlights particular issues and aims to provide the basis for advanced scholarly Twitter metrics.	cs.SI	cs	Handbook of Quantitative Science and Technology Research 1 Scholarly Twitter metrics Stefanie Haustein [email protected] 1 School for Information Studies, University of Ottawa, Ottawa, (Canada) 2 Centre interuniversitaire de recherche sur la science et la technologie (CIRST), Université du Québec à Montréal, Montréal (Canada) Abstract Twitter has arguably been the most popular among the data sources that form the basis of so-called altmetrics. Tweets to scholarly documents have been heralded as both early indicators of citations as well as measures of societal impact. This chapter provides an overview of Twitter activity as the basis for scholarly metrics from a critical point of view and equally describes the potential and limitations of scholarly Twitter metrics. By reviewing the literature on Twitter in scholarly communication and analyzing 24 million tweets linking to scholarly documents, it aims to provide a basic understanding of what tweets can and cannot measure in the context of research evaluation. Going beyond the limited explanatory power of low correlations between tweets and citations, this chapter considers what types of scholarly documents are popular on Twitter, and how, when and by whom they are diffused in order to understand what tweets to scholarly documents measure. Although this chapter is not able to solve the problems associated with the creation of meaningful metrics from social media, it highlights particular issues and aims to provide the basis for advanced scholarly Twitter metrics. 1.1 Introduction Enabled by the digital revolution, the open access and open science movement, big data and the success of social media have shaken up the scholarly metrics landscape. Academic careers are no longer shaped only by peer-reviewed papers, citation impact and impact factors, university managers and funders now also want to know how researchers perform on social media and how much their work has impacted society at large. Bibliometricians have started to adapt to the policy pull and technology push and expanded their repertoire of scholarly metrics to capture output and impact beyond the ivory tower, so far that some speak about a scientometric revolution (Bornmann, 2016). Metrics are no longer restricted to formal parts of communication but expand beyond the borders of the scholarly community (Cronin, 2013a). Similarly to how the Science Citation Index formed the field of bibliometric research and research evaluation, the altmetrics, or more precisely the social-media metrics landscape is being heavily shaped―if not entirely driven―by the availability of data, in particularly via Automated Programming Interfaces (APIs) (Haustein, 2016; Priem, 2014). Twitter has arguably been at the epicenter of the earthquake that has shaken up the scholarly metrics landscape. The majority of altmetrics research has either focused on or included Twitter (see Sugimoto, Work, Larivière, & Haustein (2017) for a review of the literature). Following a general definition of scholarly metrics which include activity on social media (Haustein, 2016), scholarly Twitter metrics are defined as "indicators based on recorded events of acts [on Twitter] related to scholarly documents […] or scholarly agents […]" (Haustein, 2016, p. 416). Although findings of an early study had suggested that tweets were a good early indicator of citations for papers published in the Journal of Medical Internet Research (Eysenbach, 2011), the generalizability of this claim was refuted by low correlations reported by more representative studies (Costas, Zahedi, & Wouters, 2015a; Haustein, Peters, Sugimoto, Thelwall, & Larivière, 2014). Low 1 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research correlations between tweets and citations did, however, spark hopes that Twitter activity was able to reflect impact on users and use beyond citing authorsa new type of previously unmeasurable impact, possibly on society at large. Twitter's popularity in the altmetrics realm has essentially been caused by two factors, which are both heavily influenced by technology and the data push and policy pull described above: a) a significant number of scholarly articles are shared on Twitter, producing a measurable signal; b) Twitter is a social media platform created for and used by a wide and general user base, which theoretically has the potential to measure impact on society at large. As tweets represent an "unprecedented opportunity to study human communication and social networks" (G. Miller, 2011, p. 1814), Twitter is being used to analyze a variety of social phenomena. Centering on either the message (i.e., the tweet, its content and associated metadata) or social connections (i.e., the network of follower-followee relations), tweets have been used to show discussions during upcoming elections, how people communicate during natural disasters, political upheaval, cultural events and conferences and have even been used to predict elections outcomes and the stock market (Rogers, 2014; Weller, Bruns, Burgess, Mahrt, & Puschmann, 2014). While only a small share of academics use Twitter for scholarly communication (Rowlands, Nicholas, Russell, Canty, & Watkinson, 2011; Van Noorden, 2014) and to diffuse scientific publications (Priem & Costello, 2010), more than one fifth of recent journal articles are being tweeted (Haustein, Costas, & Larivière, 2015), which adumbrates that it is non-academics who engage with scholarly publications on Twitter. At this point, social media-based indicators have flourished rather as vanity measuresculminating in a tongue-in-cheek metric called the Kardashian Index (Hall, 2014)than as validated indicators of societal impact. Even though altmetrics have left their mark on the scholarly publishing and metrics landscape (Desrochers et al., 2018), they have not (yet) established themselves within the reward system of science, where citations remain the only hard currency: Neither Twitter mentions nor Facebook 'likes' are, for now at any rate, accepted currencies in the academic marketplace; you are not going to get promoted for having been liked a lot, though it may well boost your ego. (Cronin, 2013b, p. 1523) Still, almost all big publishing houses now report some form of article-level metric based on social media activity, including tweets. Despite the lack of validation and a clear definition regarding the type of impact measured, the number of tweets are thus already used as scholarly metrics "in the wild" (Priem, Piwowar, & Hemminger, 2012). This chapter aims to contribute to the understanding of Twitter and Twitter-based metrics with a particular focus on their potential and limitations when applied as scholarly metrics. To provide some context for the meaning of scholarly impact measures derived from tweets and Twitter activity, this chapter describes Twitter's role in scholarly communication. It depicts how Twitter is used in academia and how scholarly contents are diffused and discussed in tweets. The chapter provides an overview of the literature of Twitter use by the scholarly community and scholarly output on Twitter. The latter part is supported by empirical results based on an analysis of 24 million tweets mentioning scientific papers captured by the data provider Altmetric.com. Both the review of the relevant literature and the patterns extracted from the Twitter data are intended to contribute to the understanding of what type of scholarly contents are diffused on Twitter, who is diffusing them, when and how. This will help to assess Twitter metrics as valid impact indicators and to interpret their meaning. 2 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research 1.2 Twitter in scholarly communication Twitter launched in 2006 as a public instant messaging service and evolved from an urban lifestyle social network, where users would update their friends about what they were doing, to a platform for communicating news and events used by 500 million users worldwide, or 23% of US adults online (Duggan, Ellison, Lampe, Lenhart, & Madden, 2015). Although other microblogging platforms (e.g., Sina Weibo, tumblr, Plurk) exist, tweeting has become a synonym (and preferred term) for microblogging and Twitter the most popular service. Twitter constraints microposts to a maximum length of 140 characters, a restriction that originates from the 160 character limit of text messages. Users can follow each other and create user lists to manage the updates they receive from other Twitter users. Similarly to regular blogs, microblogs are ordered sequentially in reverse chronological order and, due to their brevity, usually appear more frequently (Puschmann, 2014), making Twitter the "most dynamic and concise form of information exchange on social media" (Grajales, Sheps, Ho, Novak- Lauscher, & Eysenbach, 2014, p. 5). While the brevity of tweets is seen as a restriction by some, others perceive it as a particular advantage: The brevity of messages allows [tweets] to be produced, consumed, and shared without a significant amount of effort, allowing a fastpaced conversational environment to emerge. (boyd, Golder, & Lotan, 2010, p. 10) Tweets have three major specific affordances, which facilitate communication on the platform: retweets (RTs), user mentions (@mentions) and hashtags (keywords following #). All of these functions originated within the Twitter user base and were eventually adopted by Twitter, representing a co-creation of functions by users and developers. Twitter provides three main levels of communication: interpersonal communication on the micro level, meso-level exchanges of people who are directly connected through their network of followers and followees, and hashtag-centered macro- level communication which enables exchanges among all Twitter users with common interests (Bruns & Moe, 2014). In academia Twitter is used to disseminate and discuss scholarly outputs and other relevant information; maintain collaborations or find new ones; as a virtual "water cooler" (Veletsianos, 2012, p. 347) for social networking with colleagues; to increase student participation in teaching; as a backchannel at scientific conferences to foster discussions among conference attendees and those who participate remotely; as well as to increase visibility and reach wider audiences (Nentwich, 2011; Osterrieder, 2013; Pearce, Weller, Scanlon, & Ashleigh, 2010; Van Noorden, 2014; Veletsianos, 2012; Zhao & Rosson, 2009). 1.2.1 Twitter uptake Twitter is used by various stakeholders in the scholarly community, including individual researchers and academics, journals and publishers, universities and other academic institutions, as well as at scholarly conferences. From the perspective of using Twitter activity as the basis for scholarly metrics, it is essential to know Twitter uptake in academia, as it informs about biases and differences between disciplines and other user demographics, which may have a direct effect on derived metrics. 1.2.1.1 Scholars on Twitter In the scholarly context, Twitter use by academics lags behind its uptake among the general public. Although the majority of researchers are aware of the platform, most do not make use of it in a professional context, giving it the reputation of a hype medium in academia (Carpenter, Wetheridge, Tanner, & Smith, 2012; Gu & Widén-Wulff, 2011; Pscheida, Albrecht, Herbst, Minet, & Köhler, 2013; Van Noorden, 2014). A certain reluctance in academia to use Twitter might be caused by its perception as a shallow medium that is used to communicate "pointless babble" (Kelly, 2009, p. 5) rather than informative content (Rogers, 2014). Described as "phatic" (V. Miller, 2008, p. 396), Twitter is less about what people tweet rather than how they are connected. 3 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Reported Twitter uptake shows extreme variations depending on user demographics―in particular disciplinary orientation―and when a study was conducted; it usually stays behind use of other social media. For example, a survey among 2,414 researchers conducted in 2010 demonstrated that while more than three quarters used social media, less than one fifth were on Twitter (Rowlands et al., 2011). A more recent study conducted by Nature also showed that Twitter was among the least used social media platforms in academia: while almost half of 3,027 science and engineering researchers regularly used ResearchGate, only 13% regularly visited Twitter (Van Noorden, 2014). At almost one quarter of regular users, Twitter uptake was higher among the 482 social science and humanities scholars participating in the same survey (Van Noorden, 2014). Depending on the sample and when the survey was conducted, Twitter uptake varied heavily between a few percent to more than one third of surveyed scholars using Twitter, which calls the representativeness of findings into question. Moreover, surveys vary in terms of whether or not they differentiate between general Twitter uptake, Twitter use for scholarly communication and professional purposes or active vs. passive use, which further complicates comparison and generalization of findings regarding Twitter uptake in academia. Twitter represented the social media tool with the highest difference between awareness and use. Although known by 97% of university staff in Germany, as few as 15% used Twitter and 10% used in a professional context (Pscheida et al., 2013). A similar use-to-awareness ratio was found by other studies, for example, at Finnish universities (Gu & Widén-Wulff, 2011) or researchers surveyed by Nature (Van Noorden, 2014), and academic staff in Germany (Weller, Dornstädter, Freimanis, Klein, & Perez, 2010). Most studies found self-reported Twitter use in academia at around 15%; an uptake of 13-16% was reported for surveys based on 215 health services and policy researchers (Grande et al., 2014), 454 geographers (Wilson & Starkweather, 2014), 1,058 UK academic staff (Tenopir, Volentine, & King, 2013) and 3,027 scientists and engineers (Van Noorden, 2014), while uptake was lower (7-10%) for academics in Germany (Weller et al., 2010) and the UK (Procter et al., 2010; Rowlands et al., 2011). Although 18% used it, Twitter was the least popular social media tool of 345 European scholars (Ponte & Simon, 2011). The highest Twitter use was reported for a survey of 126 Finnish university staff at 23% (Gu & Widén-Wulff, 2011), 1,910 professors at US universities at 32%23% for professional purposes(Bowman, 2015b), 382 urologists attending a conference at 36% (Loeb et al., 2014) and 71 participants of the 2010 Science & Technology Indicators (STI) conference in Leiden at 44% (Haustein, Peters, Bar-Ilan, et al., 2014). As an alternative to determining Twitter uptake through self-reported use in surveys, studies have also assessed the extent of scholarly microblogging based on Twitter activity of scholars. Identifying scholars on Twitter is challenging, as the 160-character Twitter bio and the provided user name are often the only basis for identification. Most studies thus search for Twitter users based on a list of names of academics (Darling, Shiffman, Côté, & Drew, 2013; Desai et al., 2012; Desai, Patwardhan, & Coore, 2014; Priem, Costello, & Dzuba, 2012; Work, Haustein, Bowman, & Larivière, 2015) or apply snowball sampling starting from a set of known scholars on Twitter (Chretien, Azar, & Kind, 2011; Holmberg, Bowman, Haustein, & Peters, 2014; Holmberg & Thelwall, 2014). Searching for a list of 8,038 US and UK university staff, Priem, Costello, and Dzuba (2012) found Twitter accounts for 2.5% of them. Although the authors admit that their study underestimated Twitter use, it reflected that the microblogging platform is not popular in academia and thus confirms finding by most surveys. Investigating Twitter use in the scientometric community, Bar-Ilan et al. (2012) found Twitter profiles for 9 of the 57 presenters at the 2010 STI conference. Other than searching for known scholars on Twitter, some studies try to extract information from Twitter to identify scholarly users. The most common approach is to classify users based on searching for specific words in the Twitter bio. Retrieving users whose Twitter bio contained words such as 4 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research university, PhD or professor, Barthel et al. (2015) identified scientists' Twitter accounts with a precision of 88%. False positives contained university accounts or those of non-academic staff at research institutions. Recall cannot be determined in such studies as the number of false negatives, that is the scientists on Twitter who do not include any of the queried keywords in their self-descriptions, remains unknown. Altmetric also applies a keyword-based approach to categorize Twitter users as scientists, science communicators (journalists, bloggers, editors), practitioners (doctors, other healthcare professionals) and members of the general public. It should be noted that Altmetric's general public category includes all users that can not be classified as belonging to any of the other three groups and is therefore not a good indicator of how much an article has been tweeted by members of the general public. An obvious limitation of the keyword approach is that it is unable to capture scholars who do not identify themselves as such or who do not use the terminology or language covered by the list of keywords. However, many scholars seem to reveal their professional personas on Twitter. Ninety percent of doctoral students funded by the Canadian Social Sciences and Humanities Research Council (SSHRC) identified as academics on Twitter (Work et al., 2015), 87% of surveyed US university professors claimed to mention both their professional title and place of work in their Twitter profiles (Bowman, 2015b) and 78% of Twitter users who self-identified as a physicians used their full names (Chretien et al., 2011). This willingness to reveal their scholarly identities on Twitter suggests that scholars make use of the microblogging platform in a professional context at least to some extent. Ke, Ahn, and Sugimoto (2017) took advantage of crowdsourced Twitter lists to identify scholars. Based on a method introduced by Sharma, Ghosh, Benevenuto, Ganguly, and Gummadi (2012), they identified scientists on Twitter with an approach based on membership in scientific Twitter user lists. Other studies have estimated Twitter activity by scholars by analyzing users who engage with scholarly content on Twitter. Hadgu and Jäschke (2014) applied machine learning to automatically identify scholars on Twitter based on a training set of users whose tweets contained a computer science conference hashtag, while others selected users who have tweeted scientific papers (Alperin, 2015b; Haustein, Bowman, & Costas, 2015a; Haustein & Costas, 2015b; Holmberg & Thelwall, 2014; Tsou, Bowman, Ghazinejad, & Sugimoto, 2015). Since the latter type of studies focuses on categorizing who is tweeting about scholarly contents rather than estimating Twitter uptake in academia, these studies are discussed in more detail below. 1.2.1.2 Scientific conferences Twitter does particularly well in fostering communication among people participating in shared experiences (Rogers, 2014), which may be why tweeting at scholarly conferences has been one of the earliest and most popular uses of Twitter in academia. Almost every scientific conference today has a specific hashtag to connect attendees and those interested but not able to attend in person, thus expanding the conference audience to include remote participants (Bonetta, 2009; Sopan, Rey, Butler, & Shneiderman, 2012; Weller, Dröge, & Puschmann, 2011). Apart from increasing the visibility of presentations, tweeting at scientific conferences has introduced another level of communication, creating backchannel discussions online among participants complementing presentations and discussions which take place at the meeting. Conference tweets usually directly refer to presentations and discussions during sessions and sometimes summarize key take-away points (Chaudhry, Glode, Gillman, & Miller, 2012; McKendrick, Cumming, & Lee, 2012; Mishori, Levy, & Donvan, 2014). Other motivations to tweet at a scientific conferences were to share information and learn about discussions in parallel sessions, networking with others and feeling a sense of connectedness, as well as note-taking (McKendrick et al., 2012). A significant number of tweets associated with two medical conference were uninformative or promotional (Cochran, Kao, Gusani, Suliburk, & Nwomeh, 2014; Desai et al., 2012). 5 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Due to the ease of collecting tweets with a particular hashtag, as well as Twitter's relative popularity in the context of scientific conferences, there are countless studies analyzing scholarly Twitter use based on tweets with conference hashtags (Chaudhry et al., 2012; Cochran et al., 2014; Ferguson et al., 2014; Hawkins, Duszak, & Rawson, 2014; Hawkins et al., 2014; Jalali & Wood, 2013; Letierce, Passant, Breslin, & Decker, 2010; McKendrick et al., 2012; Mishori et al., 2014; Reinhardt, Ebner, Beham, & Costa, 2009; Sopan et al., 2012; Weller et al., 2011). Similar to the overall uptake among scholars, Twitter activity at conferences differs among disciplines as well as individual conferences and increased over the years. Overall, only a small share of conference participants contributed to discussions on Twitter: Less than 2% of attendees of the American Society of Nephrology's 2011 conference (Desai et al., 2012) and less than 3% of participants of the 2012 Winter Scientific Meeting of the Association of Anesthetists of Great Britain and Ireland tweeted (McKendrick et al., 2012). Another medical conference in 2013 reported higher Twitter engagement, as 13% of conference attendees tweeted using the conference hashtag (Mishori et al., 2014). Longitudinal studies also observed an increase in Twitter activity at conferences over the years (Chaudhry et al., 2012; Hawkins et al., 2014; Mishori et al., 2014). For example, 2% of conference participants tweeted at the 2010 Annual Meeting of the American Society of Clinical Oncology, while 5% contributed to conference tweets in 2011. Similarly, the number of tweets nearly doubled from 4,456 tweets in the first to 8,188 in the following year (Chaudhry et al., 2012). A similar increase was observed for the 2011 and 2012 annual meetings of the Radiological Society of North America (Hawkins et al., 2014). Conference-related discussions on Twitter are not restricted to in-person attendees. In fact, at some conferences the majority of Twitter users only participate remotely (Sopan et al., 2012). Just as with other social media and information in general, tweeting activity is usually heavily skewed with a few users contributing the majority of tweets at conferences (Chaudhry et al., 2012; Cochran et al., 2014; Mishori et al., 2014). Tweeting about a conference has been shown to lead to an increase in the number of followers regardless of attending in-person or remotely. Follower counts grew particularly for speakers and in-person attendees, while the number of followers grew least for remote participants (Sopan et al., 2012). Most organizers of scientific conferences embrace the potential of increasing visibility and outreach and thus encourage tweeting through a conference-specific hashtag. Some also specifically display conference-related tweets in real time and thus make tweeting activity visual to participants who are not on Twitter (Ferguson et al., 2014; Jalali & Wood, 2013; Sopan et al., 2012; Weller et al., 2011). 1.2.1.3 Journals and publishers Twitter's technological features afford direct connections and two-way conversations between users changing what was traditionally known as a unidirectional sender-audience relationship. Opposed to traditional publishing and mass media, Twitter has given rise to personal publics of audiences (Schmidt, 2014). This direct link between the sender and receiver has changed the relationship with audiences; for example, musicians use Twitter to market their own brand and respond to @replies from fans to seek out in-person interactions (Baym, 2014). TV audiences turn Twitter into a virtual lounge room when they connect with other users discussing TV events in real time (Harrington, 2014). Similarly, discussions of scientific publications can now happen publically, when readers share their opinions on Twitter. A specific use case are Twitter journal clubs, an adaption of small-group in-person journal clubs that are common particular in the medical sciences (Leung, Siassakos, & Khan, 2015; Mehta & Flickinger, 2014; Thangasamy et al., 2014; Topf & Hiremath, 2015; Whitburn, Walshe, & Sleeman, 2015). Twitter journal clubs are used to discuss and review recent publications and educate researchers and practitioners; in the medical sciences they also have the advantage over their offline predecessors to directly involve patients (Mehta & Flickinger, 2014). Often these journal clubs are initiated or at least supported by journals to promote their publications. A journal club initiated by a 6 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research gynecology journal showed that discussing papers and making them freely available has boosted their Altmetric scores (Leung et al., 2015). Twitter journal clubs also motivated authors of discussed papers to create Twitter accounts (Thangasamy et al., 2014). With journals and authors on Twitter, readers can get in touch directly and involve them in discussions using @mentions, tearing cracks in the wall of traditional gatekeeping, as "Twitter makes it possible to directly connect journal readers at various stages of training with authors and editors" (Mehta & Flickinger, 2014, p. 1317). Many journals and publishers have started to use Twitter as a marketing instrument to increase online visibility and promote published contents. These accounts can be used to create a personalized audience relationship and to foster interaction among readers. Similar to the mix of professional and personal interactions by academics on Twitter, the lines between scholarly communication and marketing campaign are blurred for accounts maintained by journals and publishers. Almost half of the 25 general medicine journals with the highest impact factor in 2010 had a Twitter presence (Kamel Boulos & Anderson, 2012), while Twitter uptake was lower for other sets of journals: 24% of 33 urology journals (Nason et al., 2015), 2 of the top 10 ophthalmology journals (Micieli & Micieli, 2012), 16% of 100 Web of Science (WoS) journals (Kortelainen & Katvala, 2012) and 14% of 102 journals specialized in dermatology (Amir et al., 2014) maintained an account. As most of these studies focused on the top journals according to the journal impact factor, Twitter uptake might be biased towards high-impact journals and slightly lower when including others. The variation suggests similar differences between disciplines as observed for Twitter uptake by individual scholars and conferences. While most journal accounts are used to share articles and news (Zedda & Barbaro, 2015) and often tweet the article title (Friedrich, Bowman, Stock, & Haustein, 2015; Thelwall, Tsou, Weingart, Holmberg, & Haustein, 2013), some journals have incorporated tweeting into the formal communication process. In addition to regular abstracts, they ask authors to write so-called tweetable abstracts that meet the 140-character restrictions, which are used to attract readers on Twitter (Darling et al., 2013). Twitter even interfered with the journal's role in scholarly communication, when a genomics paper was criticized and corrected results posted in a tweet, leading to a conflict with the authors of the criticized paper (Woolston, 2015). Even if a journal is not represented by a proper Twitter account, it is likely that its publisher is. Zedda and Barbaro (2015) found that Twitter adoption was particularly common among 76 publishers in science, technology and medicine; 89% had official Twitter accounts, exceeding the presence on any other social media platform, and 74% had embedded tweet buttons that allowed readers to directly share publications on Twitter. Promotion of publications by publishers seem to be welcomed by authors, as a survey by Nature Publishing Group revealed that almost one fifth of authors would consider it a very valuable service if publishers promoted papers using marketing and social media (Nature Publishing Group, 2015). As shown in the analysis of Twitter accounts diffusing scientific articles below, accounts maintained by journals and publishing houses are responsible for a significant amount of tweets mentioning scientific articles. Once Twitter metrics are being used to evaluate journal impact, these types of self- tweets might be considered as a type of gaming in a manner similar to journal self-citations and citation cartels to boost the impact factor (Seglen, 1997; Van Noorden, 2013). With publishers invested in the success of their journals, tweet cartels and tweet stacking in analogy to their citation equivalents are easily conceivable and even easier to implement. While the WoS excludes journals from the Journal Citation Reports, which have been caught increasing their impact factors artificially, companies like Altmetric.com and Plum Analytics do not (yet) intervene in such self-promotional activity. 7 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research 1.2.1.4 Universities and academic libraries Scholarly institutions are affected by Twitter's impact on academia on two levels: they exploit the microblogging platform to increase their visibility (and that of their members) and provide guidelines and recommendations for their members to navigate the new communication space. The Association of American University Professors updated their report on Academic Freedom and Electronic Communication (Association of American University Professors, 2013) in reaction to a university rescinding a tenure-track job offer to an English scholar who had made an anti-Semitic comment on Twitter (Herman, 2014). The updated report emphasized that professors enjoy academic freedom even when they comment on social media and particularly addressed the blurring of boundaries between private and professional opinions on social media. It stressed how, in this new context, comments are particularly prone to be misunderstood and misinterpreted, as they are often taken out of context: Electronic communications can be altered, or presented selectively, such that they are decontextualized and take on implicit meanings different from their author's original intent. With the advent of social media such concerns about the widespread circulation and compromised integrity of communications that in print might have been essentially private have only multiplied further. (Association of American University Professors, 2013, p. 42) The report further recommends that universities and other academic institutions, along with their staff, develop policies that address the use of social media. In general, academic institutions lack specific social media guidelines or address social media in policies. Although more and more institutions adopt specific policies (Pasquini & Evangelopoulos, 2015), only half of US doctorate-granting universities had a social media policy, while rates were even lower for other universities and colleges. At the same time, Twitter was specifically mentioned in more than 80% of policies (Pomerantz, Hank, & Sugimoto, 2015). The majority of university Twitter accounts apply a so-called megaphone model of communication, where news and information concerning the institution are broadcasted following a traditional communication model (Kimmons, Veletsianos, & Woodward, 2017; Linvill, McGee, & Hicks, 2012). Universities use Twitter for public relations, dissemination of news and events, as well as recruitment (Kimmons et al., 2017). Ninety-four percent of 474 US university admission officers reported that their institution had a Twitter account (Barnes & Lescault, 2013) and 96% of the websites of 100 US colleges linked to Twitter (Greenwood, 2012). On the departmental level, it was less common to be represented with an organizational account, as only 8% of 183 US radiology departments had a Twitter presence (Prabhu & Rosenkrantz, 2015). Twitter was also commonly used for faculty development at medical schools (Cahn, Benjamin, & Shanahan, 2013). Analyzing the Twitter activity of 29 Israeli universities and colleges, Forkosh-Baruch and Hershkovitz (2012) found significant differences between both types of institutions. Colleges were more likely than universities to post social tweets. While almost half of the tweets by universities focused on research conducted elsewhere, colleges focused more on reporting the work by its own researchers. Twitter was also frequently used by academic libraries as a marketing instrument, to communicate with patrons, to announce new resources and promote services (Boateng & Quan Liu, 2014; Hussain, 2015; Shulman, Yep, & Tomé, 2015). On par with Facebook at a 63% adoption rate, Twitter was the most commonly used social media platform among 38 surveyed academic libraries from different countries (Chu & Du, 2013), while all of the 100 US university libraries analyzed by Boateng and Quan Liu (2014) maintained a Twitter presences. In Canadian academic libraries Twitter adoption was lower at 47% (Verishagen & Hank, 2014). 8 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research 1.2.2 Twitter use Apart from Twitter uptake, scholarly Twitter metrics are further influenced by how Twitter is used. Academics use social media to share information, for impression management and to increase their visibility online, to network and establish a presence across platforms, to request and offer help, expand learning opportunities, or simply to be social (Veletsianos, 2012). Twitter specifically was used mostly to tweet work-related content, discover peers working on similar research, follow research-related discussions and get recommendations for papers (Van Noorden, 2014). One central motivation for scholars to tweet is to communicate and explain their work to lay people. As many science communicators are active on Twitter, they help to bridge the gap between the scholarly community and the general public. Science communicators were the largest user group of 518 Twitter users mentioned in tweets by 32 astrophysicists (Holmberg et al., 2014). An evolutionary biology professor valued Twitter to communicate his work to the general public: Twitter and regular blogging are more effective than anything else I do to publicize a paper, which was really surprising to me […]. If you do it right, Twitter is an effective way of telling people about your work. (Bonetta, 2009, p. 453) However, most researchers still preferred traditional media over Twitter to promote their research (Wilkinson & Weitkamp, 2013). Even when identifying professionally on Twitter, a large share of tweets by scholars are not related to their work or academia in general (Bowman, 2015b; Haustein, Bowman, Holmberg, Peters, & Larivière, 2014; Mou, 2014; Pscheida et al., 2013; Van Noorden, 2014). The Nature survey found that 21% of scientists who used Twitter regularly did not use it professionally and 28% said that they never posted content about their work (Van Noorden, 2014). Bowman (2015a, 2015b) reported that, while 29% of American university professors used Twitter strictly in a professional way and 42% used it for both for personal and professional reasons, the vast majority of tweets were coded as personal (78%) rather than professional (19%). Again, large variations can be observed between disciplines as well as individual Twitter users (Chretien et al., 2011; Holmberg & Thelwall, 2014; Loeb et al., 2014; Mou, 2014; Priem, Costello, et al., 2012; Work et al., 2015). Examining more than half a million tweets from 447 researchers, Holmberg and Thelwall (2014) found that less than 4% of tweets were classified as scholarly communication and results varied between disciplines ranging from less than 1% for sociology up to 34% for biochemistry. A study on emergency physicians' tweeting behavior found that 49% of their tweets were related to health or medical issues, 21% were personal, 12% self-promotional and 3% considered unprofessional as they contained profanity, were discriminatory or violated patient privacy (Chretien et al., 2011). In a sample of tweets by funded doctoral students in the social sciences and the humanities in Canada, 4% of tweets were related to their thesis, 21% to the discipline and 5% to academia in general, while 70% of tweets were coded as non-academic (Work et al., 2015). Personal use also prevailed among 382 urologists (Loeb et al., 2014). These findings highlight that even when scholars identify professionally on Twitter and use the platform for scholarly purposes, many tweets will be irrelevant to scholarly communication and should thus be excluded from a scholarly indicator perspective: The lack of a dividing line between scientists and non-scientists, as well as the great variety of topics that even scientists tweet about mean that Twitter is not comparable to the orderly world of science publishing, where every piece of information is assumed to be relevant. Instead, a typical user's timeline is likely to be populated both by scholarly content and personal remarks, more or less side by side. (Puschmann, 2014, p. 98) 9 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Due to their brevity and the fact that when analyzing tweets they are often taken out of context, categorizing tweet content is as difficult as it is to classify Twitter users (Bowman, 2015b). Distinguishing between scientific and non-scientific tweets is especially challenging (Holmberg & Thelwall, 2014). Large variations can also be found between individual tweeters in terms of how often they tweet. A group of astrophysicists analyzed by Haustein et al. (2014) tweeted, on average, between 0 and 58 times per day. Tweets to scientific papers have been shown to peak shortly after their publication and decay rapidly within just a few days. For example, 80% of arXiv submissions received the largest number of tweets the day after they were published (Shuai, Pepe, & Bollen, 2012). Similarly tweeted half-life was 0 days for papers published in the Journal of Medical Internet Research (Eysenbach, 2011) and 39% of a sample of tweets linking to a scholarly document referred to those published within one week before (Priem & Costello, 2010). Determining the delay between publication and first tweet as well as half-lives on Twitter is, however, challenging due to the ambiguity of publication dates (Haustein, Bowman, & Costas, 2015b). Since not all Twitter use culminates in a tweet, a large share of activity remains invisible and thus unmeasurable. In fact, passive use prevailed among UK doctoral students on Twitter (Carpenter et al., 2012) and seems to be common for scholarly use of social media in general. While most academics access and view information, only a minority actively contributes by creating content on social media (Procter et al., 2010; Tenopir et al., 2013); less than 2% of 1,078 UK researchers surveyed actively contributed daily (Tenopir et al., 2013). 1.2.2.1 Tweeting links The most frequent use of Twitter among researchers in higher education was to share information, resources or media (Veletsianos, 2012). A survey among US university professors revealed that embedding URLs was the most commonly used Twitter affordance. Half of the survey participants claimed to tweet links either sometimes, mostly or always (Bowman, 2015b). Links are a common way to send more information than 140 characters would fit. Addressing the length limitation, a scholar explained: It is a double-edged sword. The majority of my tweets are pointers to other resources, so there is a headline-an enticement in other words-and a link to the resource. You don't need more than 140 characters for that. (Bonetta, 2009, p. 453) Weller and Puschmann (2011) refer to links in tweets as "external citations" (Weller & Puschmann, 2011, p. 2). Studies about scholars on Twitter show that they make frequent use of tweeting URLs, as the share of tweets with links exceeds that observed for general Twitter users (boyd et al., 2010; Harrington, 2014). About one third of 68,232 tweets sent by 37 astrophysicists (Haustein, Bowman, Holmberg, et al., 2014) and 38% of 22,258 tweets posted by Canadian social sciences and humanities doctoral students contained links (Work et al., 2015). Tweeting links was even more popular among scholars studied by Weller and Puschmann (2011) and emergency physicians analyzed by Chretien et al. (2011), as respectively 55% and 58% of their tweets contained URLs. Links were much more common when a sample of 445 US professors tweeted professionally: 69% of professional tweets contained URLs, while only 15% of personal tweets did (Bowman, 2015b). Tweets with the #www2010 and #mla09 conference hashtags linked to a website in 40% and 27% of the cases, repeating each unique URL less than three times (Weller et al., 2011). Priem and Costello (2010) found that 6% of a sample of 2,322 of tweets by academics containing a URL mentioned a scholarly publication, 52% of which were first-order and 48% second-order links (i.e., via another website) to the document. Similarly, Holmberg and Thelwall (2014) found that scholarly tweets frequently contained a link to scholarly publications via a blog post about the paper. 10 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research First-order links were significantly more likely to refer to open access articles (Priem & Costello, 2010). Tweets containing the #iswc2009 conference hashtag linked to applications (e.g., online services or research projects; 31%), the conference website (21%), blog posts (12%), slideshows (12%) and publications (9%) (Letierce et al., 2010). Blogs were the most common linked resources in conference tweets analyzed by Weller and colleagues (2011), while news websites were a frequent link destination of tweets sent by Canadian doctoral students, even when discussing scholarly topics (Work et al., 2015). When linking to scholarly papers, tweets often contained the paper title and rarely expressed any recommendation or sentiment (Friedrich, Bowman, Stock, et al., 2015; Thelwall et al., 2013). The great majority of tweeted articles were published very recently (Eysenbach, 2011; Holmberg & Thelwall, 2014; Priem & Costello, 2010). According to surveys asking for motivations to use social media, finding relevant publications and staying up-to-date with the literature was found to be a frequent, yet passive, use of Twitter (Van Noorden, 2014). A Columbia university professor in biology and chemistry describes how they used Twitter to be alerted about the literature: Sometimes four or five people I follow will mention a paper that I did not come across and I will look it up. I think I am much more up to date on science literature since I started following Twitter. (Bonetta, 2009, p. 452) A study by Tenopir, Volentine, and King (2013) found that academics on Twitter read more scholarly publications, which seems to confirm the use of Twitter as a publication alert service. At the same time, as the analyses by Priem and Costello (2010) and Letierce et al. (2010) show, the share of tweets linking to academic papers is low, suggesting a rather passive use: scholars follow links to tweeted articles but do only infrequently distribute them themselves. This suggests that a significant part of Twitter use cannot be captured by scholarly Twitter metrics. 1.2.2.2 Retweets Retweets represent a specific form of diffusing information, as users forward messages sent by others. As such, they do not represent an original contribution by the retweeting user. Since retweets directly quote another users text, they can be seen as "internal citations" (Weller & Puschmann, 2011, p. 3) on Twitter. An analysis of retweets demonstrates how information circulates within a specific user community (Passmann, Boeschoten, & Schäfer, 2014). A common disclaimer that retweets do not equal endorsements adapted from early Twitter use by journalists, emphasizes that tweets are forwarded to increase information diffusion. Once a frequent part of Twitter bios, the disclaimer has now been established as common sense and is no longer needed (Warzel, 2014). As sharing information is one of the main motivations for scholarly Twitter use, retweeting is likely to be common among tweeting academics. Conference participants interviewed by Letierce and colleagues retweeted "tweets that are close to their interest or tweets that speak about their own work or research project" (Letierce et al., 2010, p. 7). Studies showed that retweeting is less common than other affordances used by scholars on Twitter but exceed expectations of a random sample of tweets in 2009, in which as few as 3% were retweets (boyd et al., 2010). Between 15% and 20% of tweets sent at scientific conferences were retweets (Letierce et al., 2010); similarly, 15% of 68,232 tweets by a group of 37 astrophysicists were retweets (Haustein, Bowman, Holmberg, et al., 2014). Retweeting was more common among 28 academics analyzed by Priem and Costello (2010), as 40% of tweets were retweets. Similarly, 37% of 43,176 tweets by Canadian doctoral students were retweets, while 10% of their tweets were retweeted (Work et al., 2015). At the same time, the share of retweets at a radiology conference was 60% (Hawkins et al., 2014). Being asked about Twitter affordance use, 14% of US professors said that they mostly or always and 34% that they sometimes retweeted (Bowman, 2015b). These professors were more than twice as likely to retweet when their tweets were classified 11 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research as professional rather than personal (Bowman, 2015b). These differences again demonstrate that tweeting behavior differs depending on who is tweeting and it what context. The majority of retweets sent by scholars from ten disciplines contained links, while conversational tweets (i.e., @mentions) were less likely to contain links. This clearly shows that researchers […] frequently share web content and forward information and content they have received from people they follow on Twitter, while links are not that often shared in conversations. (Holmberg & Thelwall, 2014, p. 1035) Links to papers were significantly less likely to be retweeted: while 19% of tweets with links to scholarly publications were retweeted, the retweet rate was twice as high in the overall sample of tweets analyzed by Priem and Costello (2010). This is in contrast to a random sample of more than 200,000 tweets, over half of which contained a URL (boyd et al., 2010). A random sample of 270 tweets linking to scientific journal articles found that many were modified retweets of tweets originating from the journal's own Twitter account (Thelwall et al., 2013). From the perspective of scholarly metrics, a distinction should arguably be made between tweets and retweets, as the latter reflects a rather passive act of information sharing (Holmberg, 2014). Although with each retweet the visibility of the tweet and the information it contains (e.g., the link to a publication) increases, retweets represent diffusion of information rather than impact. As retweeting requires even less effortas little as one click since the implementation of the retweet buttonthan composing an original tweet, Twitter metrics should distinguish between tweets and retweets to reflect these different levels of user engagement (Haustein, Bowman, & Costas, 2016). 1.2.2.3 Followers, @mentions and @replies Scholars on Twitter actively seek new connections and connect others (Veletsianos, 2012). Twitter is built in a way that information is spread via user networks. This means that users receive updates from those they follow and diffuse their messages to those they are followed by, creating a personal public (Schmidt, 2014). Selecting who to follow and who one is followed by are thus essential to communicating on Twitter. Twitter users build a reputation based on both their number of followers and followees. A parallel can be drawn to authors' citation identity and citation image based on which authors one cites and is cited by (White, 2000, 2001). The purpose of expanding one's social network and finding peers was apparent in a study of 632 emergency physicians on Twitter: those who included work-related information in their Twitter bios had more followers and the most influential users in the network were connected to at least 50 other emergency physicians (Lulic & Kovic, 2013). Similarly, conference participants on Twitter saw a significant increase in their number of followers, particularly when they were also presenting (Sopan et al., 2012). The average number of followers of 260 physicians analyzed by Chretien et al. (2011) was 17,217 with a median of 1,426, indicating the typically skewed distribution of followers among Twitter users. In addition to broadcasting one's message via retweets on a meso-level of communication, Twitter users can also directly address users with @mentions or @replies on an interpersonal level. Both replies and mentions thus represent a particular type of tweet that focuses on conversation rather than broadcasting (Honeycutt & Herring, 2009). Like most Twitter affordances, these types of tweets were also developed by the user base before being implemented by Twitter. While @replies happen in response to a tweet and are only visible on the timeline of the tweeter who sent the original tweet, @mentions refer to tweets that contains another user's Twitter handle, which triggers a notification to inform them about being mentioned (Bruns & Moe, 2014). About one third of tweets in 2009 included another user name (boyd et al., 2010). A random sample of tweets without an @mention were mostly 12 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research about the tweeting user's experience, while those with an @mention were more likely about the addressee. In fact, more than 90% of @mentions functioned to address another user, while 5% worked as a reference (Honeycutt & Herring, 2009). The great majority of tweets sent by social sciences and humanities doctoral students referred to other Twitter users, as 72% of the 43,176 tweets sent contained other user names (Work et al., 2015). Conversational tweets were also popular among a group of tweeting astrophysicists. Of the 68,232 tweets 46% were @replies or @mentions (61% including RTs), making it the most frequently used Twitter affordance. Conversational tweets were particularly common among those who tweeted regularly or frequently (Haustein, Bowman, Holmberg, et al., 2014). Most of these mentions referred to science communicators (24%), other astrophysicists (22%) or organizations (13%) on Twitter (Holmberg et al., 2014). Conversational tweets hardly contained links (Holmberg & Thelwall, 2014), and among tweets linking to publications, only 8% were @replies (Priem & Costello, 2010). Opposed to hashtag use, retweeting and embedding URLs, @mentions were the only affordance that were less likely to occur in professors' professional tweets (56%) than those identified as personal (67%), which suggests that when professors discuss their work, they are less likely to address or reference other users directly than when they tweet about private matters. However, mentioning other users was still more common in their professional tweets than retweeting and using hashtags (Bowman, 2015b). 1.2.2.4 Hashtags Similar to retweets and @mentions, hashtags are a user-driven Twitter affordance. Hashtags are keywords following the #-sign, which facilitate connections between users interested in the same topics. Conversations revolving around hashtags represent the macro layer of Twitter communication (Bruns & Moe, 2014). Holmberg and colleagues suggest that "hashtags may resemble the traditional function of metadata by enhancing the description and retrievability of documents" (Holmberg et al., 2014, p. 3). Hashtag use seem to be less common than that of other Twitter affordances among academics. Sixty- one percent of surveyed American university professors declared that they rarely or never used a hashtag (Bowman, 2015b). This might be because scholars are either less familiar with this Twitter- specific affordance or they do not wish to expand conversations beyond their personal publics defined by their follower networks. Although actual hashtag use was low by the US professors analyzed by Bowman (2015b), they were more likely to use hashtags in their tweets identified as professional (28%) than those coded as personal (17%). An early large-scale study found that among a random sample of 720,000 tweets, only 5% contained a hashtag (boyd et al., 2010). Almost one quarter of tweets by astrophysicists contained a hashtag (Haustein, Bowman, Holmberg, et al., 2014), but hashtag use varied significantly among different clusters of the follower network (Holmberg et al., 2014). The same share of hashtags (25%) was found for tweets of SSHRC funded doctoral students; on average each hashtag was mentioned 2.8 times (Work et al., 2015). It is problematic to infer hashtag use from most other studies on scholarly tweets, as data collection itself is often based on a specific hashtag, such as a conference hashtag (e.g., Letierce et al., 2010; Weller et al., 2011). 1.2.3 Reluctance against and negative consequences of using Twitter When using tweets as the basis to measure scholarly impact of any sorts, it is essential to consider who is not on Twitter and why academics might be reluctant to join the microblogging platform. Twitter is a platform where "content is not king" (V. Miller, 2008, p. 395) and has been perceived as "shallow media, in the sense that it favors the present, popular and the ephemeral" (Rogers, 2014, p. xiv). Early Twitter studies which identified the majority of tweets to be "pointless babble" (Kelly, 2009, p. 5) or 13 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research "daily chatter" (Java, Song, Finin, & Tseng, 2007, p. 62) casted doubt on the value of Twitter as a meaningful communication medium (Rogers, 2014). This reduction to banal content has led many in academia to consider tweeting a waste of time and therefore rejecting Twitter as a means of scholarly communication (Gerber, 2012). Particularly the 140-character limit of tweets has many scholars doubt Twitter's usefulness for research. This may be why Twitter is one of the best known and at the same time least used social media platforms in the scholarly community (Carpenter et al., 2012; Gerber, 2012; Gu & Widén-Wulff, 2011; Pscheida et al., 2013; Van Noorden, 2014). The adoption of new technologies is often met with functional and psychological barriers. Lack of time and skills as well as a negative perception of platforms have been identified as barriers to use social media in academia (Donelan, 2016). Reluctance often stems from the notion that tweeting wastes precious time and introduces challenges that come with the blurred boundaries of professional and personal communication. This mix of professional and personal identities on Twitter specifically has been revealed by many studies and has been identified as a potential reason not to use Twitter in academia. Even though in the general public, Twitter uptake is higher among young adults, it is often early career researchers who are more reluctant to tweet about their work (Bulger et al., 2011; Coverdale, 2011). Young academics feel the highest pressure to publish in high-impact journals and limit their time spent on social media, and feel more vulnerable when publicly exposing their ideas, particularly to uncertain audiences (Coverdale, 2011; Harley, Acord, Earl-Novell, Lawrence, & King, 2010; Mccrea, 2011; Veletsianos, 2013). An academic interviewed about the future of scholarly communication expressed their concern regarding the use of Facebook and Twitter for work purposes: There's this research group in my area and, for some reason, they're really into Facebook. So they want to do a lot of discussions on Facebook and that type of thing. But I really just don't have time. It's like Twittering. I just can't…There needs to be a little bit of space where I can actually think about something. And I think for some people, they're just wired in such a way that they like that constancy, and they are also able to actually say something intelligent quickly. And I'm not like that. I have to be a little bit more deliberate and think about things a little bit more. And so I can't Twitter…I need some time to reason… (Harley et al., 2010, p. 97) However, early-career researchers were often more likely to find social media useful in the context of scholarly communication and collaboration (Grande et al., 2014; Gruzd, Staves, & Wilk, 2012; Tenopir et al., 2013). Bowman (2015b) identified a u-shaped relationship between Twitter use and academic experience: US professors seven to nine years into their academic careers were more likely to use Twitter compared to those with fewer and more years of experience. A humanities scholar addressed the issue of blurred boundaries on Twitter: I think it can be distracting, especially to grad students, when they're trying to navigate, when they're needing to learn, adopt, and use these new technologies, but at the same time learn to discriminate among technologies that are more for social things but are being used in the name of research. The lines are too blurry. (Bulger et al., 2011, p. 59) The tension that arises between scholars who take to Twitter and those who are reluctant to discuss scholarly matters on social media is demonstrated by an incident where a genomics paper published in PNAS was criticized on Twitter for flaws in study design and analysis casting doubt on its conclusions. The tweet sent by a genetics researcher at the University of Chicago included charts and tables from a re-analysis of the PNAS paper's data, which he published in the open access and open-peer review journal F1000Research. The tweet and re-analysis provoked many responses on Twitter and in the comment section of the F1000Research paper, some of which demanded the retraction of the PNAS paper (Woolston, 2015), while the PNAS authors accused the critic of violating the norms of science by taking to Twitter. 14 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research The clashing of personal and professional and the fuzzy boundaries between the two has also lead to severe negative consequences for scholars. Tweets by faculty have caused outrages among students, other faculty members, university administration and the public at large. Identifying professionally on Twitter has affected the academic careers of some scholars and, in a certain case, controversial tweets have provoked death threads (Rothschild & Unglesbee, 2013). Tweeting had serious effects on the career of a tenured University of New Mexico psychology professor, who had fat-shamed students on Twitter: "Dear obese PhD applicants: if you didn't have the willpower to stop eating carbs, you won't have the willpower to do a dissertation #truth." The professor was asked to apologize and had to undergo sensitivity training, while his work was monitored and he was banned from working on the graduate students admission committee for the rest of his career (Bennett-Smith, 2013; Ingeno, 2013; Pomerantz et al., 2015). In another incident, an English professor lost his tenure-track position offer from the University of Illinois at Urbana-Champaign due to a tweet that was interpreted as anti-Semitic (Herman, 2014). In response to withdrawal of the job offer by the University of Illinois, the American Association University Professors updated their report on academic freedom to reinstate that academic freedom applies to comments made on social media (Association of American University Professors, 2013). 1.3 Scholarly output on Twitter As this chapter discusses scholarly Twitter metrics, it focuses on how scientific papers are diffused via tweets. Although, as described above, only a small amount of scholars' tweeting activity involves linking to publications (Letierce et al., 2010; Priem & Costello, 2010), tweets to journal articles represents one of―if not the―most popular altmetric. This might be due to the significance of publications in peer-reviewed journals in scholarly communication as well as the ease at which tweets that mention or link to document identifiers (e.g., the Digital Object Identifier DOI) can be retrieved. Another reason that Twitter-based altmetrics―and, in fact, all altmetrics―gravitate towards journal articles is that they aim to complement existing bibliometric measures, which reduce scholarly output in a similar manner. In the following, the altmetrics literature is reviewed to provide an overview of currently used scholarly Twitter metrics for journal articles. As the majority of available studies only scratch the surface of what can potentially be extracted from Twitter activity, the literature review is complemented by an analysis of tweets collected by Altmetric.com. This analysis goes beyond tweet counts and correlations with citations and aims to reflect the What, How, When, Where and Who of scholarly publications shared on Twitter. This includes what types of documents are tweeted, how Twitter affordances such as hashtags, retweets and @mentions, are used to share them, when and where articles are tweeted and who is tweeting them. Although this chapter is not able to provide solutions to problems associated with creating meaningful metrics from tweets, particular limitations and pitfalls are highlighted, while demonstrating the potential of available data. Together with the findings on Twitter use described above, the chapter tries to provide context to help the interpretation of different metrics. This chapter thus aims at improving the understanding of what Twitter-based scholarly metrics can and cannot reflect. 1.3.1 Data and indicators The analysis of tweets that mention scholarly documents is based on Twitter data collected by Altmetric.com until June 2016, which contains 24.3 million tweets mentioning 3.9 unique documents. Altmetric started systematically collecting online mentions of scholarly publications in 2012 and is a particularly valuable data source for tracking Twitter activity related to scholarly output, as it continuously stores tweets that mention scholarly publications with a DOI. Through accessing tweets through the Twitter API firehose, Altmetric circumvents the usual issues of Twitter data collection that researchers are confronted with when using the freely available Twitter APIs. While the Twitter Streaming API limits access to a random sample of 1% of tweets, the Representational State Transfer 15 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research (REST) API is rate-limited and the Search API restricts access to only the most recent tweets relevant to a particular query (Gaffney & Puschmann, 2014). As Altmetric started data collection in 2012, Twitter activity is incomplete for documents published earlier. Table 1 Scholarly Twitter metrics associated with scholarly documents Type of metric Tweets Retweets Users Hashtags @mentions Scholarly Twitter metric Twitter coverage Number of tweets Twitter density Twitter intensity Share of retweets Retweet density / intensity Number of users User density / intensity Mean number of followers Hashtag coverage Number of hashtags Hashtag frequency Share of hashtags Hashtag density / intensity @mention coverage Number of mentioned users @mention frequency @mention density / intensity Tweet span Time Tweet delay Twitter half-life Description Percentage of documents with at least one tweet Sum of total number of tweets Mean number of tweets per document Mean number of tweets per tweeted document Percentage of tweets that were retweets Mean number of retweets per document / tweeted document Unique number of users associated with a document Mean number of users per document / tweeted document Mean of the number of followers of users tweeting a document Percentage of documents with at least one hashtag Unique number of hashtags associated with a document Sum of total number of hashtag occurrences Percentage of tweets with at least one hashtag Mean number of hashtags per document / tweeted document Percentage of documents mentioning a user name Unique number of users mentioned in tweets associated with a document Sum of total number of @mentions Mean number of @mentions per document / tweeted document Number of days between first and last tweet Number of days between publication of a document and its first tweet Number of days until 50% of all tweets have appeared Altmetric's Twitter data is matched to bibliographic information from WoS using the DOI. This match between document metadata and tweets affords the possibility to determine the amount of scholarly output that does and does not get tweeted. The link to WoS data also provides access to cleaner and extended metadata of tweeted documents, such as the publication year, journal, authors and their affiliations, and a classification system of scientific disciplines. At the same time, the match of the two databases also excludes tweets to publications not indexed in WoS and thus comes with the known restriction and biases of WoS coverage. This is why the following analysis describes results for two datasets, containing, respectively all 24.3 million tweets covered by Altmetric (dataset A), and 3.9 million tweets mentioning documents with a DOI, covered by WoS and published in 2015 (dataset B). The number of unique documents in dataset A is based on Altmetric ID. As Altmetric.com's metadata is based on multiple sources, one publication might be treated as two documents, particularly if it has 16 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research multiple versions, such as a journal article on the publisher's website and a preprint on arXiv. Similar duplications are possible but less likely in dataset B, which is based on unique identifiers and cleaner metadata in WoS. In the following, a set of descriptive indicators are used based on tweets to scholarly documents and associated metadata. Table 1 provides an overview of each metric. As described above, the focus here is on journal articles but the metrics can, nevertheless, be applied to any scholarly document or other research object such as scholarly agents including "individual scholars, research groups, departments, universities, funding organizations and others entities acting within the scholarly community" (Haustein, Bowman, & Costas, 2016, p. 376). Similar to most bibliometrics, the metrics described in Table 1 can be applied to any aggregated set of documents, such as all documents relevant to a certain topic, published in the same journal, by the same author, institution, country or in a specific language. 1.3.2 What scholarly output is tweeted? As shown above, only a small share of tweets sent by scholars actually link to scholarly output. Tweets linking to blogs or other websites are often more frequent than those linking to scientific publications. At the same time, the great majority of altmetric studies on Twitter focus on peer-reviewed scientific journal articles with a DOI, which represent one of the main limitations of currently captured altmetrics (Alperin, 2013; Haustein, 2016; Haustein, Sugimoto, & Larivière, 2015; Taylor, 2013). In comparison to other common altmetrics, Twitter is the platform which exhibits the second largest activity related to scientific papers, following the social reference management platform Mendeley (Costas, Zahedi, & Wouters, 2015b; Haustein, Costas, et al., 2015). The disciplinary differences in Twitter uptake described above are equally visible, and likely resulting in, large difference in Twitter coverage between scientific disciplines. In the majority of studies, usually between 10% to 30% of selected documents were mentioned on Twitter at least once (Alhoori & Furuta, 2014; Andersen & Haustein, 2015; Costas et al., 2015a; Hammarfelt, 2014; Haustein, Costas, et al., 2015; Haustein, Larivière, Thelwall, et al., 2014; Haustein, Peters, Sugimoto, et al., 2014; Priem, Piwowar, et al., 2012). As the majority of tweets linking to scholarly papers occurs immediately after publication (Eysenbach, 2011; Shuai et al., 2012) and Twitter activity increased annually, Twitter coverage increases by year of publication, with the most recent papers being more likely to be tweeted and older papers hardly getting shared on Twitter (Haustein, Peters, Sugimoto, et al., 2014). For example, while more than half of 2015 PLOS papers were tweeted at least once (Barthel et al., 2015), Twitter coverage was at 12% for those published in 2012 (Priem, Piwowar, et al., 2012). Similarly, 13% of 2011(Costas et al., 2015a), 16% of 2011-2013 (Robinson-García, Torres-Salinas, Zahedi, & Costas, 2014) and 22% of 2012 WoS documents (Haustein, Costas, et al., 2015) had received at least one tweet, while Twitter coverage increased to 36% for WoS 2015 papers in dataset B (see Table 2 below). Coverage varied between disciplines and journals, but also between databases and geographic regions. Twitter is blocked in countries like Iran and China, which reflects on the visibility of their authors and papers. For example, the share of tweeted papers was low for papers published by authors from Iran (Maleki, 2014). Such geographical biases affect Twitter visibility and need to be taken into account when comparing Twitter impact of documents, authors and institutions from different countries. For example, as few as 6% of Brazilian documents published 2013 and indexed in SciELO (Alperin, 2015a) and 2% of a sample of Iranian papers covered by WoS had been tweeted (Maleki, 2014). On the contrary, at 21% Twitter coverage of Swedish publications seems to be more in line with general findings (Hammarfelt, 2014). As the sample of Iranian publication included documents published between 1997 and 2012 and 98% of tweeted papers were published between 2010 and 2012, both geographical and publication date biases influence Twitter coverage. Particularly high coverage was found for arXiv submissions (Haustein, Bowman, Macaluso, Sugimoto, & Larivière, 2014; Shuai et al., 2012). This high activity was, however, not caused by high Twitter 17 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research uptake in the physics, mathematics and computer science communities, but created by Twitter bots (Haustein, Bowman, Holmberg, et al., 2016). The extend of such automated rather than human activity and its implications for the meaning of scholarly Twitter metrics are further discussed below. While Twitter coverage describes the extent to which a set of documents gets diffused on Twitter, Twitter density (i.e., the mean number of tweets per paper) reflects the average tweeting activity per document. In general, each paper receives less than one tweet on average, with large variations between disciplines (see below). As Twitter density is influenced by Twitter coverage and thus particularly low when only a small share of papers gets tweeted, Twitter intensity reflects the average tweeting activity for tweeted papers only, excluding non-tweeted papers (Haustein, Costas, et al., 2015). The number of tweets per paper is usually heavily skewed, much more than citations. For example, 63% of tweeted biomedical papers were only tweeted once (Haustein, Peters, Sugimoto, et al., 2014). Figure 1 Number of tweets and users per tweeted document for datasets A and B. A similar distribution can be observed for the two datasets described above (Figure 1). The 24.3 million tweets in dataset A mentioned 3.9 documents (based on Altmetric ID), 43% of which were tweeted once, 19% twice, 10% three times, while only 4% of documents received more than 25 tweets. The document with the most Twitter activity was mentioned 35,135 times by 144 users, with one user tweeting 34,797 times. This highlights that the number of distinct users per document might be a better proxy of diffusion than the total number of tweets. Tweets in dataset A were sent by 2.6 million users and link to 3.9 million documents, amounting to a Twitter intensity of 6.2 tweets per tweeted paper. The number of users is based on unique Twitter handles included as Author ID on Source in Altmetric's Twitter data. Since the data does not include Twitter's unique user ID, users might be counted more than once if they changed their Twitter handle. Very likely caused by the 140-character limitation, less than 1% of tweets link to more than one document. As shown in Figure 1, the distribution of number of users per document is even more skewed with 47% of all documents tweeted by a single user only. The distribution for WoS 2015 papers (dataset B) is similar but slightly less skewed for documents (based on WoS identifier) mentioned in one tweet and by one user only; 35% and 38% of documents were tweeted once or by one user, respectively. The total number of 3.9 million tweets were sent by 601,290 users mentioning 548,841 documents, which corresponds to a Twitter coverage of 36%, intensity of 7.2 and a density of 2.6 tweets per document. 1.3.2.1 Disciplines and journals Twitter activity varies among disciplines and even journals of the same field. As there is no gatekeeping in Twitter, these variations might not be entirely due to actual impact of a particular journal but can be heavily influenced by individuals and marketing strategies of publishing houses or other stakeholders. Despite these variations, studies have shown that multidisciplinary and biomedical journals as well as social science publications are particularly visible on Twitter, while the so-called hard sciences are tweeted about less (Alhoori & Furuta, 2014; Andersen & Haustein, 2015; Barthel et al., 2015; Costas et al., 2015b; Fenner, 2013; Hammarfelt, 2014; Haustein, Costas, et al., 2015; 18 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Haustein, Larivière, Thelwall, Amyot, & Peters, 2014; Haustein, Peters, Sugimoto, et al., 2014; Knight, 2014; Priem, Piwowar, et al., 2012). Previous findings are corroborated by the analysis of discipline and journal-based tweeting activity in dataset B. As shown in Table 2, over one third of 2015 documents in WoS have been tweeted, which represents a significant increase compared to papers published in the previous years. Twitter coverage shows large variation between disciplines from 59% of publications in Biomedical Research, Health and Psychology to less than 10% in Mathematics and Engineering & Technology. This upholds previous findings that Twitter activity is particularly elevated around publications from the biomedical and social sciences (Costas et al., 2015a; Haustein, Costas, et al., 2015). Table 2 Dataset B: Twitter activity for WoS articles published in 2015. Discipline Papers 2015 Tweeted papers Tweets Users N I O D e g a r e v o c N r e t t i w T e g a r e v o c N y t i s n e D y t i s n e t n I s T R % N y t i s n e t n I All disciplines 2,014,977 76% 548,841 36% 3,960,431 2.6 7.2 50% 601,290 6.6 & s e c n e i c S l a r u t a N g n i r e e n i g n E s e c n e i c S l a i c o S s e i t i n a m u H & Biology Biomedical Res. Chemistry Clinical Medicine Earth & Space Engr. & Tech. Mathematics Physics Arts Health Humanities Prof. Fields Psychology Social Sciences 124,402 226,011 158,929 663,481 96,792 253,020 51,240 128,766 18,995 51,535 76,998 54,109 40,657 70,042 73% 84% 90% 64% 91% 88% 85% 93% 19% 74% 39% 72% 78% 76% 33,945 112,470 32,829 223,641 25,616 18,441 2,890 16,783 566 22,662 4,601 14,760 18,735 20,902 37% 59% 23% 52% 29% 8% 7% 14% 15% 59% 15% 38% 59% 39% 226,575 1,025,061 81,739 1,784,438 136,732 50,577 17,441 55,295 1,962 191,530 19,150 93,936 145,767 135,193 2.5 5.4 0.6 4.2 1.6 0.2 0.4 0.5 0.5 5.0 0.6 2.4 4.6 2.5 6.7 54% 52,235 4.3 9.1 50% 229,851 4.5 2.5 34% 13,860 5.9 8.0 51% 288,226 6.2 5.3 51% 42,641 3.2 17,235 2.9 2.7 33% 7,826 2.2 6.0 43% 3.3 30% 14,822 3.7 3.5 39% 8.5 52% 4.2 52% 6.4 47% 7.8 51% 6.5 53% 1,134 1.7 60,090 3.2 9,784 2.0 41,449 2.3 50,974 2.9 51,288 2.6 It should be emphasized that Twitter activity in the Arts and Humanities cannot be generalized as DOIs are not commonly used in these disciplines. While in general 76% of all documents had a DOI, as few as 19% of all WoS-indexed journal articles in the Arts were linked to this unique identifier. DOI use does not only differ between disciplines but also by country or language of publication, which is why results may be biased in these regards as wells. Considering National Science Foundation (NSF) specialties with more than half of its papers having a DOI, Twitter coverage was highest in Parasitology (78%), Allergy (76%) and Tropical Medicine (70%) and lowest in Metals & Metallurgy (1%), Miscellaneous Mathematics (2%), Mechanical Engineering (2%) and General Mathematics (3%). With more than 10 tweets per document, Twitter density was highest in General & Internal Medicine (13.5) and Miscellaneous Clinical Medicine (12.3). At a Twitter coverage of 100%, 198 of 9,340 tweeted journals had all of their documents diffused on Twitter, which strongly suggests a systematic and automated diffusion, possibly by a dedicated account maintained by the journal or publisher (see below). Among journals with at least 100 papers with a DOI in 2015, JAMA, and Biotechnology Advances had the highest Twitter density at 115.4 and 113.2 tweets per paper, respectively. With more than 10,000 unique Twitter users, PLOS ONE, BMJ, Nature, Science, PNAS, NEMJ, JAMA, Lancet, Scientific Report, Nature Communications, JAMA Internal Medicine, PLOS Biology, British Journal of Sports Medicine, Cell, Biotechnology Advances and BMJ Open were tweeted by the largest audience on Twitter (Table 3). These journals reflect large 19 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research multidisciplinary and biomedical journals with large readership, which have been previously identified as highly tweeted, possibly because they exhibit a particular relevance to people's everyday lives (Costas et al., 2015b). Almost all of these journals maintain official Twitter accounts appearing among the three most tweeting users, which suggests that Twitter has become an important platform for journals and publishers to promote their contents. The number of unique Twitter users, in particular for the general science journals seems to be low in comparison to global readership. For example, while Nature claimed to have 3 million unique visitors per month (Nature Publishing Group, 2012), as few as 42,365 Twitter users mentioned a 2015 paper. However, numbers seem to be in line with print circulation (Nature Publishing Group, 2010). Table 3 Twitter metrics for journals with the largest Twitter audience based on number of users (≥10,000 users) including three most active users. Journal accounts are highlighted in bold. Journal PLOS ONE BMJ Nature Science PNAS NEMJ JAMA Lancet Scientific Reports Nature communications r e t t i w T e g a r e v o c N Tweets y t i s n e D y t i s n e t n I 5.2 45.3 89.0 70.0 19.9 75.2 7.4 70% 148,494 49.7 91% 149,874 89.3 85,523 70.8 78,141 22.4 72,312 75.9 74,263 68,420 115.4 121.1 71.6 45,752 8.3 41,177 39,540 15.1 100% 99% 89% 99% 95% 98% 56% 78% 69.9 4.6 11.8 s T R % 47% 66% 56% 67% 54% 62% 64% 65% 50% 52% Users Top 3 most active users N (according to number of tweets; official journal accounts are marked in bold) sciencemagazine, PedroArtino, rkeyserling aprendedorweb, uranus_2, PLOSONE bmj_latest, bookapharmacist, npceorg 59,210 57,622 42,365 Doyle_Media, randomshandom, TheRichardDoyle 39,225 36,496 33,142 31,018 25,673 23,395 21,274 abbrabot, EcoEvoJournals, uranus_2 medicineupdate, NEJM, JebSource NatureComms, Kochi_Study, PatrickGoymer JAMA_current, robarobberlover, ehlJAMA medicineupdate, darmtag, TheLancet SciReports, uranus_2, geomatlab JAMA Internal Medicine 91% 48,882 99.8 110.1 69% 20,197 PLOS Biology BJ of Sports Medicine Cell Biotechnology Advances BMJ Open 99% 97% 96% 55% 76% 90.0 75.7 38.2 24,481 91.3 26,121 77.7 23,519 39.9 14,823 113.2 205.9 16,882 15.1 11.5 70% 69% 57% 67% 60% 12,878 11,618 11,081 10,331 10,090 JAMAInternalMed, JAMA_current, GeriatriaINNSZ PLOSBiology, PLOS, nebiogroup BJSM_BMJ, exerciseworks, SportScienceNI Brianxbio, CellCellPress, topbiopapers robinsnewswire, GrowKudos, ElsevierBiotech BMJ_Open, LS_Medical, SCPHRP Going beyond peer-reviewed journals indexed in WoS, other sources (based on Altmetric metadata) of scholarly documents are also frequently shared on Twitter. In fact, the largest number of tweeted documents in dataset A came from arXiv: a total of 319,411 arXiv submissions were tweeted 1.1 million times by 110,134 users. As shown in Figure 2, the number of tweeted documents and unique users derivate particularly for the most popular sources. Although arXiv, PLOS ONE and SSRN are the most popular platforms according to the number of tweeted documents, Nature, The Conversation and PLOS ONE are tweeted by the largest number of users. The majority of popular sources are peer-reviewed journals indexed in WoS, which also lead the ranking in dataset B (see above). Apart from these, links to repositories for documents (arXiv, SSRN, bioRxiv) or data (figshare, Dryad) as well as to website like The Conversation or ClinicalTrials.gov are also frequently tweeted. It should be mentioned that the document metadata in Altmetric is based on a variety of sources and thus shows some inconsistencies. For example, the source is unknown for 6% of tweeted documents and 4% of tweets in dataset A, and some sources appear in various spellings. Therefore, the number of distinct sources (49,379) or journal IDs (28,457) represents an overestimation of different sources. These inconsistencies can also be found in the publication year and other document metadata, which are essential to characterize what kind of scholarly output has been diffused on Twitter. 20 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Figure 2 Number of tweeted documents and unique users per source in dataset A (≥100 tweeted documents). 1.3.2.2 Document characteristics Besides disciplinary and topical differences, studies have also focused on determining what type of documents are popular on Twitter. Papers with particular high Twitter visibility often had humorous or entertaining contents rather than scientific merit (Haustein, Peters, Sugimoto, et al., 2014; Neylon, 2014). A study that coded title characteristics of 200 highly tweeted papers found that 16 included a cultural reference (i.e., proverbs, idioms, fictional characters, music) and 13 were humorous or light (Didegah, Bowman, Bowman, & Hartley, 2016). Bornmann (2014) reported that among papers recommended on F1000, those labeled as good for teaching were frequently tweeted. Andersen and Haustein (2015) found that meta-analysis and systematic reviews received significantly more tweets than other medical study types. Marking a clear distinction from citation patterns, particularly high Twitter activity was found for document types that are usually considered uncitable (Haustein, Costas, et al., 2015). For example, Twitter coverage for news items was twice as high as that for all documents. Twitter density was highest for news items (3.0), editorial material (1.6) and reviews (1.4), by far exceeding the overall average of 0.8 tweets per document. The success of these document types on Twitter suggests that "documents that focus on topical subjects, debates and opinions, which are probably presented in simpler and less technical language, are more likely to appear and become popular on Twitter" (Haustein, Costas, et al., 2015, p. 8). Publications with shorter titles, fewer pages and references tend to receive more tweets, while the opposite tendencies are usually observed for citations (Haustein, Costas, et al., 2015). 1.3.2.3 Similarity to other types of usage Since altmetrics were proposed as an alternative or complementary to traditional bibliometric indicators, most studies analyze in how far these new impact metrics correlate with citations. The motivation behind correlation studies lies in determining whether tweeting patterns are comparable to citing behavior. Positive correlations between citation and tweets would indicate that tweets measure something similar to but much earlier than citations, making tweet-based indicators predictors of impact on the scholarly community (Eysenbach, 2011). Early studies argued that low or negative correlation coefficients would indicate a different type of impact than that on citing authors, possibly impact on society in general. 21 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research The first study analyzing the relationship between tweets and citations found a significant association between highly tweeted and highly cited papers published in the Journal of Medical Internet Research, as highly tweeted publications identified 75% of those which were later highly cited (Eysenbach, 2011). However, this claim was based on 55 papers in a journal which itself maintained a strong Twitter presence. For a set of 4,606 arXiv submissions, tweets were a better predictor of early citations than downloads (Shuai et al., 2012). The generalizability of the finding that tweets were early indicators of citation impact was later refuted by a large-scale analysis based on 1.3 million documents, which found overall low correlations between tweets and citations (Haustein, Peters, Sugimoto, et al., 2014). Correlations between tweets and citations vary between datasets due to particular differences between disciplines or journals but are overall low between 0.1 and 0.2 (Barthel et al., 2015; Costas, Zahedi, & Wouters, 2015a; Haustein, Costas, et al., 2015a; Haustein, Peters, Sugimoto, et al., 2014; Priem, Piwowar, et al., 2012). It should be noted that correlations are affected by low Twitter coverage and thus differ whether untweeted papers are excluded or included from the analysis (Haustein, Costas, et al., 2015). Instead of replacing citations as a faster and better filter of relevant publications, as was suggested in the altmetrics manifesto (Priem, Taraborelli, Groth, & Neylon, 2010), the number of tweets seem to mirror visibility on other social media platforms, in particular Facebook, rather than visibility within the community of citing authors (Barthel et al., 2015; Haustein, Costas, et al., 2015a). If nothing else, the difference between tweet and citation counts as reflected in low correlations might be due to the fundamental difference between the act of citing and tweeting (Haustein, Bowman, & Costas, 2016). A moderate negative correlation was found comparing publication output and tweeting activity of a group of astrophysicists on Twitter, suggesting that researchers who tweet a lot focus their efforts on communication and outreach rather than publishing peer-reviewed articles (Haustein, Bowman, Holmberg, et al., 2014). This inverse relationship between a researchers' standing in the scholarly community and their visibility on Twitter has led to the so-called Kardashian index, a tongue-in-cheek indicator that reveals that those who tweet more publish less and vice versa (Hall, 2014). Rather than correlating tweets and citations Allen, Stanton, Di Pietro and Mosley (2013) aimed to measure the effect of promoting articles on social media (including Twitter) on usage statistics. Comparing the number of views, downloads and citations of randomly selected articles published in PLOS ONE before and after promoting them on social media, article views and downloads increased significantly but citations one year after publication and social media metrics did not. 1.3.3 How is scholarly output tweeted? Not the least due to the evaluation community's focus on counts, altmetrics research has focused much less on tweet content than on correlations and other quantitative measures. Among those looking at tweet content, the focus has been on the analysis of Twitter specific affordance use (Bowman, 2015b; Weller et al., 2011). This includes in particular the use and analysis of hashtags, retweets and @mentions, which are further described below. Analyzing 270 tweets linking to journal articles, Thelwall et al. (2013) found that 42% contained the title of the article, 41% summarized it briefly and 7% mentioned the author. As few as 5% explicitly expressed interest in the article. While sentiment was absent in the great majority of tweets, 4% of tweets were positive and none negative. Similarly a large-scale study that automatically identified sentiments in tweets using SentiStrength found that the majority of tweets were neutral and that, if sentiment was expressed, it was positive rather than negative (Friedrich, 2015). Based on 192,832 tweeted WoS documents published in 2012, 11% of 487,610 tweets were positive, 7% negative and 82% did not express any sentiment after removing the article title words from tweets. Tweets linking to chemistry papers were the least likely to express sentiments (Friedrich, 2015; Friedrich, Bowman, & Haustein, 2015). 22 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research 1.3.3.1 Retweets and @mentions As described above, retweets and @mentions represent a particular form of conversational tweets, which seemed to enjoy particular popularity among academic Twitter use. Half of the 4 million tweets linking to 2015 WoS papers were retweets (Table 2), which suggests that a significant amount of tweeting activity reflects information diffusion that does not involve much engagement. Compared to the studies investigating retweet use among general Twitter users (boyd et al., 2010) and academics (Haustein, Bowman, Holmberg, et al., 2014; Letierce et al., 2010; Priem & Costello, 2010; Work et al., 2015), the share of retweets among tweeted journal articles is rather high. The percentage of retweets tends to be lowest in disciplines with low Twitter coverage, which suggests that users in disciplines with low Twitter uptake do not use it as much for information diffusion, possibly because they are not as well connected. In 32 of 120 NSF specialties with DOI coverage above 50%, retweets exceed original tweets (Table 2): retweeting was particularly common in Miscellaneous Zoology, General & Internal Medicine, Miscellaneous Clinical Medicine and Ecology with retweet rates above 60% and low in Solid State Physics, Inorganic & Nuclear Chemistry, Chemical Physics and Applied Chemistry with less than 20%. 1.3.3.2 Hashtags Thirty-one percent of the 24.3 million tweets captured by Altmetric until June 2016 contained a hashtag, which is comparable to other studies on hashtag use by academics (Haustein, Bowman, Holmberg, et al., 2014; Work et al., 2015) but far higher than the 5% among a random sample of tweets in 2009 (boyd et al., 2010). 401,287 unique hashtags were mentioned 12.6 million times, which amounts to an average occurrence of 31 per unique term. 105,705 unique hashtags were used in tweets linking to 2015 WoS papers. While 33% tweets contained at hashtag, 46% of all articles were described with at least one hashtag. Each hashtag was mentioned on average 21 times for a total hashtag frequency of 2.2 million. Hashtag frequency is extremely skewed, as 3% and 6% of hashtags are responsible for 80% of hashtag occurrences in dataset A and B, respectively. For example, the most popular hashtag in dataset A was used 162,754 times (1.3% of all occurrences), while 169,992 hashtags only occurred once. Figure 3 demonstrates on a log-log scale the number of tweets hashtags were mentioned in, as well as the number of distinct users mentioning each hashtag. While in general, a linear relationship can be found between the number of occurrences and users, a few popular hashtags are tweeted only by limited number of users, indicating a smaller community. The most popular hashtags in dataset A were #science (Table 4), 1.3% of hashtag occurrence), #cancer (0.9%), #physics (0.8%), #openaccess, #health (0.7%), #paper, #oa and #research (0.5% each). The occurrence of #oa as well as #openaccess among the most frequent hashtags reflects the known heterogeneity of folksonomies and the need for tag gardening when trying to analyze topics (Peters, 2009). WoS 2015 papers (dataset B) were most frequently tagged as #cancer (1.0%), #health, #openaccess, #science (0.9%), #FOAMed, #Diabetes, #ornithology and #Psychiatry (0.6%). The order changes, when considering the number of unique users instead of tweets per hashtag. Among hashtags that occurred at least 1,000 times, the largest discrepancy between the number of tweets and users can be observed for #genomeregulation (1,924 tweet; 10 users), #eprompt (2,281; 17) and #cryptocurrency (4,515; 38), which were, on average, tweeted more than 100 times by the same users. On the contrary, the user-hashtag ratio was lowest for #Fit (4,818; 4,743), #StandWithPP (Stand with Planned Parenthood; 1,060; 972), #dataviz (1,010; 912), #coffee (1,517; 1,246), and #PWSYN (title of popular science book The Patient Will See You Now; 1,017; 834), which indicates a widespread adoption among Twitter users. Accordingly, these hashtags are more general and less scientific. 23 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Figure 3 Number of users and number of tweets per hashtag for all tweets captured by Altmetric (A) and for tweets to 2015 WoS papers (B). Table 4 Hashtag statistics for datasets A and B with most frequent hashtags based on number of tweets. Most frequent hashtags Statistics of hashtag frequency A t e s a t a D #science #cancer #physics #openaccess #health #paper #oa B t e s a t a D #cancer #health #openaccess #science #FOAMed #Diabetes #ornithology Mean Standard deviation Minimum Maximum 99th percentile 90th percentile 75th percentile 50th percentile Mean Standard deviation Minimum Maximum 99th percentile 90th percentile 75th percentile 50th percentile r e b m u N s r e s u f o r e p s r e s U g a t h s a h 17 237 1 1 4 1 65,334 1,425 6 2 1 1 1 2 1 274 5 2 1 1 229 15 4 2 13 111 1 12,056 186 15 5 2 f o r e b m u N s t n e m u c o d 11 209 1 54,845 129 8 2 1 7 73 1 9,765 92 7 2 1 r e b m u N s t e e w t f o 31 617 1 162,754 395 20 5 2 21 222 1 21,122 311 20 6 2 24 s l a n r u o j f o r e b m u N n a p s t e e w T n/a 233 459 0 1,825 1,746 1,036 182 0 102 192 0 1,684 1,723 831 381 130 0 4 19 1 49 6 2 1 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Table 4 shows hashtag-based stats for both datasets. As to be expected, hashtag frequency and the number of unique terms is greater for dataset A, as it covers all documents in Altmetric and the whole timespan, while dataset B is restricted to WoS documents published in 2015. On average, each hashtag occurred in 21 tweets, was used by 13 users to tag 7 documents and 4 journals indexed in WoS. As shown by the percentiles, hashtag occurrence is extremely skewed. On the individual level, the number of users, documents and journals associated with a hashtag can provide information as to how general and widespread a hashtag is, or how specific and relevant to only a small group of users. The timespan, that is, the number of days between the first and last occurrence of a hashtag, indicates its topicality or timeless relevance. For example, among hashtags that occurred at least 1,000 times, #diet, #water and #nutrition were used during the course of more than four years to describe 2015 documents, while #XmasBMJ lasted only 73 days. The first tweet linking to a 2015 WoS paper with the #diet hashtag appeared 17 November 2011. The discrepancy between tweet date and publication date can be described by the lag between online date and journal issue date (see Haustein et al. (2015b) for an analysis of the publication date problematic). 1.3.4 When is scholarly output tweeted? Tweet activity related to scholarly documents has been shown to occur shortly after publication and disappear within a few days (Eysenbach, 2011; Shuai et al., 2012). Tweeted half-lives and delay between publication and first tweet can thus be measured in hours rather than days. This short-lived attention also points to Twitter being used to diffuse new papers instead of discussing them intensely. It is, however, challenging to accurately calculate delay and decay for all publications in WoS as the publication date of the journal does not sufficiently represent when a publication was actually available. Even with the more accurate article-level information of online dates there are issues to determine the actual date of publication, as demonstrated by tweets mentioning articles before they were supposed to be published (Haustein, Bowman, et al., 2015b). Due to the inaccuracy of available publication dates, tweet delay and tweeted half-lives are not computed. Figure 4 Number of tweets per week day per year and mean of the weekly percentage per year. As shown in Figure 4, there are clear differences between weekdays and weekends, reflecting patterns of the work week, which has also been shown to have an effect on journal submissions (Cabanac & Hartley, 2013) and download patterns (Wang et al., 2012). During the week, tweeting activity increases from Monday (14% of tweets) to peak Wednesday (18%) and decrease again towards the weekend. Twitter users tweet, on average, 23% more about scholarly documents on a Wednesday and 41% less on a Sunday. Figure 4 also shows the different magnitude of Twitter activity among years, as well as an overall increase throughout each year. While a general increase from January to December can be observed for each of the four years of tweets, the general trend also reflects the academic year: activity is higher in spring and fall and drops slightly in summer and particularly during the winter break during the last two and first weeks of each year. Considering that Twitter activity often climaxes the day of 25 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research or day after publication, the season and weekday of publication might influence a document's visibility on Twitter (Haustein, Bowman, et al., 2015b). Zhang and Paxson identified Twitter bots based on tweeting patterns that were too regular to be human (Zhang & Paxson, 2011). 1.3.5 Where is scholarly output tweeted? Twitter provides the possibility to geotag each tweet with precise latitude-longitude information of a user's current location. However, since this function is not activated by default, it is only rarely used. Less than 5% of tweets contain geo coordinates (Graham, Hale, & Gaffney, 2014; Severo, Giraud, & Pecout, 2015), which is why geotags of tweets are not a reliable source to determine where scholarly output is tweeted. Another data source to determine the geographic distribution of Twitter users is to analyze the location information provided in the Twitter bio. However, since the profile location is usually not automatically generated but freely edited by users, it cannot be used without extensive data cleaning. Takhteyec, Gruzd and Wellman (2012) showed that 8% of a sample of 3,360 Twitter profiles contained specific latitude-longitude information, 57% named a location and 20% a country, while 15% used fictional places (e.g., Hogwarts) or too general descriptions to determine the users' whereabouts. Using this profile information, Altmetric is able to determine location information for two-thirds of its tweets. It becomes apparent that in many cases information is not accurate enough to determine the exact location, as remote locations in the UK and Kansas are among the most frequent tweet locations (Haustein & Costas, 2015a). Table 5 Top 10 countries by number of users for dataset A and B. Dataset A Dataset B Documents Tweets Users Number of unique items Some country information Missing country information US GB CA JP AU ES FR IN NL DE y b s e i r t n u o c 0 1 p o T s r e s u f o r e b m u n Documents Tweets Users 3,903,064 24,343,105 2,622,117 71% 69% 33.0% 21.6% 7.9% 4.3% 6.5% 7.5% 5.3% 2.6% 3.5% 4.1% 58% 42% 19.8% 11.0% 2.9% 1.8% 2.9% 2.8% 1.5% 0.7% 1.0% 1.0% 57% 44% 20.1% 8.3% 3.1% 2.3% 2.3% 2.2% 1.4% 1.1% 1.1% 0.9% 548,841 3,960,431 601,290 70% 77% 36.2% 27.4% 9.3% 9.6% 7.6% 4.4% 4.0% 7.9% 5.6% 3.4% 58% 42% 19.5% 12.3% 3.1% 3.3% 2.5% 1.5% 1.0% 1.6% 1.1% 0.7% 58% 42% 19.4% 10.6% 3.4% 2.7% 2.4% 2.0% 1.3% 1.2% 0.9% 0.9% US GB CA ES AU JP NL FR DE IN Due to these limitations, the analysis of where users tweet scholarly documents is restricted to the country level (Table 5). Altmetric provides location information for 57% of users, 58% of tweets and 71% of documents (dataset A) and 58% of both users and tweets and 70% of documents for the WoS subset (dataset B). Users from the US are overrepresented as 20% of tweets are sent by Twitter users with an identified location in the US, which is followed by the UK (8%), Canada (3%), Japan, Australia and Spain (2% each). A similar distribution can be observed for the 2015 WoS articles (dataset B), as the top 10 countries by number of users stay the same, although the UK, the Netherlands, Spain, Canada and to a lesser extent Australia and Germany gain in percentage of users, while India, France, Japan and the US lose in comparison to dataset A. While altmetrics have been marketed as democratizers of science evaluation in the terms of having the potential to correct for biases created by WoS and other US and English-centric journal databases, these results show that, when it comes to Twitter, known biases persist or are even intensified on social media. 26 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research 1.3.6 Who tweets scholarly output? One of the main motivations to consider tweets as an altmetric indicator is that Twitter is used by the general public and thus, at least theoretically, offers insight into how non-academics engage with scholarly output. In order to separate tweets by the public from those sent by members of the scholarly community, Twitter users have to be identified and classified as such. 1.3.6.1 Identifying users who tweet scholarly content One of the main challenges of determining the type of impact reflected by tweets to scientific papers is to identify who is tweeting. While Mendeley provides certain standardized user demographics such as academic status, discipline or country for users associated with a paper, the classification of tweets by user type is restricted to Twitter bios. These self-descriptions are 160-character texts, which provide users with the space to present themselves to other users of the microblogging platform. Applying a codebook to determine who tweets scientific papers based on Twitter username, bio and photo, a sample of 2,000 accounts tweeting links to articles published in Nature, PLOS ONE, PNAS and Science, Tsou et al. (2015) found that almost one quarter of accounts were maintained by an organization. Among these were mainly non-profits (42%), corporations (29%) and universities (13%), while many were also classified as news, media or outreach institutions (19%). Among the 1,520 accounts identified as individuals, two thirds were male. One third of the users were identified as having a PhD and 12% as students. This amounts to almost half of all identified individuals having completed or pursuing a doctorate degree, which stands in glaring contrast to about 1% of the US population with a PhD (Tsou et al., 2015), strongly suggesting that it is the scholarly community rather than the general public who tweets links to scientific papers. Applying a similar codebook to a random sample of 800 accounts tweeting 2012 WoS papers, 68% of accounts were maintained by an individual, 21% by an organization, while 12% could not be identified (Haustein et al., 2016), corroborating the findings by Tsou et al. (2015). Among individuals, 47% used professional terms (e.g., doctor, MD, photographer) to describe themselves, 22% identified as researchers (e.g., scientist, professor, postdoc), 13% as science communicators (e.g., writer, author, journalist, blogger) and 7% as students (e.g., grad student, PhD candidate). Reflecting the blurred boundaries between personal and professional communication, many individuals used words from more than one of these categories to describe themselves. For example, 8% of accounts were classified as researchers and professionals and 5% as professionals and science communicators (Haustein et al., 2016). Science communicators were also the largest group of Twitter users mentioned by astrophysicists (Holmberg et al., 2014). Although labor-intensive and based on little more than 160- character self-descriptions, the above studies show that it is feasible to extract members of academia from users tweeting scholarly documents. Keyword-based searches can be applied to identify scholars in larger samples, but are limited by either low recall or low precision depending on the particular query (Barthel et al., 2015). It is considerably more challenging to identify members of the general public. Although many Twitter bios contain terms depicting personal lives (e.g., father, wife, yoga lover), the presence of these terms does not necessarily mean that accounts are maintained by non-academics, because scholars often describe themselves in both personal and professional manner on Twitter (Bowman, 2015b; Haustein & Costas, 2015b; Work et al., 2015). Similarly, it is challenging to distinguish members of the public based on an indeterminate list of terms of non-academic professions (e.g., consultant, photographer), especially when also considering accounts in languages other than English. Even the comparably straight-forward identification of a researcher who strictly identifies as such on Twitter, becomes problematic when they shared a paper out of private interests. For example, a tweet by a physicist might actually reflect engagement by the public rather than scholarly communication, if they tweeted about a cancer study as a member of a patient group rather than in their academic role. 27 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research An alternative to classifying users based on publicly available information, is to approach them directly and ask them who they are. Alperin (2015b) pioneered such a survey method on Twitter, which with the help of an automated Twitter account, asked users who had tweeted a Scielo Brazil paper, whether they were affiliated with a university. Such a direct approach might also be helpful to determine the motivation of a user to tweet a specific paper, helping to quantify and distinguish different types of tweets, such as endorsement or critical discusion, diffusion or self-promotion. Author self-citations or self-tweets accounted for 7% of a sample of 270 tweets (Thelwall et al., 2013). Table 6 Number of followers, tweets, tweet span, tweets per day and tweeting activity per week for the most active users in dataset A (≥20,000 tweets per user). The 24.3 million tweets captured by Altmetric were sent by 2.6 million users. Looking at the most active users who tweeted more than 1,000 times during the whole period covered by Altmetric (Table 6), the presence of accounts automatically diffusing scholarly articles on Twitter becomes apparent (Haustein, Bowman, Holmberg, et al., 2016). In fact, 15 of the 19 most productive accounts in Table 6 with more than 25,000 tweets self-identified bots (see below). 1.3.6.2 Classifying users by Twitter activity Instead of classifying users according to their self-descriptions, accounts can also be grouped based on their activity. Dividing Twitter accounts into three groups of top 1%, 9% and 90% of users (according to number of tweets) helps to distinguish lead and highly active users from less active ones (Bruns & Stieglitz, 2014). This classification provides insights into tweeting behavior of different types of users. Separating the 601,290 users in dataset B by number of tweets linking to a 2015 WoS article, 6,016 lead users, 54,535 highly active and 540,739 least active users can be identified. Lead users contributed between 84 and 19,973 tweets, a median of 149 tweets per users, had on average 935 followers (median; mean=3,862) and tweeted the 2015 papers during an average tweet span of 598 days. Highly 28 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Twitter handleNumber of followersTweetsTweet spanTweets per dayTweeting activity per weekblackphysicists12,91492,5831,82650.7MathPaper1,88973,2391,08667.4anestesiaf1,42863,9531,71337.3oceanologia38962,5851,35346.3UIST_Papers20XX3550,2111,36036.9UIST_Papers19XX649,8381,35936.7hiv_insight14,32848,7141,82226.7russfeed2,12744,46357777.1uranus_22,51942,0531,79023.5Immunol_papers47740,72163963.7psych2evidence51340,65841099.2InorganicNews1,04436,8691,65722.3arXiv_trend16633,27245772.8hlth_literacy5,58232,3371,82317.7AstroPHYPapers3,42731,1541,08628.7cirugiaf40630,6921,71017.9ThihaSwe_dr60129,3571,16425.2semantic_bot028,908281032.4CondensedPapers53027,60391230.3libroazuln12824,9941,61315.5rnomics1,52024,9051,82113.7PhysicsPaper47622,8941,08621.1epigen_papers78821,91673929.7Handbook of Quantitative Science and Technology Research active users contributed between 9 and 83 tweets (median=16), had less followers (median=442.5; mean=2,136) and shorter tweet spans (mean=388 days), while least active users tweeted up to eight times (median=1), had 212 followers and were active for a period of 58 days. Lead (top 1% of users), highly active (9%) and least active (90%) users contributed 43%, 31% and 25% of tweets to the entire set of 2015 WoS papers, respectively (Figure 5). Interestingly, these percentages differ among NSF disciplines with least active users overrepresented among those tweeting literature from Professional Fields, Social Sciences, Psychology and Earth and Space. On the contrary, lead users were overrepresented in Chemistry, Physics, Mathematics and Engineering & Technology, which were the fields exhibiting the lowest Twitter coverage, density and number of unique users (Table 2). Assuming that the general public is least active when it comes to tweeting about scholarly papers, they are more likely to engage with articles published in journals from the Professional Fields and Social Sciences and less likely to tweet Chemical papers. The high presence of lead users in Chemistry and Physics might, at least partly, be caused by accounts promoting these papers automatically, such as @blackphysicists and @MathPaper, which were the two most active accounts in dataset A, tweeting 51 and 67 scholarly documents per day (Table 6). Figure 5 Percentage of tweets from lead users (1%), highly active users (9%) and least active users (90%) per discipline (dataset B). 1.3.6.3 Twitter bots Automated Twitter accounts have become prevalent on Twitter (Mowbray, 2014). About one fifth of tweets sent during the 2016 presidential election were estimated to be sent by bots (Bessi & Ferrara, 2016) and almost one quarter of tweets in 2009 came from accounts tweeting more than 150 times per day. Ferrara et al. (2016, p. 96) define social bots as "a computer algorithm that automatically produces and interacts with humans on social media, trying to emulate and possibly alter their behavior". Automated Twitter accounts can be further distinguished between useful bots and antisocial or spambots (Ferrara et al., 2016; Mowbray, 2014). Bots are also infiltrating academic Twitter. Among a random sample of 800 Twitter accounts captured by Altmetric, 8% seemed completely and 5% partially automated (Haustein et al., 2016), while automated accounts who self-identified as such were responsible for 9% of tweets to arXiv submissions (Haustein, Bowman, Holmberg, et al., 2016). Shuai et al. (2012) even removed half of the tweets to a sample of tweets to documents on arXiv, as they were created by bots. Regardless of whether Twitter accounts that automatically tweet scientific papers are considered useful or spam, it is safe to say that their automated tweets do not reflect impact. In the context of altmetrics and tweets to scientific papers, bot activity thus needs to be at least identified, if not entirely removed in an impact assessment. Although spammers and excessive self-promotion was identified as a challenge by Altmetric, they still considered gaming a rare and easy to identify threat in 2013 (Liu & Adie, 2013). 29 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Twitter's terms of service specifically address automation. While prohibiting spam, Twitter encourages automated tweets if they "broadcast helpful information" (Twitter, 2017). However, what is considered spam continuously evolves as users apply "new tricks and tactics" (Twitter, n.d.) to adapt to or circumvent Twitter rules. Twitter bots can be identified based on specific regularities in their tweeting behavior, such as the frequent and repetitive use of the same hashtags, URLs, tweet format and content, regular temporal activity, as well as the follower/friends ratio, @mentions to non- followers or account suspensions (Mowbray, 2014). Sixteen percent of Twitter accounts were identified as automated based on not-uniform-enough or too-uniform tweeting patterns to be stemming from a human (Zhang & Paxson, 2011). Social network indicators based on the Twitter follower and friend network are considered more robust measures, as they are harder to influence (Yang, Harkreader, & Gu, 2011). The BotOrNot algorithm additionally considers linguistic features and tweet sentiment to detect automated Twitter accounts (Davis, Varol, Ferrara, Flammini, & Menczer, 2016). However, as these spamming measures get known and integrated by Twitter to block spam accounts, bots employ more sophisticated algorithms to avoid getting caught, resulting in an arms race between those who create and those who seek to identify Twitter spam. Figure 6 Query used to identify automated Twitter accounts. Figure 7 6 Query used to identify publisher accounts. Analyzing the most active users who have tweeted at least 1,000 times (dataset A), a keyword-based query searching the Twitter bio as well as user name and handle revealed that among 2,043 accounts 248 identified themselves as automated (Figure 6) and 305 as journal or publisher accounts (Figure 7). These make up 30% and 11% of the tweets sent by the 2,043 most active users (Table 7), which correspond to 7% and 3% of the entire 24.3 million tweets to scholarly documents in dataset A. The median number of followers is significantly lower than other accounts (Table 7) and as few as 6% of the 1.8 million tweets sent by the 248 accounts contain @mentions. If scholarly bots do mention other users, they often seem to reference journals such as @hiv_insight, which frequently mentions @PLoSMedicine, @STI_BMJ and @JAMA_current. Other than social bots in the general Twittersphere, scholarly bots seem to not try to emulate human behavior or game the system. They rather resemble RSS feeds tweeting the paper title and a link, often specifying what type of information they diffuse. Some even provide instructions to create similar feeds. For example, the Twitter bio of @asthma_papers reads "RSS feed for #asthma papers in #Pubmed. Create a feed of your own using instructions here: https://github.com/roblanf/phypapers". 30 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Table 7 Twitter metrics for self-identified bots, journal and publisher accounts and other accounts based on users with at least 1,000 tweets (dataset A). Most active Twitter accounts (≥1,000 tweets, n=2,043) Followers Tweets Tweets per day Tweet span Self-identified bots n=248 30% of tweets Journal and publisher accounts n=305 11% of tweets Other accounts n=1,490 59% of tweets median mean std dev min max median mean std dev min max median mean std dev min max 212 1,014 2,781 0 25,003 3,199 21,475 116,124 3 1,448,649 1,535 5,236 14,091 0 228,224 3,479 7,339 10,390 1,001 73,239 1,670 2,249 1,874 1,001 19,256 1,599 2,408 3,670 1,000 92,583 5.1 14.4 67.0 0.6 1,032.4 1.2 1.7 1.6 0.6 14.7 1.4 2.9 9.3 0.6 297.3 845 923 477 28 1,823 1,647 1,484 369 122 1,822 1,388 1,278 496 8 1,826 Considering that scholars use Twitter to diffuse information and stay aware of relevant literature, automated accounts might be considered useful. However, bots have shown to be harmful to society, when they are used to influence public opinion and behavior such as political opinions or elections (Bessi & Ferrara, 2016; Metaxas & Mustafaraj, 2012; Ratkiewicz, Conover, Meiss, Gonçalves, Flammini, et al., 2011; Ratkiewicz, Conover, Meiss, Gonçalves, Patil, et al., 2011) or manipulation of the stock market (Ferrara et al., 2016). If Twitter impact became part of the scholarly reward system, Twitter bots might be able to similarly influence opinions or shape outcomes of certain research metrics. 1.4 Conclusion and outlook This chapter provided an overview of the use of Twitter in scholarly communication. By demonstrating who uses Twitter in academia and for what reasons, what types of scholarly outputs are diffused how, where and when, it aimed to add context and help to interpret any scholarly metrics derived from this and similar types of social media activity. Research evaluators and managers were particularly excited at the prospect of an easily accessible data source that would be able to capture traces of the societal impact of research. However, perhaps unsurprisingly, the majority of tweets stem from stakeholders in academia rather than from members of the general public, which indicates that the majority of tweets to scientific papers are more likely to reflect scholarly communication rather than societal impact. At the same time, Twitter uptake in academia lacks behind Twitter user by the general public. Twitter activity is influenced by geographical and disciplinary biases and publication date. Known biases towards US and UK sources persist, rather than democratizing scholarly communication and the reward system of science. The majority of tweets linking to scientific articles appear shortly after their publication; tweeting half- lives can be measured in hours rather than days. Moreover, one can observe weekday as well as seasonal patterns with Twitter activity peaking Wednesdays and in the fall and plummeting during the weekend and holiday season. Journal and publisher accounts as well as Twitter bots contribute significantly to tweeting activity linked to academic papers, which suggests that a significant extent of tweeting activity serves promotional purposes or is automated, which does neither reflect societal nor scholarly impact. A large share of tweets contains hashtags and mention either the title or a short summary of the paper they referred to. Half of all articles linking to 2015 WoS papers were retweets 31 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research and the majority contained no sentiments. These tweeting characteristics emphasize particular low engagement of users linking to journal articles. The main motivation for researchers to use Twitter is information diffusion, networking and to stay up-to-date with the literature. However, the sheer brevity of tweets makes intense discussions the exception rather than the rule on Twitter. As citation behavior and motivations to cite or not cite certain sources are biased and influenced by many factors other than a paper's significance, not every citation represents impact. However, each scholarly author is bound by scholarly norms to participate in the citation process. In some rare cases, where scholars tweet corrections to publications or journals provide tweetable abstracts and organize journal clubs, Twitter has started to be integrated in or even replace certain functions of formal journal publishing. However, in most fields tweeting does not yet play an important role in scholarly communication. In most disciplines, Twitter uptake is low and the platform is only used in a passive or infrequent manner, or tweets reflect only a part of informal scholarly communication, such as conference chatter. Moreover, Twitter uptake varies between disciplines, countries, journals, as well as individuals and can be easily influenced and manipulated. The presence of automated Twitter accounts which promote certain contents becomes particularly problematic when tweet counts become the basis for measures of impact. This is not to say that Twitter should be completely disregarded as a data source for scholarly metrics. Rather, the microblogging platform should be approached critically in terms of what kind of use and user populations it captures. By reviewing the role of Twitter in scholarly communication and analyzing tweets linking to scholarly documents in depths and beyond crude counts, this chapter attepmted to provide the basis for more sophisticated and well-balanced approaches to scholarly Twitter metrics. References Alhoori, H., & Furuta, R. (2014). Do altmetrics follow the crowd or does the crowd follow altmetrics? In Digital Libraries (JCDL), 2014 IEEE/ACM Joint Conference on (pp. 375–378). IEEE. Retrieved from http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6970193 Allen, H. G., Stanton, T. R., Di Pietro, F., & Moseley, G. L. (2013). Social media release increases dissemination of original articles in the clinical pain sciences. PloS One, 8(7), e68914–e68914. https://doi.org/10.1371/journal.pone.0068914 Alperin, J. P. (2013). Ask not what altmetrics can do for you, but what altmetrics can do for developing countries. Bulletin of the American Society for Information Science and Technology, 39(4), 18–21. Alperin, J. P. (2015a). Geographic variation in social media metrics: an analysis of Latin American journal articles. Aslib Journal of Information Management, 67(3), 289–304. https://doi.org/10.1108/AJIM-12-2014- 0176 Alperin, J. P. (2015b). Moving beyond counts: A method for surveying Twitter users. In altmetrics15: 5 years in, what do we know? Amsterdam, The Netherlands. Retrieved from http://altmetrics.org/wp- content/uploads/2015/09/altmetrics15_paper_3.pdf Amir, M., Sampson, B. P., Endly, D., Tamai, J. M., Henley, J., Brewer, A. C., … Dellavalle, R. P. (2014). Social networking sites: emerging and essential tools for communication in dermatology. JAMA Dermatology, 150(1), 56–60. https://doi.org/10.1001/jamadermatol.2013.6340 Andersen, J. P., & Haustein, S. (2015). Influence of study type on Twitter activity for medical research papers. In Proceedings of the 15th International Society of Scientometrics and Informetrics Conference (pp. 26–36). Istanbul, Turkey. https://arxiv.org/abs/1507.00154 Association of American University Professors. (2013). Academic Freedom and Electronic Communications. Retrieved from https://www.aaup.org/file/Academic%20Freedom%20%26%20Electronic%20Communications.pdf Bar-Ilan, J., Haustein, S., Peters, I., Priem, J., Shema, H., & Terliesner, J. (2012). Beyond citations: Scholars ' visibility on the social Web. In Proceedings of the 17th International Conference on Science and Technology Indicators Montreal Canada 58 Sept 2012 (Vol. 52900, pp. 98–109). Montreal. http://arxiv.org/abs/1205.5611 32 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Barnes, N. G., & Lescault, A. M. (2013). College presidents out-blog and out-tweet corporate CEO's. UMass Center for Marketing Research. Retrieved from http://www.umassd.edu/cmr/socialmediaresearch/collegepresidentsoutblog/ Barthel, S., Tönnies, S., Köhncke, B., Siehndel, P., & Balke, W.-T. (2015). What does Twitter Measure? Influence of Diverse User Groups in Altmetrics. In Proceedings of the 15th ACM/IEEE-CE on Joint Conference on Digital Libraries (pp. 119–128). Knoxville, Tennessee, USA: ACM Press. https://doi.org/10.1145/2756406.2756913 Baym, M. (2014). The Perils and Pleasures of Tweeting with Fans. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 221–236). New York: Peter Lang. Bennett-Smith, M. (2013, June 4). Geoffrey Miller, Visiting NYU Professor, Slammed For Fat-Shaming Obese PhD Applicants. Retrieved April 14, 2016, from http://www.huffingtonpost.com/2013/06/04/geoffrey-miller-fat-shaming-nyu-phd_n_3385641.html Bessi, A., & Ferrara, E. (2016). Social bots distort the 2016 U.S. Presidential election online discussion. First Monday, 21(11). https://doi.org/10.5210/fm.v21i11.7090 Boateng, F., & Quan Liu, Y. (2014). Web 2.0 applications' usage and trends in top US academic libraries. Library Hi Tech, 32(1), 120–138. https://doi.org/10.1108/LHT-07-2013-0093 Bonetta, L. (2009). Should you be tweeting? Cell, 139(3), 452–453. https://doi.org/10.1016/j.cell.2009.10.017 Bornmann, L. (2014). Validity of altmetrics data for measuring societal impact: A study using data from Altmetric and F1000Prime. Journal of Informetrics, 8(4), 935–950. https://doi.org/10.1016/j.joi.2014.09.007 Bornmann, L. (2016). Scientific Revolution in Scientometrics: the Broadening of Impact from Citation to Societal. In C. R. Sugimoto (Ed.), Theories of Informetrics and Scholarly Communication (pp. 347–359). Berlin: de Gruyter Mouton. Bowman, T. D. (2015a). Differences in personal and professional tweets of scholars. Aslib Journal of Information Management, 67(3), 356–371. https://doi.org/10.1108/AJIM-12-2014-0180 Bowman, T. D. (2015b, July). Investigating the use of affordances and framing techniques by scholars to manage personal and professional impressions on Twitter (Dissertation). Indiana University, Bloomington, IN, USA. Retrieved from http://www.tdbowman.com/pdf/2015_07_TDBowman_Dissertation.pdf boyd, D., Golder, S., & Lotan, G. (2010). Tweet, Tweet, Retweet: Conversational Aspects of Retweeting on Twitter (pp. 1–10). IEEE. https://doi.org/10.1109/HICSS.2010.412 Bruns, A., & Moe, H. (2014). Structural Layers of Communication on Twitter. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 15–28). New York: Peter Lang. Bruns, A., & Stieglitz, S. (2014). Metrics for Understanding Communication on Twitter. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 69–82). New York: Peter Lang. Bulger, M. E., Meyer, E. T., De la Flor, G., Terras, M., Wyatt, S., Jirotka, M., … Madsen, C. M. (2011). Reinventing research? Information practices in the humanities (Research Information Network). Retrieved from http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1859267 Cabanac, G., & Hartley, J. (2013). Issues of work-life balance among JASIST authors and editors: Work-Life Balance Issues Among JASIST Authors and Editors. Journal of the American Society for Information Science and Technology, 64(10), 2182–2186. https://doi.org/10.1002/asi.22888 Cahn, P. S., Benjamin, E. J., & Shanahan, C. W. (2013). "Uncrunching" time: medical schools' use of social media for faculty development. Medical Education Online, 1, 1–6. https://doi.org/10.3402/meo.v18i0.20995 Carpenter, J., Wetheridge, L., Tanner, S., & Smith, N. (2012). Researchers of Tomorrow. Retrieved from http://www.jisc.ac.uk/publications/reports/2012/researchers-of-tomorrow Chaudhry, B. A., Glode, L. M., Gillman, M., & Miller, R. S. (2012). Trends in Twitter Use by Physicians at the American Society of Clinical Oncology Annual Meeting , 2010 and 2011. Journal of Oncology Practice, 8(3), 173–178. https://doi.org/10.1200/JOP.2011.000483 Chretien, K. C., Azar, J., & Kind, T. (2011). Physicians on Twitter. JAMA, 305(6), 566–568. https://doi.org/10.1001/jama.2011.68 Chu, S. K.-W., & Du, H. S. (2013). Social networking tools for academic libraries. Journal of Librarianship and Information Science, 45(1), 64–75. https://doi.org/10.1177/0961000611434361 Cochran, A., Kao, L. S., Gusani, N. J., Suliburk, J. W., & Nwomeh, B. C. (2014). Use of Twitter to document the 2013 Academic Surgical Congress. The Journal of Surgical Research, 190(1), 36–40. https://doi.org/10.1016/j.jss.2014.02.029 Costas, R., Zahedi, Z., & Wouters, P. (2015a). Do "altmetrics" correlate with citations? Extensive comparison of altmetric indicators with citations from a multidisciplinary perspective. Journal of the Association for Information Science and Technology, 66(10), 2003–2019. https://doi.org/10.1002/asi.23309 33 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Costas, R., Zahedi, Z., & Wouters, P. (2015b). The thematic orientation of publications mentioned on social media: Large-scale disciplinary comparison of social media metrics with citations. Aslib Journal of Information Management, 67(3), 260–288. https://doi.org/10.1108/AJIM-12-2014-0173 Coverdale, A. (2011). Negotiating Doctoral Practices and Academic Identities Through the Adoption and Use of Social and Participative Media. In Proceedings of the European Conference on eLearning 2011 (p. 909). Brighton, UK. Cronin, B. (2013a). Metrics à la mode. Journal of the American Society for Information Science and Technology, 64(6), 1091–1091. https://doi.org/10.1002/asi.22989 Cronin, B. (2013b). The evolving indicator space (iSpace). Journal of the American Society for Information Science and Technology, 64(8), 1523–1525. https://doi.org/10.1002/asi.23041 Darling, E. S., Shiffman, D., Côté, I., & Drew, J. A. (2013). The role of Twitter in the life cycle of a scientific publication. Ideas in Ecology and Evolution, 6, 32–43. https://doi.org/10.7287/peerj.preprints.16v1 Davis, C. A., Varol, O., Ferrara, E., Flammini, A., & Menczer, F. (2016). BotOrNot: A System to Evaluate Social Bots (pp. 273–274). ACM Press. https://doi.org/10.1145/2872518.2889302 Desai, T., Patwardhan, M., & Coore, H. (2014). Factors that contribute to social media influence within an Internal Medicine Twitter learning community. F1000Research, 3, 120–120. https://doi.org/10.12688/f1000research.4283.1 Desai, T., Shariff, A., Shariff, A., Kats, M., Fang, X., Christiano, C., & Ferris, M. (2012). Tweeting the Meeting : An In-Depth Analysis of Twitter Activity at Kidney Week 2011. PLoS ONE, 7(7), 1–9. https://doi.org/10.1371/journal.pone.0040253 Desrochers, N., Paul-Hus, A., Haustein, S., Costas, R., Mongeon, P., Quan-Haase, A., Bowman, T.D., Pecoskie, J., Tsou, A., & Larivière, V. (2018). Authorship, citations, acknowledgments and visibility in social media: Symbolic capital in the multifaceted reward system of science. Social Science Information, 1–26. https://doi.org/10.1177/0539018417752089 Didegah, F., Bowman, T. D., Bowman, S., & Hartley, J. (2016). Comparing the characteristics of highly cited titles and highly alted titles. In 21st International Conference on Science and Technology Indicators (pp. 1190–1195). Valencia, Spain. Retrieved from http://ocs.editorial.upv.es/index.php/STI2016/STI2016/paper/viewFile/4543/2327 Donelan, H. (2016). Social media for professional development and networking opportunities in academia. Journal of Further and Higher Education, 40(5), 706–729. https://doi.org/10.1080/0309877X.2015.1014321 Duggan, M., Ellison, N. B., Lampe, C., Lenhart, A., & Madden, M. (2015, January 9). Demographics of Key Social Networking Platforms. Retrieved from http://www.pewinternet.org/2015/01/09/demographics-of-key- social-networking-platforms-2/ Eysenbach, G. (2011). Can tweets predict citations? Metrics of social impact based on Twitter and correlation with traditional metrics of scientific impact. Journal of Medical Internet Research, 13(4), e123–e123. https://doi.org/10.2196/jmir.2012 Fenner, M. (2013). What can article-level metrics do for you? PLoS Biology, 11(10), e1001687–e1001687. https://doi.org/10.1371/journal.pbio.1001687 Ferguson, C., Inglis, S. C., Newton, P. J., Cripps, P. J. S., Macdonald, P. S., & Davidson, P. M. (2014). Social media: A tool to spread information: A case study analysis of Twitter conversation at the Cardiac Society of Australia & New Zealand 61st Annual Scientific Meeting 2013. Collegian, 21(2), 89–93. https://doi.org/10.1016/j.colegn.2014.03.002 Ferrara, E., Varol, O., Davis, C., Menczer, F., & Flammini, A. (2016). The rise of social bots. Communications of the ACM, 59(7), 96–104. https://doi.org/10.1145/2818717 Forkosh-Baruch, A., & Hershkovitz, A. (2012). A case study of Israeli higher-education institutes sharing scholarly information with the community via social networks. The Internet and Higher Education, 15(1), 58–68. https://doi.org/10.1016/j.iheduc.2011.08.003 Friedrich, N. (2015, September 3). Applying sentiment analysis for tweets linking to scientific papers (Bachelor's thesis). Heinrich-Heine-Universität Düsseldorf, Düsseldorf. Friedrich, N., Bowman, T. D., & Haustein, S. (2015). Do tweets to scientific articles contain positive or negative sentiments? In altmetrics15: 5 years in, what do we know? Amsterdam, The Netherlands. Retrieved from http://altmetrics.org/altmetrics15/friedrich/ Friedrich, N., Bowman, T. D., Stock, W. G., & Haustein, S. (2015). Adapting sentiment analysis for tweets linking to scientific papers. In Proceedings of the 15th International Society of Scientometrics and Informetrics Conference (pp. 107–108). Istanbul, Turkey. https://arxiv.org/abs/1507.01967 Gaffney, D., & Puschmann, C. (2014). Data Collection on Twitter. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 55–67). New York: Peter Lang. 34 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Gerber, A. (2012). Online trends from the first German trend study on science communication. In Science and the Internet (pp. 13–18). Düsseldorf University Press. Graham, M., Hale, S. A., & Gaffney, D. (2014). Where in the World Are You? Geolocation and Language Identification in Twitter. The Professional Geographer, 66(4), 568–578. https://doi.org/10.1080/00330124.2014.907699 Grajales, F. J., Sheps, S., Ho, K., Novak-Lauscher, H., & Eysenbach, G. (2014). Social media: a review and tutorial of applications in medicine and health care. Journal of Medical Internet Research, 16(2), e13–e13. https://doi.org/10.2196/jmir.2912 Grande, D., Gollust, S. E., Pany, M., Seymour, J., Goss, A., Kilaru, A., & Meisel, Z. (2014). Translating Research For Health Policy: Researchers' Perceptions And Use Of Social Media. Health Affairs, 33(7), 1278–1285. https://doi.org/10.1377/hlthaff.2014.0300 Greenwood, G. (2012). Examining the Presence of Social Media on University Web Sites. Journal of College Admission, 216, 24–28. Gruzd, A., Staves, K., & Wilk, A. (2012). Connected scholars: Examining the role of social media in research practices of faculty using the UTAUT model. Computers in Human Behavior, 28(6), 2340–2350. https://doi.org/10.1016/j.chb.2012.07.004 Gu, F., & Widén-Wulff, G. (2011). Scholarly communication and possible changes in the context of social media. The Electronic Library, 29(6), 762–776. https://doi.org/10.1108/02640471111187999 Hadgu, A. T., & Jäschke, R. (2014). Identifying and analyzing researchers on twitter (pp. 23–32). ACM Press. https://doi.org/10.1145/2615569.2615676 Hall, N. (2014). The Kardashian index: a measure of discrepant social media profile for scientists. Genome Biology, 15(7), 424–424. https://doi.org/10.1186/s13059-014-0424-0 Hammarfelt, B. (2014). Using altmetrics for assessing research impact in the humanities. Scientometrics, 101, 1419–1430. https://doi.org/10.1007/s11192-014-1261-3 Harley, D., Acord, S. K., Earl-Novell, S., Lawrence, S., & King, C. J. (2010). Assessing the future landscape of scholarly communication: an exploration of faculty values and needs in seven disciplines. Berkeley: The Center for Studies in Higher Education, Univ Of California Press. Retrieved from http://escholarship.org/uc/item/15x7385g Harrington, S. (2014). Tweeting about the Telly: Live TV, Audiences and Social Media. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 237–247). New York: Peter Lang. Haustein, S. (2016). Grand challenges in altmetrics: heterogeneity, data quality and dependencies. Scientometrics. https://doi.org/10.1007/s11192-016-1910-9 Haustein, S., Bowman, T. D., & Costas, R. (2015a). "Communities of attention" around scientific publications: who is tweeting about scientific papers? Presented at the Social Media & Society 2015 International Conference, Toronto, Canada. Haustein, S., Bowman, T. D., & Costas, R. (2015b). When is an article actually published? An analysis of online availability, publication, and indexation dates. In Proceedings of the 15th International Society of Scientometrics and Informetrics Conference (pp. 1170–1179). Istanbul, Turkey. https://arxiv.org/abs/1505.00796 Haustein, S., Bowman, T. D., & Costas, R. (2016). Interpreting "altmetrics": viewing acts on social media through the lens of citation and social theories. In C. R. Sugimoto (Ed.), Theories of Informetrics and Scholarly Communication. A Festschrift in Honor of Blaise Cronin (pp. 372–405). Berlin: De Gruyter. http://arxiv.org/abs/1502.05701 Haustein, S., Bowman, T. D., Holmberg, K., Peters, I., & Larivière, V. (2014). Astrophysicists on Twitter: An in-depth analysis of tweeting and scientific publication behavior. Aslib Journal of Information Management, 66(3), 279–296. https://doi.org/10.1108/AJIM-09-2013-0081 Haustein, S., Bowman, T. D., Holmberg, K., Tsou, A., Sugimoto, C. R., & Larivière, V. (2016). Tweets as impact indicators: Examining the implications of automated "bot" accounts on Twitter. Journal of the Association for Information Science and Technology, 67(1), 232–238. https://doi.org/10.1002/asi.23456 Haustein, S., Bowman, T. D., Macaluso, B., Sugimoto, C. R., & Larivière, V. (2014). Measuring Twitter activity of arXiv e-prints and published papers. In altmetrics14: expanding impacts and metrics, Workshop at Web Science Conference 2014. https://doi.org/10.6084/m9.figshare.1041514%20 Haustein, S., & Costas, R. (2015a). Determining Twitter audiences: Geolocation and number of followers. In altmetrics15: 5 years in, what do we know? Amsterdam, The Netherlands. Retrieved from http://altmetrics.org/altmetrics15/haustein/ 35 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Haustein, S., & Costas, R. (2015b). Identifying Twitter audiences: who is tweeting about scientific papers? Presented at the Metrics 2015 ASIS&T SIG/MET Workshop. Retrieved from https://www.asist.org/SIG/SIGMET/wp-content/uploads/2015/10/sigmet2015_paper_11.pdf Haustein, S., Costas, R., & Larivière, V. (2015). Characterizing social media metrics of scholarly papers: The effect of document properties and collaboration patterns. PLoS ONE, 10(3), e0120495. https://doi.org/10.1371/journal.pone.0120495 Haustein, S., Larivière, V., Thelwall, M., Amyot, D., & Peters, I. (2014). Tweets vs. Mendeley readers: How do these two social media metrics differ? It - Information Technology, 56(5), 207–215. https://doi.org/10.1515/itit-2014-1048 Haustein, S., Minik, V., Brinson, D., Hayes, E., Costas, R., & Sugimoto, C. R. (2016). Identifying Twitter user communities in the context of altmetrics. In 3rd Altmetrics Conference 3:AM. Bucharest, Romania. Haustein, S., Peters, I., Bar-Ilan, J., Priem, J., Shema, H., & Terliesner, J. (2014). Coverage and adoption of altmetrics sources in the bibliometric community. Scientometrics, 101(2), 1145–1163. https://doi.org/10.1007/s11192-013-1221-3 Haustein, S., Peters, I., Sugimoto, C. R., Thelwall, M., & Larivière, V. (2014). Tweeting biomedicine: an analysis of tweets and citations in the biomedical literature. Journal of the Association for Information Science and Technology, 65(4), 656–669. https://doi.org/10.1002/asi.23101 Haustein, S., Sugimoto, C. R., & Larivière, V. (2015). Guest editorial: social media in scholarly communication. Aslib Journal of Information Management, 67(3). https://doi.org/10.1108/AJIM-03-2015- 0047 Hawkins, C. M., Duszak, R., & Rawson, J. V. (2014). Social media in radiology: early trends in Twitter microblogging at radiology's largest international meeting. Journal of the American College of Radiology : JACR, 11(4), 387–90. https://doi.org/10.1016/j.jacr.2013.07.015 Herman, B. (2014, September 5). Steven Salaita Twitter Scandal: University Offers Settlement, But Free Speech Questions Linger. International Business Times. Retrieved from http://www.ibtimes.com/steven- salaita-twitter-scandal-university-offers-settlement-free-speech-questions-linger-1678854 Holmberg, K. (2014). The impact of retweeting. Figshare. Retrieved from http://dx.doi.org/10.6084/m9.figshare.1037555 Holmberg, K., Bowman, T. D., Haustein, S., & Peters, I. (2014). Astrophysicists' conversational connections on twitter. PloS One, 9(8), e106086–e106086. https://doi.org/10.1371/journal.pone.0106086 Holmberg, K., & Thelwall, M. (2014). Disciplinary differences in Twitter scholarly communication. Scientometrics, 101(2), 1027–1042. https://doi.org/10.1007/s11192-014-1229-3 Honeycutt, C., & Herring, S. C. (2009). Beyond Microblogging: Conversation and Collaboration via Twitter (pp. 1–10). IEEE. https://doi.org/10.1109/HICSS.2009.89 Hussain, A. (2015). Adoption of Web 2.0 in library associations in the presence of social media. Program, 49(2), 151–169. https://doi.org/10.1108/PROG-02-2013-0007 Ingeno, L. (2013, June 4). Fat-Shaming in Academe. Inside Higher Ed. Retrieved from https://www.insidehighered.com/news/2013/06/04/outrage-over-professors-twitter-post-obese-students Jalali, A., & Wood, T. (2013). Analyzing Online Impact of Canadian Conference of Medical Education through Tweets. Education in Medicine Journal, 5(3). https://doi.org/10.5959/eimj.v5i3.162 Java, A., Song, X., Finin, T., & Tseng, B. (2007). Why we twitter: understanding microblogging usage and communities (pp. 56–65). ACM Press. https://doi.org/10.1145/1348549.1348556 Kamel Boulos, M. N., & Anderson, P. F. (2012). Preliminary survey of leading general medicine journals' use of Facebook and Twitter. Journal of the Canadian Health Libraries Association, 33(2), 38–47. Ke, Q., Ahn, Y.-Y., & Sugimoto, C. R. (2017). A systematic identification and analysis of scientists on Twitter. PLOS ONE, 12(4), e0175368. https://doi.org/10.1371/journal.pone.0175368 Kelly, R. (2009). Twitter study (White Paper). San Antonio, TX: Pear Analytics. Retrieved from http://pearanalytics.com/wp-content/uploads/2012/12/Twitter-Study-August-2009.pdf Kimmons, R., Veletsianos, G., & Woodward, S. (2017). Institutional Uses of Twitter in U.S. Higher Education. Innovative Higher Education, 42(2), 97–111. https://doi.org/10.1007/s10755-016-9375-6 Knight, S. R. (2014). Social media and online attention as an early measure of the impact of research in solid organ transplantation. Transplantation, 98(5), 490–6. https://doi.org/10.1097/TP.0000000000000307 Kortelainen, T., & Katvala, M. (2012). "Everything is plentiful-Except attention". Attention data of scientific journals on social web tools. Journal of Informetrics, 6(4), 661–668. https://doi.org/10.1016/j.joi.2012.06.004 Letierce, J., Passant, A., Breslin, J., & Decker, S. (2010). Understanding how Twitter is used to spread scientific messages. Retrieved from http://journal.webscience.org/314/ 36 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Leung, E. Y. L., Siassakos, D., & Khan, K. S. (2015). Journal Club via social media: authors take note of the impact of #BlueJC. BJOG: An International Journal of Obstetrics & Gynaecology, 122(8), 1042–1044. https://doi.org/10.1111/1471-0528.13440 Linvill, D. L., McGee, S. E., & Hicks, L. K. (2012). Colleges' and universities' use of Twitter: A content analysis. Public Relations Review, 38(4), 636–638. https://doi.org/10.1016/j.pubrev.2012.05.010 Liu, J., & Adie, E. (2013). Five Challenges in Altmetrics : A Toolmaker' s Perspective. Bulletin of Associatioon for Infomation Science and Technology, 39(4), 31–34. Loeb, S., Bayne, C. E., Frey, C., Davies, B. J., Averch, T. D., Woo, H. H., … Eggener, S. E. (2014). Use of social media in urology: data from the American Urological Association (AUA). BJU International, 113(6), 993–998. https://doi.org/10.1111/bju.12586 Lulic, I., & Kovic, I. (2013). Analysis of emergency physicians' Twitter accounts. Emergency Medicine Journal : EMJ, 30(5), 371–6. https://doi.org/10.1136/emermed-2012-201132 Maleki, A. (2014). Networking in science tweets linking to scholarly articles of Iran. Presented at the SIG/MET Workshop, ASIS&T 2014 Annual Meeting, Seattle. Retrieved from http://www.asis.org/SIG/SIGMET/data/uploads/sigmet2014/maleki.pdf Mccrea, P. (2011). Reflections on Academic Blogging as a Vehicle for Professional Development. In Proceedings of the European Conference on eLearning (p. 997). Brighton, UK. McKendrick, D. R., Cumming, G. P., & Lee, A. J. (2012). Increased Use of Twitter at a Medical Conference: A Report and a Review of the Educational Opportunities. Journal of Medical Internet Research, 14(6), e176. https://doi.org/10.2196/jmir.2144 Mehta, N., & Flickinger, T. (2014). The Times They Are A-Changin': Academia, Social Media and the JGIM Twitter Journal Club. Journal of General Internal Medicine, 29(10), 1317–1318. https://doi.org/10.1007/s11606-014-2976-9 Metaxas, P. T., & Mustafaraj, E. (2012). Social Media and the Elections. Science, 338(6106), 472–473. https://doi.org/10.1126/science.1230456 Micieli, R., & Micieli, J. a. (2012). Twitter as a tool for ophthalmologists. Canadian Journal of Ophthalmology. Journal Canadien D'ophtalmologie, 47(5), 410–3. https://doi.org/10.1016/j.jcjo.2012.05.005 Miller, G. (2011). Social Scientists Wade Into The Tweet Stream. Science, 333(6051), 1814–15. https://doi.org/10.1126/science.333.6051.1814 Miller, V. (2008). New Media, Networking and Phatic Culture. Convergence: The International Journal of Research into New Media Technologies, 14(4), 387–400. https://doi.org/10.1177/1354856508094659 Mishori, R., Levy, B., & Donvan, B. (2014). Twitter Use at a Family Medicine Conference : Analyzing #STFM13. Family Medicine, 46(8). Mou, Y. (2014). Presenting professorship on social media: from content and strategy to evaluation. Chinese Journal of Communication, 7(4), 389–408. https://doi.org/10.1080/17544750.2014.938669 Mowbray, M. (2014). Automated Twitter Accounts. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 183–94). New York: Peter Lang. Nason, G. J., O'Kelly, F., Kelly, M. E., Phelan, N., Manecksha, R. P., Lawrentschuk, N., & Murphy, D. G. (2015). The emerging use of Twitter by urological journals: Use of Twitter by urological journals. BJU International, 115(3), 486–490. https://doi.org/10.1111/bju.12840 Nature Publishing Group (2010). Media Options. Retrieved from: https://web.archive.org/web/20110513233735/http://www.nature.com/advertising/resources/pdf/2010nature mediakit.pdf?prod_code=NATURE&prod=Nature Nature Publishing Group. (2012). Announcement: A new iPad app for Nature readers. Nature, 492(7428), 154–154. https://doi.org/10.1038/492154a Nature Publishing Group. (2015). Author Insights 2015 survey. Figshare. Nentwich, M. (2011). Das Web 2.0 in der wissenschaftlichen Praxis. In Digitale Wissenschaftskommunikation–Formate und ihre Nutzung (pp. 35–53). Giessen. Retrieved from http://geb.uni-giessen.de/geb/volltexte/2011/8227/pdf/DigitaleWissenschaftskommunikation_2011.pdf Neylon, C. (2014, October 3). Altmetrics: What are they good for? Retrieved January 10, 2015, from http://blogs.plos.org/opens/2014/10/03/altmetrics-what-are-they-good-for/ Osterrieder, A. (2013). The value and use of social media as communication tool in the plant sciences. Plant Methods, 9(1), 26–26. https://doi.org/10.1186/1746-4811-9-26 Pasquini, L. A., & Evangelopoulos, N. (2015). Organizational identity, meaning, and values: analysis of social media guideline and policy documents (pp. 1–5). ACM Press. https://doi.org/10.1145/2789187.2789198 37 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Passmann, J., Boeschoten, T., & Schäfer, M. T. (2014). The gift of the gab. Retweet cartels and gift economies on Twitter. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 331–344). New York: Peter Lang. Pearce, N., Weller, M., Scanlon, E., & Ashleigh, M. (2010). Digital Scholarship Considered: How New Technologies Could Transform Academic Work. In Education, 16(1), 33–44. Peters, I. (2009). Folksonomies indexing and retrieval in Web 2.0. Berlin, Allemagne: De Gruyter/Saur. Pomerantz, J., Hank, C., & Sugimoto, C. R. (2015). The State of Social Media Policies in Higher Education. PLOS ONE. https://doi.org/10(5): e0127485. doi:10.1371 Ponte, D., & Simon, J. (2011). Scholarly Communication 2.0: Exploring Researchers' Opinions on Web 2.0 for Scientific Knowledge Creation, Evaluation and Dissemination. Serials Review, 37(3), 149–156. https://doi.org/10.1016/j.serrev.2011.06.002 Prabhu, V., & Rosenkrantz, A. B. (2015). Enriched Audience Engagement Through Twitter: Should More Academic Radiology Departments Seize the Opportunity? Journal of the American College of Radiology, 12(7), 756–759. https://doi.org/10.1016/j.jacr.2015.02.016 Priem, J. (2014). Altmetrics. In B. Cronin & C. R. Sugimoto (Eds.), Beyond bibliometrics: harnessing multidimensional indicators of performance (pp. 263–287). Cambridge, MA: MIT Press. Priem, J., Costello, K., & Dzuba, T. (2012). Prevalence and use of Twitter among scholars. Presented at the Metrics 2011 Symposium on Informetric and Scientometric Research. https://doi.org/http://dx.doi.org/10.6084/m9.figshare.104629 Priem, J., & Costello, K. L. (2010). How and why scholars cite on Twitter. Proceedings of the American Society for Information Science and Technology, 47(1), 1–4. https://doi.org/10.1002/meet.14504701201 Priem, J., Piwowar, H., & Hemminger, B. M. (2012). Altmetrics in the wild: Using social media to explore scholarly impact. arXiv Print, 1–17. Priem, J., Taraborelli, D., Groth, P., & Neylon, C. (2010, October 26). Altmetrics: A manifesto. Retrieved from http://altmetrics.org/manifesto/ Procter, R. N., Williams, R., Stewart, J., Poschen, M., Snee, H., Voss, A., & Asgari-Targhi, M. (2010). If you build it, will they come? How researchers perceive and use Web 2.0. London, UK: Research Network Information. Retrieved from http://wrap.warwick.ac.uk/56246 Pscheida, D., Albrecht, S., Herbst, S., Minet, C., & Köhler, T. (2013). Nutzung von Social Media und onlinebasierten Anwendungen in der Wissenschaft. ZBW–Deutsche Zentralbibliothek für Wirtschaftswissenschaften–Leibniz-Informationszentrum Wirtschaft. Retrieved from http://www.qucosa.de/fileadmin/data/qucosa/documents/13296/Science20_Datenreport_2013_PDF_A.pdf Puschmann, C. (2014). Opening Science. Opening Science, 89–106. https://doi.org/10.1007/978-3-319-00026-8 Ratkiewicz, J., Conover, M. D., Meiss, M., Gonçalves, B., Flammini, A., & Menczer, F. (2011). Detecting and tracking political abuse in social media. In Proceedings of the Fifth International AAAI Conference on Weblogs and Social Media (pp. 297–304). Ratkiewicz, J., Conover, M., Meiss, M., Gonçalves, B., Patil, S., Flammini, A., & Menczer, F. (2011). Truthy: mapping the spread of astroturf in microblog streams (p. 249). ACM Press. https://doi.org/10.1145/1963192.1963301 Reinhardt, W., Ebner, M., Beham, G., & Costa, C. (2009). How People are using Twitter during Conferences. In Creativity and Innovation Competencies on the Web, Proceeding of 5th EduMedia conference (pp. 145– 156). Salzburg: Hornung-Prähauser, V., Luckmann, M. (Ed.). Retrieved from http://lamp.tu- graz.ac.at/~i203/ebner/publication/09_edumedia.pdf Robinson-García, N., Torres-Salinas, D., Zahedi, Z., & Costas, R. (2014). New data, new possibilities: Exploring the insides of Altmetric.com. El Profesional de La Información, 23(4), 359–366. https://doi.org/10.3145/epi.2014.jul.03 Rogers, R. (2014). Debanalising Twitter. The Transformation of an Object of Study. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. ix–xxvi). New York: Peter Lang. Rothschild, S., & Unglesbee, B. (2013, September). Kansas University professor receiving death threats over NRA tweet. The Dispatch. Retrieved from http://www.shawneedispatch.com/news/2013/sep/24/kansas- university-professor-receiving-death-threat/ Rowlands, I., Nicholas, D., Russell, B., Canty, N., & Watkinson, A. (2011). Social media use in the research workflow. Learned Publishing, 24(3), 183–195. https://doi.org/10.1087/20110306 Schmidt, J.-H. (2014). Twitter and the Rise of Personal Publics. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. 3–12). New York: Peter Lang. 38 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Seglen, P. O. (1997). Citations and journal impact factors: questionable indicators of research quality. Allergy, 52(11), 1050–1056. https://doi.org/10.1111/j.1398-9995.1997.tb00175.x Severo, M., Giraud, T., & Pecout, H. (2015). Twitter data for urban policy making: an analysis on four European cities. In C. Levallois (Ed.), Handbook of Twitter for Research. EMLYON. Sharma, N. K., Ghosh, S., Benevenuto, F., Ganguly, N., & Gummadi, K. (2012). Inferring who-is-who in the Twitter social network (p. 55). ACM Press. https://doi.org/10.1145/2342549.2342563 Shuai, X., Pepe, A., & Bollen, J. (2012). How the scientific community reacts to newly submitted preprints: article downloads, Twitter mentions, and citations. PloS One, 7(11), e47523–e47523. https://doi.org/10.1371/journal.pone.0047523 Shulman, J., Yep, J., & Tomé, D. (2015). Leveraging the Power of a Twitter Network for Library Promotion. The Journal of Academic Librarianship, 41(2), 178–185. https://doi.org/10.1016/j.acalib.2014.12.004 Sopan, A., Rey, P. J., Butler, B., & Shneiderman, B. (2012). Monitoring Academic Conferences: Real-Time Visualization and Retrospective Analysis of Backchannel Conversations. 2012 International Conference on Social Informatics, (SocialInformatics), 62–69. https://doi.org/10.1109/SocialInformatics.2012.20 Sugimoto, C. R., Work, S., Larivière, V., & Haustein, S. (2017). Scholarly use of social media and altmetrics: a review of the literature. Journal of the Association for Information Science and Technology, 68(9), 2037– 2062. https://doi.org/10.1002/asi.23833 Takhteyev, Y., Gruzd, A., & Wellman, B. (2012). Geography of Twitter networks. Social Networks, 34(1), 73–81. https://doi.org/10.1016/j.socnet.2011.05.006 Taylor, M. (2013). Towards a common model of citation : some thoughts on merging altmetrics and bibliometrics. Research Trends, (35), 1–6. Tenopir, C., Volentine, R., & King, D. W. (2013). Social media and scholarly reading. Online Information Review, 37(2), 193–216. https://doi.org/10.1108/OIR-04-2012-0062 Thangasamy, I. A., Leveridge, M., Davies, B. J., Finelli, A., Stork, B., & Woo, H. H. (2014). International Urology Journal Club via Twitter: 12-Month Experience. European Urology, 66(1), 112–117. https://doi.org/10.1016/j.eururo.2014.01.034 Thelwall, M., Tsou, A., Weingart, S., Holmberg, K., & Haustein, S. (2013). Tweeting Links to Academic Articles. Cybermetrics: International Journal of Scientometrics, Informetrics and Bibliometrics, 17(1), 1–8. Topf, J. M., & Hiremath, S. (2015). Social media, medicine and the modern journal club. International Review of Psychiatry, 27(2), 147–154. https://doi.org/10.3109/09540261.2014.998991 Tsou, A., Bowman, T. D., Ghazinejad, A., & Sugimoto, C. R. (2015). Who tweets about science? In Proceedings of the 2015 International Society for Scientometrics and Informetrics (pp. 95–100). Istanbul, Turkey. Retrieved from https://pdfs.semanticscholar.org/81fe/8b63188cf25648a7c592bc6b5457fee3c101.pdf?_ga=1.184338726.126 4550827.1478885332 Twitter. (2017, April 6). Automation rules. Retrieved from https://support.twitter.com/articles/76915 Twitter. (n.d.). The Twitter Rules. Retrieved from https://support.twitter.com/articles/18311 Van Noorden, R. (2013). Brazilian citation scheme outed. Nature, 500(7464), 510–511. https://doi.org/10.1038/500510a Van Noorden, R. (2014). Online collaboration: Scientists and the social network. Nature, 512(7513), 126–129. https://doi.org/10.1038/512126a Veletsianos, G. (2012). Higher education scholars' participation and practices on Twitter. Journal of Computer Assisted Learning, 28(4), 336–349. https://doi.org/10.1111/j.1365-2729.2011.00449.x Veletsianos, G. (2013). Open practices and identity: Evidence from researchers and educators' social media participation. British Journal of Educational Technology, 44(4), 639–651. https://doi.org/10.1111/bjet.12052 Verishagen, N., & Hank, C. (2014). Are there birds in the library? The extent of Twitter adoption and use by Canadian academic libraries. First Monday, 19(11). https://doi.org/10.5210/fm.v19i11.4945 Wang, X., Xu, S., Peng, L., Wang, Z., Wang, C., Zhang, C., & Wang, X. (2012). Exploring scientists' working timetable: Do scientists often work overtime? Journal of Informetrics, 6(4), 655–660. https://doi.org/10.1016/j.joi.2012.07.003 Warzel, C. (2014, April 15). Meet The Man Behind Twitter's Most Infamous Phrase. BuzzFeed News. Retrieved from https://www.buzzfeed.com/charliewarzel/meet-the-man-behind-twitters-most-infamous- phrase?utm_term=.jwWokRoZw#.fnElNylG4 Weller, K., Bruns, A., Burgess, J., Mahrt, M., & Puschmann, C. (2014). Twitter and Society: An Introduction. In K. Weller, A. Bruns, J. Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (pp. xxiv– xxxviii). New York: Peter Lang. 39 Glänzel, Moed, Schmoch, & Thelwall (Eds.) Handbook of Quantitative Science and Technology Research Weller, K., Dornstädter, R., Freimanis, R., Klein, R. N., & Perez, M. (2010). Social software in academia: Three studies on users' acceptance of Web 2.0 Services. Retrieved from http://journal.webscience.org/360/2/websci10_submission_62.pdf Weller, K., Dröge, E., & Puschmann, C. (2011). Citation Analysis in Twitter : Approaches for Defining and Measuring Information Flows within Tweets during Scientific Conferences. In Proceedings of the ESWC2011 Workshop on "Making Sense of Microposts": Big things come in small packages (Vol. 718, pp. 1–12). Heraklion, Greece. Retrieved from http://ceur-ws.org/Vol-718/msm2011_proceedings.pdf Weller, K., & Puschmann, C. (2011). Twitter for Scientific Communication : How Can Citations / References be Identified and Measured ? In Proceedings of the ACM WebSci'11. Koblenz, Germany. Whitburn, T., Walshe, C., & Sleeman, K. E. (2015). International Palliative Care Journal Club on Twitter: Experience so far. BMJ Supportive & Palliative Care, 5(1), 120.2-120. https://doi.org/10.1136/bmjspcare- 2014-000838.48 White, H. D. (2000). Towards ego-centered citation analysis. In B. Cronin & H. B. Atkins (Eds.), The Web of Knowledge: A Festschrift in Honor of Eugene Garfield (pp. 475–496). Medford, NJ: Information Today. White, H. D. (2001). Authors as citers over time. Journal of the American Society for Information Science and Technology, 52(2), 87–108. https://doi.org/10.1002/1097-4571(2000)9999:9999<::AID-ASI1542>3.0.CO;2-T Wilkinson, C., & Weitkamp, E. (2013). A case study in serendipity: environmental researchers use of traditional and social media for dissemination. PloS One, 8(12), e84339–e84339. https://doi.org/10.1371/journal.pone.0084339 Wilson, M. W., & Starkweather, S. (2014). Web Presence of Academic Geographers: A Generational Divide? The Professional Geographer, 66(1), 73–81. https://doi.org/10.1080/00330124.2013.765290 Woolston, C. (2015). Potential flaws in genomics paper scrutinized on Twitter. Nature, 7553(521), 397. https://doi.org/10.1038/521397f Work, S., Haustein, S., Bowman, T. D., & Larivière, V. (2015). Social Media in Scholarly Communication. A Review of the Literature and Empirical Analysis of Twitter Use by SSHRC Doctoral Award Recipients (Study commissioned by the Social Sciences and Humanities Research Council). Yang, C., Harkreader, R. C., & Gu, G. (2011). Die Free or Live Hard? Empirical Evaluation and New Design for Fighting Evolving Twitter Spammers. In R. Sommer, D. Balzarotti, & G. Maier (Eds.), Recent Advances in Intrusion Detection (Vol. 6961, pp. 318–337). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-23644-0_17 Zedda, M., & Barbaro, A. (2015). Adoption of web 2.0 tools among STM publishers. How social are scientific journals? Journal of the EAHIL, 11(1), 9–12. Zhang, C. M., & Paxson, V. (2011). Detecting and analyzing automated activity on Twitter. In Proceedings of the 12th International Conference on Passive and Active Measurement PAM'11 (pp. 102–111). Berlin, Heidelberg: Springer. Retrieved from http://www.icir.org/vern/papers/pam11.autotwit.pdf Zhao, D., & Rosson, M. B. (2009). How and why people Twitter: the role that micro-blogging plays in informal communication at work (p. 243). ACM Press. https://doi.org/10.1145/1531674.1531710 40 Glänzel, Moed, Schmoch, & Thelwall (Eds.)
1606.03347	2	1606	2017-03-07T15:26:56	Evaluating balance on social networks from their simple cycles	[ "cs.SI", "cs.DM" ]	Signed networks have long been used to represent social relations of amity (+) and enmity (-) between individuals. Group of individuals who are cyclically connected are said to be balanced if the number of negative edges in the cycle is even and unbalanced otherwise. In its earliest and most natural formulation, the balance of a social network was thus defined from its simple cycles, cycles which do not visit any vertex more than once. Because of the inherent difficulty associated with finding such cycles on very large networks, social balance has since then been studied via other means. In this article we present the balance as measured from the simple cycles and primitive orbits of social networks. We specifically provide two measures of balance: the proportion $R_\ell$ of negative simple cycles of length $\ell$ for each $\ell\leq 20$ which generalises the triangle index, and a ratio $K_\ell$ which extends the relative signed clustering coefficient introduced by Kunegis. To do so, we use a Monte Carlo implementation of a novel exact formula for counting the simple cycles on any weighted directed graph. Our method is free from the double-counting problem affecting previous cycle-based approaches, does not require edge-reciprocity of the underlying network, provides a gray-scale measure of balance for each cycle length separately and is sufficiently tractable that it can be implemented on a standard desktop computer. We observe that social networks exhibit strong inter-edge correlations favouring balanced situations and we determine the corresponding correlation length $\xi$. For longer simple cycles, $R_\ell$ undergoes a sharp transition to values expected from an uncorrelated model. This transition is absent from synthetic random networks, strongly suggesting that it carries a sociological meaning warranting further research.	cs.SI	cs	IMA Journal of Complex Networks (2018) Page 1 of 26 doi:10.1093/comnet/xxx000 Evaluating balance on social networks from their simple cycles PIERRE-LOUIS GISCARD∗ University of York, Department of Computer Science, UK ∗[email protected] Universit´e de Nantes, Laboratoire de Math´ematiques Jean Leray, France PAUL ROCHET AND RICHARD C. WILSON University of York, Department of Computer Science, UK [Received on 17 October 2018] Signed networks have long been used to represent social relations of amity (+) and enmity (-) between individuals. Group of individuals who are cyclically connected are said to be balanced if the number of negative edges in the cycle is even and unbalanced otherwise. In its earliest and most natural formulation, the balance of a social network was thus defined from its simple cycles, cycles which do not visit any vertex more than once. Because of the inherent difficulty associated with finding such cycles on very large networks, social balance has since then been studied via other means. In this article we present the balance as measured from the simple cycles and primitive orbits of social networks. We specifically provide two measures of balance: the proportion R(cid:96) of negative simple cycles of length (cid:96) for each (cid:96) (cid:54) 20 which generalises the triangle index, and a ratio K(cid:96) which extends the relative signed clustering coefficient introduced by Kunegis. To do so, we use a Monte Carlo implementation of a novel exact formula for counting the simple cycles on any weighted directed graph. Our method is free from the double-counting problem affecting previous cycle-based approaches, does not require edge-reciprocity of the underlying network, provides a gray-scale measure of balance for each cycle length separately and is sufficiently tractable that it can be implemented on a standard desktop computer. We observe that social networks exhibit strong inter-edge correlations favouring balanced situations and we determine the corresponding correlation length ξ . For longer simple cycles, R(cid:96) undergoes a sharp transition to values expected from an uncorrelated model. This transition is absent from synthetic random networks, strongly suggesting that it carries a sociological meaning warranting further research. Keywords: Social networks; simple cycles; primitive orbits; and Monte-Carlo 2000 Math Subject Classification: 05C38, 90B10, 91D30, 91G60 1. Introduction 1.1 Balance in networks Relations of amity and enmity between individuals are well represented by signed networks, where an edge is assigned a positive value if two individuals are acquainted and in good terms, and a negative one if they are instead enemies [24, 9, 19, 37, 21]. Such networks provide a natural setting to study inter- personal relationships and their correlations. For example, one could expect that people are friendly towards the friends of their friends, a situation that is said to be "balanced". More generally, on signed c(cid:13) The author 2018. Published by Oxford University Press on behalf of the Institute of Mathematics and its Applications. All rights reserved. 2 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON networks, a group of individuals who are cyclically connected -- i.e. forming a triangle, a square, a pentagon etc. -- are said to be balanced if the number of negative edges in the cycle is even. Otherwise the cycle is said to be unbalanced. Sociologists have suggested that such negative cycles are the cause of tension and thus, that social networks should evolve into a state where balanced cycles are largely predominant [24, 9, 2, 3]. The question of whether this holds for real-social networks and if not, by how much this fails to be true, arose from these considerations in the 1940s [24]. Mathematically speaking, this sociological question translates into the following problem: on a signed network G, determine for all (cid:96) the percentage of negative simple cycles of length (cid:96). This prob- lem remains largely unsolved owing to its natural formulation in terms of simple cycles -- cycles which do not visit any vertex more than once. Unfortunately, enumerating all the simple cycles of a network exactly is computationally intractable. Indeed, this includes counting all the Hamiltonian cycles of the graph, a problem known to be #P-complete [8]. For this reason, we need to seek more efficient meth- ods, which inevitably lead to some approximations. Two strategies are implemented in this work: i) approximate the balance of the network to within any desired accuracy by evaluating the balance on a large sample of subgraphs of the network; or ii) compute the balance exactly from objects which are not simple cycles, but should carry a similar information. We successfully implemented the first strategy thanks to a novel exact formula for counting simple cycles on any (weighted directed) graph in conjunction with a Monte Carlo approach. This method is presented in Section 2. It effectively solves the mathematical problem enunciated earlier since the quality of the obtained approximation is controlled and can be improved at will. For the second strategy, we relied on the primitive orbits of the graph, cycles which contain no backtracking steps or tail, and are not the multiple of any other cycle. This is presented in Section 3. The results produced by both approaches on four social networks are discussed and compared in Section 4. 1.2 Notation Throughout this article, we consider signed directed networks G = (V ; E ), of which undirected net- works are a special case. The adjacency matrix of G is denoted AG or simply A. Each edge of the network is weighted with a value +1 or -1 indicating a positive or negative interaction. A cycle is posi- tive if the product of its edge values is positive, and otherwise it is negative. A cycle is simple if it does not visit any vertex more than once. The starting point of a simple cycle is irrelevant but its orientation is retained. For example, v0v1v2v0 and v1v2v0v1 represent the same triangle, which is however distinct from v0v2v1v0. The number of positive and negative simple cycles of length (cid:96) on G are designated by (cid:96) and N− N+ When discussing the balance of a network, we refer to the ratio R(cid:96) of the number of negatively signed simple cycles of length (cid:96) to the total number of simple cycles of length (cid:96), i.e. (cid:96) , respectively. R(cid:96) := N− N− (cid:96) + N+ (cid:96) (cid:96) , (cid:96) (cid:62) 1. In particular, R(cid:96) = 0 when the network is perfectly balanced for length (cid:96), while R(cid:96) = 1 indicates a totally unbalanced situation. To facilitate comparisons with existing results, we will also provide the ratio of negative to positive simple cycles U(cid:96) := N− (cid:96) /N+ (cid:96) and the relative signed clustering coefficient K(cid:96) := (N+ In the next section, we will see that, for (cid:96) = 3, the coefficients R(cid:96) and K(cid:96) are related to two measures of balance in social networks: the triangle index and the relative signed clustering coefficient (for triads), respectively. (cid:96) ) in Appendix C. (cid:96) + N+ (cid:96) − N− (cid:96) )/(N− EVALUATING BALANCE ON SOCIAL NETWORKS 3 of 26 1.3 Existing approachs The study of balance in social networks has a long history and has been studied via a variety of approaches and we can thus only mention the latest works on the subject here. Of fundamental impor- tance was the introduction of grayscale measures of balance, which quantify how balanced any given social network is [34, 31, 10]. Many such measures have been proposed, of which we shall briefly review a few important ones here. The signed clustering coefficient and algebraic conflict evaluate the balance of a network locally -- at the level of triangles -- and globally, respectively [31, 30]. The former is defined as the number T + of positive triangles minus the number T− of negative ones divided by the total number of triplets of vertices connected by at least two edges. In other terms, the signed clustering coefficient takes into account the sign and the tendency to form small clusters (the triangles), putting both considerations on an equal footing. The relative signed clustering coefficient separates the two contributions, being the ratio of T + − T− to the total number of triangles, that is exactly the coefficient K3 introduced earlier. The values of K(cid:96) for (cid:96) (cid:54) 20 provided in this study can thus be seen as an extension of the relative signed clustering coefficient to larger clusters of up to 20 individuals. In general, the relative signed clustering coefficient K3 is exactly accessible via the signed adjacency matrix, see Appendix C. The algebraic conflict is defined on Harary's basic premise that "in a balanced network, all vertices can be divided into two groups such that all positive edges connect vertices within the same group, and all negative edges connect vertices of the two different groups" [18]. Kunegis [31], observed that this condition was fulfilled if and only if the smallest eigenvalue of the graph signed Laplacian is nil. This eigenvalue, called the algebraic conflict, thus carries information about balance and its distance to zero measures the failure of the network to be perfectly balanced. Cycle-based approaches have also been devised, which evaluate various ratios involving the positive and negative cycles of a signed graph. These include the degree of balance and its weighted coun- terpart, arising either from weights attached to the graph edges or from additional length-dependent weights attached to the cycles, see e.g. [13]. The triangle index is the fraction of positive triangles on a graph to the total number of triangles. This index is 1−R3 and can therefore be seen as a special case of the balance ratios studied in this work. Formally, obtaining the triangle index is a tractable task taking at most O(N3) time, see Appendix C for a formula giving R3. This computational cost may nonetheless be too high on very large real-life networks. Instead, recent research has shown it could be reliably approximated by computing only the k most dominant eigenvalues of the signed adjacency matrix of the network, thereby reducing the time needed to O(N2k) [43]. Alternatively, the triangle index can be approximated via the Monte Carlo approach presented here. In a 2011 study, Facchetti, Iacono, and Altafini [14] studied the global balance of social networks using the number of edges whose sign must be changed so that all the simple cycles -- and hence all the cycles -- be positive. This measure and its close variants are known as the frustration index [26, 33] and line index of balance [19, 20]. Although these were believed to be computationally intractable to obtain exactly [14], recent research shows otherwise [39], putting these quantities within reach on networks with a few thousands nodes. By studying the frustration index, Facchetti et al. found that social networks are indeed strongly balanced.Their conclusions were called into question in a recent work by Estrada and Benzi [13] who rather concluded that social networks are not so well balanced using a completely different cycle-based measure of balance, while still other studies employing the frustration index have instead confirmed the conclusion of Facchetti et al. [3]. As noted earlier, since counting all the simple 4 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON cycles of a large graph exactly is intractable, one may instead count objects which are not simple but carry a similar information when it comes to balance. In this vein, Estrada and Benzi [13] proposed the use of D = Tr exp(A) = TrA(cid:96), (1.1) ∞ ∑ (cid:96)=0 1 (cid:96)! as a method of counting the number of balanced and unbalanced cycles, also known as closed walks, in a network. They show that, by computing the ratio K = D/D+ with D+ := Tr exp(A), called the weighted degree of balance, then the ratio of negative to positive cycles can be obtained as U walks := 1− K 1 + K . This is an extremely efficient method which simply requires the evaluation of the eigenvalues of the adjacency matrix. In practice, this could be expensive on very large networks, but it is sufficient to compute only a few of the largest eigenvalues of A, as D and D+ are dominated by those [13]. Expression (1.1) counts all closed walks (weighted by a factor 1/(cid:96)! for a walk of length (cid:96)). Because of this, backtracking steps and multiple cycles are counted. For example, the non-simple cycle v0v1v2v0v1v2v0 = (v0v1v2v0)2 is part of the sum. Such cycles are positive, and so they do not upset the balance of an already balanced network, but they do have an effect on the (global) balance ratio(s) for an unbalanced network. In other words, the cycle-sum embodied in (1.1) contains non-simple cycles from the order 2 onwards, effectively mixing the balance ratios R(cid:96) at all lengths. In addition, the global signature of balance Uwalks further mixes the contributions of the various cycles lengths. Figure 1 illustrates this issue in a triad. Whilst the network is completely unbalanced for triangles, we get U walks = 0.19, a number that is not easy to interpret. These difficulties cannot be resolved easily using walks. In an attempt to better account for the := + := TrA(cid:96) and Kwalks +, U walks := D(cid:96)/D(cid:96) (cid:96) (cid:96) )/(1 + Kwalks length dependency of the balance, we define D(cid:96) := TrA(cid:96), D(cid:96) (1− Kwalks + − D(cid:96) D(cid:96) 2D(cid:96) + Rwalks (cid:96) ) and := (cid:96) (cid:96) . (1.2) (cid:96) These quantities only take walks of length (cid:96) into account when calculating the balance. Yet, since short cycles and their multiples are typically much more abundant than long cycles, the values of U walks and Rwalks are still largely dominated by the contributions from self-loops, backtracks and triangles. Conse- (cid:96) quently Rwalks overestimates the proportion of balanced cycles. We demonstrate this concretely in Section 4, Figures (2) and (3), where we compare Rwalks will be depressed as compared to the true balance ratio R(cid:96), that is, Rwalks with R(cid:96) calculated from the simple cycles on two social networks. While one may empirically argue that long cycles are less relevant than short ones in real social networks [47, 13], it seems better to offer as detailed a mathematical analysis as possible before deciding this issue. For these reasons, we found it necessary to abandon the use of walks and rather recur either to the simple cycles themselves or to primitive orbits. (cid:96) (cid:96) (cid:96) 1.4 Motivating our approach 1.4.1 On using simple cycles Earlier this year, Aref and Wilson [3] have produced a global study of existing measures of balance. In particular, they recommended the use of the frustration index as a EVALUATING BALANCE ON SOCIAL NETWORKS 5 of 26 FIG. 1. An unbalanced network of three vertices. The dotted line represents a negative relationship. The exponential unbalance ratio is Uwalks = 0.19 while in fact R2 = 0 and R3 = 1. measure of balance and raised several issues affecting cycle-based methods, which we briefly discuss below. First, Aref and Wilson make an essential point concerning directed networks: many of the existing cycle-based studies do not take into account the fact that some edges are not reciprocated. It could be that an individual feels in good terms with someone who, in fact, feels negatively towards him. Or it could be that someone has passed a (positive or negative) judgment about someone else, whose attitude towards the first person is not reported in the data. This makes it difficult to interpret the results of approaches assuming edge-reciprocity, which should thus be, strictly speaking, limited to undirected networks. Second, Aref and Wilson write [3] "cycle-based measures [...] are difficult to compute and do not provide a proper range of values, whether weighted or not." Finally, they note that such measures must be improved to avoid or diminish the influence of closed-walks with repeated edges or cycles -- which they call cycle double-counting -- as these are always balanced. These observations are central to the present study, which addresses all of them. First, our approach is valid on directed signed networks, i.e. we do not assume nor need reciprocity of the edges1. Remark that an undirected graph is a special case of a directed one, where each directed edge ei j from i to j has a corresponding edge e ji of opposite orientation. It follows that any method for evaluating some indicator of balance that is valid on directed graphs gives rise to a valid approach for undirected graphs. Furthermore, a simple cycle on an undirected graph corresponds to two cycles with identical signs on the directed version of the same graph. Consequently, the balance ratios R(cid:96) can always be calculated from the directed form of an undirected graph. Second, our approach offers a proper range of values, since 0 (cid:54) R(cid:96) (cid:54) 1 for all (cid:96), which permits a gray-scale distinction between balanced or unbalanced situations on the one hand and the null hypothesis on the other. In the results, we will see that R(cid:96) typically varies between 0 and 0.6, exploiting much of its theoretical range. Third, our method is sufficiently tractable that it can be implemented on a standard desktop computer (see below), even on large networks, thanks to a recent breakthrough in the combinatorics of cycles. Finally, our simple-cycle based method is rigorously free from the double-counting problem. 1.4.2 On studying cycle-lengths separately Existing approaches to social balance tend to fall into two categories, both of which provide useful but incomplete information: (cid:73) Methods based on small structures in social networks, in particular triads, are only sensitive to 1In other terms, the situations described by Aref and Wilson are exactly taken into account in our analysis. This motivated us to define the balance ratios R(cid:96), for (cid:96) (cid:62) 1 rather than (cid:96) (cid:62) 3. And indeed, we will see below that R2 > 0 in three of the four networks analysed, indicating that the lack of edge-reciprocity does occur in practice -++6 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON local interactions in small groups of individuals. For this reason they are unable to detect long range correlations in relationships as well as characterise the dynamics of larger groups of individuals. (cid:73) Global measures of balance such as the degree of balance and algebraic conflict, are sensitive to social tensions occurring on all scales (that is group-sizes) but do not yield much information on where those tensions might be, as well as on the groups of individuals typically involved in balanced or unbal- anced relations. The approach recently taken by Facchetti and coworkers [14] is a step towards bridging this divide, as it provides both global and local information on the network balance. Yet, Estrada and Benzi have observed that the frustration index seems to be mostly determined by triads, leaving the influence of larger groups in the final result in doubt. In this work we have taken another route by extending two measures that have been developed for triads, the relative signed clustering coefficient and triangle index, to longer cycles. These longer cycles are sensitive to long range correlations between edge signs, and, by studying each length separately, provide a wealth of information on where social tension occurs by disentangling the influence of the various cycle-lengths on the overall balance. 2. Balance from simple cycles 2.1 Core combinatorial result One possible strategy to estimate the balance ratios R(cid:96) consists of approximating them from a large sample of subgraphs of the network under study. The main novelty permitting this straightforward approach in practice is a recently developed mathematical formula for counting simple cycles of any length on weighted directed graphs. Rather than sampling the simple cycles directly, the formula allows for a rapid and exact evaluation of the balance ratios on subgraphs of the original network. We show below that this strategy is much better from a computational standpoint than sampling the simple cycles themselves. 2.1.1 Formula for counting simple cycles Let P(z) be the ordinary generating function of the simple cycles of any weighted directed graph G, that is P(z) := ∑ c: simple cycle w(c)z(cid:96)(c), where w(c) is the weight of c, that is the product of the weights of its edges, z is a formal variable and (cid:96)(c) is the length of c. By exploiting algebraic structures associated with walks on graphs, P(z) can be shown to be (cid:16)(cid:0)zAH (cid:1)H(cid:0)I− zAH (cid:1)N(H)(cid:17) dz. (2.1) (cid:90) 1 z ∑ H≺G H connected Tr P(z) = In this expression H is a connected2 induced subgraph of G, AH its adjacency matrix, H the number of vertices in H and N(H) the number of neighbours of H in G. A neighbour of H in G is a vertex v of G which is not in H and such that there exists at least one edge, possibly directed, from v to a vertex of H or from a vertex of H to v. The result of Eq. (2.1), as well as further exact formulas for P(z), is presented in [16] and shall not be proven here. 2If G is directed, then the subgraphs H should be weakly connected induced subgraphs of G. Recall that a digraph is said to be weakly connected if replacing all its directed edges by undirected edges produces a connected undirected graph. EVALUATING BALANCE ON SOCIAL NETWORKS 7 of 26 2.1.2 Computational cost Let nH (G) be the number of connected induced subgraphs of a graph G and let nc(G) be the total number of simple cycles on G. The main advantage of Eq. (2.1) is that, if we use a direct algorithm for finding all the connected induced subgraphs of a graph, for exam- ple using reverse-search, the time complexity is O(cid:0)N +E + N2nH (G)(cid:1), with N and E the number O(cid:0)(N + E)(nc(G) + 1)(cid:1), and on undirected graphs there exists an algorithm achieving the optimal O(cid:0)N(nc(G) + 1)(cid:1) time complexity [6]. of vertices and of edges of the graph, respectively [4, 45, 11]. In comparison a direct search of the simple cycles themselves must have a time complexity scaling with nc(G). The best general purpose algorithm for finding such cycles, Johnson's 1975 landmark algorithm [29, 35], has time complexity In spite of the fact that these algorithms can find the sim- ple cycles, they are also commonly used when counting them is sufficient. In this situation however, Eq. (2.1) with its time complexity scaling with nH (G) represents a substantial speed-up since in gen- eral nH (G) (cid:28) nc(G). This is best demonstrated on the complete graph KN on N vertices -- which is the worst case scenario. On KN we have nc(G) (cid:39) e× (N − 1)! where e (cid:39) 2.71828 is the base of the natural logarithm, while nH (G) = 2N is "only" exponential3 Furthermore, in practice, the computational cost is much smaller as real networks are typically sparse. Besides Johnson's algorithm, the most efficient general purpose4 algorithm for counting simple cycles is that of Bax and Franklin [5]. This algorithm evaluates a formula counting the simple cycles that involves a sum over all induced subgraphs of the graph, as opposed to only the connected ones appearing in Eq. (2.1). On sparse graphs, there are far more induced subgraphs than connected induced subgraphs and evaluating Eq. (2.1) yields a significant speed-up as compared to Bax and Franklin's algorithm. Most importantly for applications, Eq. (2.1) is well suited to truncations: only those connected induced subgraphs H of G for which H (cid:54) (cid:96) (cid:54) H +N(H) can possibly contribute to the coefficient of z(cid:96) in P(z). This means that if one is interested in the first (cid:96) terms of P(z) -- that is in the simple cycles of length up to (cid:96) -- it suffices to consider those connected induced subgraphs of G with H (cid:54) (cid:96). Using reverse search, this has time complexity O(N +E + (cid:96)2nH:H(cid:54)(cid:96)(G)) (cid:54) O(N2 + (cid:96)2nH:H(cid:54)(cid:96)(G)) [45, 11]. Since furthermore only the small adjacency matrices AH enter Eq. (2.1), each term of the equation costs O(H3) (cid:54) O((cid:96)3) to evaluate. Thus, getting the first (cid:96) terms of P(z) from Eq. (2.1) has time complexity bounded by O((cid:96)3N(cid:96)) in the worst case scenario of the complete graph, and far less on sparse graphs. Instead, the computational cost on sparse graphs can be bounded as follows: let ∆ be the maximum vertex-degree on the network. Then the number of connected induced subgraphs on at most (cid:96) vertices is bounded by O(cid:0)N∆ (cid:96)/((∆ − 1)(cid:96)2)(cid:1) [45]. In addition, the time complexity of reverse search in this case order (cid:96) in O(cid:0)N∆ (cid:96)/(cid:96)(cid:1) time.5 For a more detailed analysis of the algorithmic implementation of Eq. (2.1) is O(N(∆ + 1) + N∆ (cid:96)) [4, 45, 11]. Consequently Eq. (2.1) can be shown to produce P(z) exactly up to together with comparison with other algorithms for the same task, see [38]. To give concrete examples, with an Intel Core i7-4790 CPU @ 3.60 GHz desktop computer, evaluat- ing Eq. (2.1) on the complete graph on 15 vertices took on average ∼0.7sec, yielding 255,323,504,932(cid:39) 3One should keep in mind that since P(z) determines the existence and number of Hamiltonian cycles on G, under the expo- nential time hypothesis [28] this exponential cost is, in principle, the best possible. 4In particular, the remarkable algorithm of Alon, Yuster and Zwick [1], which is extremely efficient for counting simple cycles, is not a general purpose one. Rather, it is limited to undirected graphs and cycle lengths of at most 7. 5The properties of Eq. (2.1) contrast it with other analytical formulas for counting simple cycles of small lengths [22, 36, 1]. First, these formulas work only on undirected graphs. Second, these formulas comprise a large number of terms, e.g. 160 terms for simple cycles of length 10, all of which involve the adjacency matrix of the full graph. P.-L. GISCARD, P. ROCHET AND R. C. WILSON 8 of 26 2.5× 1011 for the total number of simple cycles, which we verify analytically to be exact. This went up to ∼25sec for the complete graph on 20 vertices where 349,096,664,728,623,336 (cid:39) 3.5×1017 simple cycles were counted, a number that is, once again, exact. In both cases about half of the computer time was spent looking for the connected induced subgraphs and the other half implementing Eq. (2.1). For the real-world networks analysed below, a randomly chosen induced subgraph on 30 vertices is typically analysed in 0.001− 0.3seconds on the same computer, depending on its sparsity. 2.2 Monte Carlo implementation When the size of the network to study is large -- what "large" means here strongly depends on the graph sparsity -- an exact calculation of the desired terms of P(z) from Eq. (2.1) becomes intractable and the core combinatorial result must be supplemented by a Monte Carlo approach. The reader may have noticed upon close inspection of Eq. (2.1) that P(z) results from subtle cancel- lations between the contributions of the various connected induced subgraphs H of G. For this reason, Eq. (2.1) is not directly amenable to a Monte Carlo method which would consist of randomly selecting a sample of connected induced subgraphs H of the whole network G and estimating P(z) from this sam- ple. Eq. (2.1) can however be evaluated very quickly on graphs of "reasonable" size -- once again this depends on the sparsity of the underlying graph and the available computational resources. Our strategy is therefore to sample N induced subgraphs of the network under study and to calculate the balance ratios R(cid:96) exactly up to the desired length (cid:96) on each of these samples via Eq. (2.1). The average value of all the R(cid:96) then converges to that of the whole network as N grows. The quality of this approximation is appraised by repeating the whole procedure N(cid:48) times and extracting the standard deviation on the averaged R(cid:96). If this deviation is too large, the number N of samples is increased and the standard deviation is reevaluated. Once the deviation is below the desired accuracy, the method is deemed to have converged. We also systematically tested the method against bias by comparing it with exact results whenever available, see Appendix A. Firstly, we decided to limit our study to simple cycles of length up to 20. Longer cycles are both rare and largely irrelevant since, as will see, at such lengths (cid:96) > 20 all the social networks follow the null hypothesis. There remains three important parameters controlling the Monte Carlo approach which we can choose using a qualitative analysis: 1) Number of samples N: it is the total number of subgraphs on which the simple cycles are counted. The larger this number, the more simple cycles are found and the smaller the variance var(R(cid:96)) of the R(cid:96) ratios, that is the better the results. Following well established principles, we expect var(R(cid:96)) ∝ 1/(cid:112)Cycle(cid:96), Cycle(cid:96) being the total number of simple cycles of length (cid:96) found in all samples. 2) Size Nvertices of each sample: it is the number of vertices in each sampled subgraph. Since the longest simple cycle on a graph on Nvertices has length (cid:96) (cid:54) Nvertices, we must have Nvertices (cid:62) 20. Letting f(cid:96)(Nvertices) be the average number of simple cycles of length (cid:96) found on all sampled subgraphs, we have Cycle(cid:96) = Nsample f(cid:96)(Nvertices). In addition, we can expect f(cid:96) to be monotonous increasing function of Nvertices, in particular f(cid:96)(Nvertices) is at most O(N(cid:96) vertices) on complete graphs. 3) Time τ(cid:96)(Nvertices) spent per sample: we have shown in the preceding section that this time obeys τ(cid:96)(Nvertices) = O(cid:0)Nvertices∆ (cid:96)/(cid:96)(cid:1) , where ∆ is the average maximum degree of any vertex in the EVALUATING BALANCE ON SOCIAL NETWORKS 9 of 26 sampled subgraph. On average, ∆ (cid:54) Nvertices is expected to be a monotonous increasing function of Nvertices. Taking these three quantities into account, we end up with the following trade-off between accuracy (cid:32) (cid:33) , ⇐⇒ Total time = O N N(cid:96)+1 vertices (cid:96) var(R(cid:96)) ∝ of the results and time taken to reach them 1(cid:112)N f(cid:96)(Nvertices) diminish faster than(cid:112) f(cid:96)(Nvertices) (cid:54) N(cid:96)/2 from which we deduce that N should be large while Nvertices should be as small as possible. Indeed, the time taken to count simple cycles of length (cid:96) can scale as badly as N(cid:96)+1 vertices while the variance cannot vertices. In consequence increasing N is a better way to improve the accuracy of the results than increasing the size of each sampled subgraph, since the overall compu- tation time scales only linearly with N. Empirically, we confirmed the qualitative analysis outlined above as we found that a large number N of samples of size Nvertices ∼ 20− 30 vertices offered the best trade-off between accuracy and compu- tation time. Note, in practice τ20(Nvertices) is found to be between a millisecond and a tenth of a second time. 3. Balance from primitive orbits 3.1 Background A primitive orbit 6 on a network is a cycle which contains no backtracking steps or tail, and is not a recurrence of any other cycle, e.g. (v0v1v2v0)2 [42]. It is important to note that this is not the same as a simple cycle -- for example if c1 and c2 are simple cycles sharing an edge, then c1c2 is a primitive orbit. However, primitive orbits are identical to simple cycles up to order five 7. The (inverse) Ihara zeta function of a graph is given by [27] (cid:16) 1− z(cid:96)(c)(cid:17) ζ−1G (z) = ∏ c∈[C] , (3.1) where (cid:96)(c) is the length of the primitive orbit and G denotes the unsigned version of G. The notation [C] designates the set of equivalence classes of primitive orbits, i.e. the set of primitive orbits where all starting points on the same cycle are considered equivalent. The zeta function ζ−1G (z) can be expanded into terms representing each primitive orbit; to first order the expansion is ζ−1G (z) = 1− ∑c∈[C] z(cid:96)(c) + O(z6). Therefore, up to order 5, the zeta function corresponds to a sum over simple cycles, and non- simple cycle contamination only occurs at order 6 and higher. Because of this, the Ihara zeta function will provide a more precise measure of balance than a walk-based one. 6Unfortunately, primitive orbits are also known as "prime cycles", which is, strictly speaking, a misnomer. Indeed, primitive orbits do not satisfy the fundamental definition of a prime element, namely p is prime if and only if pa.b ⇐⇒ pa or pb for all a, b. Ironically, the only objects obeying this definition on a graph are the simple cycles, see [15]. them by replacing A by A− Diag(A) in the calculations of R(cid:96)(cid:62)3. 7This is true if and only if the graph has no self-loops, i.e. length 1 cycles. In the presence of such loops, we can simply remove 10 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON To approximate the balance ratios R(cid:96) using primitive orbits, we begin by introducing a modified version of the Ihara zeta function for signed networks (cid:16) 1− s(c)z(cid:96)(c)(cid:17) , ζ−1 G (z) = ∏ c∈[C] where s(c) is the sign of the primitive orbit c. We now need an efficient way of evaluating the zeta function. This can be achieved using its determinant form ζ−1 G (z) = det (I− zT) , where T is the Hashimoto matrix of the network, also known as edge adjacency matrix [23]. Let us recall the standard definition of T here. We designate by ei, 1 (cid:54) i (cid:54) E, an edge of the graph equipped with a randomly chosen orientation, and by e−1 the edge ei with the opposite orientation. Then T is the i 2E× 2E matrix defined as [42]: (cid:40) Ti j := 1, 0, otherwise. if and only if the end vertex of edge ei is the start vertex of edge e j and e j (cid:54)= e−1 i In other words, T is the adjacency matrix of the oriented line graph (OLG), that is its vertices correspond to the directed edges of the original graph. In the OLG, vertices are connected if there is an allowed two- step walk along the corresponding edges in the original graph and backtracking steps are not allowed (so ab,bc is allowed if ab and bc are edges, but ab,ba is not allowed). To incorporate the edge signs into the matrix, we use forward sign assignment. Since all terms in the zeta function are cycles, we can uniquely place the sign of edge ab into any edge in T which begins from ab, and the sign of the cycle in T will be the same as the sign of the original cycle. 3.2 Computation In principle, the number of primitive orbits of any length is easily determined from traces of powers of T. We have the following result, which we prove in Appendix B. PROPOSITION 3.1 Let G be a signed directed graph, T its Hashimoto adjacency matrix and N+ N− ob; (cid:96) be the number of positive and negative primitive orbits of length (cid:96) on G, respectively. Then ob; (cid:96) and ob; (cid:96) − N− N+ ob; (cid:96) = 1 (cid:96) ∑ d(cid:96) µ((cid:96)/d)TrTd, where µ(.) is the number-theoretic Mobius function. A similar result holds for N+ replacing T by T. ob; (cid:96) + N− ob; (cid:96) upon This formula is particularly revealing as to the connection between the Hashimoto matrix and the primitive orbits. Traces of powers of T count all cycles in the OLG, i.e. the backtrackless closed walks, including the so-called power orbits, e.g. (c1)2 and (c1c2)2, which are not primitive. The set of cycles of length (cid:96) contains such power orbits if and only if there are orbits whose length is a divisor of (cid:96). These divisors are removed via a Mobius inversion in the above sum. For moderately sized graphs, it is straightforward to compute T and its spectrum, and so to compute the number of primitive orbits of any length. However, since T is the adjacency of the OLG, its size is equal to the number of edges in the EVALUATING BALANCE ON SOCIAL NETWORKS 11 of 26 original network. In practice, it can therefore become difficult to compute the eigenvalues of this matrix. Since TrT = TrT2 = 0, we need only concern ourselves with third and higher powers of the eigenvalues. The spectrum can therefore be effectively truncated, considering only the largest magnitude eigenvalues, thereby reducing the computational burden. Nevertheless, this can be computationally demanding for large networks. For undirected unsigned graphs, Stark and Terras [41] provide a way to compute the number of all orbits (i.e. TrT(cid:96)), which we adapted to count the number of positive and negative primitive orbits in a signed but undirected network. Let D be the diagonal degree matrix of G and let Q = D− I. Further let A+ be the adjacency of only the positive edges in G and similarly A− be the adjacency of only the negative edges in G (coded as -1). Then the following iteration counts the number of positive and negative backtrackless closed walks (orbits) between any two vertices, denoted W + ob; (cid:96) and W− ob; (cid:96), 2 = A+A+ + A−A− − (Q + I), A+ 2 = A−A+ + A+A−, A− (cid:96)−1A+ + A− (cid:96)−1A− − A+ A+ (cid:96) = A+ (cid:32) A− (cid:96) = A− (cid:96)−1A− − A− (cid:96)−1A+ + A+ ((cid:96)−1)/2 (cid:32) ∑ j=1 ((cid:96)−1)/2 ∑ A− (cid:96) − (Q− I) (cid:96) − (Q− I) A+ ob; (cid:96) = Tr W− ob; (cid:96) = Tr W + A− (cid:96)−2 j j=1 (cid:96)−2Q, (cid:96)−2Q, A+ (cid:96)−2 j (cid:33) (cid:33) , . ob; (cid:96) −W− ob; (cid:96) +W− ob; (cid:96) = TrT(cid:96), using these results in conjunction with Propo- Since W + sition 3.1 counts the primitive orbits. This method is very efficient thanks to its use of A rather than T, but remains limited to undirected networks. ob; (cid:96) = TrT(cid:96) and W + Armed with the number of positive and negative primitive orbits of length (cid:96), we compute the ob; (cid:96)). We will see that these provide better approx- primitive-orbits-based ratios Rob imations of the true balance ratios R(cid:96) than walk-based ones Rwalks (cid:96)(cid:62)3 := N− ob; (cid:96)/(N− ob; (cid:96) +N+ . (cid:96) 4. Results In this section we present the evolution of the balance ratios R(cid:96) with (cid:96) on several social networks. As detailed in §1.4, we aim, on each network, at determining the evolution of R(cid:96) with (cid:96). We will see that on all the large social networks studied here, R(cid:96) exhibits a sharp transition around (cid:96) ∼ 10− 12, which is absent from synthetic networks. We discuss the relation between this observation and past studies in §4.7, concluding with a possible structural interpretation for it. Tables with full numerical results for R(cid:96), K(cid:96) and U(cid:96) are included in Appendix C. 4.1 Data sets Following the precedent studies by Facchetti et al. [14], Estrada and Benzi [13] and Aref and Wilson [3], we have analysed four social networks: i) Gahuku-Gama with 16 vertices [44]; ii) WikiElections 12 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON with 8297 vertices [46]; iii) Slashdot with 82,144 vertices [40]; and iv) Epinions with 131,828 vertices [12]. Note that among these, only the Gahuku-Gama network is undirected. 4.2 Null-hypothesis In order to meaningfully determine if social networks are balanced, we compare our results to the balance that would be obtained on a graph with the same proportion p of negative directed edges than the real network under study, but where the sign of any directed edge is negative with probability p. In particular, in the null-hypothesis model, the signs of any two directed edges are independent random variables. Then the probability that a simple cycle c of length (cid:96) be negative is Prob(c negative) = (cid:100)(cid:96)/2(cid:101)−1 ∑ i=0 (cid:18)(cid:96) (cid:19) i p2i+1(1− p)(cid:96)−2i−1. (4.1) Supposing for simplicity that the signs of any two simple cycles are independent random variables then the probability distribution for N− (cid:96) ) in the null-hypothesis is a binomial law with expectation value Rnull given by Eq. (4.1). Consequently, in this simple model the null-hypothesis is compatible up to a near 95% confidence level with any value of R(cid:96) within the 2σ interval (cid:96) /(N− (cid:96) +N+ (cid:96) (cid:113) (cid:96) ± 2 Rnull (cid:113) (1− Rnull Rnull (cid:96) (cid:96) N− (cid:96) + N+ (cid:96) ) . (4.2) The assumption that the signs of any two simple cycles are independent random variables is not true on real social networks. Calculating the null-hypothesis without this assumption is very difficult in practice however. Indeed, a more accurate null model is given by evaluating the average balance ratios of all lengths over all random shufflings of the edges-signs from the social network under study. We implemented this more accurate model on the WikiElections network and found it to yield null balance ratios that are up to 9% lower than the values predicted by Eq. (4.1) when (cid:96) (cid:46) 10, while differences diminish for longer simple cycles. Yet, all the conclusions that can be drawn from comparing the simple null model Eqs. (4.1, 4.2) with the computed balance ratios are unchanged, since the relative positions of the two are unaltered by the more accurate model. 4.3 Gahuku-Gama network The Gahuku-Gama network represents the relation between sixteen tribes living in the eastern central highlands of New-Guinea [17]. Since the network is very small, the Monte Carlo approach is not necessary and we obtained the exact balance from simple cycles of all lengths thanks to Eq. (2.1). The results are shown on Fig. (2). Observe how the balance ratios Rwalks respectively calculated from the walks and primitive orbits overestimate the proportion of balanced cycles. and Rob (cid:96) The exact results show that up to length (cid:96) = 7, the actual ratio R(cid:96) is well below that of the null- hypothesis, indicating strong inter-edges correlation in favour of balanced cycles. This observation can be made more precise on noting that the balance is well fitted by a simple exponential model (cid:96) =(cid:0)1− e−((cid:96)−2)/2ξ(cid:1), Rmodel (cid:96) (4.3) EVALUATING BALANCE ON SOCIAL NETWORKS 13 of 26 FIG. 2. Exact percentage R(cid:96) of negatively signed simple cycles on the Gahuku-Gama network calculated from Eq. (2.1) (red squares) as compared to the null-hypothesis (blue circles). The dashed black curve is a simple exponential fit of R(cid:96) yielding the correlation length ξ (cid:39) 2.08. Also shown are the percentages Rwalks;(cid:96) (blue triangles) and Rob;(cid:96) (black diamonds) calculated from the walks and primitive orbits, respectively. where ξ (cid:39) 1.04 is the correlation length. In this context, ξ characterises the distance between two edges of the graph such that their signs are correlated. Note, the maximum distance between any two vertices on a cycle of length (cid:96) is (cid:98)(cid:96)/2(cid:99), hence Eq. (4.3) fits 2ξ . This indicates that tribes of the Gahuku-Gama network are mostly sensitive to the relations with all their first degree neighbours. Furthermore, while the network is less balanced than might seem to be the case when considering only the triangles, Fig. (2) suggests that much of the imbalance is shifted to long-length simple cycles. In particular, the rebound of R(cid:96) above 50% for 7 < (cid:96) (cid:46) 13 shows that long negative simple cycles are over-represented as compared to a totally uncorrelated model (the null hypothesis). This means that negatively signed situations are slightly favoured in cyclic groups involving 7 < (cid:96) (cid:46) 13 tribes. We will discuss the significance of this observation in the Section 4.7. 4.4 WikiElections network The WikiElections network represents the votes of Wikipedia users during the elections of other users to adminship. The network is obtained as follows: when a user votes against the candidate, an edge with a negative weight is created from the voting user to the candidate. If instead the user is neutral or supports the candidate, a positive weight is given to this edge. The network counts 8,297 vertices and is thus too large for a direct exact calculation of the balance ratios beyond (cid:96) = 6 and we employed a Monte Carlo approach in tandem with Eq. (2.1). In total we evaluated the balance on 1,800,000 graphs on 20 vertices. The results are shown on Fig. (3). We find the balance ratio R(cid:96) to evolve with (cid:96) in three major phases, which we will also observe on NullhypothesisExactRℓRℓobRℓwalksFitRℓmodel234567891011121314151601020304050CyclelengthRℓ(%)14 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON FIG. 3. Computed percentage of negatively signed simple cycles on the WikiElection network for cycle length up to 17 (red line and error bars). The blue shaded region bordered by dashed blue lines shows the values of R(cid:96) compatible with the null-hypothesis, as determined by Eqs. (4.1, 4.2). Also shown are the percentages Rwalks (blue triangles) and Rob (black diamonds) calculated from (cid:96) the walks and primitive orbits, respectively. In inset, a subgraph of the WikiElections network. (cid:96) the Slashdot and Epinions networks. For short simple cycles (cid:96) (cid:46) 12, R(cid:96) increases slowly and smoothly with (cid:96) and is also well approximated by the primitive orbits results. In addition, within this range of cycle lengths, R(cid:96) is always much smaller than predicted by the null-hypothesis, witnessing a strong inter-edge correlation in favour of balance. A sharp transition to R(cid:96) values circa 50% then occurs around (cid:96) ∼ 12−14. This transition is not an artefact of our algorithm as it is not observed on synthetic networks, see Appendix A for details. The transition demonstrates that at (cid:96) ∼ 12− 14, the simple cycle length becomes longer than twice the inter-edge correlation length ξ , which must thus be around 6−7.8 Following the sharp transition, R(cid:96) is reliably found to be over 50%, only to slowly decay to results consistent with null-hypothesis 9. This last behaviour, which is also present on the Gahuku-Gama network, suggests that much of the imbal- ance is shifted to long simple cycles for which edges signs appear to be weakly correlated in favour of imbalance. 8Indeed, although the network is directed, many of its edges are bidirectional so that the maximum distance between any two vertices on a cycle of length (cid:96) is around (cid:98)(cid:96)/2(cid:99). We emphasise that ξ = 6− 7 does not mean that individuals participating in the WikiElections network are sensitive to all the relations between their neighbours up the 6th or 7th degree. Rather, ξ only provides an upper bound on the depth of the correlation. This is because simple cycles of length (cid:96) typically sustain shortcuts which lower the average distance between the individuals participating in the cycle. The inset of Fig. 3 illustrates this phenomenon with a subgraph of the WikiElections network sustaining an octagon, but where the average distance between any two vertices is only ∼ 2.5. 9This effect becomes even more pronounced when R(cid:96) is compared with the more accurate null model that takes the structural correlations between simple cycles into account. NullhypothesisRℓRℓobRℓwalk12345678910111213141516170102030405060SimplecyclelengthRℓ(%)EVALUATING BALANCE ON SOCIAL NETWORKS 15 of 26 FIG. 4. Computed percentage of negatively signed simple cycles on the Slashdot network for cycle length up to 18 (red line and error bars). The blue shaded region bordered by dashed blue lines shows the values of R(cid:96) compatible with the null-hypothesis, as determined by Eqs. (4.1, 4.2). Also shown is the percentage Rob (cid:96) determined from the primitive orbits of the graph (black diamonds). Slashdot network 4.5 The Slashdot network is a large directed graph on 82,144 vertices representing relations of amity/enmity between the users of the Slashdot website [32, 31]. For the Monte Carlo implementation of Eq. (2.1), we sampled 20,000,000 graphs on 20 vertices from this network. We present the balance ratio R(cid:96) up to (cid:96) = 20 on Fig. (4). The balance ratio on this network exhibit a behaviour similar to that observed on the WikiElection network: at first R(cid:96) increases smoothly with (cid:96) (cid:54) 10. Then R(cid:96) undergoes a rapid transition to higher values broadly consistent with the null hypothesis. This indicates a correlation length between the edges of ξ (cid:39) (cid:98)11/2(cid:99) = 5 and thus a correlation depth of 5 or less. It is also remarkable that the balance ratios for 14 (cid:54) (cid:96) (cid:54) 16 are once more notably higher than 50%, indicating, as in the Gahuku-Gama and WikiElections networks, that much of the imbalance is shifted to long-length cycles. 4.6 Epinions The Epinions network is a large directed graph on 131,828 vertices representing relations between the users of the consumer review website Epinions.com. For the Monte Carlo implementation of Eq. (2.1), we sampled 1,000,000,000 (one billion) graphs on 30 vertices from this network. We present the result- ing balance ratio R(cid:96) up to (cid:96) = 15 on Fig. (5). We were not able to compute the balance ratio Rob (cid:96) using primitive orbits, owing to the very large size of this network. Indeed, recall that in order to compute the number of primitive orbits, one needs the spectrum of the Hashimoto edge adjacency matrix of the net- work. This matrix has size E×E, with E the number of (directed) edges of the graph. Counting the primitive orbits thus takes O(E3) time and on Epinions E = 841372. Preliminary calculations indi- cate that it would take well over a week of computation to get sufficiently many eigenvalues to obtain a NullhypothesisRℓRℓob234567891011121314151617180102030405060SimplecyclelengthRℓ(%)16 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON reliable approximation of Rob case. (cid:96) . For this reason, we decided not to seek the primitive orbits results in this Broadly speaking, the balance ratio R(cid:96) behaves similarly on this network as it does on the WikiElec- tions and Slashdot ones. The transition of R(cid:96) from small to high values indicates a correlation length ξ circa 10/2 = 5, and thus a correlation depth of 5 or less. Strikingly, for 4 (cid:54) (cid:96) (cid:54) 9, R(cid:96) is almost constant around 15% witnessing a very strong, almost length-independent, inter-edge correlation. 4.7 Comparison with past studies and interpretation of the results Past studies of the WikiElections, Slashdot and Epinions networks have shown them to be very well balanced or quite unbalanced, depending on the measure studied. Kunegis [30] found that these three networks were globally well balanced, with algebraic conflicts on the order of 10−3 that is a hundred time smaller than in the case of Gama-Gahuku. Kunegis further confirmed this with the signed clus- tering and relative signed clustering coefficients, which are sensitive to local structures at the level of triads, and were found to be at least five times smaller on WikiElections, Slashdot and Epinions than on Gama-Gahuku. Similar conclusions had been reached by Facchetti et al. [14], who showed using the frustration index that the percentage of perfectly balanced nodes was well over 50% on the three online social networks. At the opposite, Estrada and Benzi [13] who studied the degree of balance, argued that these net- works are less-well balanced than concluded so far.10 In particular they found highly unbalanced sub- networks of individuals in WikiElections, Slashdot and Epinions, which we can put in relation with the results presented here. Indeed, Estrada and Benzi predict "the existence of many negative cycles respon- sible for the global lack of balance" in these networks. They further remark that finding and studying these negative cycles would be extremely expensive computationally. It is tempting to relate the negative cycles predicted by Estrada and Benzi and the cyclic groups of over circa 10 individuals responsible for the overrepresentation of unbalanced simple cycles at such lengths as visible in our results. This suggests a possible structural interpretation for the sharp transition of R(cid:96) and the ensuing over- representation of unbalanced cycles that we have observed. Consider a network comprising well bal- anced, relatively dense clusters of individuals with negative edges mostly located in between these clusters. Then most short simple cycles will exist inside the dense clusters, consequently appearing strongly balanced. At the opposite, simple cycles whose lengths exceed the typical cluster size, being simple, are forced to cross cluster boundaries, thereby appearing mostly unbalanced. As a corollary, the highly unbalanced subnetworks of individuals are those that include individuals from different well balanced clusters and are visited mostly by long simple cycles. This kind of cluster structure can be confirmed in the case of the Gama-Gahuku network. There, Kunegis' in-depth analysis has shown that the network comprises three dense, perfectly balanced clusters with negative edges only between them. The average size of these clusters is 5.3 with the largest one regrouping 7 tribes, in line with our analysis which shows that R(cid:96) (cid:38) 50% from (cid:96) = 7 onwards. 10The difference of conclusion might originate from that the degree of balance is sensitive to cycles of all lengths, when the signed clustering and relative signed clustering coefficients are only influenced by triads. Estrada and Benzi argued that a similar observation could be made for the frustration index, which appeared to follow closely what could be predicted from triads only. EVALUATING BALANCE ON SOCIAL NETWORKS 17 of 26 FIG. 5. Computed percentage of negatively signed simple cycles on the Epinions network for cycle length up to 15 (red line and error bars). The blue shaded region bordered by dashed blue lines shows the values of R(cid:96) compatible with the null-hypothesis, as determined by Eqs. (4.1, 4.2). 5. Conclusion 5.1 Balance Ever since relations of amity and enmity between individuals have been modelled as positive and nega- tive signs attached to the edges of a network, the possibility of finding patterns in these signs has arisen. In one of the earliest formulation of the theory of balance, Heider proposed that most groups of individ- uals whose relationships form a cycle, should evolve into a state where this cycle has a positive overall sign [24]. While much research on the subject has appeared since then, altering the concept of social balance and of Heider's initial conjecture as well as questioning its validity, this very earliest statement remained largely unverified owing to its natural formulation in terms of simple cycles. Contrary to closed walks, that is "ordinary" cycles, simple cycles may not visit any vertex more than once and most problems concerning them are difficult, indeed NP-hard and #P-complete. We thus saw Heider's earliest conjecture as a well posed challenge of network analysis and while its relevance to modern sociology is open for debate, simple cycles undeniably offer several advantages over other cycle-based studies. They are indeed free from the double-counting problem, work on undi- rected and directed networks alike, and the resulting balance indicators K(cid:96) and R(cid:96) take on a range of values, offering a gray-scale measure of the state of the network. Furthermore, distinguishing the con- tributions of the cycles based on length bridges the divide between local and global balance indices, since e.g. doing so probes edge-correlations on a wide range of length scales. Consequently we sought, as network analysts, to provide the most detailed possible study of social balance using hitherto unac- cessible objects, the simple cycles, via novel combinatorial techniques. We hope that our results will stimulate further research into simple cycles on the one hand, and will be exploited by sociologists in the light of the latest developments in this field on the other. NullhypothesisRℓ1234567891011121314150102030405060SimplecyclelengthRℓ(%)18 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON We have studied the proportion R(cid:96) of negative simple cycles of length (cid:96) (cid:54) 20 and its variants K(cid:96) and U(cid:96); so as to be as close as possible to Heider's original conjecture, while also relating our results to the latest research. For exemple, 1− R3 equals the triangle index and K3 is the relative signed clustering coefficient. Our results show that R(cid:96) is greatly depressed as compared to an independent sampling sce- nario (the null hypothesis), typically up to lengths of circa (cid:96) ∼ 10. This means that balanced situations are strongly favoured when short simple cycles are considered. It is interesting that on the three large networks analysed here (Wikielections, Slashdot and Epinions), a rapid transition from balance (small ) occurs around (cid:96) ∼ 10. This is a signature of strong inter-edges cor- R(cid:96) values) to random (R(cid:96) ∼ Rnull relations with correlation length ξ (cid:39) 10/2 = 5. The correlation depth, which quantifies the degree up to which individuals are correlated with their neighbours, is thus less than or equal to 5. A rebound of the balance ratio over 50% following the transition is also clearly detectable in the data. This means that unbalanced situations are over-represented for long-length simple cycles as compared to the null- hypothesis. (cid:96) While the relevance of cyclic clusters of individuals in social networks beyond triads is currently controversial, we believe that these do contain sociological information: we observe there is no mathe- matical reason why R(cid:96) should undergo a sharp transition as seen in the data. In particular, this transition does not exist in the null-hypothesis model, where edge signs are attributed randomly to a real social network, nor does it exist on synthetic networks where both edges and their signs are random. We there- fore hope that researchers with a more sociological background than ours will consider the results of this work. 5.2 Functions on simple cycles and simple paths Technically, the approach presented here to study the balance in networks is generally applicable to estimate any function of the simple cycles of a graph. Furthermore, the core combinatorial result of our method immediately extend to vertex-specific questions, e.g. for evaluating the balance of the simple cycles passing through some specified vertex. It also remains valid when asking questions pertaining to simple paths (also known as self-avoiding walks). Both of these observations stem from a matrix extension of Eq. (2.1) which is presented in [16]. This extension provides a matrix P(z) whose entry P(z)i j is the ordinary generating function of the simple paths from i to j (i (cid:54)= j) or of the simple-cycles from i to itself (i = j). Note, the matrix extension is not obtained upon just removing the trace from Eq. (2.1). This matrix formulation should permit the calculation of such functions as the all-paths kernel [7] for data-mining algorithms. The all-paths kernel is a positive semi-definite function on pairs of graphs aimed at measuring their degree of similarity via the similarity between their simple paths: kall paths(G,G(cid:48)) = ∑ p simple path on G ∑ p(cid:48) simple path on G(cid:48) k(p, p(cid:48)), with k(p, p(cid:48)) a kernel on pairs of paths. Another example is the cyclic pattern kernel [25] which evaluates the similarities between the simple cycles of a pair of graphs. Both of these examples currently suffer from efficiency problems which one should overcome (at least partially) with Eq. (2.1) and its matrix extension, possibly in conjunction with a Monte Carlo approach as effected here. EVALUATING BALANCE ON SOCIAL NETWORKS 19 of 26 Acknowledgment This work was supported by Royal Commission for the Exhibition of 1851 [research fellowship to P.-L. G.] We thank two anonymous referees for their numerous and detailed comments which have helped improve this paper. A. Checking the method against bias While the convergence of our calculations can be assessed via the standard deviation of the results, Monte Carlo approaches are not immune to bias. It is thus necessary to verify the quality of the results independently of the method itself: i) We computed the exact balance ratios R1, R2 and R3 via conventional means and verified our results to be consistent with these, see Appendix C, Tables A.4, A.7 and A.10. In the case of WikiElections, we also computed R4, R5 and R6 exactly using Eq. (2.1). ii) When available, we used the primitive orbits results to verify that the balance ratios R4 and R5 predicted by the algorithm were consistent with the exact results. Indeed, recall that up to (cid:96) = 5, Rob (cid:96) = R(cid:96) exactly. iii) On the Gahuku-Gama network, we verified that the Monte Carlo results are consistent with the exact balance ratios at all lengths. In addition, we found the WikiElections, Slashdot and Epinions networks to exhibit sharp transitions of their balance ratios from low values R(cid:96) ∼ 10− 30% up to values consistent with the null-hypothesis circa 50%. Given the importance of this observation, it is necessary to check that it is not an artifact of the algorithm we employed: i) In the case of the WikiElections network, we reallocated the edge signs randomly and ran our approach on the resulting signed graph. The balance ratios did not exhibit any sharp transition anymore but rather were consistent with the more accurate null model. This indicates that the transition is not an artifact of our method. ii) We should expect a transition of R(cid:96) to values consistent with the null hypothesis as the cycle length becomes longer than the correlation length. Indeed, if there was no sharp transition, a simple extrapolation of the trend exhibited by the first five (exactly known) balance ratios, suggests that R(cid:96) would not be reach 50% until at least (cid:96) (cid:38) 50. This conservative estimate would mean that ξ (cid:38) 25 or more, a number that is far too large to be plausible. B. Proof of Proposition 3.1 The proposition results from equating the product and determinant forms of the Ihara zeta function. Recall that 1− z(cid:96)(cid:17)Nob; (cid:96) (cid:16) , = ∏ (cid:96) 1− z(cid:96)(c)(cid:17) (cid:16) (cid:16) 1− z(cid:96)(cid:17)Nob; (cid:96) ζ−1G (z) = ∏ c∈[C] ζ−1G (z) = ∏ (cid:96) where Nob; (cid:96) = N+ ob; (cid:96) + N− ob; (cid:96) is the total number of primitive orbits of length (cid:96). Thus we have = det (I− zT) . (A.1) 20 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON Now, on taking the logarithm on both sides we obtain ∞ ∑ i=1 1 i zi TrTi = ∞ ∑ j ∞ ∑ k=1 Nob; j zk j k . Equating the coefficient of z(cid:96) on left and right hand sides then gives TrT(cid:96) = 1 (cid:96) ∞ ∑ k(cid:96) 1 k Nob; (cid:96)/k, and a Mobius inversion finally provides Nob;(cid:96) ob; (cid:96) + N− N+ ob; (cid:96) = µ((cid:96)/k)TrTk. 1 (cid:96) ∑ k(cid:96) The proof is entirely similar on signed networks, where T replaces T and N+ instead of N+ ob; (cid:96) + N− ob; (cid:96). ob; (cid:96) − N− ob; (cid:96) is obtained C. Full numerical results In this section we present the full numerical results obtained on the four networks mentioned earlier, which include the balance ratios, R(cid:96) = (N− (cid:96) ), ratios of negative to positive simple cycles U(cid:96) = N− (cid:96) ) up to (cid:96) = 20. We also give the exact values of R(cid:96), U(cid:96) and K(cid:96) for the self-loops, backtracks and triangles ((cid:96) = 1,2,3). These are respectively given by (cid:96) and the relative signed clustering coefficient K(cid:96) := (N+ (cid:96) )/(N− (cid:96) − N− (cid:96) )/(N− (cid:96) + N+ (cid:96) + N+ (cid:96) /N+ R1 = U1 = K1 = Tr(A− A) 2TrA Tr(A− A) Tr(A + A) Tr(A) Tr(A) , , R2 = , U2 = K2 = where A = A− Diag(A). Tr( A2 − A2) 2Tr A2 Tr( A2 − A2) Tr( A2 + A2) Tr( A2) Tr( A2) , , R3 = , U3 = K3 = , , Tr( A3 − A3) 2Tr A3 Tr( A3 − A3) Tr( A3 + A3) Tr( A3) Tr( A3) , EVALUATING BALANCE ON SOCIAL NETWORKS 21 of 26 Exact Exact Exact Exact Self-loops R1 Backtracks R2 0% R7 48.64% R12 51.77% R17 R6 44.47% R11 52.43% R16 49.74% Triangles R3 13.24% Squares R4 27.92% Pentagons R5 37.93% R8 51.10% R13 51.03% R18 R9 52.33% R14 50.46% R19 R10 52.7% R15 50.09% R20 Table A.1. Exact balance ratios R(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 obtained on the Gama network using Eq. (2.1). Exact Exact Exact Exact Self-loops U1 Backtracks U2 0% U7 94.7% U12 107.3% U17 U6 80.1% U11 110.2% U16 99% Triangles U3 15.3% Squares U4 38.7% Pentagons U5 61.1% U8 104.5% U13 104.2% U18 U9 109.8% U14 101.9% U19 U10 111.4% U15 100.3% U20 Table A.2. Exact ratio of negative to positive simple cycles U(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 obtained on the Gama network using Eq. (2.1). Exact Exact Exact Exact Self-loops K1 Backtracks K2 1 K7 0.027 K12 −0.035 K17 K6 0.111 K11 −0.049 K16 0.005 Triangles K3 0.735 Squares K4 0.442 Pentagons K5 0.241 K8 −0.022 K13 −0.021 K18 K9 −0.047 K14 −0.009 K19 K10 −0.054 K15 −0.002% K20 Table A.3. Exact relative signed cluster coefficient K(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 obtained on the Gama network using Eq. (2.1). Recall that −1 (cid:54) K(cid:96) (cid:54) 1 and that K(cid:96) = 1 designates a totally balanced situation, while K(cid:96) = −1 characterises a totally unbalanced one. 22 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON Monte Carlo Exact Monte Carlo Exact Monte Carlo Monte Carlo Self-loops R1 45.488± 0.048% 45.455% Backtracks R2 3.436± 0.006% 3.438% Triangles R3 13.075± 0.014% 13.068% Squares R4 16.862± 0.098% 16.875% Pentagons R5 20.421± 0.052% 20.393% R6 23.43± 0.05% 23.401% R11 33.95± 0.9% R7 26.22± 0.03% R8 28.84± 0.19% R9 31.2± 0.31% R10 32.82± 0.54% R12 35.89± 0.9% R13 40.96± 0.9% R14 50.16± 0.19% R15 57.3± 0.36% R16 54.38± 3.41% R17 53.85± 4.57% R18 R19 R20 Table A.4. Computed balance ratios R(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the WikiElections network together with twice the standard deviation exhibited by the Monte Carlo results. We found no simple cycle of length 18, 19 or 20 on this network. Monte Carlo Exact Monte Carlo Monte Carlo Monte Carlo Self-loops U1 83.45± 0.17% 83.33% Backtracks U2 3.56± 0.006% 3.56% Triangles U3 15.04± 0.02% 15.03% Squares U4 20.3± 0.14% Pentagons U5 25.7± 0.08% U6 30.7± 0.05% U7 35.5± 0.06% U8 40.5± 0.37% U9 45.3± 0.7% U10 48.9± 1.2% U15 134.2± 2% U11 51.4± 2% U16 119± 15% U12 56± 2.2% U17 116± 20% U13 69.4± 2.5% U14 100.7± 0.8% U18 U19 U20 Table A.5. Computed ratio of negative to positive simple cycles U(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the WikiElections network together with twice the standard deviation exhibited by the Monte Carlo results. We found no simple cycle of length 18, 19 or 20 on this network. Monte Carlo Exact Monte Carlo Monte Carlo Self-loops K1 0.0902± 0.0014 0.0909 Backtracks K2 0.9312± 0.0001 0.9312 Triangles K3 0.7385± 0.0004 0.7386 Squares K4 0.663± 0.003 Pentagons K5 0.592± 0.001 K6 0.53± 0.001 K11 0.321± 0.025 K7 0.476± 0.001 K12 0.282± 0.025 K8 0.423± 0.005 K13 0.181± 0.025 K9 0.376± 0.009 K10 0.344± 0.015 K14 −0.003± 0.005 K15 −0.146± 0.01 Monte Carlo K16 −0.088± 0.096 K17 −0.077± 0.129 K18 K19 K20 Table A.6. Computed relative signed cluster coefficient K(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the WikiElections network together with twice the standard deviation exhibited by the Monte Carlo results. EVALUATING BALANCE ON SOCIAL NETWORKS 23 of 26 Monte Carlo Exact Self-loops R1 Backtracks R2 3.9995± 0.001% 4.0003% Triangles R3 6.352± 0.017% 6.3608% Squares R4 7.108± 0.013% Pentagons R5 8.192± 0.013% Monte Carlo R6 9.37± 0.06% R7 10.55± 0.12% R8 11.82± 0.15% R9 13.3± 0.09% R10 15.32± 0.46% Monte Carlo Monte Carlo R11 16.5± 0.8% R16 54.5± 4.3% R12 21.3± 2.6% R17 50.0± 6.6% R13 32.7± 1.5% R18 50.0± 9.4% R14 52.9± 3.6% R15 58.4± 2.6% R19 R20 Table A.7. Computed balance ratios R(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Slashdot network together with twice the standard deviation exhibited by the Monte Carlo results. We found no simple cycle of length 19 or 20 on this network. Monte Carlo Exact Self-loops U1 Backtracks U2 4.166± 0.001% 4.167% Triangles U3 6.783± 0.019% 6.793% Squares U4 7.552± 0.015% Pentagons R5 8.923± 0.015% Monte Carlo U6 10.34± 0.07% U7 11.79± 0.15% U8 13.40± 0.19% Monte Carlo Monte Carlo U11 19.8± 1.1% U16 120± 20% U12 27.1± 4.1% U17 100.0± 25% U13 48.6± 3.2% U18 100± 31% U9 15.3± 0.12% U14 112.3± 6.7% U19 U10 18.09± 0.64% U15 140± 15% U20 Table A.8. Computed ratio of negative to positive simple cycles U(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Slashdot network together with twice the standard deviation exhibited by the Monte Carlo results. We found no simple cycle of length 19 or 20 on this network. Monte Carlo Exact Monte Carlo Monte Carlo Monte Carlo Self-loops K1 K6 0.813± 0.002 K11 0.67± 0.023 K16 −0.09± 0.122 Backtracks K2 0.92001± 0.00003 0.92000 Triangles K3 0.8730± 0.0005 0.8728 Squares K4 0.8578± 0.0004 Pentagons R5 0.8362± 0.0004 K7 0.789± 0.003 K12 0.574± 0.074 K17 0± 0.187 K8 0.764± 0.004 K13 0.346± 0.042 K18 0.001± 0.266 K9 0.734± 0.003 K10 0.694± 0.013 K14 −0.058± 0.102 K15 −0.168± 0.074 K19 K20 Table A.9. Computed relative signed cluster coefficient K(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Slashdot network together with twice the standard deviation exhibited by the Monte Carlo results. 24 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON Monte Carlo Exact Monte Carlo Monte Carlo Monte Carlo Self-loops R1 6.1082± 0.0002% 6.1082% Backtracks R2 2.0857± 0.0003% 2.0858% Triangles R3 11.2355± 0.0016% 11.2343% Squares R4 14.17± 0.002% Pentagons R5 15.61± 0.007% R6 16.07± 0.02% R7 15.99± 0.12% R8 15.97± 0.58% R9 17.30± 1.58% R10 21.83± 2.84% R11 43.8± 2.3% R16 R12 48.1± 0.5% R17 50% R13 52.9± 1.2% R18 R14 62.2± 0.4% R15 47.6± 0.4% R19 44.4% R20 Table A.10. Computed balance ratios R(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Epinions network together with twice the standard deviation exhibited by the Monte Carlo results. We found no simple cycle of length 16, 18 or 20 on this network. Furthermore, we were unable to determine the standard deviation of the balance for R17 and R19. Regardless of the accuracy on these results, the small numbers of simple cycles of such lengths that we found imply that the null-hypothesis is compatible with all values 15% (cid:46) R(cid:96) (cid:46) 85%, as per Eq. (4.2). Monte Carlo Exact Monte Carlo Monte Carlo Monte Carlo Self-loops U1 6.5056± 0.0003% 6.5056% Backtracks U2 2.1301± 0.0004% 2.1302% Triangles U3 12.658± 0.0028% 12.656% Squares U4 16.51± 0.004% Pentagons U5 18.50± 0.014% U6 19.15± 0.04% U11 78.2± 10.3% U16 U7 19.03± 0.24% U12 92.7± 2.6% U17 100% U8 19.01± 1.16% U13 112.5± 7.7% U18 U9 20.96± 3.27% U10 28.09± 5.58% U14 164.6± 4.0% U15 90.9± 2.1% U19 79.9% U20 Table A.11. Computed ratio of negative to positive simple cycles U(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Epinions network together with twice the standard deviation exhibited by the Monte Carlo results. Monte Carlo Exact Monte Carlo Monte Carlo Monte Carlo Self-loops K1 0.87784± 0.00001 0.87784 Backtracks K2 0.95829± 0.00001 0.95828 Triangles K3 0.77529± 0.00005 0.77531 Squares K4 0.7166± 0.0001 Pentagons K5 0.6878± 0.0002 K6 0.679± 0.001 K11 0.124± 0.065 K16 K7 0.68± 0.003 K12 0.038± 0.014 K17 0 K8 0.681± 0.016 K9 0.654± 0.045 K10 0.563± 0.08 K13 −0.058± 0.034 K14 −0.244± 0.011 K15 0.048± 0.011 K18 K19 0.112 K20 Table A.12. Computed relative signed cluster coefficient K(cid:96) for 1 (cid:54) (cid:96) (cid:54) 20 on the Epinions network together with twice the standard deviation exhibited by the Monte Carlo results. EVALUATING BALANCE ON SOCIAL NETWORKS 25 of 26 REFERENCES [1] Alon, N., Yuster, R. & Zwick, U. (1997) Finding and counting given length cycles. Algorithmica, 17, 209 -- 223. [2] Antal, T., Krapivsky, P. L. & Red, S. (2006) Social balance on networks: The dynamics of friendship and enmity. Physica D: Nonlinear Phenomena, 224, 130 -- 136. [3] Aref, S. & Wilson, M. C. (2015) Measuring Partial Balance in Signed Networks. arXiv:1509.04037. [4] Avis, D. & Fukuda, K. (1996) Reverse search for enumeration. Discrete Applied Mathematics, 65, 21 -- 46. [5] Bax, E. & Franklin, J. (1996) A finite-difference sieve to count paths and cycles by length. Information Processing Letters, 60(4), 171 -- 176. [6] Birmel´e, E., Ferreira, R., Grossi, R., Marino, A., Pisanti, N., Rizzi, R. & Sacomoto, G. (2013) Optimal listing of cycles and st-paths in undirected paths. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1884 -- 1896. [7] Borgwardt, K. M. & Kriegel, H.-P. (2005) Shortest-path kernels on graphs. In Proceedings of the 5th IEEE International Con- ference on Data Mining (ICDM 2005), 27-30 November 2005. [8] Burgisser, P., Clausen, M. & Shokrollahi, M. A. (1997) Algebraic Complexity Theory, volume 315 of Grundlehren der mathematischen Wissenschaften. Springer, Berlin; New York. [9] Cartwright, D. & Harary, F. (1956) Structural balance: a generalization of Heider's theory. Psychological Review, 63, 277 -- 293. [10] Chiang, K.-Y., Hsieh, C.-J., Natarajan, N. & Dhillon, I. S. (2014) Prediction and Clustering in Signed Net- works: A Local to Global Perspective. Journal of Machine Learning Research, 15, 1177 -- 1213. [11] Elbassioni, K. (2015) A Polynomial Delay Algorithm for Generating Connected Induced Subgraphs of a Given Cardinality. Journal of Graph Algorithms and Applications, 19, 273 -- 280. [12] Epinions social network (2016) Epinions social network. http://snap.stanford.edu/data/ soc-sign-epinions.html. [13] Estrada, E. & Benzi, M. (2014) Walk-based measure of balance in signed networks: Detecting lack of balance in social networks. Physical Review E, 90, 042802. [14] Facchetti, G., Iacono, G. & Altafini, C. (2011) Computing global structural balance in large-scale signed social networks. Proceedings of the National Academy of Sciences, 108, 20953 -- 20958. [15] Giscard, P.-L. & Rochet, P. (2016a) Algebraic Combinatorics on Trace Monoids: Extending Number Theory to Walks on Graphs. arXiv:1601.01780. [16] Giscard, P.-L. & Rochet, P. (2016b) Enumerating simple paths from connected induced subgraphs. arXiv:1606.00289. [17] Hage, P. (1973) A graph theoretic approach to the analysis of alliance structure and local grouping in highland New Guinea. Anthropological Forum: A Journal of Social Anthropology and Comparative Sociology, 3, 280 -- 294. [18] Harary, F. (1953) On the notion of balance of a signed graph. Michigan Math. J., 2, 143 -- 146. [19] Harary, F. (1959) On the measurement of structural balance. Behavioral Science, 4, 316 -- 323. [20] Harary, F. (1960) A matrix criterion for structural balance. Naval Research Logistics, 7, 195 -- 199. [21] Harary, F. & Kabell, J. A. (1980) A simple algorithm to detect balance in signed graphs. Mathematical Social [22] Harary, F. & Manvel, B. (1971) On the number of cycles in a graph. Matematick´y casopis, 21, 55 -- 63. [23] Hashimoto, K. (1989) Zeta functions of finite graphs and representations of p-adic groups. Stud. Pure Math., Sciences, 1, 131 -- 136. 15, 211 -- 280. [24] Heider, F. (1946) Attitudes and cognitive organization. The Journal of Psychology, 21, 107 -- 112. [25] Horv´ath, T., Gartner, T. & Wrobel, S. (2004) Cyclic pattern kernels for predictive graph mining. In Proceed- ings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. [26] Iacono, G., Ramezani, F., Soranzo, N. & Altafini, C. (2010) Determining the distance to monotonicity of a biological network: a graph-theoretical approach. IET Systems Biology, 4, 223 -- 235. [27] Ihara, Y. (1966) On discrete subgroups of the two by two projective linear group over p-adic fields. Journal 26 of 26 P.-L. GISCARD, P. ROCHET AND R. C. WILSON of the Mathematical Society of Japan, 18, 219 -- 235. [28] Impagliazzo, R. & Paturi, R. (2001) On the Complexity of k-SAT. Journal of Computer and System Sciences, 62, 367 -- 375. [29] Johnson, D. B. (1975) Finding all the elementary circuits of a directed graph. SIAM J. Comput., 4, 77 -- 84. [30] Kunegis, J. (2014) Applications of Structural Balance in Signed Social Networks. arXiv:1402.6865. [31] Kunegis, J., Schmidt, S., Lommatzsch, A. S., Lerner, J., DeLuca, E. W. & Albayr, S. (2010) Spectral analysis of signed graphs for clustering, prediction and visualization. SIAM Conference on Data Mining (Society for Industrial and Applied Mathematics, Philadelphia), pages 559 -- 570. [32] Lampe, C. A., Johnston, E. & Resnick, P. (2007) Follow the reader: Filtering comments on Slashdot. Proceed- ings of Computer/Human Interaction 2007 Conference (Association for Computing Machinery, New York), pages 1253 -- 1262. [33] Lange, C., Liu, S., Peyerimhoff, N. & Post, O. (2015) Frustration index and Cheeger inequalities for discrete and continuous magnetic Laplacians. Calculus of variations and partial differential equations, 54, 4165 -- 4196. [34] Leskovec, J., Huttenlocher, D. & Kleinberg, J. (2010) Signed networks in social media. Conference on Human Factors in Computing Systems. [35] Mateti, P. & Deo, N. (1976) On Algorithms for Enumerating all Circuits of a Graph. SIAM J. Comput., 5, 90 -- 99. [36] Movarraei, N. & Boxwala, S. A. (2016) On the Number of Cycles in a Graph. Open Journal of Discrete Mathematics, pages 41 -- 69. [37] Norman, R. Z. (1972) A derivation of a measure of relative balance for social structures and a characterization of extensive ratio systems. Journal of Mathematical Psychology, 9, 66 -- 91. [38] P.-L. Giscard, N. Kriege, R. C. W. (2016) A general purpose algorithm for counting simple cycles and simple paths of any length. arXiv:1612.05531. [39] S. Aref, A. J. Mason, M. C. W. (2016) An exact method for computing the frustration index in signed networks using binary programming. arXiv:1611.09030. [40] Slashdot social network, February 2009 (2016) Slashdot social network, February 2009. http://snap. stanford.edu/data/soc-sign-Slashdot090221.html. [41] Stark, H. M. & Terras, A. A. (1996) Zeta functions of finite graphs and coverings. Advances in Mathematics, 121, 124 -- 165. [42] Terras, A. A. (2011) Zeta Functions of Graphs: A Stroll through the Garden. Cambridge University Press, Cambridge, 1st edition. [43] Terzi, E. & Winkler, M. (2011) A spectral algorithm for computing social balance. Proceedings of the 8th international conference on Algorithms and models for the web graph, pages 1 -- 13. [44] UCINET IV Datasets (2016) UCINET IV Datasets, Read Highland Tribes. http://vlado.fmf. uni-lj.si/pub/networks/data/ucinet/ucidata.htm. [45] Uehara, R. (1999) The number of connected components in graphs and its applications. IEICE Technical Report, COMP99-10. [46] Wikipedia adminship election data (2016) Wikipedia adminship election data. http://snap. stanford.edu/data/wiki-Elec.html. [47] Zajonc, R. B. & Burnstein, E. (1965) Structural balance, reciprocity, and positivity as sources of cognitive bias. Journal of Personality, 33, 570 -- 583.
1912.00176	1	1912	2019-11-30T10:24:28	Generalized Reputation Computation Ontology and Temporal Graph Architecture	[ "cs.SI", "cs.CY" ]	The problem of reliable democratic governance is important for survival of any community, and it will be more and critical over time communities with levels of social connectivity in society rapidly increasing with speeds and scales of electronic communication. In order to face this challenge, different sorts of rating and reputation systems are being develop, however reputation gaming and manipulation in such systems appears to be serious problem. We are considering use of advanced reputation system supporting "liquid democracy" principle with generalized design and underlying ontology fitting different sorts of environments such as social networks, financial ecosystems and marketplaces. The suggested system is based on "weighted liquid rank" algorithm employing different sorts of explicit and implicit ratings being exchanged by members of the society. For the purpose, we suggest "incremental reputation" design and graph database used for implementation of the system. Finally, we present evaluation of the system against real social network and financial blockchain data.	cs.SI	cs	Знания-Онтологии-Теории (ЗОНТ-19) Generalized Reputation Computation Ontology and Temporal Graph Architecture Anton Kolonin1,2 1Stichting SingularityNET, Barbara Strozzilaan 362 (Eurocenter II), 1083, HN, Amsterdam. 2Novosibirsk State University, 1 Pyrogova str., Novosibirsk, 630090, Russia. [email protected] Annotation. The problem of reliable democratic governance is important for survival of any community, and it will be more and critical over time communities with levels of social connectivity in society rapidly increasing with speeds and scales of electronic communication. In order to face this challenge, different sorts of rating and reputation systems are being develop, however reputation gaming and manipulation in such systems appears to be serious problem. We are considering use of advanced reputation system supporting "liquid democracy" principle with generalized design and underlying ontology fitting different sorts of environments such as social networks, financial ecosystems and marketplaces. The suggested system is based on "weighted liquid rank" algorithm employing different sorts of explicit and implicit ratings being exchanged by members of the society. For the purpose, we suggest "incremental reputation" design and graph database used for implementation of the system. Finally, we present evaluation of the system against real social network and financial blockchain data. Keywords: graph database, liquid democracy, system, social system ontology multi-agent environment, reputation 1 Introduction The entire history of human communities shows that no reliable solution for reaching truly long- term democratic consensus in society has been invented so far [1]. Different sorts of social organizations and governance policy have been tried, starting with completely centralized governance in from of "monarchy" and ending with completely distributed "anarchy". In any case, the crucial part of social organization design remains principles of reaching social consensus recognized and accepted by entire society. One form of consensus is known to be based on brute force in animal groups and ancient societies, serving the minority having the access to the force. The same solution is reproduced in nowadays distributed computing systems such as blockchains and called Proof-of-Work (PoW). More advanced form of consensus employed by human race now is based of financial capabilities of members of society. It is known to lead to the situation when "reacher become richer" and gain more and more power. This is is also replicated with the same phenomena observed in latest developments of blockchain systems relying on Proof-of-Stake (PoS). Further, in some of the latest blockchain systems, the employed solution is so-called Delegated Proof-of-Stake (DPoS). The latter solution still implies the rule on basis of financial capabilities being implemented indirectly, by means of manually controlled voting process to selected delegates who conduct the governance of the system. The limited list of the options above leads to the situation that consensus in any community or a distributed computing system may be easily taken over by a group which concentrates substantial amount of power (be it physical, military or computational one) of financial resources. Obviously, the latter group may be minority organized towards the goals hostile to the majority of the community. So, the better and more reliable and fair forms of consensus are in demand. The one of them is so-called Proof-of-Reputation (PoR) [2] consensus being discussed further. 2 Reputation System Concept The suggested Reputation Consensus [2] implementing Proof-of-Reputation (PoR) principle, opposing power of brute force (PoW), power of money (PoS or DPoS), as shown on Fig. 1. We anticipate that the Proof-of-Reputation can make it possible to implement system of Liquid Democracy to cure known issues of representative democracy, influenced by power of money. In this case the governing power of member of a human or artificial society depends on Reputation of the member earned on basis of the following principles.  Reputation may be computed by means of different measures, called "ratings", performed explicitly or implicitly by all members of community, called "raters", in respect to ones who Reputation is being computed for, called "ratees", with account to Reputations of the "raters" themselves.  Reputation computation is time-scoped, so that measures collected by a ratee in the past are less contributing to its current Reputation than the latest ones, which have more impact.  Data used to compute the Reputations of all raters and ratees with the ratings that they issue or receive is open so that audit of Reputations and the historical measures over the history can be performed in order to prevent Reputation manipulation and gaming. Fig. 1. Types of consensus in distributed systems such as Proof-of-Work, Proof-of-Stake and Proof-of-Reputation. There may be many kinds of explicit and implicit measures contributing to evaluation of Reputation considered, depending on a kind of practical use case and implementation of a given Reputation system [3, 4, 5, 6, 7]. For instance, it could be system serving social network [8] or a marketplace [9]. Applicability of the measures or ratings may depend on accuracy and reliability that they may provide as well as resistance to attack vectors targeting takeover of the consensus by means of reputation cheating and gaming for specific case. In the current work we are trying to come up with generic purpose architecture so are going to enumerate all possible options. Respectively, following the work [2] we consider such measures as: a) members explicitly staking financial values on other members; b) members explicitly providing ratings in respect to transactions committed with other members; c) implicit ratings computed from the financial values of transactions between the members; d) evaluation of textual, audial and video reviews or mentions made by members in respect to other members or transactions between them. 3 Reputation System Ontology Generic purpose ontology for a Reputation System serving an online community may be depicted on Fig. 2 with the following entities identified (some of them show on the Fig.2 and some omitted).  Account -- primary entity of a Reputation System to play a "rater" or "ratee" role or both, can be impersonating a physical person, business or governmental entity, robotic system etc.; Smart Contract -- secondary entity specific to blockchain environments which may belong to some of the Accounts; Product/Service -- secondary entity specific to marketplace environments identifying products or services provided by some of the Accounts; Post/Comment -- secondary entity specific to social networks or messaging environments representing any sort of textual, audial or any other sensible non-financial communication;     Word -- tertiary entity specific to social networks or messaging environments identifying a word used in a post/comment and carrying positive or negative sentiment which can be purposed to assess its impact on reputation; Tag/Category -- any classification of any of the entities identified above.  Fig. 2. Simplified ontology of the generic-purpose Reputation System for online environments. Besides the entities above, there are relationship connecting them, such as the following.  Votes -- any "Vote", "Like" or "Rate" event providing assessment of any communication on behalf of a "rater" Account;  Authors -- authorship of a communication;  Mentions -- mentioning of an Account in a communication;  Uses -- use of a Word in a communication;  Relates -- relevance of a Post/Comment or a Product/Service (not shown on Fig.1) to specific Tag/Category;      Provides -- provision of a Product/Service (not shown on Fig.1) on behalf of an Account; Follows -- following of one Account by another in a social network; Creates -- creation of a Smart Contract by an Account; Calls -- call of a Smart Contract by an Account; Pays (Transfers) -- transfer of a funds for a Service/Product (not shown on Fig.1), which may be involving Smart contract or not, assuming that each Pay or Transfer may also have multiple properties such as amount of the payment and currency, inventory of services or products being ordered and paid and the optional rating "Rate" event, if applies. 4 Incremental Implementation and Temporal Graph Database Given the high volumes of data being involved in real-world financial networks and requirements for low response time on behalf of production systems, we focus on "incremental" design option for Reputation System described in the earlier work [2]. The implementation employs in-memory graph database of Aigents project available in Java as open source https://github.com/aigents/aigents-java. The first key feature of the Aigents graph database is its ability to store labeled temporal graphs with possibility to attach any value to an edge between two vertices, so the value can be either numeric value indicating weight or strength of the relationship in a weighted graph or a compound truth value in probabilistic logic network [10]. In case of the Reputation System implementation, it may keep ei- ther single rating assessment or a financial transaction value and currency, or the combination of them all so he ratings may be weighted by financial values as presented in the earlier work [9]. Notably, the edge may contain entire probabilistic distribution or the list of associated transactions along with rat- ings and financial values. The second feature of the Aigents graph database is its ability to slice graphs on temporal basis so that each time period is stored in separate subgraph while the subgraphs can be arbitrary merged, or the subgraphs of multiple temporal subgraphs can be extracted and merged over time. In the case of Repu - tation System implementation, it has appeared logical to keep all transactions segmented in temporal subgraphs specific to what is called recalculation period [2] or observation period [9], which is 1 day by default in the current implementation. The indexing of the data is primarily temporal and secondary based on vertices and types of the re - lationships. During the run-time, temporal ranges of subgraphs being processed on time scale are bound to memory resources available as well, so that amount of storage space limits time range that can be processed simultaneously. However, the incremental nature of the reputation recalculation given the "incremental" design option needs only few subgraphs to be present in memory at one time. Specifically, the two types of graphs are purposed for the Reputation System implementation. First, there is "reputation evidence data" with historical data accumulated for each of the observation period such as single day by default. Second, there is "reputation state" graph keeping current "reputation bal- ance" for each of the periods. That is, for each of the reputation update process per observation period accordingly to the algorithm specification [9], the only three subgraphs should be present in physical memory -- one "reputation evidence data" for the current period and two "reputation state" subgraphs for current and previous periods. 5 Application to real Social Network and Blockchain Data The illustration of how the Reputation System can work has been evaluated with use of real world data extracted from public blockchain Steemit as discussed in the earlier work [8]. We have tried to evaluate how the level of reputation computed by the Reputation System for participants of Steemit social network corresponds to the evaluations of trust and credibility given to them manually. For the purpose, we were computing reputations for entire network for long period of time based on both social and financial data involving voting for posts, commenting on posts and sending financial payments as well. Further, the computed reputation ranks were compared against the "black lists" maintained by network administrators and volunteers as well as against the lists of "whales", called so for listing well-known publicly available participants. The extra study has been performed to see how the reputation changes over time for accounts of different kinds ("black-listed" or "whales"), assuming every account starts with default reputation of 0.5 which may get changed to higher or lower over time, as it is shown on Fig. 3. The observation made given the reputation dynamics is that "expectedly highly reputable" (from the "whales" list) accounts are given longer "tails" spanning over time so reputation either does not de - cay or decays slowly. On the opposite, the "expectedly low reputable accounts" (from the "black lists") are present with fast decay of the reputation values. Fig. 3. Temporal dynamics of reputation values for randomly selected 5 ac- counts from "black-lists" and 5 accounts from "whales" list. Horizontal axis corresponds to time period of 3 months from left to right, vertical axis indi- cates reputation value in range from 0.0 to 1.0, labeled on scale from 0 to 100 on the chart. 6 Conclusion We have came up with generalized ontology capable to describe possible interactions in a wide range of online environments such as social networks, marketplaces and financial ecosystems including blockchains. We have designed and implemented Reputation System available as part of the open source Aigents project at https://github.com/aigents/aigents-java. As a part of the implementation, the temporal Aigents graph database has been successfully evaluated for the purpose of storage and processing of the reputation data. We were able to extract data described by the ontology from the real world social network and financial ecosystem based on public blockchain and have successfully evaluated the performance of the computations agains known reference data. References [1] Hazin M., Shheglov S. Stairway to Heaven. Dialogues about power, careers and the world elite. Book on Demand Ltd. February 24, 2018. https://www.amazon.com/Stairway-Heaven-Dialogues- careers-Russian/dp/5519547254 [2] Kolonin A., Goertzel D., Duong D., Ikle M. A Reputation System for Artificial Societies. arXiv:1806.07342 [cs.AI], 2018. https://arxiv.org/abs/1806.07342 [3] Swamynathan G., Almeroth K., Zhao B. The design of a reliable reputation system. Electron Commer Res 10: 239 -- 270, pp.239-270, 2010. DOI 10.1007/s10660-010-9064-y, 31 August 2010 http://people.cs.uchicago.edu/~ravenben/publications/pdf/reputation-ecrj10.pdf [4] Sänger, J. & Pernul G. Interactive Reputation Systems. Business & Information Systems Engineering, August 2018, Volume 60, Issue 4, pp 273 -- 287, 2018 60: 273, 2018. https://doi.org/10.1007/s12599-017-0493-1 [5] Androulaki E., Choi S.G., Bellovin S.M., Malkin T. Reputation Systems for Anonymous Networks. In: Borisov N., Goldberg I. (eds) Privacy Enhancing Technologies. PETS 2008. Lecture Notes in Computer Science, vol 5134. Springer, Berlin, Heidelberg, 2008. https://link.springer.com/chapter/10.1007/978-3-540-70630-4_13 [6] Gupta M., Judge P., Ammar M. A Reputation System for Peer-to-Peer Networks. NOSSDAV'03, June 1 -- 3, 2003, Monterey, California, USA, ACM 1-58113-694-3/03/0006, 2003). https://www.cs.indiana.edu/~minaxi/pubs/reputation.pdf [7] Blömer J., Juhnke J. & Kolb C. Anonymous and Publicly Linkable Reputation Systems. Financial Cryptography and Data Security 2015, Cryptology ePrint Archive: Report 2014/546, 2015. https://eprint.iacr.org/2014/546 [8] Kolonin A. Reputation Systems for Human-Computer Environments" In: Complexity, Informatics and Cybernetics: IMCIC 2019, conference materials, Orlando, Florida, USA, 2019. https://aigents.com/papers/2019/ReputationSystemsForHumanComputerEnvironmentsIMCIC201 9.pdf [9] Kolonin A., Goertzel D., Pennachin C., Duong D., Argentieri M., Znidar N. A Reputation System for Marketplaces - Viability Assessment. arXiv: 1902.03857 [cs.SI], 2019. https://arxiv.org/abs/1902.03857 [10]Goertzel B., Ikle M., Goertzel I., Heljakka A. Probabilistic Logic Networks: A Comprehensive Framework for Uncertain Inference, Springer Publishing Company, Incorporated ©2008 ISBN:0387768718 9780387768717. https://dl.acm.org/citation.cfm?id=1468044
1304.3747	1	1304	2013-04-12T22:48:15	The Social Maintenance of Cooperation through Hypocrisy	[ "cs.SI", "physics.soc-ph" ]	Cooperation is widespread in human societies, but its maintenance at the group level remains puzzling if individuals benefit from not cooperating. Explanations of the maintenance of cooperation generally assume that cooperative and non-cooperative behavior in others can be assessed and copied accurately. However, humans have a well known capacity to deceive and thus to manipulate how others assess their behavior. Here, we show that hypocrisy - claiming to be acting cooperatively while acting selfishly - can maintain social cooperation because it prevents the spread of selfish behavior. We demonstrate this effect both theoretically and experimentally. Hypocrisy allows the cooperative strategy to spread by taking credit for the success of the non-cooperative strategy.	cs.SI	cs	The Social Maintenance of Cooperation through Hypocrisy Todd J Bodnar, Marcel Salathé Department of Biology, Penn State University University Park, PA 18062, USA *Corresponding authors: Marcel Salathé, [email protected] W-251 Millennium Science Complex, University Park, PA 18062, USA Tel: (814) 867-4431 Cooperation is widespread in human societies, but its maintenance at the group level remains puzzling if individuals benefit from not cooperating. Explanations of the maintenance of cooperation generally assume that cooperative and non-cooperative behavior in others can be assessed and copied accurately. However, humans have a well known capacity to deceive and thus to manipulate how others assess their behavior. Here, we show that hypocrisy - claiming to be acting cooperatively while acting selfishly - can maintain social cooperation because it prevents the spread of selfish behavior. We demonstrate this effect both theoretically and experimentally. Hypocrisy allows the cooperative strategy to spread by taking credit for the success of the non-cooperative strategy. The maintenance of cooperation is perhaps one of the greatest enigmas in all of biology. In systems where non-cooperation is more beneficial to an individual than cooperation, why does cooperation persist? Among numerous theories that have been proposed in the past few decades, reciprocity (both direct and indirect)(1, 2), kin selection(3) and group selection(4), and spatial dynamics(5, 6) are the ones most frequently employed to account for this puzzling phenomenon.(7) In addition, punishment and reward have been identified as important contributors to the maintenance of cooperation in human societies.(8-13) The explanatory power for the maintenance of cooperation is usually assessed in game-theoretic models that assume faithful transmission (barring random mutation) of the cooperative or selfish behavior. If the behavior is transmitted genetically, faithful transmission is a valid assumption. However, if the behavior is adopted by imitation, as is generally the case in human affairs, then faithful 2 transmission requires that the behavior can be accurately inferred.(11, 14) Unless the behavior can be directly observed, it must be inferred through other means such as communications, thus allowing for the possibility of deception. In the context of the game theoretic model of cooperation without direct observation, an individual may act selfishly while claiming to be cooperative. Such hypocritical behavior is indeed observed commonly in psychological studies, where it is argued that hypocrites benefit not only from reaping the rewards of selfish behavior, but also “garner the social and self- rewards of being seen and seeing oneself as upstanding and moral”(15). Here, we argue that hypocrisy can promote the maintenance of cooperation in the population: even if a hypocrite acts selfishly, hypocrisy masks the selfish behavior, preventing it from spreading. We demonstrate this effect with a mathematical model, a computational agent-based simulation model, and an online experiment carried out on Amazon Mechanical Turk(16, 17). To assess the effect of hypocrisy on the maintenance of cooperation, we use a common model of cooperation, the public goods game(1, 11, 16, 18, 19). In the standard public goods game, players have the option of cooperating by donating to a pool. The donations are then multiplied by a constant and evenly divided between the other players regardless of whether they cooperated or not. In this game, non-cooperating individuals will on average have a higher payoff than cooperating individuals because they don’t pay the cost of donating. If strategies with higher payoffs are more often replicated, then the cooperative strategy will become less common over time despite the fact that the population would be best off if everyone would cooperate. We introduce a modification that concerns the copying process of strategies. It is typically assumed that strategies are faithfully copied (barring mutation), either through biological 3 transmission(20) (e.g. genetic inheritance), or through cultural transmission(14, 16, 21) (e.g. observational learning). This faithful copying process rests on the assumption that the information transmitted from sender (individual to be copied) to receiver (copying individual) is accurate. The modification that we introduce here is that the information does not need to be accurate. Specifically, we allow for selfish individuals to deceive others about their previous behavior during the imitation process, i.e. to falsely claim that they have been cooperative. We can thus describe an individual by two attributes with respect to cooperation and transmission. First, an individual can be described by the strategy played (cooperative or selfish). Second, and individual can also be described by the strategy transmitted. Standard models of cooperation don’t need to make this distinction because the assumed faithful copying process described above - however, in our model, the uncoupling of the strategy played from the strategy transmitted is essential. In order to study the effect of hypocrisy, we assume that there are two types of individuals, non-hypocritical and hypocritical. A non-hypocritical individual will always transmit the strategy that it has just played. A hypocritical individual will always transmit the cooperative strategy, regardless of the strategy that it has just played (Figure 1). We will later modify this definition and assume that a hypocritical individual will always transmit the opposite of the strategy that it has just played. While this latter scenario is perhaps not in line with the common definition of the word hypocrisy, it will provide a conservative baseline for the maintenance of cooperation because it forces hypocritical individuals to transmit the non- cooperative strategy even if it has just played the cooperative strategy (see supplementary material). We develop a mathematical model (see supplemental 2 for details) that calculates the equilibrium frequency of cooperation c given a constant frequency of hypocrisy h in the 4 population. It can be shown that for a population of any size, c will go to an equilibrium defined by: (1) Low levels of random strategy exploration does not fundamentally modify the outcome. These results were confirmed by computational agent-based simulations. A detailed description of the mathematical and computational model is given in the supplementary material. We tested these predictions experimentally by setting up an online experiment on Amazon Mechanical Turk (AMT), a widely used online platform(16, 17) to recruit participants from a diverse pool. We implemented a public goods game and recruited 414 adults on AMT to participate. Initially, each participant was assigned a random strategy (cooperative or noncooperative), but after each round, participants were allowed to switch strategies. Each round, the participants were shown their payoff in that round, and the payoff and strategy of another player chosen at random. Importantly, participants were not informed that they were playing a public goods game, and the strategies shown were simply labeled A and B. Thus, participants had no prior knowledge of the underlying game design, but were only informed at the beginning that they would be paid according to their performance. In order to test the effect of hypocrisy on cooperation, we compared 4 different levels of hypocrisy h = 0%, 25%, 50% or 75%, and ran 10 independent trials per hypocrisy level, where each trial is a game with maximally 12 participants playing 10 rounds (participants were prohibited from playing more than one game). If a participant played the noncooperative strategy, her strategy was falsely shown to others as cooperative with probability h. This allowed 5 us to implement hypocrisy without letting participants know that we did, preventing any bias that such knowledge might have introduced. Because the experimental setup of each trial is identical with the exception of h, any difference in the observed frequency of cooperation can be attributed to the effect of hypocrisy. The results of the online experiment are in agreement with the predictions made by the mathematical and computational models (Figure 2). After random initiation, the cooperation frequency follows the predicted trajectories, attaining near-equilibrium value given by equation (1) after 10 rounds. Using Tukey's honest significance test, we find that the frequency of cooperation is significantly different among all tested levels of hypocrisy (see supplemental table 1). We’ve shown here that hypocrisy allows for the maintenance of cooperation in a public goods game. In the standard public goods game without hypocrisy, cooperation cannot be maintained because non-cooperative individuals fare better than cooperative individuals, and their non-cooperative strategy is adopted by others wanting to maximize their payoffs. It’s important to note that the addition of hypocrisy to the public goods game does not affect any of these underlying assumptions: selfish individuals are still doing better, and individuals are still adopting the strategies of those around them with the highest payoffs. The only difference is that hypocrisy makes non-cooperative individuals with high payoffs appear cooperative, leading to the adoption of cooperation: Hypocrisy allows the cooperative strategy to spread by taking credit for the success of the non-cooperative strategy. Punishment of non-cooperative behavior is a well documented solution to problem of cooperation in human communities(8-13, 22). We may hypothesize that hypocrisy has evolved as a means to avoid punishment, providing obvious benefits in addition to those gained by social 6 or self-rewards(15, 23). Our model makes no claim as to why hypocrisy exists, and at what frequencies - it is only concerned with its effect on the maintenance of cooperation. Since hypocrisy is a form of deception, we would expect evolutionary pressure to detect hypocrisy and consequent counter-pressure to avoid detection, leading to an ongoing arm’s race. Low levels of hypocrisy detection abilities however do not seem to change our results (see supplementary material). Supplementary material is available at www.salathegroup.com REFERENCES 1. U. Fischbacher, S. Gächter, E. Fehr, Are people conditionally cooperative? Evidence from a public goods experiment, Economics Letters 71, 397–404 (2001). 2. T. Neugebauer, J. Perote, U. Schmidt, M. Loos, Selfish-biased conditional cooperation: On the decline of contributions in repeated public goods experiments, Journal of Economic Psychology 30, 52–60 (2009). 3. W. D. Hamilton, R. Axelrod, The Evolution of Cooperation, Science 211, 1390–1396 (1981). 4. M. A. Nowak, C. E. Tarnita, E. O. Wilson, The evolution of eusociality, Nature 466, 1057– 1062 (2010). 7 5. E. Lieberman, C. Hauert, M. A. Nowak, Evolutionary dynamics on graphs, Nature 433, 312– 316 (2005). 6. T. Bodnar, M. Salathé, Governing the global commons with local institutions, PLoS ONE 7, e34051 (2012). 7. M. A. Nowak, Five Rules for the Evolution of Cooperation, Science 314, 1560–1563 (2006). 8. H. Brandt, C. Hauert, K. Sigmund, Punishing and abstaining for public goods, Proc. Natl. Acad. Sci. U.S.A. 103, 495–497 (2006). 9. D. Helbing, A. Szolnoki, M. Perc, G. Szabó, Punish, but not too hard: how costly punishment spreads in the spatial public goods game, New Journal of Physics 12, 083005 (2010). 10. K. Sigmund, Punish or perish? Retaliation and collaboration among humans, Trends Ecol. Evol. (Amst.) (2007). 11. K. Sigmund, H. De Silva, A. Traulsen, C. Hauert, Social learning promotes institutions for governing the commons, Nature 466, 861–863 (2010). 12. E. Fehr, S. Gächter, Altruistic punishment in humans, Nature 415, 137–140 (2002). 13. E. Fehr, Cooperation and punishment in public goods experiments, (2000). 8 14. L. Rendell et al., Why copy others? Insights from the social learning strategies tournament, Science 328, 208–213 (2010). 15. C. D. Batson, D. Kobrynowicz, J. L. Dinnersteir, H. C. Kampf, A. D. Wilson, In a very different voice: unmasking moral hypocrisy, Journal of Personality and Social Psychology 72, 1335–1348 (1997). 16. D. G. Rand, S. Arbesman, N. A. Christakis, Dynamic social networks promote cooperation in experiments with humans, Proc. Natl. Acad. Sci. U.S.A. 108, 19193–19198 (2011). 17. D. G. Rand, The promise of Mechanical Turk: How online labor markets can help theorists run behavioral experiments, J. Theor. Biol. (2012). 18. G. Hardin, The tragedy of the commons, New York (1968). 19. E. Ostrom, Governing the Commons (1990). 20. M. A. Nowak, Evolutionary dynamics: exploring the equations of life (2006). 21. J. H. Fowler, N. A. Christakis, Cooperative behavior cascades in human social networks, Proceedings of the National Academy of Sciences 107, 5334–5338 (2010). 9 22. E. Fehr, B. Rockenbach, Detrimental effects of sanctions on human altruism, Nature 422, 137–140 (2003). 23. K. Sigmund, Reward and punishment, Proceedings of the National Academy of Sciences 98, 10757–10762 (2001). Acknowledgements MS acknowledges support from Society in Science - Branco Weiss Fellowship. Author Contributions TJB and MS conceived the project, designed the experiment, analyzed the data, and wrote the paper. TJB developed the models and implemented the AMT experiments. Competing Interests The authors declare no competing financial interests. Human Subject Experiments The AMT experiments were approved by the Penn State Institutional Review Board on October 1, 2012 (Protocol ID 41024). 10 FIGURE CAPTIONS Figure 1 A visual representation of the reported behavior (background) and the actual behavior (circle) for each combination of game and reporting strategies. Honest individuals report the same strategy that they actually did while hypocrites always report that they cooperated. (Key: Red = Defect, Blue = Cooperate) Figure 2 Dynamics predicted by the mathematical (solid lines) and agent based (dashed lines) models compared to the observed frequency of cooperation observed in the AMT experiments (circles). Transparent circles are the frequencies from each of the experimental groups, while the solid circles are the means of all of the experiments per level of hypocrisy and round. Horizontal jittering added for better visibility. 11 FIGURES Figure 1 12 Figure 2 13
1511.07337	1	1511	2015-11-23T17:40:18	Harnessing Mobile Phone Social Network Topology to Infer Users Demographic Attributes	[ "cs.SI", "physics.soc-ph" ]	We study the structure of the social graph of mobile phone users in the country of Mexico, with a focus on demographic attributes of the users (more specifically the users' age). We examine assortativity patterns in the graph, and observe a strong age homophily in the communications preferences. We propose a graph based algorithm for the prediction of the age of mobile phone users. The algorithm exploits the topology of the mobile phone network, together with a subset of known users ages (seeds), to infer the age of remaining users. We provide the details of the methodology, and show experimental results on a network GT with more than 70 million users. By carefully examining the topological relations of the seeds to the rest of the nodes in GT, we find topological metrics which have a direct influence on the performance of the algorithm. In particular we characterize subsets of users for which the accuracy of the algorithm is 62% when predicting between 4 age categories (whereas a pure random guess would yield an accuracy of 25%). We also show that we can use the probabilistic information computed by the algorithm to further increase its inference power to 72% on a significant subset of users.	cs.SI	cs	Harnessing Mobile Phone Social Network Topology to Infer Users Demographic Attributes Jorge Brea Grandata Labs Buenos Aires, Argentina [email protected] Javier Burroni Grandata Labs Buenos Aires, Argentina [email protected] ∗ Carlos Sarraute Grandata Labs Buenos Aires, Argentina Martin Minnoni Grandata Labs Buenos Aires, Argentina [email protected] [email protected] ABSTRACT We study the structure of the social graph of mobile phone users in the country of Mexico, with a focus on demographic attributes of the users (more specifically the users’ age). We examine assortativity patterns in the graph, and observe a strong age homophily in the communications preferences. We propose a graph based algorithm for the prediction of the age of mobile phone users. The algorithm exploits the topology of the mobile phone network, together with a sub- set of known users ages (seeds), to infer the age of remaining users. We provide the details of the methodology, and show experimental results on a network GT with more than 70 mil- lion users. By carefully examining the topological relations of the seeds to the rest of the nodes in GT , we find topological metrics which have a direct influence on the performance of the algorithm. In particular we characterize subsets of users for which the accuracy of the algorithm is 62% when pre- dicting between 4 age categories (whereas a pure random guess would yield an accuracy of 25%). We also show that we can use the probabilistic information computed by the algorithm to further increase its inference power to 72% on a significant subset of users. 1. INTRODUCTION Over the past decade, mobile phones have become preva- lent in all parts of the world, across all demographic back- grounds. Mobile phones are used by men and women across a wide age range in both developed and developing countries. They have become one of the most important mechanisms for social interaction within a population, making them an increasingly important source of information to understand human behavior, human demographics, and the correlations ∗Corresponding author. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. SNAKDD’14, August 24-27 2014, New York City, NY, USA Copyright is held by the owner/author(s). Publication rights licensed to ACM. 978-1-4503-3192-0/14/08 ...$15.00. http://dx.doi.org/10.1145/2659480.2659492. between them. In particular, mobile phone usage informa- tion is being used to perform quantitative analysis on the demographics of users regarding their age group, gender, education, and socioeconomic status [2, 3, 9]. In this work we combine two sources of information: com- munication logs from a major mobile operator in Mexico, and information on the demographics of a subset of the users population. This allowed us to first perform an ob- servational study of mobile phone usage, differentiated by categories such as gender and age groups [14]. This study is interesting in its own right, since it provides knowledge on the structure and demographics of the mobile phone market in Mexico. Here we focus our study on one variable: the age of the users. Furthermore, we tackle the problem of inferring demo- graphic attributes, such as their age. The ability to predict such features has numerous applications for social organiza- tions (for example, enabling them to perform specific health campaigns for women [9]) and for businesses, allowing them to gain insights on their customers and on the open mar- ket, to segment (current and potential) customers by de- mographic attributes, and to perform directed marketing campaigns. A semantic description of the network may consist in char- acterizing particular user attributes (e.g. gender, age or credit risk), or link attributes describing relationships be- tween users (e.g. friendship, family or working relationship). One can also characterize individual users based on more abstract social attributes derived from their communication patterns [9] or their geographic mobility [10, 13], and study how these features relate to other social and behavioral prop- erties [5]. Methods such as logistic regression, readily used in this approach, emphasize finding similarity patterns be- tween nodes given their known features and mostly rely on the assumption that the selected node features have a func- tional relationship with the particular attribute one wants to infer. A complementary approach is to focus on the complex relationships between individuals given by their social net- work, and exploit both short and long range interactions to infer feature relationships among all individuals in the net- work. Methods such as Laplacian spectral decomposition and graph diffusion algorithms, which emphasize the topo- logical relationship between nodes within the network, have been widely used [4]. These methods rely on the existence of social or group properties such as homophily and other com- munity structures prevalent in the network for a successful implementation. Our main contribution in this work is to present an ex- clusively graph based inference method, relying only on the topological structure of the mobile users network, together with a topological analysis of the performance of the algo- rithm. The equations of our algorithm can be described as a diffusion process with two added properties: (i) memory of its initial state, and (ii) the information is propagated as a probability vector for each node attribute (instead of the value of the attribute itself). We show that these two additions to the diffusion process significantly improve its in- ference power in the mobile network studied. Our algorithm can successfully infer different age groups within the network population given known values for a subset of nodes (the seed set). Most interestingly, we show that by carefully an- alyzing the topological relationships between correctly pre- dicted nodes and the seed nodes, we can characterize par- ticular subsets of nodes for which our inference method has significantly higher accuracy, going up to 62% (when pre- dicting between 4 age categories). We then show that we can use the information in the probability vector of each node to again increase our predictive power (up to 72%). The remainder of this paper is organized as follows. In section 2 we provide an overview of the datasets used in this study. We also describe the observations performed on the social graph, in particular concerning age homophily among users. In section 3 we describe our graph based algorithm and the topological metrics used in the analysis. In sec- tion 4 we present the results of the inference power of our algorithm. We focus on understanding how the performance of the algorithm depends on age groups and the topological properties of correctly inferred nodes. We then reexamine our results on the subnetwork composed solely of the mobile operator’s clients. We further study how the model parame- ters and the added information contained in the probability vectors can enhance the inference power of the algorithm. Finally in section 5 we conclude this work remarking the most significant findings and ideas for future work. 2. DATA SOURCE AND OBSERVATIONS 2.1 Dataset Description The dataset used in this work consists of cell phone calls and SMS (short message service) records (CDRs) for a three month period collected by a mobile phone operator for the whole country of Mexico. Each record contains detailed in- formation about the anonymized phone numbers of both the caller and the callee, whether it is an incoming or outgoing call, the duration and time stamp of the call, and the lo- cation of the antenna servicing the client involved in the communication. The detailed information provided by this dataset allows for the construction of large scale complex dynamic networks with added node (client) mobility based on the antenna location, which can be used for the study of human mobility [13]. A milestone of this work is to study how the “bare bones” structure of the mobile phone network contains useful in- formation for the inference of the users age group. For our purposes, we summarize each call/SMS record as a tuple (i, j) representing a call/SMS between users i and j. We then aggregate these tuples for all calls/SMS for a three month period into an edge list (ni, nj, wi,j) where nodes ni and nj represent users i and j respectively and wi,j is a boolean value indicating whether these two users have com- municated at least once within the three month period. This edge list will represent our mobile graph G = (cid:104)N ,E(cid:105) where N denotes the set of nodes (users) and E the set of com- munication links. We note that only a subset NC nodes in N are clients of the mobile operator, the remaining nodes N \ NC are users that communicated with users in NC but themselves are not clients of the mobile operator. This dis- tinction implies that our network G is an incomplete network in the sense that communications among members in N \NC are not observed in G. Together with the CDR records, we had access to several demographic variables for over 500,000 clients. These de- mographic variables include their actual age. We therefore distinguish three sets of nodes in our graph G: the whole set N , the set of client nodes NC , and the set of client nodes with known age NGT which we use as the ground truth nodes (NGT ⊂ NC ⊂ N ). Lastly we split our set NGT into a set of seed nodes NS consisting of approximately 75% of NGT which we use as seeds to generate age predictions for the remaining nodes in G, and the remaining nodes NV which we use as a validation set to evaluate the performance of our predictions. Figure 1: Population pyramid. In Fig. 1 we plot the population pyramid of the seed nodes discriminating by gender. We can see a bimodal structure for both genders with peeks centered around the ages of 30 and 45 years. This double peak in the distribution is not observed in the Mexico age population census, thus it is possibly indicative of some particular feature of the popula- tion of mobile phone clients arising from special offers such as family plans. The last step in our data preprocessing consisted in parti- tioning users ages into four age groups: below 25 years, from 25 to 34 years, from 35 to 49 years and 50 years and above. The motivation for the chosen partition is grounded in mar- keting purposes of the mobile phone operator. The predic- tive algorithm will therefore infer one of these age groups for each of the nodes in G. 15000100005000050001000015000population152030405060708085age(years)femalemale(a) Communications matrix C. (b) Random links matrix R. (c) Difference between C and R. (d) Number of links as function of age difference. Figure 2: Age homophily plots showing the communications matrix C, the random links matrix R, the difference between C and R, and the number of links according to the age difference. 2.2 Age Homophily Homophily is defined as “the principle that contacts be- tween similar people occurs at a higher rate than between dissimilar people” [11]. This basic observation about human relations can be traced back to Aristotle, who wrote that people “love those who are like themselves” [1]. In partic- ular, age homophily in social networks has been observed in the sociological literature, for instance in the context of friendship in urban setting [8] and other social structures [7]. These historical studies were limited to hundreds of individ- uals, due to the difficulty and cost of gathering data. A re- cent study on the structure of the Facebook social graph [17] shows that a strong age homophily can be observed in a so- cial graph with hundreds of millions of users. Our dataset allowed us to verify that the age homophily principle is valid in the mobile phone communication network on a nation wide scale (the country of Mexico in our case). As we mentioned in the introduction, graph based meth- ods like the one we present in this work rely strongly on the ability of the graph topology to capture correlations between node attributes we are aiming to predict. A most fundamen- tal structure is that of correlations between nodes and their neighborhoods. Figure 2(a) shows the correlation matrix C where Ci,j is the number of links between users of age i and age j for the nodes in the the ground truth NGT . Though we can observe some smaller off diagonal peaks, we can see that it is most strongly peaked along the diagonal, showing that users are much more likely to communicate with users of their same age. Performing a linear regression between the ages of linked users yields a regression coefficient r = 0.99 and gradient 1.06, confirming this observation. To account for the population density bias shown in Fig. 1, we compute a surrogate correlation matrix R as the expected number of edges between ages i and j under the null hypoth- esis of independence, ploted in Fig. 2(b): Ri,j = NGT × NGT (j) NGT (i) NGT × EGT (1) 051015202530354045age difference (years)102103104105106number of linkswhere NGT (i) = {x ∈ NGT : age(x) = i} and EGT = {(x, y) ∈ E : x ∈ NGT ∧ y ∈ NGT}. This matrix repre- sents a graph with the same nodes as the original, but with random edges (while maintaining the same number of edges as the original). Both C and R are represented with a loga- rithmic color scale. If we subtract the logarithm of R to the logarithm of the original matrix C, we can isolate the “social effect” (i.e. homophily) from the pure random connections, as can be seen in Fig. 2(c). Figure 2(d) summarizes the number of links as a function of the age difference δ, showing a clear peak when the dif- ference is δ = 0. The number of links decreases with the age difference, except around the value δ = 25, where an inter- esting inflection point can be observed; possibly relating to different generations (e.g. parents and children). This phe- nomenon can also be seen in Fig. 2(c), in the off diagonal bands at distance 25 years from the diagonal. 3. METHODS The strong correlations shown in Fig. 2 indicate that the topological information of the network can be used to suc- cessfully infer the nodes age group (knowing the age group for a small subset of seed nodes). We describe below our algorithm that performs such task. 3.1 Reaction-Diffusion Algorithm For each node x in G we define an initial state probability vector gx,0 ∈ RC representing the initial probability of the nodes age belonging to one of the C = 4 age categories defined in section 2.1. More precisely, each component of gx,0 is given by (gx,0)i = δi,a(x) 1/C if x ∈ NS if x (cid:54)∈ NS (2) (cid:40) (cid:80) (cid:80) (cid:80) where δi,a(x) is the Kronecker delta function and a(x) the age category assigned each seed node x. For non seed nodes, equal probabilities are assigned to each category. These probability vectors are the set of initial conditions for the algorithm. The evolution equations for the probability vectors g are then as follows x∼y wy,x gy,t−1 gx,t = (1 − λ) gx,0 + λ (3) where x ∼ y is the set of x’s neighbours and wy,x is the weight of the link between nodes x and y. For our purposes all link weights are set to one, thus the second term simplifies to a normalized average over the values of the neighborhood nodes for x: x∼y wy,x gx,t = (1 − λ) gx,0 + λ x∼y gy,t−1 {y : x ∼ y} . (4) The evolution equation is made up of two distinctive terms: the first term can be thought of as a memory term where in each iteration the node remembers its initial state gx,0, while the second term is a mean field term of the average values of the nodes neighborhood for x. The model parameter λ ∈ [0, 1] defines the relative importance of each of these terms. The influence of λ on the performance of the algorithm will be analyzed in section 4.5. We note that the evolution of each term in the probability vectors are totally decoupled from one another. This last assumption in the algorithm relies on the strong homophily suggested by the correlation plots of Fig. 2. Next, we briefly outline our motivation for calling it a reaction-diffusion algorithm. A reaction–diffusion model ex- plains how the concentration of one or more (typically chem- ical) substances changes under the influence of two pro- cesses: local reactions and diffusion across the medium [12]. Let A ∈ RN×N be the adjacency matrix of G, and D ∈ RN×N the diagonal matrix with the degree of each node in G, that is Ai,j = wi,j and Di,i = deg(i). Furthermore let L ∈ RN×N be the matrix defined as L = D − A. We can now rewrite equation (4) as t = (1 − λ)ga ga −1/2[D − L]D −1/2ga 0 + λD (5) where ga ∈ RN now runs over all nodes, for each age prob- ability category a. We are able to do this since from equa- tion (4), the probabilities of each category evolve indepen- dently. We now rewrite the above equation using the graph Laplacian for G defined as L = D−1/2LD−1/2 [6]: t−1 t = (1 − λ)ga ga 0 − λLga t−1 + λga t−1 finally reordering terms we get t−1 = (1 − λ)(ga t − ga ga 0 − ga t−1) − λLga t−1 (6) (7) t−1), is the reactive term which drives ga 0 − ga 0 and the second term λLga where on the left hand side we have the discrete deriva- tive of ga with respect to time, the first term on the right, (1 − λ)(ga t to- wards ga t−1 can be thought of as a diffusive term where the Laplacian operator in continuous space has been replaced by the Laplacian of the graph G over which diffusion is taking place. From equation (7), we can see that λ tunes the relative importance of the seed nodes information g0 and the topological properties of G given by the Laplacian L. We have therefore gone from a local de- scription of the model in equation (3), to a global description in equation (7). 3.2 Topological Metrics One of the main goals of this work is to uncover topolog- ical properties for a given node that increase its chances of being correctly predicted, and characterize (based on these properties) a subset of nodes in G for which our algorithm works particularly well. For this, we describe three topo- logical metrics used to characterize each node in G. The first metric is the number of seeds in a node’s neighbor- hood which we denote SIN (seeds in neighborhood). We expect SIN to be a relevant metric given the strong correla- tions shown in Fig. 2. The second metric is the topological distance of a node to the seed set which we denote DTS (distance to seeds), where distance is defined as the shortest path from the node to its closest seed. The third metric is the degree d of a node in the network, that is, the size of its neighborhood. Nodes with high degree are more likely to be strongly connected to the seed nodes, therefore we expect their response to the reaction diffusion algorithm, all else equal, to be significantly different than that of a node with low degree. 4. RESULTS In this section we first present the results for the pre- dictive power of the reaction-diffusion algorithm over the whole network G. To construct the mobile phone network we start with a directed edge list consisting of 252, 248, 440 edges where an edge can represent outgoing communica- tions from client x to user y and incoming communications from user y to client x distinctively. This set of edges in- volves 93, 797, 349 users. After pruning users with high de- gree (above 100), or not belonging to any connected com- ponent containing seed nodes, we end up with a resultant graph GT with ET = 125, 433, 585 edges (where edges have been symmetrized) and NT = 71, 949, 815 nodes, of which NS = 493, 871 are seed nodes and NV = 164, 768 are validation nodes for which we know their actual age group. These are used to test the inference power of our algorithm. As stated in equation (2), the initial conditions for each node’s probability vector gx,0 is that of total certainty for seed nodes and no information for the remaining nodes in GT . At each iteration, each gx,t in (4) updates its state as a result of its initial state and the mean field resulting from the probability vector of its neighbors. To reduce diffusion noise, in every time step we consider the mean field generated only by the neighbors of x that have received some information from the seed nodes. After the last iteration tend we assign the age group with highest probability in gx,tend to each node in GT . Unless stated otherwise, we set the value of λ = 0.5 and the number of iterations to 30 (which is 10 iterations more than the maximum distance from a node to the seed set: DTS max = 20). 4.1 Performance by Age Group In Table 1 we present the results for the performance of the reaction diffusion algorithm discriminating by age group. The age group demographics predicted by the algorithm, both for the whole graph NT and the validation set NV are similar to that of the seed set NS, that is, the algorithm does a good job at preserving the age group demographics of the seed set for all nodes in G. The last column shows the fraction of correctly validated nodes for each age group (hits) we refer to as the performance or inference power of the algorithm. The age group between 35 and 50 years is significantly better predicted than the other three age groups with a performance of 52.3%. Age group %NT <25 5.4% 25-34 35-50 50+ %NS %NV %hits 4.6% 26.6% 8.6% 32.7% 27.7% 32.3% 44.6% 37.6% 35.9% 35.8% 52.3% 24.4% 27.8% 27.3% 45.0% Table 1: Performance by age group. A possible explanation for the increased performance for the age group 35-50 can be found by looking at the correla- tion plot in Fig. 2 where we observe not only strong correla- tions and high population densities for this age group, but it is also less fooled by the offside peaks mostly concentrated within the age group 25-34. Again we see the importance of the correlation structure among neighboring nodes. Fi- nally, the overall performance for the entire validation set was 46.6% and we note that a performance based on ran- dom guessing without prior information would result in an expected performance of 25%, or an expected performance of ∼ 36% if we set all nodes age goup to the most probable category (35 − 50). Figure 3: Performance and population as function of SIN (seeds in network). 4.2 Performance Based on Topological Met- rics We next focus our attention on how our algorithm per- forms for different subsets of GT selected according to the topological metrics described in section 3.2. The main pur- pose of this section is to show that, given a set of seed nodes, there are specific topological features characterizing each node in the network that allows us to, a priori, select a set of nodes with optimal expected performance. We label each node in our validation set with the three topological metrics described in section 3.2. Figure 3 plots the performance (hits) of the algorithm over a set of nodes as a function of SIN (number of seeds in the nodes neighborhood). The algorithm performs worst for nodes with no seeds in the immediate neighborhood with hits = 41.5%, steadily rising as the amount of seeds increase with a performance of hits = 66% for nodes with 4 seeds in their neighborhood. We also see that the amount of nodes decreases exponentially with the amount of seeds in their neighborhood. An interesting feature we observed is that it is the total amount of seed nodes and not the proportion of seed nodes that correlate most with the performance of the algorithm (this last observation is not shown in the figure). We next examine how the algorithm performs for nodes in GT that are at a given distance to the seed set (DTS ). In Fig. 4 we plot the population size of nodes as a function of their DTS. The most frequent distance to the nearest seed is 2, and almost all nodes are at distance less than 4. This implies that after four iterations of the algorithm, the seeds information have spread to most of the nodes in GT . This figure also shows that the performance of the algorithm decreases as the distance of a node to NS increases. Both SIN and DTS, although topological quantities them- selves, are defined in terms of the distribution of the seeds set, and are not intrinsic properties the graph GT . We there- fore look at a basic intrinsic topological property of a node x in GT , namely its degree d(x). In Fig. 5 we see that the performance of the algorithm is lowest for nodes with small degree and gradually increases as the degree increases reaching a plateau for nodes with d(x) > 10. 0.00.51.01.52.02.53.03.54.0seeds in neighborhood0.00.10.20.30.40.50.60.7hits%population%Degree [1, 2] (2, 29] (29, 48] (48, 66] (66, 100] Distance to Seeds 1 2 3 0.415057 0.581540 0.618067 0.588592 0.566038 0.465881 0.445736 0.435747 0.438933 0.441583 0.313735 0.334968 0.299578 0.288136 0.206897 Table 2: Performance as function of degree and dis- tance to seeds. nodes at distance 2 from the seed set. Most interestingly, we find that selecting a subset based only on the nodes de- gree and distance to seeds, the optimal set is at distance 1 from the seed set and with degree between 29 and 48 for which the performance of the algorithm increases to hits = 61.8% with 20,050 nodes in NV satisfying these set of con- ditions. This last result indicates that it is not sufficient to look at each metric independently to select the subset of nodes for optimal performance of the algorithm. We also looked at the performance of the algorithms as a function of SIN and degree, and also found subsets of nodes for whom performance increased to similar values (table not shown for space reasons). 4.3 Preserving Age Demographics A salient feature of our algorithm is that the demographic information being spread is not the age group itself, but a probability vector for each age group. In each iteration, the algorithm does not collapse the information in each node to a preferred value; instead it allows the system to evolve as a probability state over the network, and only at the end of the simulation, when we “observe” the system, are we required to collapse the probability vectors to specific age groups. Choosing how to perform this collapse is not an obvious matter, in the previous sections we selected for each node the category with highest probability. This is a natural approach, and as we have shown, the algorithm’s performance is satisfactory. A draw back of this approach is that, when collapsing the probability vector at each node, we ignore the resultant probabilities for all other nodes in the network which could inform us in making a better se- lection choice. We might want to collapse the probability state of the system as a whole and not for each node in- dependently. For instance, if we want to impose external constraints on our solution, namely that the age group dis- tribution for the whole network be that of the seed set, we can do this optimally using the following algorithm, which we call Population Pyramid Scaling (PPS). Note that the population to predict has size NT. For each age group a we compute the number of nodes Na that should be allocated to group a in order to satisfy the dis- tribution constraint (the age distribution of NGT ), and such a=1 Na = NT (where C is the number of age groups). that(cid:80)C The PPS procedure is described in Algorithm 1. The performance of our algorithm constraining the de- mographic population of GT to that of the seed nodes NS is 46.5%, almost identical to 46.6% obtained without PPS. The reason for this is that the algorithm (without PPS) was able to preserve a distribution over the whole graph very similar to that of the seed nodes as we showed in Table 1. Figure 4: Performance and population as function of DTS (distance to seeds set). Figure 5: Performance and population as function of the nodes degree. Understanding what makes degree values close to 10 an inflection region in the performance of the reaction-diffusion algorithm is something we will address in a future work, but it is here worth remarking that GT is an essentially incom- plete graph, where we are missing all the communication links among users that are not clients of the mobile oper- ator. An important topological consequence of this is that whereas the average degree over the whole graph GT is 3.48, the average degree for client nodes is 37.53. Having shown the strong correlations between the three topological metrics and the algorithm’s performance – which allows us to select optimal values for each metric indepen- dently – we next examine whether we can find an optimal combination of these metrics. Table 2 shows the perfor- mance of our algorithm for different values of their DTS and their degree. Even though we have shown that perfor- mance increases with smaller distance to seeds and higher degree, here we find that this rule can be violated for spe- cific ranges of the nodes degree – as is shown for nodes with degree between 1 and 2 where performance is optimal for 1.01.52.02.53.03.54.04.55.0distance to seed0.00.10.20.30.40.50.6hits%population%100101102degree0.00.10.20.30.40.50.6hits%population%Algorithm 1: Population Pyramid Scaling 1 foreach node i and group a do 2 compute the probability pi,a that node i belongs to group a using our (unconstrained) algorithm; 3 end 4 Create an ordered list T of tuples (i, a, pi,a); 5 Sort list T in descending order by the column pi,a. The list T will be iterated starting with the element with the highest probability; 6 foreach element (i, a, pi,a) ∈ T do 7 8 9 10 if node i has not been assigned to a group then if less than Na nodes assigned to group a then assign node i to group a; end end 11 12 end Therefore in this case, the demographic constraint has not significantly changed our results. 4.4 Analysis of the Clients Subgraph In section 4.2 we mentioned that the graph GT has two distinct types of nodes, client and non client nodes with considerably different average degree. Motivated by this difference, and also considering the distinctive importance from a business point of view of understanding the predictive power of our algorithm on the client nodes, we performed the same study as above for the graph GC consisting only of client nodes. This graph is made up of EC = 6, 484, 571 edges and NC = 3, 225, 538 nodes of which 349,542 nodes are seed nodes and 143,240 nodes are validation nodes. Note the average degree of the client nodes has droped from 37.53 in GT to 4.02 in GC . Age <25 25-34 35-50 50+ %NT 5.3% %NS %NV %hits 4.6% 27.6% 7.1% 28.7% 29.9% 28.4% 50.9% 41.7% 34.6% 37.2% 53.4% 24.3% 28.4% 29.8% 48.0% Table 3: Performance by age group on the clients subgraph GC . In Table 3 we summarize the performance of the reaction- diffusion algorithm discriminated by age group. As for the full graph GT , the algorithm did a good job in preserving the age group demographics of the seed nodes over all the nodes in GC , in particular for the validation nodes. The last column indicates the performance of the algorithm for each age group, where two observations are in order: first we notice that the overall performance of 49.9% is significantly better than the performance of 46.6% for the full graph GT . Second the relative performance of each age group is quite similar for both graphs, plausibly indicating that it is mainly governed by the correlation structure described in Fig. 2. Following the same analysis done for the full graph GT we looked at how the performance of the reaction-diffusion algorithm correlates to the three topological metrics. These results are shown in Figures 6 and 7. Comparing these plots with those for graph GT we can notice several topological Figure 6: Performance as function of SIN (seeds in egonetwork) on the clients subgraph GC . Figure 7: Performance as function of DTS (distance to seeds) on the clients subgraph GC . differences between them. First we notice that while in GT only ∼ 30% of the validation nodes have at least one seed in their neighborhood, over 65% of nodes in GC have at least one seed in their neighborhood. This is not surprising given that both validation and seed nodes are client nodes. Most interestingly, even though for both graphs the performance of the algorithm increases as their SIN increases, for a fixed number of seeds the algorithm does better in the full graph than in the client graph. This is quite surprising given that the overall performance over the full graph is worse than in the client graph. In Fig. 7 we see that DTS for client nodes peaks at DTS = 1, falling exponentially as DTS increases (whereas in the full graph the peak was observed at DTS = 2). Performance of the algorithm again correlates strongly with distance to the seed set. Performance as function of the degree increased as in the study for the whole graph GT , though less abruptly (figure not shown for space reasons). 0.00.51.01.52.02.53.03.54.0seeds in neighborhood0.00.10.20.30.40.50.6hits%population%1.01.52.02.53.03.54.04.55.0distance to seed0.00.10.20.30.40.50.60.7hits%population%4.5 Sensitivity to Model Parameters So far we have analyzed the performance of the reaction- diffusion algorithm without addressing the issue of model parameters selection, namely the parameter λ presented in section 3.1 – which tunes the relative importance of the re- action and diffusion terms – together with the number of iterations of the algorithm, which determines to which ex- tent the age information in the seed set spreads across the network. In this section we examine these two parameters for the client network GC . Parameter λ. For λ = 0.0 (no diffusion) the performance is 28.1% but as soon as the diffusion term is slightly turned on, λ = 10−7, the performance of the algorithm jumps to 48.3% and re- mains almost constant for 0 < λ < 1 and drops suddenly to a performance of 42.5% for λ = 1.0 (total diffusion with no reactive term). We conclude from these results two things. First, the reactive term is important, increasing the perfor- mance from 42.5% (no reactive term) to 48.8% (optimally choosing the reactive term). Second, inside the boundary values for λ the performance remains almost unchanged and thus the algorithm is very robust to parameter perturba- tions. In other words, our algorithm needs some diffusion and reactive terms but the relative weights of these are unim- portant. Convergence. An important property of the model we address here is that of the model’s time scale, that is, how long it takes the system to reach a stationary state. Stationarity anal- ysis in complex networks such as the one we study here is a hard problem, thus for the purpose of this work we limit ourselves to the notion of performance. We therefore say the algorithm reaches a stationary state after t iterations if its performance remains unchanged for further iterations. We have seen that most nodes in the graph are not directly con- nected to a seed node, and that the diffusion term in Equa- tion (4) spreads from neighbor to neighbor one iteration at a time. We therefore expect that as iterations increase, the seed nodes’ information will spread further across the net- work. Our experiments show that after a single iteration, the performance of the reaction-diffusion algorithm is already at 45.5% gradually increasing its performance to a stationary value of ∼ 50% after 5 iterations. As conclusion, setting the number of iterations as tend = 30 gives us a good margin to ensure convergence of the algorithm. 4.6 Optimal Node Subset using Probability Vec- tor Information The focus of our exploratory analysis to find an opti- mal subset of nodes (where our algorithm works best) has been centered on looking at the network topology. Here we present a conceptually different strategy which exploits the information in the probability vector for the age group on each node. Namely, we examine the performance of the reaction-diffusion algorithm as we restrict the performance to a subset of nodes whose selected category satisfies a min- imal threshold value τ in its probability vector. In Fig. 8 we observe a monotonic increase in the performance as the threshold is increased and simultaneously a monotonic de- crease of the validation set. We note that for τ = 0.5 the Figure 8: Performance as function of τ . performance increased to 72% with 3,492 out of the 143,240 (2.4%) of the validation nodes remaining. The performance of the algorithm increases to ∼ 81% for τ = 0.55 but the number of validation nodes remaining has decreased to 201 (0.1%) making this estimate very unreliable. 5. CONCLUSION AND FUTURE WORK Large scale social networks such as those emerging from mobile phone communications are rapidly increasing in size and becoming more ubiquitous around the globe. Under- standing how these community structures forming complex network topologies can inform mobile phone operators, as well as other organizations, on unknown attributes of their users is of vital importance if they are to better understand, not only their clients’ interests and behavior, but also their social environment including the users that are not clients of the mobile operator. A better description of this ecosystem can give the mobile operators a business edge in areas such as churn prediction, targeted marketing, and better client service among other benefits. In this work we focused our attention on the bare bones topology of the mobile network, in other words, we aimed at uncovering the potential of the topology of the network itself to inform us on user attributes, in particular on user’s age groups. We proposed an algorithm where the only prior knowledge for inference is the age group of a subset of the network nodes (the seed nodes), and the network topology itself. Our algorithm achieved a performance of 46.6% for the entire network and of 53% for the age group 35−50 where pure random selections would have achieved a performance of 25%, or at most ∼ 36% by labeling each node with the most frequent age group. To understand why the communi- cation topology is allowing for this increase in performance we first studied the age correlations of users communicat- ing between each other where we included the labeled set. As shown in Fig. 2, we can observe a strong age homophily between callers which is being harnessed by our algorithm. We next searched for basic topological features to guide us in finding a subset of nodes where our algorithm per- forms best. Optimizing over all three metrics described in section 4.2, we found a subset of 20, 050 nodes where the performance of our algorithm increased to 62%. 0.250.300.350.400.450.500.55τ0.450.500.550.600.650.700.750.800.85hits%hits%validations020000400006000080000100000120000140000160000validationsGiven the particular nature of the mobile graph, where client nodes and non clients nodes stand on a different foot- ing, we tested the reaction-diffusion algorithm on the client only subgraph to examine differences in performance. Even though performance increased somewhat, relations of the topological metrics with performance behaved quite simi- larly and we were able to again find a sweet spot subset of nodes in NC , characterized by the topological metrics, for which the performance of our algorithm increased to ∼ 67%. Next we showed that the performance of the reaction- diffusion algorithm is robust to changes in the parameter λ which governs the relative importance of the diffusion and reactive term, as long as both these terms are present. We also saw the performance reaches a stationary value after 5 iterations. We found that restricting the prediction over nodes with stronger inference given by their probability vec- tor, we could achieve an increase of up to 72% in perfor- mance for a significant subset of nodes. In conclusion, in this work we have presented a novel al- gorithm that can harness the bare bones topology of mobile phone networks to infer with significant accuracy the age group of the network’s users. We have shown the importance of understanding nodes topological properties, in particular their relation to the seed nodes, in order to fine grain our expectation of correctly classifying the nodes. Though we have carried out this analysis for a specific network using a particular algorithm, we believe this approach can be useful to study graph based prediction algorithms in general. As future work, one direction that we are investigating is to combine the graph based inference approach presented here with classical machine learning techniques based on node features [14]. We are also interested in applying our methodology to predict variables related to the users’ spend- ing behavior. In [16] the authors show correlations between social features and spending behavior for a small population (52 individuals). We are currently tackling the problem of predicting spending behavior characteristics on the scale of millions of individuals. Another research direction is to study the correlations be- tween the user’s mobility and their demographic attributes. The geolocation information contained in the CDRs allows the analysis of users’ mobility patterns, e.g. their mobility motifs [15]. Recent studies have focused on the mobility patterns related to cultural events, such as sports [13, 18], where differences between genders and age groups may nat- urally be found. We expect that the interplay between mo- bility patterns and demographic attributes (such as gender and age) will provide relevant features to feed our predic- tion algorithms, and more generally will contribute to gain further insights on the behavior of different segments of the population. 6. REFERENCES [1] Aristotle. Complete works of Aristotle. Ed. J. Barnes, Princeton, NJ, 1995. [2] J. Blumenstock and N. Eagle. Mobile divides: gender, socioeconomic status, and mobile phone use in Rwanda. In Proceedings of the 4th ACM/IEEE International Conference on Information and Communication Technologies and Development, page 6. ACM, 2010. [3] J. E. Blumenstock, D. Gillick, and N. Eagle. Who’s calling? demographics of mobile phone use in Rwanda. Transportation, 32:2–5, 2010. [4] O. Chapelle, B.Scholkopf, and A. Zien. Semi-supervised learning. MIT Press, 2006. [5] E. Cho, S. A. Myers, and J. Leskovec. Friendship and mobility: user movement in location-based social networks. In ACM SIGKDD, pages 1082–1090. ACM, 2011. [6] F. R. K. Chung. Spectral graph theory. CBMS, 1997. [7] S. L. Feld. Social structural determinants of similarity among associates. American Sociological Review, pages 797–801, 1982. [8] C. S. Fischer, C. Stueve, L. M. Jones, R. M. Jackson, K. Gerson, and M. Baldassare. Networks and places: Social relations in the urban setting. Free Press New York, 1977. [9] V. Frias-Martinez, E. Frias-Martinez, and N. Oliver. A gender-centric analysis of calling behavior in a developing economy using call detail records. In AAAI Spring Symposium: Artificial Intelligence for Development, 2010. [10] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi. Understanding individual human mobility patterns. Nature, 453(7196):779–782, 2008. [11] M. McPherson, L. Smith-Lovin, and J. M. Cook. Birds of a feather: Homophily in social networks. Annual review of sociology, pages 415–444, 2001. [12] G. Nicolis and I. Prigogine. Self-organization in non equilibrium systems. Wiley, 1977. [13] N. Ponieman, A. Salles, and C. Sarraute. Human mobility and predictability enriched by social phenomena information. In Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pages 1331–1336. ACM, 2013. [14] C. Sarraute, P. Blanc, and J. Burroni. A study of age and gender seen through mobile phone usage patterns in Mexico. In 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining. ACM, 2014. [15] C. M. Schneider, V. Belik, T. Couronn´e, Z. Smoreda, and M. C. Gonz´alez. Unravelling daily human mobility motifs. Journal of The Royal Society Interface, 10(84):20130246, 2013. [16] V. K. Singh, L. Freeman, B. Lepri, and A. S. Pentland. Predicting spending behavior using socio-mobile features. In Social Computing (SocialCom), 2013 International Conference on, pages 174–179. IEEE, 2013. [17] J. Ugander, B. Karrer, L. Backstrom, and C. Marlow. The anatomy of the Facebook social graph. Structure, 5:6, 2011. [18] F. H. Xavier, C. H. Malab, L. Silveira, A. Ziviani, J. Almeida, and H. Marques-Neto. Understanding human mobility due to large-scale events. In Third International Conference on the Analysis of Mobile Phone Datasets (NetMob), 2013.

Subsets and Splits